ArXiv Domain 2026-09-22

数据来源：ArXiv Domain

LLM Domain Papers

1. Do small language models know what they don’t know?

Abstract:We explore whether entropy-based confidence signals can be leveraged to improve the accuracy of Small Language Models (SLMs) with fewer than 3 billion parameters, running entirely on consumer hardware. We evaluate seven distinct approaches, including token-level entropy early stopping, semantic entropy estimation, and uncertainty-aware routing to larger expert models, across 7 model pairs and 5 standard NLU benchmarks. Our key finding is that token-level entropy is effectively blind in SLMs: in 91% of dataset-model combinations, mean token entropy is near zero regardless of answer correctness, rendering token-based confidence signals unusable at this scale. We demonstrate that semantic entropy, computed by generating multiple samples, clustering answers by meaning, and measuring distributional uncertainty, recovers a viable confidence signal. Using semantic entropy to selectively route uncertain queries to a larger expert model yields accuracy improvements of up to +50 percentage points. Notably, cross-family routing (e.g., SmolLM 360M to Phi-3.5-mini) averages +22.0% improvement compared to +6.8% for same-family routing, revealing that expert model quality matters more than architectural compatibility. Our results suggest that the value proposition for entropy-based methods in SLMs is not computational savings but intelligent compute allocation: spending more tokens where they matter most.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.20824 (HTTP 429)

Authors: Prashant Mudgal

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.20824.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.20824

Published: 2026-09-22T01:51:39.637Z

2. HERMES: Contrast-Aware Knowledge Graph Reasoning from Clinical Notes for Patient Outcome Prediction

Abstract:Clinical predictive models often rely on structured Electronic Health Record data, such as time-series and procedure codes. While recent approaches have begun leveraging unstructured clinical notes, they typically encode them as flat sequences, which may lose explicit relational and temporal structure present in clinical narratives. In response, we propose HERMES, a graph-based framework that operates exclusively on clinical text while preserving clinical relationships. This approach builds on two key ideas. First, personalized Knowledge Graphs (KGs) are constructed through Large-Language-Model-guided extraction from clinical notes with Contrastive Logic Modeling that explicitly captures temporal dynamics and treatment failures and changes in outcomes. Second, a Graph Attention Network synthesizes patient representations through graph-based learning over the KGs. Experiments on MIMIC-III and MIMIC-IV for in-hospital mortality and 30-day readmission prediction show that HERMES consistently outperforms strong text-only baselines. Our findings demonstrate that explicit relational modeling with Contrastive Logic Modeling significantly advances predictive performance.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.20825 (timeout of 15000ms exceeded)

Authors: Gia-Bach Nguyen, Hoang-Ha Nguyen, Tuan-Cuong Vuong, Trang Mai Xuan, Duy Quoc Ngo, Tien-Cuong Nguyen, Huan Vu, Thien Van Luong

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.20825.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.20825

Published: 2026-09-22T01:51:39.637Z

3. TALON: A Temporally Aware Longitudinal Framework for Radiology Report Generation

Abstract:Current radiology report generation (RRG) models usually produce descriptive reports based on a single examination or only the most recent prior examination, limiting their ability to perform accurate and meaningful longitudinal comparisons and detect subtle interval changes. Although recent approaches have begun to incorporate multiple prior examinations, they usually aggregate a fixed-length history without explicitly modeling the role-dependent relevance of each prior examination before fusion. To address this, we propose TALON, a Temporally Aware LONgitudinal RRG framework that adaptively integrates variable-length patient histories. The underlying Dual-Channel Temporal Fusion Module (DCTFM) compares the current examination with each prior examination through complementary similarity and change channels to capture persistent findings and interval changes, respectively. The specially designed channel-specific attention estimates the relevance of each prior examination, while a learned prior-specific gate adaptively integrates informative longitudinal evidence and suppresses redundancy. Experiments on MIMIC-CXR show that TALON outperforms the current state-of-the-art method on various clinical efficacy and graph-based metrics. When more prior examinations become available, TALON’s performance on these metrics improves even further, emphasizing the strength of TALON’s DCTFM in modeling longitudinal RRG across longer and more complex patient histories than existing approaches.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.20826 (HTTP 429)

Authors: Nien-Tsyr Sun, Min-Chen Chen, Hui Nien Hung, Vincent S. Tseng

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.20826.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.20826

Published: 2026-09-22T01:51:39.637Z

4. From Discharge Notes to Patient Understanding: Persona-Grounded, Open-Ended Simulation of LLMs as Discharge Educators

Abstract:Hospital discharge education is an interactive teaching task: a clinician adapts a discharge plan to a patient’s literacy, recall, and personality. Existing LLM evaluations target static or artifact-generation tasks and do not measure patient understanding under open-ended dialogue. We introduce DischargeBench, a persona-grounded simulation in which a candidate LLM educator conducts a multi-turn session with a Virtual Patient, while an Education Monitor Agent regulates patient realism without modifying the educator, protecting the evaluation signal. We curate MIMIC-IV-Ext-DischargeBench, 477 cases over 24 ICD chapters with persona axes (personality, education level, health literacy, past-medical-history recall) for stratified analysis. Each simulation is scored on four axes — Conversation Quality, Topic Checklist, Comprehension, and Factual Consistency — by an LLM-as-a-Judge aligned against physician annotations. Across closed- and open-source LLMs, aggregate scores conceal clinically relevant variation across ICD chapters and patient personas; difficult personas expose coverage failures, comprehension gaps, and reduced source-answer agreement. LLM evaluation for discharge education should center patient understanding, not text quality or answer accuracy alone.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.20827 (HTTP 429)

Authors: Won Seok Jang, Zonghai Yao, Hong Yu

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.20827.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.20827

Published: 2026-09-22T01:51:39.637Z

5. Beyond WER: Entity and Disfluency Recall in Accented Conversational ASR

Abstract:ASR systems optimised for Word Error Rate (WER) often miss named entities and filled pauses in accented conversational English, both critical for language-learning feedback. We present a three-stage pipeline for speakers from India, Indonesia, and Latin America: (1) heuristic SQL filters curating entity-rich training data at 2.8x the entity density of random sampling, (2) regional LoRA adapters fine-tuned on Qwen2.5-Omni-3B producing both verbatim and corrected transcripts in a single forward pass, and (3) a six-category error taxonomy validated by an LLM-based judge (83.8% agreement, 210 human-labelled samples). The pipeline achieves 80-85% entity recall (up from 53-55%), 76-86% filler recall (up from <5%), and 6-10% WER across 6k test utterances, outperforming Whisper and a commercial ASR on entity recall while matching a zero-shot 30B model with 10x fewer parameters. Paired bootstrap tests confirm that curation alone accounts for 2.8-4.2 pp of entity recall gain (p<0.0001).

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.20828 (HTTP 429)

Authors: Fiza Husain, Ankit Pandey, Yash Singh

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.20828.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.20828

Published: 2026-09-22T01:51:39.637Z

6. SAGE: Schema-Guided LLMs for Grant Review

Abstract:Grant reviewers must apply detailed criteria to application forms, budgets, and supporting documents while producing assessments that colleagues can inspect. We present SAGE, Schema-Guided Aspect-Based Grant Evaluation, a system that translates a grant rubric into structured checks and links its judgements to evidence from the application package. We evaluate SAGE in two stages on 35 nonprofit grant applications. A post-factum comparison with 105 reviews from the original competition shows fair ordinal agreement (kappa = 0.29). The foundation then conducted a criterion-level re-review after inspecting SAGE, producing 202 assessments. In this assisted round, SAGE reached kappa = 0.58 and outperformed a one-prompt-per-criterion baseline (kappa = 0.33 on the common subset), with higher rank correlation and lower error. A claim-level audit further identifies confirmed, disputed, and unaddressed parts of the structured draft. SAGE operationalizes the review methodology by producing a detailed, evidence-linked, and auditable draft for expert correction.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.20829 (HTTP 429)

Authors: Erik Varapaev, Andrei Chetvergov, Stepan Ukolov, Timofei Sivoraksha, Alexander Evseev, Sergey Bolovtsov

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.20829.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.20829

Published: 2026-09-22T01:51:39.637Z

7. Reviser: Revision-Capable Text Generation via Autoregressive Cursor Actions

Abstract:Revision-capable generation is appealing because it can insert or revise earlier content, but many non-autoregressive and edit-based approaches obtain this flexibility through repeated sequence-level computation. We propose Reviser, a decoder-only Transformer that generates a response as a sequence of cursor-relative actions on a mutable canvas. At each step, Reviser predicts exactly one action token: INSERT(token), MOVE($\Delta$), or STOP, and is autoregressive over edit-history actions rather than final text order. This design enables genuinely non-monotonic generation while preserving a simple next-action interface. On a continuation benchmark, Reviser is strongly preferred to SEDD and MDLM in our arena evaluations, and trajectory statistics confirm that the model performs frequent backward moves and mid-canvas insertions rather than merely emulating end-append decoding. Against size-matched autoregressive baselines, Reviser is competitive at both the 100M and 300M scales. Under our shared FLOPs convention, Reviser also requires substantially less inference compute than representative multi-pass refinement and diffusion-style baselines.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.20830 (HTTP 429)

Authors: Sean Diab

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.20830.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.20830

Published: 2026-09-22T01:51:39.637Z

8. Recursive Language Models Generalize Out of Domain

Abstract:We study when limiting what a language model can see improves learning. We compare standard CoT, the more general learner that reads the full trace, with recursive language models, which restricts itself by solving each subtask in an isolated context. In-distribution, this generality comes for free: CoT can efficiently simulate the recursive rule, so the IID generalization guarantee changes only by a constant factor, and recursion does not offer much. But out of domain, CoT can fit training by relying on context outside the current subtask, i.e. a shortcut that breaks once those tokens change; recursive context isolation rules out this failure mode. Even though CoT’s class still covers the recursive rule, simplicity bias picks the shortcut over the truth. Thus, to go beyond distributional accuracy and truly reason, covering the right rule is not enough; this contrasts with classical learning theory.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.20831 (HTTP 429)

Authors: Chenxiao Yang, Zhiyuan Li, David McAllester, Nathan Srebro

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.20831.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.20831

Published: 2026-09-22T01:51:39.637Z

9. TatBLiMP: A Benchmark of Linguistic Minimal Pairs for Tatar

Abstract:We introduce TatBLiMP, the first benchmark of linguistic minimal pairs for Tatar (tt, ISO 639-3 tat), a Qypchaq Turkic language written in Cyrillic. To our knowledge it is the first grammaticality evaluation for Tatar language models of any kind, since even the 101-language MultiBLiMP does not include Tatar. TatBLiMP covers 16 morphosyntactic phenomena in 1248 sentence pairs. Each pair differs by a single morpheme, one grammatical and one ungrammatical. A model passes a pair when it assigns higher probability to the grammatical member. Scoring compares probabilities the model already assigns, so the benchmark needs no text generation and no parser, and it runs on base models and on mid-training checkpoints. TatBLiMP adapts the phenomenon inventory and single-morpheme breaking operations of TurBLiMP to Tatar and adds one phenomenon specific to Tatar, bare-noun number after numerals and quantifiers. The grammatical member of every pair is an attested sentence from Tatar literary prose. The ungrammatical member is produced by a deterministic single-morpheme perturbation with the apertium-tat transducer. Every pair is ratified by a native speaker. A plausibility principle governs construction, so the ungrammatical member is a plausible real-world error rather than an arbitrary corruption. Across from-scratch Tatar models, cross-lingual adaptations, and frontier multilingual LLMs, the benchmark tracks focused Tatar training rather than parameter scale. A 478M from-scratch model and a 125M monolingual model lead near 0.97, a 7B adaptation trails, frontier LLMs of 30-120B parameters fall to 0.80-0.92, and a lightly tuned multilingual model is weakest. We close with the benchmark’s main limitation. Its inherited taxonomy omits the morphophonology, vowel harmony and consonant assimilation, that is most salient to native speakers, and we sketch a native second layer that would add it.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.20832 (HTTP 429)

Authors: Ilshat Saetov, Dmitry Gaynullin

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.20832.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.20832

Published: 2026-09-22T01:51:39.637Z

10. Transsion’s Speaker-Attributed Multilingual ASR System for the MLC-SLM 2026 Challenge

Abstract:This paper presents the Transsion Speech Team submission to Task 1 of the MLC-SLM 2026 Challenge, which focuses on speaker-attributed transcription for multilingual conversational speech. We propose a cascaded framework consisting of three components: a speaker diarization module, a long-form multilingual ASR module, and a speaker-transcription fusion module. The diarization module is built upon DiariZen and produces speaker-homogeneous segments through local speaker activity estimation and global speaker clustering. The ASR module is based on Qwen3-Omni and generates multilingual transcriptions, while an external CTC-based alignment model provides precise word- and character-level timestamps. Finally, the fusion module combines diarization outputs with timestamped transcriptions to generate speaker-attributed STM outputs. Experimental results on the official evaluation set demonstrate the effectiveness of the proposed framework. The submitted system achieves a tcpMER of 15.41% and ranks second among all participating teams.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.20833 (HTTP 429)

Authors: Zhecheng Ren, Xuanji He, Xiaoxiao Li, Zhichen Han, Gaoyang Dong, Gaosheng Zhang, Minchuan Chen, Fengjie Zhu

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.20833.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.20833

Published: 2026-09-22T01:51:39.637Z

Agent Domain Papers

1. RBS-Attention: Radius-Bounded Sparse Prefill for Long-Context Large Language Models

Abstract:Long-context large language model inference is increasingly limited by prefill, where dense self-attention processes the entire prompt before generation begins. Sparse block selection can reduce this cost, but a block centroid may hide a highly relevant token among many irrelevant ones. We call this failure mode mean dilution and propose RBS-Attention, a training-free sparse-prefill method with two complementary selection branches. A centroid base branch captures average relevance, while a rescue branch uses the maximum key-block radius and its prompt-, layer-, and head-dependent distribution to identify blocks at risk of underestimation. Independently thresholding the two branches and combining their masks controls the contribution of rescue blocks while preserving regular block-sparse FlashAttention execution. On H100 GPUs, RBS-Attention achieves 20.65$\times$ standalone prefill-attention speedup, 11.92$\times$ vLLM prefill-attention speedup, and 5.97$\times$ end-to-end time-to-first-token speedup at 128K on Qwen3-30B-A3B-Instruct-2507-FP8. On the dense Qwen3-32B model, it obtains 88.65 overall RULER accuracy versus 89.52 for dense attention; LongBench-v2, InfiniteBench, and Video-MME provide additional quality evaluation. Supporting experiments measure actual retention, compare selectors at matched density, and characterize block-size, threshold, and memory behavior. Together, these results support radius-adaptive dual-branch selection as an effective approach to long-context prefill.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.20971 (HTTP 429)

Authors: Chuxu Song, Jiuqi Wei, Zhencan Peng

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.20971.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.20971

Published: 2026-09-22T01:52:31.318Z

2. Attention-Aware Routing: Coupling Routing and Attention in MoEs

Abstract:In Mixture-of-Experts language models, the router typically selects and weights experts based on the token’s hidden state, utilizing limited contextual information. We propose Attention-Aware Routing (AAR), which augments the router with temporal and spectral features extracted from a sliding window of attention weights that represent a summary of the model’s contextual state, disentangled from the hidden state. Keeping the base transformer entirely frozen, we train only the routing parameters, isolating routing as the sole variable. AAR improves GSM8K by +3.37 pp over a routing-only SFT baseline on OLMoE. Beyond performance, we show that routing and attention form a coupled circuit: routing changes at layer l propagate through the residual stream to amplify attention sinks at layer l+1, reshaping attention without any direct update to the attention mechanism itself. Further, AAR reduces long diverging generation, with incorrect answers getting shorter, while correct answers remain unchanged in length. Finally, AAR is strongly depth-sensitive: applying it indiscriminately across layers can degrade factual retrieval, whereas mathematical reasoning gains persist when it is introduced deeper in the network. This sensitivity exposes a retrieval—reasoning tension across depth and makes layer-selective AAR a controlled probe of the routing-relevant information carried by attention at different layers.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

Q: 这篇论文试图解决什么问题？这篇论文针对\*\*标准混合专家（MoE）模型路由器仅依赖隐藏状态进行专家选择所带来的局限性\*\*展开研究。具体而言，标准路由器以token的隐藏状态作为唯一输入，导致其只能利用有限且与内容信息高度纠缠的上下文信号，难以充分捕捉序列内部的结构关系。为解决上述问题，论文提出\*\*注意力感知路由（Attention-Aware Routing, AAR）\*\*，其核心思路与解决目标可概括为： - \*\*上下文与内容的解耦\*\*：通过向路由器引入从滑动窗口注意力权重中提取的时间与频谱特征，为路由决策提供一种与隐藏状态分离的、压缩的序列内关系信号，从而将“上下文关系信息”显式地注入专家选择过程。 - \*\*提升推理性能\*\*：在仅训练路由参数、保持主干 transformer 完全冻结的条件下，AAR显著提升了数学推理能力（如在 OLMoE 上将 GSM8K 提高了 +3.37 个百分点），且效果可迁移至不同规模的 MoE 模型。 - \*\*揭示路由与注意力的耦合机制\*\*：论文证明路由决策并非孤立的调度行为——第 l 层的路由改变会通过残差流传播，在无需直接修改注意力机制的情况下放大第 l+1 层的注意力汇聚（attention sinks），从而影响模型的内部信息处理模式。 - \*\*改善生成行为\*\*：AAR能够减少模型在回答错误时的冗长发散生成（过长推理链），使错误答案更短，同时不改变正确答案的长度，缓解“过度思考”现象。 - \*\*暴露深度相关的功能张力\*\*：论文发现AAR的效果具有强烈的深度敏感性——在早期层应用会干扰事实检索，而在中深层应用则能保持数学推理增益。这揭示了注意力信号在不同深度承载着“检索”与“推理”的不同功能，为理解MoE模型的分层信息处理提供了可控探针。 Q2: 有哪些相关研究？该论文在第2节及正文讨论中涉及的相关研究可归纳为以下六个方向： \*\*1. MoE 路由机制\*\* - \*\*标准 token-choice 路由\*\*：Shazeer 等人（2017）提出以 token 表示经线性投影后计算 softmax，再选取 top-k 专家的方式，这构成了当前 MoE 层的主流路由范式。 - \*\*负载均衡\*\*：Fedus 等人（2022）引入辅助损失以鼓励专家利用率均匀，缓解路由坍塌。 - \*\*Expert-choice 路由\*\*：Zhou 等人（2022）将分配逻辑反转，由专家选择 token，而非 token 选择专家。 \*\*2. 路由器的多样化改进\*\* - \*\*层间循环路由器\*\*：Qiu 等人（2025）在每层使用共享的循环单元，其输入除当前层隐藏状态外，还包含先前路由决策的隐状态。 - \*\*超网络生成路由参数\*\*：Do 等人（2023）提出 HyperRouter，利用固定的随机初始化超网络基于可训练路由嵌入生成路由器参数，以降低路由熵并限制推理时激活的专家数量。 - 这些工作的共同动机与 AAR 一致：均试图突破“路由仅依赖单一隐藏状态”的限制，引入额外条件信号以改进专家分配。 \*\*3. 注意力汇聚（Attention Sinks）\*\* - \*\*现象发现\*\*：Xiao 等人（2024）观察到 Transformer 会将不成比例的注意力质量集中于序列首 token（即 attention sink），该机制在流式上下文中起到稳定表示的作用。 - \*\*功能影响\*\*：Barbero 等人（2025）进一步证明 attention sink 的存在会减缓信息混合，并提升模型对提示扰动的鲁棒性。AAR 的实验表明，路由改变可通过残差流间接放大下一层的 attention sink。 \*\*4. 层间功能特化\*\* - \*\*深度与任务分工\*\*：Yang 等人（2026）与 Song 等人（2026）的研究显示，浅层大语言模型层主导似然估计与检索任务，而中深层对推理与生成至关重要。 - \*\*推理模式与事实捷径\*\*：Yang 等人（2025）区分了深层中的潜在推理（latent reasoning）与浅层的事实捷径（factual shortcuts）。AAR 的深度敏感性实验与这些发现形成呼应，并提供了基于注意力信号的深度探针视角。 \*\*5. 面向推理的注意力感知干预\*\* - \*\*推理时干预\*\*：Nguyen 等人（2026）在推理阶段基于经验性硬编码规则实施注意力感知干预，以改进思维链推理。 - \*\*与 AAR 的区别\*\*：AAR 并非在推理时注入外部规则，而是在训练期间将注意力信号嵌入路由机制，使该效应成为模型架构的原生、可学习且持续存在的特性。 \*\*6. 思维链中的过度思考（Overthinking）\*\* - \*\*长度与错误率的关系\*\*：Li 等人（2026）与 Su 等人（2025）发现较长的推理链往往与更高的错误率相关。 - \*\*与 AAR 的关联\*\*：AAR 在 GSM8K 上的实验显示，错误答案的生成长度显著缩短，而正确答案长度几乎不变，从而将路由层面的改变与“过度思考”这一外部行为现象联系起来。 Q3: 论文如何解决这个问题？论文通过提出\*\*注意力感知路由（Attention-Aware Routing, AAR）\*\*来解决标准 MoE 路由器仅依赖隐藏状态、缺乏显式序列结构信号的问题。该方法在不修改主干 Transformer 参数的前提下，将注意力权重作为与隐藏状态解耦的上下文信号引入路由决策。具体实现分为以下步骤： ### 1. 提取局部注意力窗口对于位于第 l 层的当前 token i ，从该 token 对前 M 个 token 的注意力权重中，跨注意力头取平均，组装成一个固定长度的一维滑动窗口：

w_i^l ∈ R^M
若序列前端不足 M 个 token，则以零从左侧填充，确保窗口长度恒定。 ### 2. 构造增强型注意力窗口原始窗口 w_i^l 可进一步经固定的非学习特征映射 φ 变换，以提取多视角结构信息：

w_i^l = φ(w_i^l) ∈ R^(M’)
映射 φ 包含两种视图： - **时域视图**：直接保留原始窗口（AART 设定）； - **频域视图**：计算离散傅里叶变换（DFT）的幅度谱，取前 K = lfloor M/2 rfloor + 1 个幅度分量，并以 log(1 + ·) 压缩动态范围。二者拼接后得到：
φ(w_i^l) = [ w_i^l ,|, log(1 + |F(w_i^l)|) ]
其中 F 表示 DFT。若仅使用时域视图，则 M’ = M ；若拼接频域视图，则 M’ = M + K 。 ### 3. 注意力路由器生成专家 logits 增强窗口经一个线性层投影到专家空间，产生注意力路由 logits：

g_i^(attn) = A_l^top w_i^l, quad A_l ∈ R^(M’ × E)
矩阵 A_l 被设为**块对角**形式，使时域与频域视图各自拥有独立的线性块，避免跨视图权重耦合。 ### 4. 与标准路由器的插值融合为控制注意力信号的插入强度，AAR 为每个视图引入一个由隐藏状态和该视图窗口共同决定的门控置信度：

α_i = σ( G_l [ h_i^l ,|, w_i^l ] ) ∈ R^(2E)
其中 G_l 同样为块对角矩阵， σ 为 sigmoid 激活。令 $B = (1) / (2)
I_E ,|, I_E
，并定义标准路由器的残差权重 β_0 = 1 - B α_i$，则最终路由 logits 为：
g_i = β_0 odot g_i^(std) + B ( α_i odot g_i^(attn) )
当 α_i equiv 0 时，系统退化为标准路由器；当 α_i > 0 时，注意力信号以自适应、逐专家的方式调制路由决策。 ### 5. 仅训练路由参数的隔离式微调在整个训练过程中，**主干 Transformer 的全部参数保持冻结**，仅优化以下路由相关参数： - 原始路由器权重 W_r^l ； - 注意力路由器权重 A_l ； - 门控网络权重 G_l 。这种设置将路由机制隔离为唯一的变量，使得任何性能或行为变化均可归因于注意力信号的引入与路由决策的改变，从而精确验证 AAR 的因果效应。 Q4: 论文做了哪些实验？论文中的实验体系围绕**性能验证**、**机制解析**与**深度探针**三个层面展开，具体包括以下七个方面： ### 1. 实验设置与基线构造 - **模型**：主实验基于 OLMoE-1B-7B-SFT（16 层 MoE，64 专家，top- k=8 ），并迁移验证至 Qwen1.5-MoE-A2.7B-Chat 与 Qwen3.6-MoE（40 层，3B active / 35B total）。 - **训练协议**：仅微调路由参数（原始路由器 W_r^l 、注意力路由器 A_l 、门控 G_l ），主干 Transformer 完全冻结；优化器为 AdamW，batch size 1024，序列长度 512，学习率 1 × 10^(-4) ，warmup 比例 0.03。 - **数据**：使用 Tulu3 的 25% 子采样（约 220K 样本）进行监督微调。 - **基线**：对比原始预训练模型、仅训练路由参数的 routing-only SFT 基线，以隔离注意力信号带来的因果效应。 - **评估基准**：GSM8K、BBH、MMLU、MATH-500、HumanEval、IFEval。 ### 2. 主性能对比实验在 OLMoE 与 Qwen1.5 上系统比较标准路由、routing-only SFT 与多种 AAR 配置： - **AARS**（时域+频域拼接）：应用于 OLMoE 的层集合 S = 9, dots, 15 ，窗口 M=40 时，GSM8K 达到 55.36 ± 0.74 ，相较 routing-only SFT（ 51.99 ）提升 **+3.37 pp**。 - **AART**（仅时域）： M=20 时 GSM8K 为 54.57 ± 1.10 （+2.58 pp）。 - 对 Qwen1.5（层集合 S_Q = 14, dots, 17 ），AAR 同样将 GSM8K 从 53.48 提升至 55.36 。 - 在 BBH、HumanEval、IFEval 上表现与基线持平，说明增益具有任务选择性。 ### 3. 跨基准与跨规模泛化实验验证数学推理增益是否局限于 GSM8K： - **MATH-500**：在 OLMoE 上，AAR（L9–15）将准确率从 14.00 ± 1.83 提升至 16.64 ± 0.91 （+2.64 pp）。 - **大模型迁移**：在 Qwen3.6-MoE（40 层）的中深层全注意力层 23, 27, 31 应用 AAR（ M=40 ），仅训练 12.8K 样本，即将 MATH-500 从 78.47 ± 1.30 提升至 79.60 ± 1.51 （+1.13 pp），证明增益可迁移至更强基线与更大规模。 ### 4. 注意力窗口大小的消融与信号结构分析 - **窗口大小消融**：对 AART（应用于所有层）测试 M ∈ 10, 20, 128, 512 。GSM8K 在 M=20 时达到峰值（ 54.10 ± 0.52 ）， M=10 与基线持平， M=128 与 M=512 则衰减。全拼接的 AARS 则在 M=40 时最优。 - **自相关分析**：通过计算注意力窗口 w_i^l(t) 的归一化自相关函数（ACF），发现无论窗口设定多大，信号的有效相关长度在约 20 个 token 后即衰减至 0.1 以下。该结构特性解释了为何中等窗口（20–40）即可捕获全部线性可预测信息，过大的窗口反而引入噪声。 ### 5. 内部机制：路由对注意力的级联重塑为解析 AAR 为何能改善推理，论文采集 GSM8K prompt 上的注意力权重进行内部对比： - **Attention Sink 放大**：AART 模型在各层对序列首 token 的绝对注意力质量显著高于基线；且该效应存在**严格的一层延迟**——当 AAR 应用于第 l 层时，第 l+1 层的 sink 才开始被放大。 - **机制解释**：第 l 层的路由改变通过专家加权和残差流修改了传入第 l+1 层的表示，从而间接重塑了该层注意力子模块的输出，无需直接更新任何注意力参数。这证明路由与注意力构成一个**耦合电路**。 ### 6. 外部行为：生成长度与错误答案发散性对 GSM8K 上 AART 与基线的答案长度进行分层统计： - **双错情况**：两模型均回答错误时，AART 的平均长度缩短 **−11.8%**，P99 长度缩短 **−14.9%**；冗长回答率（>1000 字符）从 5.63% 降至 4.29%。 - **双对情况**：均回答正确时，长度几乎不变（均值 −0.8%，中位数 0%）。 - **纠错情况**：被 AART 纠正的错误答案，其原始长度平均缩减 **−36.25%**，尾部（P99）缩减 **−85.8%**。 - 结论：AAR 抑制了错误推理链的**过度发散**，使模型在犯错时更早停止，而非均匀变短。 ### 7. 深度敏感性：检索–推理张力探测通过改变 AAR 的起始层，将其作为网络深度的功能探针： - **全层应用的副作用**：当 AART 应用于所有层时，MMLU 整体下降 −2.1 pp，但不同科目呈现分离效应——数学（Math）与计算机科学（Comp. Sci.）等推理科目保持或提升，而生物学（Biology）、历史（History）、哲学（Philosophy）等检索密集型科目显著恶化（最高 −4.5 pp）。 - **起始层扫描**：逐层改变 AAR 的起始深度后，检索型科目的准确率呈“钟形”恢复——起始层越早，损伤越大；起始层后移后性能回升。推理型科目的增益则不受起始深度影响，始终维持。 - **张力区域**：层 7–11 被识别为检索与推理张力的主要集中区。在此之后，检索所需结构已巩固，引入 AAR 不再破坏事实查找，同时保留数学增益。该实验将 AAR 转化为一种**可控的深度探针**，直接读取不同层注意力信号的功能角色。 Q5: 有什么可以进一步探索的点？基于论文的局限性与核心发现，以下几个方面值得进一步探索： **1. 与高效注意力实现的兼容性与工程优化** 当前 AAR 要求显式物化注意力权重，导致在 AAR 激活层无法使用 FlashAttention 等 IO 优化的注意力核函数，引入随序列长度和 AAR 层数线性增长的推理开销。未来工作可探索： - 设计无需完整物化注意力矩阵的近似提取方案（如低秩投影或采样估计）； - 开发支持注意力侧输出（side-output）的高效内核，使 AAR 的窗口提取与 FlashAttention 的前向计算融合，降低内存与延迟惩罚。 **2. 从嫁接式微调扩展到全生命周期训练** 现有实验均在 post-SFT 阶段将 AAR 嫁接至已训练的路由参数上，导致最优层范围受模型先前训练轨迹的约束，跨架构迁移时需手动重新搜索。可探索： - 在预训练或完整 SFT 阶段从头联合训练标准路由器与 AAR 模块，检验是否涌现出更稳定、跨模型一致的层敏感模式； - 研究 AAR 参数与主干网络协同训练时的梯度交互，验证其是否缓解早期层的事实检索干扰。 **3. 自动化层选择与动态深度分配** 论文揭示了 AAR 强烈的深度敏感性，但当前层集合 S 依赖人工网格搜索。未来可探索： - 基于可微架构搜索（DAS）或强化学习，自动为不同任务学习最优 AAR 层子集； - 设计输入自适应的动态门控，根据序列内容或任务类型实时决定是否在特定层启用 AAR，从而在单个模型内实现检索与推理的动态调和。 **4. 注意力窗口的特征变换与自适应窗口** 论文使用时域与 DFT 幅度谱作为固定特征映射。可进一步探索： - 可学习的特征映射 φ （如小型 MLP 或卷积核），替代手工设计的频域变换； - 任务自适应或内容自适应的窗口大小 M ，替代全局固定值，以匹配不同推理任务固有的注意力相关长度。 **5. 路由–注意力双向耦合的机理与反向利用** 论文证明路由改变可通过残差流重塑下一层注意力（sink 放大）。反向问题尚未被探索： - 是否可通过在损失函数中引入注意力结构正则项（如显式强化或抑制 sink），反向诱导路由器收敛至更优的专家组合？ - 该耦合电路在训练动态中的稳定性条件（如梯度传播、临界深度）有待理论刻画。 **6. 向非 MoE 架构与多模态场景迁移** AAR 的核心思想——用注意力信号指导计算路径分配——未必局限于 MoE 的路由器： - 在密集 Transformer 中，类似机制可用于指导动态深度（early exiting）或动态宽度（token-level adaptive computation）； - 在多模态 MoE（如视觉-语言模型）中，可将空间注意力或跨模态注意力权重引入路由，探究模态间的计算分配策略。 **7. 长上下文与流式生成中的行为** 论文使用中等长度窗口（ M=20 – 40 ），验证了局部注意力结构的信息量。在超长上下文或流式对话中： - 注意力 sink 与窗口特征的交互可能呈现新的时间动态，需检验 AAR 是否仍能稳定解码并抑制错误答案的发散； - 结合 sink 锚定机制，AAR 或可成为流式 KV cache 压缩与计算路径选择联合优化的接口。 **8. 检索–推理张力的显式解耦机制** 实验表明，早期层 AAR 损害事实检索而中后期层增强推理。未来可探索显式解耦架构： - 在浅层冻结或屏蔽 AAR，保护事实检索所需的静态注意力模式； - 设计具有显式”检索层”与”推理层”功能标签的混合路由策略，使模型在不同深度自动切换信息处理模式，而非依赖后验的层选择搜索。 Q6: 总结一下论文的主要内容这篇论文研究了**混合专家（MoE）语言模型中路由机制的上下文感知增强**。标准路由器仅依赖 token 的隐藏状态 h_i^l 进行专家选择，导致其只能获取与内容高度纠缠的有限上下文信号。为此，论文提出**注意力感知路由（Attention-Aware Routing, AAR）**，将滑动窗口注意力权重作为与隐藏状态解耦的显式序列结构信号引入路由决策。 ### 核心方法 AAR 在选定 MoE 层并行于标准路由器运行。对于当前 token，其提取最近 M 个位置的注意力权重，跨头平均后得到一维窗口 w_i^l ∈ R^M 。该窗口经固定特征映射 φ 增强：

w_i^l = φ(w_i^l) = [ w_i^l ,|, log(1 + |F(w_i^l)|) ]
其中 F 为离散傅里叶变换，提供时域与频域双视图。增强窗口通过线性注意力路由器映射为专家 logits：
g_i^(attn) = A_l^top w_i^l
最终通过门控机制 $α_i = σ(G_l
h_i^l ,|, w_i^l
) 将标准路由 logits g_i^(std)$ 与注意力路由 logits 插值融合：
g_i = β_0 odot g_i^(std) + B(α_i odot g_i^(attn))
**关键训练设定**：保持主干 Transformer 完全冻结，**仅训练路由相关参数**（原始路由器、注意力路由器及门控），从而将路由作为唯一变量，精确隔离因果效应。 ### 主要实验与发现 **1. 数学推理性能提升** - 在 OLMoE-1B-7B-SFT 上，AAR（层 S=9,dots,15 ， M=40 ）将 GSM8K 从 routing-only SFT 的 51.99 提升至 55.36 （+3.37 pp）。 - 在 MATH-500 上提升 +2.64 pp，且在 Qwen1.5-MoE 与更大规模的 Qwen3.6-MoE（40 层）上复现了类似增益，证明其跨基准、跨规模泛化能力。 **2. 路由与注意力的耦合电路** - 内部机制分析表明，第 l 层的路由改变会通过残差流传播，在**无需修改任何注意力参数**的情况下，放大第 l+1 层的 attention sink（对首 token 的注意力集中）。 - 外部生成行为上，AAR 显著缩短**错误答案**的长度（双错情况下均值 −11.8%，P99 −14.9%，冗长回答率 −23.9%），而**正确答案**长度几乎不变。这表明 AAR 减少了错误推理链的过度发散。 **3. 深度敏感性与检索–推理张力** - AAR 并非在所有层都有益：应用于全层时，MMLU 整体下降，且损害集中在检索密集型科目（如生物学、历史、哲学），而推理型科目（数学、计算机科学）保持或提升。 - 通过扫描 AAR 起始层发现，早期层对 AAR 引入的扰动极为敏感（事实检索恶化），而中深层（约 7–11 层之后）引入 AAR 既能保持数学推理增益，又不破坏已巩固的检索结构。 - 这一深度敏感性使**层选择性 AAR 成为可控探针**，揭示了不同深度注意力信号在检索与推理中的功能分化。 ### 结论该研究表明，**路由不仅仅是专家调度机制，而是与注意力、生成行为紧密耦合的关键自由度**。通过将路由决策从孤立的、仅依赖隐藏状态的局部选择，拓展为受模型自身注意力模式指导的上下文感知过程，AAR 在不更新主干参数的情况下即可改善数学推理、重塑内部注意力分布、并规范生成长度。论文最终强调了在 MoE 架构中显式利用注意力–路由交互的潜力与必要性。 Q7: 想要进一步了解论文以上只是了解一篇论文的几个基本FAQ。如果你还想与Kimi进一步讨论该论文，请点击这里为你跳转Kimi AI网页版，并启动一个与该论文相关的新会话。

Authors: Despoina Kosmopoulou, Anastasios Tsetsilas, Efthymios Georgiou, Giannis Karamanolakis, Swastik Roy, Alexandros Potamianos

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.20974.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.20974

Published: 2026-09-22T01:52:31.318Z

3. CaLR: Causal Latent Revision for Robust Diffusion Reasoning

Abstract:Autoregressive (AR) models suffer from local greediness, while diffusion language models (DLMs) often lack the strict causal structure required for reasoning. To combine the advantages and overcome the drawbacks of the dual, we propose Causal Latent Revision (CaLR), a framework that reformulates reasoning as constrained latent optimization. By adopting a causal topology matrix (CTM) from an expert model and implicit differentiation, CaLR performs gradient-guided ``thought revision” to enforce logical consistency, enabling dynamic self-correction of intermediate steps during parallel generation. Empirically, CaLR achieves SOTA DLM performance on complex benchmarks, surpassing strong AR baselines and demonstrating superior robustness in constrained tasks like Sudoku.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.20981 (HTTP 429)

Authors: Wei Cai, Jian Zhao, Yuchen Yuan, Xuelong Li

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.20981.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.20981

Published: 2026-09-22T01:52:31.318Z

4. LoRA Enhanced Contrastive Learning with SAS Vision Transformers

Abstract:Automatic target recognition (ATR) with synthetic aperture sonar (SAS) supports advanced naval capabilities, but deep learning is constrained by scarce target imagery, background clutter, and human-in-the-loop assessment. We adapt DINOv3 Vision Transformer (ViT) models to underwater SAS ATR using a three-stage parameter-efficient framework. Stage 1 uses Low-Rank Adaptation (LoRA) while freezing the ViT backbone, bridging the gap between natural-image pretraining and underwater acoustic propagation. Stage 2 uses hard-negative mining to strengthen the decision boundary against acoustic mimics, including rocks and sediment formations resembling man-made targets. Stage 3 uses Supervised Contrastive Learning (SupCon) to separate target and clutter representations. We evaluate at-sea SAS data using a mission-level geographic split, compare all arms at 85 percent test recall, and repeat each comparison over three random seeds. LoRA accounts for the primary effect, increasing area under the precision-recall curve (AUPRC) from 0.300 to 0.679 +/- 0.027 using the same frozen backbone. Rank 4 achieves this result while training only 0.26 percent of weights. Neither refinement stage exceeds its matched control: hard-negative mining changes AUPRC by -0.0045 +/- 0.0119 versus an equal-size random curriculum, and SupCon changes AUPRC by +0.0002 +/- 0.0096 versus the preceding stage. These null results indicate that mining occurred on data the encoder had already fit and that supervised stages had already imposed most target-clutter geometry. One efficient adaptation stage is sufficient; stacked refinement is not.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.21061 (HTTP 429)

Authors: Dan Zimmerman, Frank E. Bobe III, Amelia L. McCormack, Matthew Cook, Gregory D. Vetaw

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.21061.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.21061

Published: 2026-09-22T01:52:31.318Z

Abstract:In this work, we examine the topology of information flow patterns within attention graphs to effectively distinguish hallucinated from non-hallucinated responses. We analyze the Forman-Ricci curvature to identify structural patterns indicating information bottlenecks in attention graphs. We then introduce a method that captures both semi-local and global information-flow characteristics of attention heads associated with hallucinated responses. We evaluate our approach extensively across several LLMs and established benchmarks. Empirical results demonstrate that our proposed single-pass approach provides consistent improvements over existing attention-based and multi-response baselines across two hallucination-detection benchmarks, while achieving competitive performance across diverse LLM architectures. Further analysis reveals that impaired context sharing among tokens during causal generation is strongly associated with hallucination occurrences in LLMs. In particular, hallucinated responses are consistently characterized by an over-reliance on self-attention, diffused context retrieval from earlier tokens, or information over-squashing, especially in the final transformer layer.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.21096 (HTTP 429)

Authors: Amir Jalilifard, Anderson Rocha, Eric Wong, Marcos Medeiros Raimundo

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.21096.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.21096

Published: 2026-09-22T01:52:31.318Z

6. Decoupling Internal Representational Changes and Causal Importance in Fine-Tuned Large Language Models

Abstract:Fine-tuning has emerged as a widely adopted approach for adapting LLMs to a variety of downstream tasks. However, how it reshapes their internal mechanisms remains poorly understood. To address this, we investigate how fine-tuning alters internal representations in LLMs, including attention patterns and layer-wise activations, and examine whether these changes are linked to task-relevant components identified by EAP (e.g., attention heads and logit-level activations) that drive task performance. We find that EAP-identified components are concentrated within specific layers, indicating a degree of functional localisation in how models internalise task-specific behavior. Notably, the distribution of these components across layers is largely uncorrelated with the layers undergoing the most substantial representational changes during fine-tuning. Furthermore, we observe that overlap in EAP-identified components across tasks does not translate into cross-task performance transfer if the tasks are different in nature (e.g. classification vs. generative tasks). More specifically, fine-tuning on one task can lead to a degradation of performance on another when the two tasks exhibit a high degree of overlap in their EAP-identified components.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.21113 (HTTP 429)

Authors: Lingfang Li, Procheta Sen, Shubham Das, Danushka Bollegala

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.21113.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.21113

Published: 2026-09-22T01:52:31.318Z

7. TinyCeNN-LM: Quality-Gated Conversion of Pretrained Attention with CeNN-Inspired Cellular-Recurrent Layers

Abstract:Replacing attention in a pretrained language model is a compatibility problem: a plausible substitute may alter representations expected by later layers. TinyCeNN-LM introduces a \emph{quality-gated post-training conversion} framework using CeNN-inspired cellular-recurrent layers with bounded local processing, compact recurrent memory, routing, fusion, and accept-or-rollback validation. Three implementations are studied: Integrated Memory, MemoryFusion, and PDelta3-GDN2-CLVR+Local32. Strict PDelta3 conversion accepts a layer only when representation and NLL criteria pass fixed thresholds. On SmolLM2-135M, layers 0-2 are accepted with cumulative $\Delta\mathrm{NLL}=+0.01209$, while layer 3 is rejected despite acceptable NLL because representation fidelity fails. On Qwen3.5-0.8B, full-attention layers 3, 7, and 11 are accepted with final $\Delta\mathrm{NLL}=+0.02073$. Integrated Memory keeps perplexity within $-0.07\%$ to $+0.93\%$ while reducing total cache by up to $6.01\%$. A sampled 200-item downstream sanity check gives $28.5\%$—$32.0\%$ overall accuracy for converted Qwen releases. The results support conservative, quality-gated structural conversion rather than universal attention replacement or speedup.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.21139 (HTTP 429)

Authors: Kabeh Mohsenzadegan, Vahid Tavakkoli, Kyandoghere Kyamakya

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.21139.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.21139

Published: 2026-09-22T01:52:31.318Z

8. Clinician-Grounded Quality Assurance for AI-Assisted Psychiatric Intake

Abstract:Before patients can use AI-assisted psychiatric intake systems, health systems need practical ways to routinely evaluate these tools against their clinical standards for quality assurance. Because clinicians may use different intake styles, evaluation for this task must (1) support comparison across interviewing approaches, (2) minimize clinician burden, and (3) measure clinically relevant performance for health systems deploying these technologies. We present a clinician-grounded evaluation platform built around a memory-augmented patient simulator for open-ended AI interviewing, InterviewPlayground. We created interactive patients using InterviewPlayground with our expert-authored vignettes, constructed a simulated intake platform for the interviews, and designed evaluation modalities relevant to intake. In a pilot of 6 clinicians in a 25-minute assessment compared to a GPT-based LLM intake interviewer, the LLM recovered more of the clinically relevant items embedded in the patient vignettes (88.0% vs. 38.9%), but made more clinical inferences not based on the interview (56.8% vs. 27.8%), and characterized identified safety concerns less often (33.3% vs. 66.7%), setting the stage for deployed quality assurance for this task.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.21149 (HTTP 429)

Authors: King Shi, Amanda Li, Jonathan Ivey, Synthia Qia Wang, Guan Gui, Hyunseo Kim, Peter Zandi, Jason Straub, Jacob Taylor, Ananya Joshi

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.21149.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.21149

Published: 2026-09-22T01:52:31.318Z

9. Can Agents Design Better Chips with a Higher Level Abstraction?

Abstract:Large Language Model (LLM) agents are increasingly being explored for chip design, but most existing approaches operate directly at RTL. We ask whether agents can design better chips by leveraging higher-level abstractions. We compare Direct RTL Design, Agent-based HLS Design, Post-Compiler HLS Refinement, and Post-HLS RTL Refinement, and combine Agent-based HLS Design with Post-HLS RTL Refinement as Agent-based HLS with RTL Refinement (AHRR). We use FPGAs as a practical, easy-to-deploy platform for end-to-end evaluation, but note that the design-flow tradeoffs we study are largely independent of the target technology. Across a diverse 11-tasks benchmark suite, AHRR achieves a 2.6$\times$ geometric-mean speedup over Direct RTL Design across our benchmark suite. Case studies show that HLS distills design knowledge into abstractions that agents can leverage, while RTL refinement recovers lower-level optimization opportunities. Together, these results make AHRR a promising workflow for agentic chip design. The code and evaluation artifacts are available at this https URL.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

Q: 这篇论文试图解决什么问题？这篇论文试图解决的核心问题是：\*\*基于大型语言模型（LLM）的智能体在芯片设计任务中，是否能通过更高层次的抽象（如高级综合/HLS）获得比直接寄存器传输级（RTL）设计更优的设计质量，以及如何实现这种优势。\*\* 具体而言，论文聚焦于以下几个层面的研究问题： - \*\*抽象层次的选择困境\*\*：现有智能体芯片设计工作大多直接在RTL层面操作，赋予智能体完全的周期级控制权，但同时也要求其实现所有微架构细节。论文质疑这种低层次抽象是否为智能体的最优接口，并探索将抽象层次提升至C/C++（通过HLS）或领域特定编译器是否更有利。 - \*\*不同设计范式的系统性比较\*\*：论文识别出四种可能的智能体芯片设计流程： 1. \*\*直接RTL设计\*\*（Direct RTL Design） 2. \*\*基于智能体的HLS设计\*\*（Agent-based HLS Design） 3. \*\*HLS后RTL优化\*\*（Post-HLS RTL Refinement） 4. \*\*编译器后HLS优化\*\*（Post-Compiler HLS Refinement）论文试图通过端到端评估，量化比较这些流程在性能、时序和资源利用等方面的差异。 - \*\*人机设计知识的有效结合\*\*：HLS工具与领域特定编译器内部蕴含了人类专家长期积累的设计知识（如调度、流水线、存储划分等）。论文探索智能体如何利用这些已被“蒸馏”到编译器中的知识，同时在RTL层面进行精细化调整，以弥补高层抽象可能丢失的低层优化机会。 - \*\*提出并验证AHRR流程\*\*：基于上述比较，论文提出 \*\*Agent-based HLS with RTL Refinement (AHRR)\*\* 流程，即智能体先在高层次（HLS C/C++）进行架构探索，再对HLS生成的RTL进行精细化优化。论文试图验证该组合流程能否在基准测试集上显著优于直接RTL设计。 Q2: 有哪些相关研究？根据论文第2节（Background）及参考文献列表，相关研究可分为以下几个方向： ### 1. LLM智能体用于RTL设计与验证该方向探索智能体直接在寄存器传输级（RTL）生成、调试与验证硬件： - \*\*VerilogCoder\*\*

10
：基于图的规划与AST波形追踪的自主Verilog编码智能体。 - **VerilogEval**
13
：针对大语言模型Verilog代码生成能力的评估基准。 - **VeriGen**
22
：面向Verilog代码生成的大语言模型。 - **MAGE**
34
：多智能体引擎，用于自动化RTL代码生成。 - **FVDebug**
1
：基于LLM的形式化验证失败调试助手。 - **Comprehensive Verilog Design Problems**
17
：下一代用于评估大语言模型与智能体在RTL设计和验证方面的基准数据集。 ### 2. LLM智能体用于HLS设计与优化近期研究开始将智能体应用于高级综合（HLS）层面，包括pragma插入、代码重构与设计空间探索： - **C2HLSC**

6
：探索LLM能否弥合软件到硬件的设计鸿沟。 - **LIFT**
19
：基于GNN监督微调的LLM驱动HLS pragma插入方法。 - **LLM-DSE**
23
：利用LLM智能体搜索加速器参数。 - **HLSPilot**
25
：基于LLM的高级综合工具。 - **HLSRewriter**
26
：利用LLM对C/C++代码进行高效重构与优化以适配HLS。 - **AgRefactor**
35
：自进化的智能体工作流，用于HLS兼容性与性能优化。 ### 3. 领域特定HLS编译器与加速器框架这些工作通过更高层次的抽象（如参数化架构模板）生成优化硬件，为智能体提供编译器辅助的起点： - **AutoSA**

24
：基于多面体模型的高性能脉动阵列编译器。 - **StreamHLS**
2
：面向自动数据流加速的高级综合框架。 - **SODA**
5
：面向Stencil计算的数据流架构优化。 - **Allo**
4
：可组合加速器设计的编程模型。 - **ScaleHLS**
30
：基于多级中间表示的可扩展HLS框架。 - **HIDA**
31
：层次化数据流编译器。 - **StreamTensor**
29
：面向LLM的数据流加速器张量流优化。 - **Holistic Optimization Framework**
18
：FPGA加速器的整体优化框架。 ### 4. 通用HLS方法论 - **FPGA HLS Today**

7
与 **High-Level Synthesis for FPGAs**
8
：综述了HLS从原型到部署的发展历程、成功经验与挑战，为本文的HLS流程提供了方法论基础。 ### 5. 智能体框架与自动化算法发现 - **pi**

14
：开源AI智能体工具包，提供统一LLM API与编码智能体循环，本文直接采用其作为实验 harness。 - **AlphaEvolve**
15
：面向科学与算法发现的编码智能体。 - **Reflexion**
21
：具备言语强化学习的语言智能体。 - **ReAct**
28
：将推理与行动协同的语言模型智能体框架。 ### 6. 应用背景与芯片设计智能体范式 - **从Pragmas到Partners**

33
：探讨智能体HLS的共生演进，与本文提出的AHRR范式观点相近。 - **OpenAI Jalapeño**
16
：OpenAI于2026年8月宣布的芯片设计项目，代表了业界对AI智能体+HLS设计方法的采纳趋势。 Q3: 论文如何解决这个问题？论文通过**系统性地构建对比框架、提出分层优化流程，并在真实FPGA平台上进行端到端实验验证**来解决该问题。具体解决方案可分为以下四个层面： —- ### 1. 建立四种设计流程的统一比较框架为严格隔离“抽象层次”对智能体设计质量的影响，论文构建了覆盖从高层到低层的完整流程谱系，并在统一条件下进行评估： - **(i) Direct RTL Design**：智能体直接从任务规范生成Verilog。 - **(ii) Agent-based HLS Design**：智能体生成HLS C/C++，由Vitis HLS编译生成RTL。 - **(iii) Post-HLS RTL Refinement**：智能体在HLS生成RTL的基础上进一步优化Verilog。 - **(iv) Post-Compiler HLS Refinement**：由领域特定编译器（如AutoSA、StreamHLS）先生成HLS C++，再由智能体进行精炼。所有流程均采用相同的**迭代式编码智能体循环**（基于开源工具 π

14
），配备相同的任务包（含参考C++实现、测试激励与接口契约），并通过**隐藏测试台**验证正确性。评估指标统一为：
执行时间 = 模拟周期数 × max(布线后时钟周期, 目标时钟周期)
以及BRAM、DSP、FF、LUT等资源利用率。 —- ### 2. 提出AHRR分层优化流程论文将流程 (ii) 与 (iii) 结合，提出 **Agent-based HLS with RTL Refinement (AHRR)**。该流程分为两个阶段，形成“粗调+精修”的分工： - **阶段一：Agent-based HLS Design** 智能体在C/C++层面通过循环变换、pragma插入（如`#pragma HLS PIPELINE II=1`、`#pragma HLS ARRAY_PARTITION`）快速表达并行性与存储划分意图。HLS工具（如Vitis HLS）负责将其转换为正确的流水线、调度和RTL实现。这使得智能体无需显式构造复杂的周期级控制逻辑。 - **阶段二：Post-HLS RTL Refinement** 智能体检查HLS工具生成的RTL，识别仅在门级/结构级才暴露的优化机会，如： - 将常量乘法器替换为查找表（LUT-based lookup table）； - 重构算术表达式以消除关键路径上的减法器（如将 a+b-∫ersection < 2·∫ersection 改写为 a+b < 3·∫ersection ）； - 调整存储绑定（`bind_storage`）与FIFO深度； - 根据操作数符号位提前进行min/max选择，从而减少乘法器数量。 —- ### 3. 引入领域特定编译器作为更强的初始起点为测试智能体在“已编码人类专家知识”的基础上的改进能力，论文在流程 (iv) 中引入了**领域特定HLS编译器**： - 使用 **AutoSA**

24
生成脉动阵列GEMM设计的初始HLS代码； - 使用 **StreamHLS**
2
生成数据流风格的卷积/GEMM设计。智能体在此基础上进行**Post-Compiler HLS Refinement**，实验表明即使在编译器已优化的设计上，智能体仍能通过调整存储绑定、重构控制逻辑和PE数据通路实现额外 1.2× – 3.3× 的加速。这验证了智能体能够**补足编译器启发式策略留下的空白**。 —- ### 4. 端到端实验验证与量化评估论文在包含11个任务的基准套件上（涵盖LLM推理、注意力机制、GEMM、机器人感知、内存管理等）进行实验，使用 **Gemini 3.1 Pro** 与 **GPT-5.6-sol** 作为底层模型，目标平台为 **AMD Alveo U55C**。核心结果包括： - **Agent-based HLS Design** 相比 **Direct RTL Design** 实现 2.31× 的几何平均加速； - **AHRR**（HLS + RTL Refinement）将加速比进一步提升至 ** 2.62× **； - **Post-HLS RTL Refinement** 单独带来 1.17× 的几何平均性能提升； - 在小规模内核（如ROB1、ROB3）上，智能体通过RTL重写可获得 2× – 4× 的额外加速。案例研究进一步揭示：HLS抽象层将“256路并行性”的实现压缩为几条pragma，而直接RTL设计在相同智能体下仅能达成32/8路并行；同时，较弱的模型通过HLS可获得接近强模型的表现（相对性能从直接RTL的 0.40× 提升至 0.95× ），证明**高层抽象显著降低了设计优化的认知负担**。 —- ### 5. 总结：AHRR的核心优势论文最终论证，AHRR之所以有效，是因为它实现了**机器智能与人类设计知识在工具中的互补**： - **HLS工具与领域编译器**将人类专家的调度、流水线、存储优化知识“蒸馏”为可复用的编译策略，使智能体避免在RTL层面重复造轮子； - **RTL精细化**则赋予智能体对时序、面积和门级结构的精确控制权，修复编译器启发式无法覆盖的特例低效。这一组合使智能体既能利用高层抽象的广度进行架构探索，又能通过低层精细化实现案例特定的极限优化，从而系统性解决“纯RTL设计难以兼顾效率与开发复杂度”的问题。 Q4: 论文做了哪些实验？论文通过多组端到端实验，系统比较了不同抽象层次下LLM智能体的芯片设计能力。实验设计、基准测试与主要结果如下： —- ### 1. 实验平台与配置 - **目标设备**：AMD Alveo U55C FPGA（`xcu55c-fsvh2892-2L-e`） - **工具链**：Vitis 2025.2，out-of-context 模式（不含shell） - **智能体模型**：Gemini 3.1 Pro（high reasoning effort）与 GPT-5.6-sol（extra-high effort） - **智能体框架**：基于开源工具 π

14
的标准迭代编码循环，每轮最多60分钟，最多5轮，总超时10小时 - **评估超时**：RTL仿真75分钟，综合与布局布线90分钟 - **正确性验证**：公开测试台提供单一激励；隐藏测试台覆盖分支覆盖率100%的规格派生边界情况，结合C仿真与RTL仿真验证 - **有效性判定**：通过隐藏测试台、成功布线、且资源利用率不超过上限（BRAM、DSP、FF、LUT、URAM各60%） - **执行时间计算**：
执行时间 = 模拟周期数 × max(布线后时钟周期, 目标时钟周期)
目标时钟周期设定为 3.33,ns 。 —- ### 2. 基准测试套件实验采用11个任务的基准集，涵盖LLM推理、矩阵乘法、存储管理与机器人感知等领域： | 任务 | 类别 | 接口 | 算术类型 | 动态维度 | 缩放处理 | |:—-|:—-|:—-|:—-|:—-|:—-| | LLM1 | LLM decode attention + KV cache | m_axi | fixed | seq_len | 无 | | LLM2 | RMSNorm + quantize + INT8 linear | m_axi | mixed | seq_len | 缩放至12 tokens | | GEM1 | INT8 GEMM (3/12 AXI变体) | m_axi | fixed | 无 | 无 | | MEM1 | BM25检索排序 | ap_memory | fixed | 无 | 无 | | MEM2 | Paged-KV内存准备 | ap_memory | fixed | seq_len | 无 | | MEM3 | LServe风格页选择 | ap_memory | fixed | 无 | 无 | | ROB1 | 贪婪非极大值抑制 | ap_memory | fixed | num_boxes | 无 | | ROB2 | PointPillars柱特征编码 | ap_memory | fixed | n_points | 粗化至0.25m柱 | | ROB3 | 立体块匹配SAD | ap_memory | fixed | 无 | 图像宽高缩放 | | SHL1 | StreamHLS生成的GEMM | ap_memory | float32 | 无 | 无 | | SHL2 | StreamHLS生成的depthwise卷积 | ap_memory | float | 无 | 无 | —- ### 3. 四大设计流程对比实验（第4节） #### 3.1 流程 (i) vs. 流程 (ii)：直接RTL设计与基于智能体的HLS设计为隔离抽象层次本身的影响，同一智能体在相同任务规范下分别直接生成Verilog或HLS C/C++。关键发现包括： - **基于智能体的HLS设计**在15组有效对比中达到相对于直接RTL设计的 ** 2.31× 几何平均加速** - HLS的优势主要体现在大规模并行 kernel 上：例如ROB1（NMS）通过两条pragma（`ARRAY_PARTITION complete` 与 `PIPELINE II=1`）实现256路并行，而直接RTL设计仅实现32/8路并行，周期数从 127K 对比 1.86M - 直接RTL设计仅在小规模寄存器型kernel（ROB1、ROB3）上因可直接操控寄存器而取得 1.78× 和 1.85× 的优势 #### 3.2 流程 (iii)：HLS后RTL精细化（Post-HLS RTL Refinement）以流程 (ii) 生成的最佳Verilog作为起点，让智能体继续在同一任务上直接优化RTL： - 在16个有效案例中，Post-HLS RTL Refinement 相对其HLS生成起点实现 ** 1.17× 几何平均性能提升** - **增量优化**：在LLM1、MEM1/2/3、ROB2上，智能体通过局部编辑（通常周期变化 <1.15× ）逐步改进 - \*\*重写优化\*\*：在ROB1与ROB3上，智能体完全重写实现，获得 2× – 4× 加速 - \*\*资源优化\*\*：在LLM2与MEM3上，GPT与Gemini通过少量RTL编辑显著降低资源占用 #### 3.3 流程 (iv)：编译器后HLS精细化（Post-Compiler HLS Refinement）以领域特定编译器生成的HLS C++为起点，测试智能体在“已优化基线”上的改进能力： - \*\*SHL2\*\*（StreamHLS生成的depthwise卷积）：Gemini提升 3.3× ，GPT提升 1.7× - \*\*GEM1\*\*（AutoSA生成的脉动阵列）：Gemini提升 1.3× ，GPT提升 1.2× ，周期变化 <2% ，收益主要来自时序改善（如存储绑定改为 \`ram\_2p impl=lutram\`、FIFO深度调整、PE数据通路重构） - \*\*SHL1\*\*：无提升，因编译器生成设计的仿真时间已接近评估超时极限，留给迭代探索的空间极小 #### 3.4 综合评估：AHRR整体性能将流程 (ii) 与 (iii) 串联为 \*\*AHRR（Agent-based HLS with RTL Refinement）\*\*，并与直接RTL设计比较： - \*\*AHRR实现 2.62× 几何平均加速\*\*（基于13组有效对比，排除2个GEM1的3-AXI变体） - 该结果高于单独的Agent-based HLS Design（ 2.31× ），证明HLS层面的架构探索与RTL层面的精细化具有互补性 --- ### 4. 逐轮优化追踪实验（图3）对Post-HLS RTL Refinement的每一轮进行细粒度追踪，记录： - 累计最佳吞吐量/延迟改进 - 每轮LUT与DSP的节省结果显示： - 多数成功轮次为\*\*局部编辑\*\*，周期变化幅度有限 - 所有显著延迟改进（ >1.15× ）均来自**完全重写** - 不同模型（GPT vs. Gemini）在相同起点上展现出不同的优化轨迹与资源权衡策略 —- ### 5. 案例研究实验（第5节）通过三个典型案例，揭示抽象层次如何影响智能体的设计决策： #### 5.1 ROB1：高层抽象如何“蒸馏”设计知识（图4） - **HLS路径**：智能体仅用2行pragma（256路数组划分与流水线）表达意图，Vitis HLS自动生成含对数深度比较树的43914行Verilog，运行 127K 周期 - **RTL路径**：同一智能体手写378行Verilog，构建32路argmax与8路抑制扫描，运行 1.86M 周期 - **跨模型对比**：较弱模型通过HLS可达到较强模型HLS性能的 0.95× ，而直接RTL仅达 0.40× ，证明HLS显著降低了设计优化的认知门槛 #### 5.2 SHL2：智能体弥补编译器启发式缺陷（图5） - **Post-Compiler HLS精细化**：StreamHLS生成4个总计405K浮点数的本地数组与填充阶段；智能体将其重写为2行line buffer加 3×3 窗口（共1842浮点），周期从 540K 降至 209K （ -61% ），BRAM从224降至2 - **Post-HLS RTL精细化**：智能体发现RTL中两个乘法器的第二操作数为常量，将其替换为16项查找表，时钟周期从 4.20,ns 降至 3.82,ns （ -9% ），周期数不变 #### 5.3 MEM3与ROB1：RTL级时序与面积恢复（图6） - **MEM3**：Vitis HLS为32个元素各生成2个乘法器实现 max(q·min, q·max) ，共64个DSP。Gemini发现仅需根据 q 的符号位前置选择min/max，即可将乘法器减半至32个，LUT从4868降至2228（ 2.2× ），时钟周期从 3.14,ns 改善至 2.83,ns - **ROB1**：生成RTL将50% IoU条件实现为 union < 2·∫ersection ，关键路径含33位减法器。GPT将其改写为 a+b < 3·∫ersection ，消除减法器，时钟周期从 3.60,ns 降至 3.31,ns ，LUT减少4% —- ### 6. 实验结论支撑上述实验共同支撑了论文的核心主张：AHRR流程通过**高层抽象获取架构收益**与**低层精细化获取实现收益**的互补分工，使智能体能够系统性超越纯RTL设计方法；同时证明即使在领域特定编译器优化后的设计上，智能体仍具备识别并修复残余低效的能力。 Q5: 有什么可以进一步探索的点？基于论文的实验结果与方法论局限，以下方向值得进一步探索： —- ### 1. 向ASIC与非FPGA平台的迁移验证论文以FPGA为评估平台，利用Vitis HLS与Vivado工具链验证了AHRR的有效性。然而，HLS与RTL refinement的设计流程权衡在多大程度上独立于目标技术仍需检验： - **ASIC流程中的AHRR**：商用ASIC HLS工具（如Cadence Stratus、Siemens Catapult）与FPGA HLS的调度、存储综合、时钟约束机制存在差异。智能体在ASIC语境下是否能保持 2.6× 量级的优势，以及是否需要针对标准单元库、物理效应（如线延迟、时钟树综合）调整RTL refinement策略，仍是开放问题。 - **后摩尔技术节点**：在先进工艺下，物理设计（布局、布线、信号完整性）对RTL微架构极度敏感。需要验证Post-HLS RTL Refinement能否有效修复HLS生成代码在物理实现层面的时序违例。 —- ### 2. 更高效的评估与反馈机制实验表明，仿真与综合时间构成了智能体迭代优化的关键瓶颈： - **快速性能估计**：论文指出，SHL1因编译器生成设计的仿真时间接近超时极限，导致智能体缺乏迭代空间。未来可构建混合评估机制：在优化阶段使用HLS延迟估计（latency estimation）与静态资源模型提供即时反馈，仅将完整的RTL仿真与布局布线保留用于最终验证。 - **增量式验证**：当前每轮修改均需重新运行完整仿真。开发面向智能体的增量式形式化等价性检查或差异测试（differential testing）框架，可显著压缩验证开销。 —- ### 3. 中间表示（IR）与多抽象层协同论文比较了C++/HLS与RTL两个端点，但编译器基础设施中的中间表示可能提供新的优化空间： - **MLIR/CIRCT层级的智能体优化**：在HLS C++与RTL之间存在多级IR（如LLVM IR、FIRRTL）。智能体在IR层面操作，既可避免HLS前端解析的复杂性，又能比RTL更紧凑地表达架构变换（如存储合并、操作数重排序）。 - **多智能体分层协作**：将HLS Agent与RTL Agent解耦为专门化角色。例如，HLS Agent专注于循环变换与并行度探索，RTL Agent专注于门级时序与面积修复，二者通过标准化接口（如优化日志、约束文件）协作，而非由同一智能体串行执行。 —- ### 4. 编译器与智能体的深度协同当前Post-Compiler HLS Refinement采用“编译器生成→智能体编辑”的松耦合模式： - **智能体驱动的编译器参数搜索**：除编辑生成的C++代码外，智能体可直接控制编译器的调度启发式、分块（tiling）参数与存储绑定策略。这要求将AutoSA/StreamHLS等工具暴露为可微分或可通过自然语言/代码配置的环境。 - **编译器反馈引导的智能体学习**：将编译器的依赖分析、多面体模型（polyhedral model）结果作为智能体的上下文，帮助其理解为何特定循环变换合法或非法，减少无效探索。 —- ### 5. 验证、安全性与可解释性 - **形式化方法集成**：论文依赖隐藏测试台进行正确性验证，但测试台覆盖率（100%分支覆盖）仍无法保证无漏洞。将智能体与形式化验证工具（如JasperGold、SymbiYosys）结合，使Agent在收到反例（counter-example）后进行调试与修复，是提升可信度的关键。 - **可解释的优化决策**：当前智能体的编辑（如为何将乘法器替换为查找表）缺乏可解释性。构建能输出自然语言优化理由的Agent，有助于人类设计师审计与信任自动化流程。 —- ### 6. 模型专业化与知识沉淀 - **硬件专用模型的微调**：论文使用通用大模型（Gemini、GPT）。利用AHRR过程中积累的大量（规范, HLS代码, RTL补丁, 性能反馈）轨迹，对模型进行监督微调或强化学习，可训练出专门用于AHRR流程的硬件设计智能体。 - **设计模式库**：将成功案例中的pragma组合、RTL重写模式（如位宽截窄、常量传播）形式化为可复用的设计模式库，供跨任务检索与少样本提示（few-shot prompting）。 —- ### 7. 更大规模系统级设计 - **多核与内存层次**：当前11个任务多为单加速器kernel。AHRR能否扩展至包含多核互联、一致性协议、DDR/HBM控制器与片上网络（NoC）的SoC级设计，需要新的抽象接口（如Chisel/SpinalHDL或系统级DSL）。 - **软硬件协同设计**：将Host代码（驱动、调度）与FPGA/加速器设计纳入同一Agent优化循环，探索通信开销与计算并行度的联合优化。 —- ### 8. 多目标优化与成本意识设计 - **功耗与热约束**：论文以执行时间（周期 × 时钟周期）为主要目标，资源为次要约束。未来需显式引入动态功耗、静态功耗与热分布作为优化目标，尤其是在数据中心与边缘设备场景中。 - **经济性分析**：量化智能体迭代过程中的API调用成本、计算开销与最终芯片性能收益之间的投入产出比，建立成本感知的早期终止策略。 Q6: 总结一下论文的主要内容这篇论文围绕**大语言模型（LLM）智能体在芯片设计中应如何有效利用抽象层次**这一核心问题展开，系统比较了从寄存器传输级（RTL）到高级综合（HLS）及领域特定编译器的多种设计流程，并提出了一种分层协同的智能体设计范式。 ### 1. 研究背景与核心问题现有LLM智能体芯片设计研究大多直接在RTL层面操作，要求智能体显式实现周期级控制与微架构细节。论文质疑这一做法是否为最优选择，并提出根本性问题：**智能体能否通过更高层次的抽象（如HLS C/C++）设计出更优质的芯片？** 高层抽象（HLS及领域特定编译器）将人类专家的调度、流水线、存储优化等知识“蒸馏”为编译策略，可能降低智能体的认知负担；但同时，高层抽象也可能丢失低层优化机会。 ### 2. 四种设计流程与AHRR框架论文构建了四种统一评估的设计流程： - **(i) 直接RTL设计（Direct RTL Design）**：智能体直接从规范生成Verilog。 - **(ii) 基于智能体的HLS设计（Agent-based HLS Design）**：智能体编写HLS C/C++，由HLS工具生成RTL。 - **(iii) HLS后RTL精细化（Post-HLS RTL Refinement）**：智能体在HLS生成的RTL基础上进一步优化。 - **(iv) 编译器后HLS精细化（Post-Compiler HLS Refinement）**：由领域特定编译器（如AutoSA、StreamHLS）生成初始HLS，再由智能体精炼。论文将流程 (ii) 与 (iii) 相结合，提出 **AHRR（Agent-based HLS with RTL Refinement）** 工作流。该流程分为两个阶段： - **HLS阶段**：智能体通过C/C++代码与pragma快速探索并行架构，利用HLS工具完成调度与RTL生成； - **RTL精细化阶段**：智能体检查生成的RTL，针对时序、面积与门级结构进行案例特定的精确修复。 ### 3. 实验设计与基准实验在 **AMD Alveo U55C FPGA** 上进行，使用 **Vitis 2025.2** 工具链，底层模型包括 **Gemini 3.1 Pro** 与 **GPT-5.6-sol**，基于开源智能体框架 π 实现迭代式编码循环。基准测试集包含 **11个任务**，涵盖LLM推理（如KV Cache注意力、RMSNorm）、矩阵乘法（GEMM）、存储管理（如Paged-KV、BM25）及机器人感知（如PointPillars、NMS、立体匹配）。执行时间统一量化为：

执行时间 = 模拟周期数 × max(布线后时钟周期, 目标时钟周期)
正确性通过隐藏测试台验证，资源约束设定为各类型资源（BRAM、DSP、FF、LUT、URAM）不超过60%。 ### 4. 核心实验结果 - **Agent-based HLS Design** 相对于 **Direct RTL Design** 实现 ** 2.31× 几何平均加速**。HLS抽象使智能体能够通过简洁的pragma表达复杂并行结构（如256路数组划分与流水线），而直接RTL设计在相同智能体下往往只能实现更窄的数据通路。 - **AHRR**（HLS + Post-HLS RTL Refinement）将加速比进一步提升至 ** 2.62× **。 - **Post-HLS RTL Refinement** 单独贡献 ** 1.17× ** 的几何平均性能提升，典型优化包括：将常量乘法器替换为LUT查找表、重构关键路径上的算术表达式、前置min/max选择以减少乘法器用量，以及调整存储绑定。 - **Post-Compiler HLS Refinement** 实验表明，即使在AutoSA与StreamHLS等编译器已优化的设计上，智能体仍能实现 ** 1.2× – 3.3× ** 的额外加速，证明智能体能够弥补编译器启发式策略的残余低效。 ### 5. 关键洞察 - **抽象层次的知识蒸馏**：HLS工具将硬件设计知识封装为编译器能力，使较弱模型通过HLS即可接近较强模型的性能（相对性能从直接RTL的 0.40× 提升至 0.95× ），显著降低了高质量设计的门槛。 - **高层与低层的互补分工**：HLS负责“大步”架构变换（并行度、流水线、存储划分），RTL精细化负责“小步”案例级修复（时序路径、位宽优化、门级映射）。二者协同使智能体同时获得设计效率与实现精度。 ### 6. 结论论文提出并验证了 **AHRR** 作为智能体芯片设计的有前景范式。该工作流结合了HLS的正确性保障与架构抽象能力，以及智能体对RTL实现细节的推理与修复能力。研究进一步指出，AI智能体有望克服传统HLS推广中的两大障碍——既有RTL设计惯性与HLS生成代码难以人工修改的问题——从而加速高层次设计方法论的普及。论文预期，利用更高抽象层次工具（如HLS与领域特定编译器）来增强智能体设计能力，将成为芯片设计自动化领域的重要发展方向。 Q7: 想要进一步了解论文以上只是了解一篇论文的几个基本FAQ。如果你还想与Kimi进一步讨论该论文，请点击这里为你跳转Kimi AI网页版，并启动一个与该论文相关的新会话。

Authors: Zijian Ding, Yang Zou, Yizhou Sun, Jason Cong

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.21157.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.21157

Published: 2026-09-22T01:52:31.318Z

10. SpecOpt: Contact-Diff Reasoning for Agentic Molecule Optimization Toward Binding Specificity

Abstract:Off-target protein binding is a major source of adverse effects for small-molecule drugs, yet most structure-based molecular design methods focus on generating selective compounds de novo rather than improving the selectivity of existing, well- characterized drugs. We introduce specificity optimization (SpecOpt), a molecular design task that seeks constrained structural modifications to an existing compound that increase its binding preference for an intended target over known off-targets while preserving its structural identity and drug-like properties. To enable systematic evaluation, we construct a ChEMBL-derived benchmark from compound-target interaction data, identifying intended targets through curated drug-mechanism annotations and off- targets through measured activities. We then develop an agentic framework that docks each compound against its intended target and off-targets, compares the resulting poses through residue-aware atom-protein contacts, and provides these differential interactions to a large language model to propose targeted structural modifications. Candidates are retained only if they satisfy molecular similarity, ADMET, and target-off-target docking selectivity criteria. On 915 compounds, the agent improves the target- off-target binding gap for 84.8% of compounds, shifting the mean gap from -0.72 to +0.47 kcal/mol while maintaining a mean Tanimoto similarity of 0.72 to the starting compounds. Ablation studies identify residue-specific contact information as the critical optimization signal: replacing residue identities with binary contact indicators eliminates improvement on all 29 ablation compounds. These results establish SpecOpt as a distinct molecular design problem and demonstrate residue-aware differential interactions as an effective signal for improving the specificity of existing compounds.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.21165 (HTTP 429)

Authors: Thao Nguyen, Heng Ji

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.21165.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.21165

Published: 2026-09-22T01:52:31.318Z

Evaluation Domain Papers

1. RBS-Attention: Radius-Bounded Sparse Prefill for Long-Context Large Language Models

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.20971 (HTTP 429)

Authors: Chuxu Song, Jiuqi Wei, Zhencan Peng

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.20971.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.20971

Published: 2026-09-22T01:52:44.430Z

2. Attention-Aware Routing: Coupling Routing and Attention in MoEs

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

Authors: Despoina Kosmopoulou, Anastasios Tsetsilas, Efthymios Georgiou, Giannis Karamanolakis, Swastik Roy, Alexandros Potamianos

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.20974.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.20974

Published: 2026-09-22T01:52:44.430Z

3. CaLR: Causal Latent Revision for Robust Diffusion Reasoning

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.20981 (HTTP 429)

Authors: Wei Cai, Jian Zhao, Yuchen Yuan, Xuelong Li

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.20981.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.20981

Published: 2026-09-22T01:52:44.430Z

4. LoRA Enhanced Contrastive Learning with SAS Vision Transformers

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.21061 (HTTP 429)

Authors: Dan Zimmerman, Frank E. Bobe III, Amelia L. McCormack, Matthew Cook, Gregory D. Vetaw

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.21061.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.21061

Published: 2026-09-22T01:52:44.430Z

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.21096 (HTTP 429)

Authors: Amir Jalilifard, Anderson Rocha, Eric Wong, Marcos Medeiros Raimundo

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.21096.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.21096

Published: 2026-09-22T01:52:44.430Z

6. Decoupling Internal Representational Changes and Causal Importance in Fine-Tuned Large Language Models

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.21113 (HTTP 429)

Authors: Lingfang Li, Procheta Sen, Shubham Das, Danushka Bollegala

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.21113.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.21113

Published: 2026-09-22T01:52:44.430Z

7. TinyCeNN-LM: Quality-Gated Conversion of Pretrained Attention with CeNN-Inspired Cellular-Recurrent Layers

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.21139 (HTTP 429)

Authors: Kabeh Mohsenzadegan, Vahid Tavakkoli, Kyandoghere Kyamakya

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.21139.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.21139

Published: 2026-09-22T01:52:44.430Z

8. Clinician-Grounded Quality Assurance for AI-Assisted Psychiatric Intake

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.21149 (HTTP 429)

Authors: King Shi, Amanda Li, Jonathan Ivey, Synthia Qia Wang, Guan Gui, Hyunseo Kim, Peter Zandi, Jason Straub, Jacob Taylor, Ananya Joshi

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.21149.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.21149

Published: 2026-09-22T01:52:44.430Z

9. Can Agents Design Better Chips with a Higher Level Abstraction?

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

Authors: Zijian Ding, Yang Zou, Yizhou Sun, Jason Cong

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.21157.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.21157

Published: 2026-09-22T01:52:44.430Z

10. SpecOpt: Contact-Diff Reasoning for Agentic Molecule Optimization Toward Binding Specificity

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.21165 (HTTP 429)

Authors: Thao Nguyen, Heng Ji

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.21165.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.21165

Published: 2026-09-22T01:52:44.430Z

VLM Domain Papers

1. TAPe+ML: A Compact Structured Representation for Multi-Task Computer Vision

Abstract:We present TAPe+ML v3, a compact computer vision system based on TAPe (Theory of Active Perception), a structured representation that encodes relations among perceptual elements before recognition. Instead of operating directly on pixel tensors, the system uses a shared TAPe representation and a modular recognition architecture for image classification, object detection, and instance segmentation. TAPe+ML v3 combines background and contour processing, local object localization, prototype-based classification, and a coordinator for specialized submodels. Across the reported experiments, it uses fewer than 100,000 parameters. On COCO object detection, it obtains 84.7 mAP50 and 65.3 mAP50-95. On COCO instance segmentation, it obtains 80.7 mask mAP50 and 58.4 mask mAP50-95. In classification experiments, it reaches 92 percent validation accuracy on Imagenette under an identical-training comparison with a raw-pixel baseline, and 89.9 percent Top-1 accuracy on ImageNet-Real. We also evaluate compactness in video scene detection and adaptation under distribution shift in an industrial pilot. The results suggest that shifting part of the modeling burden from network parameters to a structured input representation can support compact multi-task vision systems with reduced data, memory, and compute requirements.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.20869 (HTTP 429)

Authors: Sergey Kurinov, Alexey Upatov

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.20869.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.20869

Published: 2026-09-22T01:52:57.262Z

2. MemeTAG: Keyword-Driven Meme Classification through Tag Embedding Reconstruction

Abstract:The proliferation of harmful internet memes poses a significant societal threat, yet their automated classification remains a formidable algorithmic challenge due to the nuanced, multimodal nature of their content. To address this, we introduce MemeTAG, a novel dual-objective framework that pioneers a keyword-aware approach to meme classification. Our core innovation is a two-part semantic guidance mechanism: first, we leverage a pretrained Vision-Language Model to generate a set of descriptive keywords, that capture the high-level semantics. Second, we introduce the Aggregated Tag Inference Network (ATIN), an attention-based module that distills these keywords into a single, rich semantic embedding. This embedding serves as a target for a novel auxiliary reconstruction loss, which compels the model to learn deeply aligned visual and textual features. This approach, combined with an efficient three-stage training strategy, establishes a new state-of-the-art on the HarMeme, Hateful Memes Challenge (HMC), and PrideMM datasets, decisively outperforming existing state-of-the-art methods.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.20962 (HTTP 429)

Authors: Akshit Sharma, Prashant W. Patil

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.20962.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.20962

Published: 2026-09-22T01:52:57.262Z

3. Image-Derived PM10 Estimation in Cattle Feedlot Using Machine Learning: Addressing Concentration Ranges Beyond Existing Digital Imaging Methods

Abstract:Affordable dust monitoring remains a pressing need for the cattle feedlot industry, yet camera-based PM estimation, despite its growing body of research in urban air quality settings, has not been evaluated under the extended concentration ranges characteristic of intensive livestock operations. This study developed an image-based approach using contrast panel features and machine learning to estimate PM10 concentrations in a commercial cattle feedlot, where hourly average PM10 ranged from 250 to 1,000 ug/m^-3 and instantaneous concentrations reached 5,000 to 20,000 ug/m^-3. Grayscale images were captured during the evening dust peak period, and features including panel contrast, black and white panel pixel values, and overall image brightness were extracted. The model also incorporated recent past values from preceding images and solar zenith angle as predictors. Among the candidate models evaluated, XGBoost achieved the highest predictive performance, with an R^2 of 0.792 and a median absolute error of 103 ug/m^-3. Feature importance analysis revealed that (a) panels positioned farthest from the camera contributed most strongly to predictions and (b) that black panel pixel values were more sensitive than white panel values to changes in PM10 concentration. Prediction accuracy during the sunset transition, which coincides with the onset of the feedlot evening dust peak, remains an area for further refinement. These findings demonstrate the feasibility of image-based PM10 estimation across PM concentration ranges substantially exceeding those reported in prior urban studies and provide practical guidelines for future deployment in feedlot environments.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.20975 (HTTP 429)

Authors: Sirapoom Peanusaha, Greg B. Ferguson, K. Jack Bush, Peiyang Li, Brent W. Auvermann

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.20975.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.20975

Published: 2026-09-22T01:52:57.262Z

4. Fragment-Aware Vision Transformers for Fresco-Fragment Style Classification

Abstract:Artistic style classification is usually studied on complete artworks, where models can exploit global composition, spatial organisation, and iconographic structure. In archaeological settings, however, artworks often survive only as fragmented remains, forcing recognition from incomplete, irregular, and context-limited visual evidence. We study fresco-fragment style classification using a progressive transformer-based framework. Starting from a ViT-B/16 baseline, we introduce foreground-guided masking to suppress background-only tokens, inpainting-based geometric regularisation to align irregular fragment supports with the ViT patch grid, and a supervised contrastive objective that operates on predictive distributions through a Kullback-Leibler similarity and consistently improves every branch. We combine the branches with a deliberately simple learnable logit ensemble. Experiments on CLEOPATRA and POMPAAF show that fragment-aware modelling improves over the standard ViT baseline, with the ensemble increasing accuracy from 0.604 to 0.656 and macro-F1 from 0.596 to 0.648 on CLEOPATRA, and outperforming the best single branch in four of six fragmentation settings on POMPAAF. We additionally evaluate a more complex graph-fusion variant and find that it matches the simple ensemble on POMPAAF while offering only a small, dataset-specific gain on CLEOPATRA, which does not justify its added complexity. Beyond these empirical gains, our contribution is twofold: a distribution-level contrastive objective that consistently sharpens single-branch recognition, and an interpretability analysis that verifies the models exploit genuine painted evidence, while quantifying that the inpainting-based branch draws part of its attribution from the synthesised surround.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.21012 (HTTP 429)

Authors: Sara Miketek, Biagio Barchielli, Nadeem Iqbal Kajla, Sinem Aslan

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.21012.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.21012

Published: 2026-09-22T01:52:57.262Z

5. MAGIC: Marginal-Guided Compression with Optimal Transport for Efficient Visual Document Retrieval

Abstract:Recent visual document retrieval (VDR) systems such as ColPali use multi-vector page embeddings, in which patch-level vectors enable fine-grained evidence matching but incur substantial index storage and MaxSim scoring overhead. Post-hoc merging offers a practical route to efficient VDR by reducing this cost without retraining the retriever, but its uniform reconstruction objectives are poorly aligned with the sparse, non-uniform patch usage induced by late-interaction retrieval. Under aggressive compression, this misalignment can preserve rarely used patches while concentrating retrieval activity on too few retained representatives. To address this misalignment, we propose Marginal-Guided Compression with Optimal Transport (MAGIC), a training-free post-hoc compressor for efficient retrieval with frozen multi-vector embeddings. MAGIC derives a MaxSim-induced compression surrogate and optimizes it through a two-marginal entropic optimal-transport formulation, where a retrieval-demand source marginal prioritizes high-use patches and a balanced target marginal regularizes retained-facet usage. Across ViDoRe benchmarks, keep ratios, and retrieval backbones, MAGIC consistently outperforms strong post-hoc compressors, with particularly large gains in the aggressive-compression regime; component ablations verify the complementary effects of its two marginals. We release the code at: this https URL.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.21018 (HTTP 429)

Authors: Xu Yuan, Hua Liu, Wenqi Fan, Qing Li

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.21018.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.21018

Published: 2026-09-22T01:52:57.262Z

6. MarsFM: Shading-Regularized Flow Matching for Martian Relief Estimation

Abstract:We present MarsFM, an image-conditioned latent flow-matching model for local Martian relief estimation from single-band HiRISE RED orthoimagery. The method combines a pretrained generative prior with stereo-derived geometric supervision and a differentiable Lunar—Lambert shading objective. Relief, normal, gradient, curvature, and ordinal terms constrain complementary aspects of terrain structure, while a positive-affine-invariant image comparison constrains rendered appearance. An evaluation comprising 2024 gathered patch records per integration-step count yields mean affine-aligned RMSE between 0.0935 and 0.0957 in normalized signed-log relief space for one to twenty Euler steps. These scores measure agreement with VAE-reconstructed references on positive-reference support. Their narrow range supports low-step inference under this protocol. Spatial, differential, and spectral diagnostics show broad terrain correspondence alongside smoothing, amplitude compression, and boundary mismatch. MarsFM provides a framework for combining learned terrain priors with image-based constraints; establishing improved physical terrain resolution requires matched baselines and independent high-resolution reference data. Data: this https URL code: this https URL.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.21095 (HTTP 429)

Authors: Marius F. R. Juston

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.21095.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.21095

Published: 2026-09-22T01:52:57.262Z

Abstract:This paper addresses the challenges of dynamic scene synthesis from sparse-view videos. Existing methods employ geometric priors, adaptive optimization, or density-control strategies to improve 4D Gaussian modeling under sparse observations. However, they cannot fundamentally resolve the ill-posed problem caused by insufficient observations and missing scene information. Moreover, sparse-view 4D Gaussian Splatting (4DGS) often suffers from poor geometric initialization: with only a few input views, COLMAP typically reconstructs sparse and incomplete point clouds, leaving large scene regions without sufficient Gaussian support and making them difficult to recover through subsequent optimization. To address these limitations, we propose a novel iterative refinement framework based on a video diffusion model to improve the completeness and consistency of dynamic 4D scenes. Specifically, we first estimate multi-view depth maps and fuse them into dense point clouds to provide more complete geometric initialization for a dynamic 4DGS representation. We then employ a pretrained video restoration model to refine sequences rendered along novel camera trajectories at different time steps. The restored sequences serve as pseudo-supervision to regularize and iteratively refine the 4DGS representation. Experiments on a widely used benchmark dataset demonstrate that our method substantially outperforms existing baselines, achieving nearly a 2 dB PSNR improvement over the previous best-performing method.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

Q: 这篇论文试图解决什么问题？该论文旨在解决\*\*稀疏视角动态场景重建\*\*中的核心挑战，具体可归纳为以下三个方面： --- ### 1. 观测不足导致的不适定问题（Ill-Posed Problem） - 现有稀疏视角4D高斯溅射（4DGS）方法依赖几何先验、自适应优化或密度控制策略来缓解稀疏观测的影响。 - 然而，这些方法仍受限于输入视角中可用的视觉证据，\*\*无法恢复未观测区域缺失的场景信息\*\*，导致动态视图合成在本质上仍是严重不适定的。 ### 2. 几何初始化稀疏且不完整 - 在仅有少量输入视角（如四个同步相机）时，传统COLMAP重建往往产生\*\*稀疏、不完整且偏向高纹理区域\*\*的点云。 - 由于高斯密化（densification）主要在现有基元周围扩展，初始化中缺失的大片区域在后续优化中\*\*难以恢复\*\*，造成重建空洞和几何覆盖不足。 ### 3. 生成先验在动态场景中的局限 - 虽然生成式先验（如扩散模型）为弥补缺失观测提供了可行路径，但现有方法多聚焦于\*\*静态场景重建\*\*。 - 针对动态的生成式方法（如Diffuman4D）则通常局限于\*\*特定领域\*\*（例如人体），缺乏对一般动态场景的适用性。 --- 为解决上述问题，论文提出 \*\*4DGS-Fixer\*\*，通过一个两阶段框架实现互补式改进： - \*\*第一阶段\*\*：利用多视角立体（MVS）深度估计与融合，提供\*\*密集点云初始化\*\*，弥补COLMAP的稀疏性； - \*\*第二阶段\*\*：基于预训练的视频修复扩散模型，对沿新视角轨迹渲染的视频进行\*\*迭代式伪监督精修\*\*，引入生成先验以恢复弱观测区域的信息，从而提升动态场景重建的完整性与时空一致性。 Q2: 有哪些相关研究？该论文涉及的相关研究主要集中在\*\*稀疏视角动态场景重建\*\*领域，可归纳为以下三条技术脉络： --- ### 1. 动态场景表示：从 Dynamic NeRF 到 4D Gaussian Splatting - \*\*Dynamic NeRFs\*\*：早期方法通过变形场（deformation fields）或显式分解的时空表示来建模动态场景，例如利用 K -Planes 将辐射场分解到时空平面中

Fridovich-Keil et al. 2023
。 - **4D Gaussian Splatting (4DGS)**：近期研究采用时空高斯基元实现高效的动态重建与实时渲染，代表性工作包括 Spacetime Gaussian Splatting (STGS)
Li et al. 2024
和 4DGS
Yang et al. 2024
。这类方法通常依赖密集且同步的相机阵列，在稀疏视角下性能显著下降。 —- ### 2. 稀疏视角动态重建：正则化与优化策略 - **基于几何/神经正则化的方法**：为降低对密集相机的依赖，现有稀疏视角方法引入神经正则化或几何约束进行优化，例如 4C4D

Zhou et al. 2026
。然而，这些方法本质上仍受限于输入视角中可用的视觉证据，**无法推断未观测区域缺失的信息**，导致重建在遮挡、弱纹理区域存在固有缺陷。 —- ### 3. 生成先验辅助的重建：扩散模型作为监督信号 - **静态场景重建**：近期研究尝试利用图像或视频扩散模型的生成先验，通过合成新视角并将其蒸馏到 3D 表示中，以弥补稀疏观测的不足。代表性工作包括 Generative Sparse-View Gaussian Splatting

Kong et al. 2025
和基于参考引导视频扩散的 GSFixer
Yin et al. 2026
。 - **动态/人体特定重建**：在动态领域，Diffuman4D
Jin et al. 2025
利用时空扩散模型进行 4D 一致的人体视角合成。然而，该方法局限于**人体中心场景**，难以推广到一般动态环境。 —- ### 4. 对比基线（Experimental Baselines）在实验评估中，该论文还与以下稀疏或密集视角的 4D 重建方法进行了定量与定性对比： - **STGS**

Li et al. 2024

**4DGS**
Yang et al. 2024
**4DGaussians**
Wu et al. 2024
**4C4D**
Zhou et al. 2026
Q3: 论文如何解决这个问题？该论文提出 **4DGS-Fixer**，一个两阶段生成式稀疏视角 4D 高斯溅射框架，通过**密集几何初始化**与**视频扩散引导的迭代精修**的互补设计，系统性地解决稀疏观测与初始化缺陷问题。 —- ### 1. 总体框架给定四视角同步输入视频，框架分为两个阶段： - **Stage 1：密集几何初始化** —— 利用多视图立体（MVS）深度估计替代 COLMAP，生成密集点云以初始化 4D 高斯表示，并在观测视角上优化得到初始动态场。 - **Stage 2：扩散引导迭代精修** —— 沿采样得到的新视角轨迹渲染视频，通过预训练的视频修复扩散模型生成高质量的伪监督信号，迭代优化 4D 高斯场。 —- ### 2. 密集几何初始化（Dense Geometric Initialization）传统 STGS 依赖 COLMAP 重建的稀疏点云，但在四视角稀疏设置下，COLMAP 因特征匹配不可靠而产生稀疏、不完整且有偏的初始化。该论文的改进如下： - **相机参数估计**：仅在首帧使用 COLMAP 估计相机内参与外参，由于相机固定，后续所有时刻共享同一套相机参数。 - **逐帧深度预测**：在每个时刻 t ，将四视角同步帧及对应相机参数输入 **MVSAnywhere**

Izquierdo et al. 2025
，为每个视角预测深度图 D(i,t)(u) 。 - **深度反投影与融合**：将有效深度像素反投影至世界坐标系并融合为密集彩色点云。反投影公式为：
x(i,t)(u) = T(w arrow c_i) D(i,t)(u) Ki^(-1) u 1 , quad u = u & v & 1 ^top
其中 K_i 为相机内参矩阵， T(w arrow ci) 为相机到世界坐标系的变换。时刻 t 的密集点云由四视角反投影点合并得到：
P_t = F( ∪(i=1)^(4) ( x(i,t)(u), I(i,t)(u) ) mid u ∈ Omegai, M(i,t)(u) = 1 )
其中 I(i,t)(u) 为对应点颜色， M(i,t)(u) ∈ 0,1 为有效深度掩码， F 表示点云滤波与融合操作。相比 COLMAP 稀疏点云，该初始化提供了**更完整的区域几何覆盖**，使 4DGS 优化效率与质量均显著提升。 —- ### 3. 视频扩散引导的迭代精修（Video-Diffusion-Guided Iterative Refinement）尽管密集初始化改善了观测区域的覆盖，弱观测或未观测区域仍可能存在模糊与空洞。第二阶段引入生成先验进行精修。 #### 3.1 新视角轨迹采样（Novel-View Trajectory Sampling）在四个输入相机之间构建六组配对轨迹。对于相机对 (i,j) ，通过线性插值与球面线性插值（SLERP）分别插值相机中心与旋转：

c(ij)(α) = (1-α)c_i + αc_j + A sin(πα) n(ij)

q(ij)(α) = SLERP(q_i, q_j; α), quad α ∈ [0,1]
其中 n(ij) 为垂直于相机基线的方向， A 为随机采样的小幅扰动幅值。六条直线轨迹设 A=0 ，对角相机对额外采用 A>0 以扩展空间覆盖。 #### 3.2 参考引导视频修复（Reference-Guided Video Restoration）利用冻结的 **CogVideoV2V**

Yin et al. 2026
模型进行修复。对于时刻 t 与相机对 (i,j) ，沿插值轨迹从当前 4DGS 渲染得到视频序列 V^(ren)(t,ij) ，并将其连同两个端点真实观测 I^i_t, I^j_t 输入模型：
V(t,ij) = Fφ ( V^(ren)(t,ij), I^it, I^j_t )
端点观测作为**锚点（anchors）**，促进跨视角一致性。修复后的序列 V(t,ij) 被缓存为伪监督目标。 #### 3.3 缓存策略与损失函数为避免每轮优化均调用计算昂贵的扩散模型，并降低生成随机性带来的伪目标不一致，论文采用**缓存机制**：每 K 次迭代刷新一次修复视频，中间步骤复用缓存。对于从缓存中采样的小批量 B_(gen) ，定义生成监督损失：

L(gen) = (1) / (|mathcalB)(gen)| ∑((t,n) ∈ B)(gen) | R(G; t, π(t,n)) - I(t,n) |1
总优化目标为：
L(total) = L(STG) + λ(gen) L(gen)
其中 R(G; t, π(t,n)) 表示从 4D 高斯场 G 在时刻 t 、位姿 π(t,n) 下渲染的图像， I(t,n) 为对应的 CogVideoV2V 生成伪目标， L(STG) 为原始 STGS 重建损失， λ(gen) = 0.5 用于平衡生成监督。 #### 3.4 完整算法流程 ALGORITHM 1: Iterative refinement with generative video supervision. Input: Four-view sparse videos V; camera parameters P; CogVideoV2V model F_φ; refinement iterations N; cache interval K. Output: Refined 4DGS representation G. X ← DensePointInit(V, P); G ← OptimizeInitSTGS(V, P, X); C ← ∅; for r ← 1 to N do if r = 1 or r mod K = 0 then C ← ∅; foreach timestamp t and camera pair (i, j) do Π_(t,ij) ← SampleTrajectory(P_i, P_j); V^(ren)_(t,ij) ← Render(G, t, Π_(t,ij)); V_(t,ij) ← F_φ(V^(ren)_(t,ij), I^i_t, I^j_t); C ← C ∪ Π_(t,ij), V_(t,ij); end end B_obs ← SampleObsViews(V); B_gen ← SampleCache(C); L_STG ← STGLoss(G, B_obs); L_gen ← GenLoss(G, B_gen); L_total ← L_STG + λ_gen * L_gen; G ← Update(G, L_total); end return G; —- ### 4. 方法总结 | 阶段 | 核心问题 | 解决手段 | 作用 | |—-|—-|—-|—-| | Stage 1 | COLMAP 稀疏/有偏初始化 | MVSAnywhere 多视图深度估计 + 密集点云融合 | 提升观测区域几何覆盖，加速优化收敛 | | Stage 2 | 未观测/弱观测区域信息缺失 | 参考引导视频扩散模型 + 缓存伪监督迭代优化 | 引入生成先验填补空洞，增强时空一致性 | 两个阶段相互补充：密集初始化确保观测区域有充分的基元支撑；扩散引导精修则通过新视角合成，将生成先验蒸馏到 4D 表示中，从而在根本上缓解稀疏视角下的不适定性。 Q4: 论文做了哪些实验？该论文的实验围绕 Neural3DV 数据集 展开，涵盖定量对比、定性可视化与消融分析三部分。 —- ### 1. 实验设置 - 数据集：Neural3DV，包含 6 个动态场景，由 18–21 台相机以 2704 × 2028 分辨率、30 FPS 采集。 - 训练/测试划分：遵循 4C4D 的评估协议，选取空间分布最分散的 4 个相机视角（ID 为 1、10、13、20）作为训练输入，其余视角用于测试。 - 对比基线：STGS

Li et al. 2024
、4DGS
Yang et al. 2024
、4DGaussians
Wu et al. 2024
、4C4D
Zhou et al. 2026
。 —- ### 2. 定量结果（Quantitative Comparison）论文采用 PSNR（↑）、DSSIM1（↓）、DSSIM2（↓）与 LPIPS（↓）作为评价指标。结果如表 1 所示： | Method | PSNR ↑ | DSSIM1 ↓ | DSSIM2 ↓ | LPIPS ↓ | |—-|—-|—-|—-|—-| | STGS | 17.70 | 0.158 | 0.107 | 0.325 | | 4DGS | 20.60 | 0.143 | 0.094 | 0.244 | | 4DGaussians | 20.82 | 0.117 | 0.077 | 0.190 | | 4C4D | 22.29 | 0.098 | 0.062 | 0.146 | | **4DGS-Fixer** | **24.18** | **0.086** | **0.053** | **0.128** | 4DGS-Fixer 在所有指标上均达到最优，相较先前最佳方法 4C4D，PSNR 提升 **1.89 dB**。 —- ### 3. 定性结果（Qualitative Visualization）论文在图 2 中给出了与 4C4D 及真实值（GT）的定性对比。结果表明，该方法在保留更锐利纹理细节的同时，有效抑制了模糊与伪影。 —- ### 4. 消融实验（Stage-wise Ablation）论文通过表 2 系统分析了密集初始化和扩散引导精修各自的贡献： | Method | PSNR ↑ | DSSIM1 ↓ | DSSIM2 ↓ | LPIPS ↓ | Prep. Time ↓ | |—-|—-|—-|—-|—-|—-| | STGS w/ COLMAP | 17.70 | 0.158 | 0.107 | 0.325 | ~20 min | | STGS w/ MVS | 21.67 | 0.109 | 0.071 | 0.165 | ~3 min | | CogVideoV2V restoration (init. 4DGS) | 23.56 | 0.105 | 0.071 | 0.107 | — | | **Full** | **24.18** | **0.086** | **0.053** | **0.128** | — | | CogVideoV2V restoration (refined 4DGS) | 24.33 | 0.096 | 0.065 | 0.099 | — | 消融结论如下： - **密集初始化的有效性**：以 MVSAnywhere 替代 COLMAP 后，PSNR 提升近 **4 dB**，且点云预处理时间由约 20 分钟缩短至约 3 分钟。 - **扩散修复的增益**：将 CogVideoV2V 修复后的视频直接作为监督蒸馏，可使初始 4DGS 渲染的 PSNR 从约 21.67 dB 提升至 23.56 dB。 - **迭代精修的正向效应**：完整训练后 4DGS 表示达到 **24.18 dB**；在此基础上再次对渲染结果进行扩散修复，可进一步获得 **24.33 dB**，验证了迭代框架的持续优化能力。 Q5: 有什么可以进一步探索的点？基于论文结论与方法局限，以下方向值得进一步探索： —- ### 1. 提升计算效率当前精修阶段需周期性调用预训练的 CogVideoV2V 视频修复模型，该过程计算开销显著，限制了方法的可扩展性。未来可探索： - 更轻量级的视频生成/修复模型作为替代先验； - 降低伪监督刷新频率或采用**单向蒸馏**（one-time distillation）而非迭代缓存策略； - 开发端到端可微的生成监督模块，避免频繁离屏渲染与模型推理。 —- ### 2. 增强生成监督的可靠性与鲁棒性视频扩散模型的输出可能包含**不一致或失真的内容**（尤其在高频细节或复杂非刚性运动区域），这些不完美生成会直接污染伪监督信号。可进一步研究： - **不确定性估计机制**：为扩散模型的输出像素赋予置信度权重，在损失函数中抑制高不确定性区域的影响； - **异常检测与滤波**：在缓存伪监督前，自动识别并剔除明显不合理的帧或区域； - **多模型集成**：利用多个生成模型的共识（consensus）降低单一模型的随机偏差。 —- ### 3. 向多样化真实世界场景泛化论文实验主要在 **Neural3DV** 这一受控环境下的多相机数据集上进行。面向更广泛的实际应用（如手持相机、无人机采集或室内外混合场景），需解决： - 相机标定误差与未同步视频的处理； - 极端光照变化、运动模糊及大尺度场景下的深度估计漂移； - 动态场景中**快速运动与严重遮挡**导致的 MVS 深度失效问题。 —- ### 4. 极端稀疏视角下的扩展论文采用四视角输入，仍存在一定程度的视场重叠。进一步探索**双视角甚至单目动态视频**的 4D 重建具有重要价值，但这要求： - 更强的单目/双目深度先验与时空约束； - 结合显式物理约束（如刚性结构、关节运动模型）以补偿几何歧义。 —- ### 5. 长时间序列的时间一致性当前方法以逐帧或短时窗口方式处理动态场景。对于**长视频序列**，累积的时序误差与生成模型的逐段随机性可能导致全局不一致。未来可引入： - 长程时序传播机制（如循环状态或全局运动基元）； - 层级化时间表示，平衡局部细节与长程动态稳定性。 —- ### 6. 生成先验与几何/光度约束的深度融合现有框架中，生成监督（ L(gen) ）与原始重建损失（ L(STG) ）以固定权重 λ(gen) 简单相加。更可取的方式是： - **自适应融合策略**：根据视角稀疏程度或重建置信度动态调整生成先验的权重； - 在优化过程中显式引入**深度、法向或光流一致性约束**，抑制扩散模型的几何幻觉，确保生成内容符合多视图几何一致性。 Q6: 总结一下论文的主要内容该论文提出 **4DGS-Fixer**，一种面向稀疏视角动态场景重建的生成式 4D 高斯溅射框架，通过**密集几何初始化**与**视频扩散引导迭代精修**的两阶段设计，显著提升了仅由四视角输入重建一般动态场景的质量与完整性。 —- ### 1. 研究动机与问题稀疏视角动态场景重建面临两个根本性挑战： - **不适定的观测不足**：现有稀疏 4DGS 方法受限于输入视角中有限的视觉证据，无法恢复未观测区域的缺失信息，导致重建模糊、出现空洞。 - **稀疏且有偏的几何初始化**：传统 COLMAP 在稀疏视角下特征匹配不可靠，产生稀疏、不完整且偏向高纹理区域的点云，使后续高斯优化难以填补缺失区域。此外，现有生成先验方法多针对静态场景或特定领域（如人体），缺乏对一般动态场景的适用性。 —- ### 2. 核心方法论文提出一个两阶段互补框架，如图 1 所示： #### Stage 1：密集几何初始化 - **相机标定**：仅在首帧使用 COLMAP 估计相机内参与外参，因相机固定，后续时刻共享同一套参数。 - **多视图深度估计**：对每个时刻 t ，将四视角同步帧输入 **MVSAnywhere** 预测逐像素深度图 D(i,t)(u) 。 - **深度反投影与融合**：将有效深度像素反投影至世界坐标并融合为密集彩色点云 P_t ：

x(i,t)(u) = T(w arrow ci) D(i,t)(u) K_i^(-1) u 1 , quad u = u & v & 1 ^top

Pt = F( ∪(i=1)^(4) ( x(i,t)(u), I(i,t)(u) ) mid u ∈ Omegai, M(i,t)(u) = 1 )
其中 Ki 为内参矩阵， T(w arrow ci) 为相机外参， M(i,t) 为有效深度掩码， F 为融合滤波操作。所得密集点云初始化 4DGS 并在观测视角上优化，得到初始动态高斯场 G 。 #### Stage 2：视频扩散引导迭代精修 - **新视角轨迹采样**：在六对相机间通过线性插值（中心）与 SLERP（旋转）构造平滑轨迹，对对角相机引入小幅正弦扰动以扩展空间覆盖：

c(ij)(α) = (1-α)c_i + αc_j + A sin(πα) n(ij)

q_(ij)(α) = SLERP(q_i, q_j; α), quad α ∈ [0,1]

**参考引导视频修复**：沿轨迹从当前 G 渲染视频 V^(ren)(t,ij) ，将其与两端真实帧 I^i_t, I^j_t 一并输入冻结的 **CogVideoV2V** 模型，以端点观测为锚点促进跨视角一致性：
V(t,ij) = Fφ ( V^(ren)(t,ij), I^i_t, I^j_t )
**缓存式伪监督优化**：每 K 次迭代刷新一次修复视频缓存，其余步骤复用缓存以降低计算成本并抑制生成随机性。总损失为：
L(total) = L(STG) + λ(gen) L(gen)
其中 L(gen) 为渲染图与对应伪目标之间的 ell_1 损失， λ(gen) = 0.5 。 —- ### 3. 主要贡献 - 提出面向**一般动态场景**的生成式稀疏 4DGS 框架，突破了现有生成方法对静态或人体特定领域的限制； - 引入基于 MVS 的**密集点云初始化策略**，将预处理时间从约 20 分钟缩短至 3 分钟，同时显著提升几何覆盖； - 开发**参考引导视频扩散迭代精修机制**，利用修复后的新视角视频作为伪监督，持续优化时空高斯表示； - 在 Neural3DV 四视角基准上达到最优性能，PSNR 较先前最佳方法 4C4D 提升 **1.89 dB**。 —- ### 4. 实验结果 - **定量对比**：在 Neural3DV 数据集上，4DGS-Fixer 的 PSNR 达到 **24.18 dB**，在 PSNR、DSSIM 与 LPIPS 上均优于 STGS、4DGS、4DGaussians 与 4C4D。 - **消融分析**： - 以 MVSAnywhere 替代 COLMAP 初始化，PSNR 提升近 **4 dB**； - 扩散引导精修将初始 4DGS 从约 21.67 dB 提升至 24.18 dB； - 对精修后的渲染再次进行扩散修复，PSNR 可进一步升至 24.33 dB，验证了迭代优化的正向效应。 —- ### 5. 局限与未来方向论文指出，视频修复模型的调用使精修阶段计算成本较高，且生成结果的不完美可能引入噪声。未来工作可探索更高效的监督方式、不确定性加权机制，以及向更稀疏视角（如双目或单目）和多样化真实场景的泛化。 Q7: 想要进一步了解论文以上只是了解一篇论文的几个基本FAQ。如果你还想与Kimi进一步讨论该论文，请点击这里为你跳转Kimi AI网页版，并启动一个与该论文相关的新会话。

Authors: Haitao Huang, Shenghao Zhao, Boyuan Tian, Shin-Fang Chng, Songlin Yang, Sheila Lim, Huangying Zhan, Yi Xu, Anyi Rao, Frank Guan

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.21176.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.21176

Published: 2026-09-22T01:52:57.262Z

8. OnomatoBridge: Onomatopoeia Translation and Rendering Pipeline in Manga

Abstract:Manga is a comic drawn by black and white paints gaining popularity around the world. Onomatopoeia in Manga specifically appeals to the audience with its unique visual styles, which convey sound, motion, and emotion. Visual onomatopoeia translation requires the clean replacement of Japanese onomatopoeia with onomatopoeia in the other language while preserving their visual style. Existing approaches often produce residual artifacts or style inconsistency when removing the Japanese onomatopoeia and rendering stylized English onomatopoeia. To approach these problems, we present OnomatoBridge, a filtering pipeline for visual onomatopoeia translation. We evaluate OnomatoBridge from Japanese to English on the Manga109 onomatopoeia dataset and compare it with baseline image editing models. Experimental results show that the filtered outputs by the proposed method outperform those of conventional methods. OnomatoBridge improves English text correctness by roughly 10 to 25 points and reduces residual Japanese text by about 20 to 50% in relative terms.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.21199 (HTTP 429)

Authors: Takara Taniguchi, Wataru Shimoda, Kota Yamaguchi, Hideki Nakayama

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.21199.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.21199

Published: 2026-09-22T01:52:57.262Z

9. Hand-Aware Transition Modeling for Bimanual Procedural Anomaly Detection

Abstract:Procedural anomaly detection in bimanual assembly requires judging each hand action against the execution so far. A corrective action may look unusual in isolation, while a visually plausible action can violate the order of the procedure. We present HACT, a transition model over predicted per-hand events. A role-preserving history keeps the concurrent responsibilities of both hands, and a marked temporal point process assigns each observed transition a semantic and temporal surprisal. A supervised evidence head and a two-state filter convert these surprisals into per-hand anomaly posteriors. A recovery-aware protocol on predicted events and participant-disjoint folds reports the recovery false-positive rate at an operating point selected on validation participants. On two bimanual power-tool procedures HACT has the highest AUPRC and F1 among the compared methods and the fewest recovery alarms. Applied without retraining to a different assembly order of the same product, it retains the highest AUPRC and F1. The source code is available at this https URL.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.21207 (HTTP 429)

Authors: Di Wen, Jimmy Weissert, Luc Maria Scherrer, Cedric Zöllner, Kailun Yang, Ruiping Liu, Yufan Chen, Jiale Wei, Junwei Zheng, Kunyu Peng

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.21207.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.21207

Published: 2026-09-22T01:52:57.262Z

10. Multi-viewpoint Geo-localization with Event Cameras

Abstract:Robot localization is an ongoing challenge that demands mapping and positioning systems that are tolerant to viewpoint change. Event cameras are attracting increasing interest and adoption in robotics; however, dealing with viewpoint variance is an under-investigated problem in existing event-based localizers. In addition, event-based datasets that emphasize viewpoint variance for challenging localization situations are scarce. Here, we introduce an event-based visual place recognition (VPR) system that performs robustly under viewpoint changes. We converted five large-scale geo-tagged datasets, conventionally used to train frame-based localization systems, into synthetic event streams using Image-to-Event (I2E) conversion, and used them to fine-tune a pre-trained event-based vision transformer backbone with a multi-loss function, yielding a system we call MegaEvent that learns viewpoint-robust features for place recognition. We achieved an average Recall@1 of 82% across three existing event-based localization datasets, leading the next best event-based method by 20 recall points, and frame-based VPR models applied directly to event frames by 8 to 26 recall points. We introduce a new, challenging dataset - Springfield-Event-VPR - which features a 3.7km walking route recorded in three camera orientations for a total of 11.1km, which MegaEvent outperforms the strongest baseline by 9 recall points. The code for MegaEvent is available at this https URL.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.21219 (HTTP 429)

Authors: Adam D. Hines, Michael Milford, Tobias Fischer

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.21219.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.21219

Published: 2026-09-22T01:52:57.262Z

ArXiv Domain 2026-09-22

LLM Domain Papers

1. Do small language models know what they don’t know?

2. HERMES: Contrast-Aware Knowledge Graph Reasoning from Clinical Notes for Patient Outcome Prediction

3. TALON: A Temporally Aware Longitudinal Framework for Radiology Report Generation

4. From Discharge Notes to Patient Understanding: Persona-Grounded, Open-Ended Simulation of LLMs as Discharge Educators

5. Beyond WER: Entity and Disfluency Recall in Accented Conversational ASR

6. SAGE: Schema-Guided LLMs for Grant Review

7. Reviser: Revision-Capable Text Generation via Autoregressive Cursor Actions

8. Recursive Language Models Generalize Out of Domain

9. TatBLiMP: A Benchmark of Linguistic Minimal Pairs for Tatar

10. Transsion’s Speaker-Attributed Multilingual ASR System for the MLC-SLM 2026 Challenge

Agent Domain Papers

1. RBS-Attention: Radius-Bounded Sparse Prefill for Long-Context Large Language Models

2. Attention-Aware Routing: Coupling Routing and Attention in MoEs

3. CaLR: Causal Latent Revision for Robust Diffusion Reasoning

4. LoRA Enhanced Contrastive Learning with SAS Vision Transformers

5. Detecting Hallucination in LLMs: Tracing the Topological Signatures of Impaired Context Sharing

6. Decoupling Internal Representational Changes and Causal Importance in Fine-Tuned Large Language Models

7. TinyCeNN-LM: Quality-Gated Conversion of Pretrained Attention with CeNN-Inspired Cellular-Recurrent Layers

8. Clinician-Grounded Quality Assurance for AI-Assisted Psychiatric Intake

9. Can Agents Design Better Chips with a Higher Level Abstraction?

10. SpecOpt: Contact-Diff Reasoning for Agentic Molecule Optimization Toward Binding Specificity

Evaluation Domain Papers

1. RBS-Attention: Radius-Bounded Sparse Prefill for Long-Context Large Language Models

2. Attention-Aware Routing: Coupling Routing and Attention in MoEs

3. CaLR: Causal Latent Revision for Robust Diffusion Reasoning

4. LoRA Enhanced Contrastive Learning with SAS Vision Transformers

5. Detecting Hallucination in LLMs: Tracing the Topological Signatures of Impaired Context Sharing

6. Decoupling Internal Representational Changes and Causal Importance in Fine-Tuned Large Language Models

7. TinyCeNN-LM: Quality-Gated Conversion of Pretrained Attention with CeNN-Inspired Cellular-Recurrent Layers

8. Clinician-Grounded Quality Assurance for AI-Assisted Psychiatric Intake

9. Can Agents Design Better Chips with a Higher Level Abstraction?

10. SpecOpt: Contact-Diff Reasoning for Agentic Molecule Optimization Toward Binding Specificity

VLM Domain Papers

1. TAPe+ML: A Compact Structured Representation for Multi-Task Computer Vision

2. MemeTAG: Keyword-Driven Meme Classification through Tag Embedding Reconstruction

3. Image-Derived PM10 Estimation in Cattle Feedlot Using Machine Learning: Addressing Concentration Ranges Beyond Existing Digital Imaging Methods

4. Fragment-Aware Vision Transformers for Fresco-Fragment Style Classification

5. MAGIC: Marginal-Guided Compression with Optimal Transport for Efficient Visual Document Retrieval

6. MarsFM: Shading-Regularized Flow Matching for Martian Relief Estimation

7. 4DGS-Fixer: Generative Sparse-View 4D Gaussian Splatting with Iterative Refinement Guided by Video Diffusion Priors

8. OnomatoBridge: Onomatopoeia Translation and Rendering Pipeline in Manga

9. Hand-Aware Transition Modeling for Bimanual Procedural Anomaly Detection

10. Multi-viewpoint Geo-localization with Event Cameras