hackcv · 实时 AI 资讯聚合

自然语言处理 · 大模型、推理与对齐

全部生成式 AI Agent / 智能体多模态计算机视觉自然语言处理强化学习 AI 安全编程助手公司与产品政策与监管研究进展 AI 视频

论文精选 82

Low-Bit N-Gram Embeddings for Language Models· 大模型低精度 n-gram 嵌入

Many recent open-weight language models use n-gram embedding tables to improve performance…

领域：cs.CL作者：Stephen Chung、Wenyu Du

#cs.CL

📎 arXiv🕒 10-23 01:56🔗 arxiv.org

论文精选 82

Can an Agent Debug a Web Corpus? Agentic Auditing of Slovak FineWeb2· LLM 能否调试网络语料库？

Multilingual pipelines such as FineWeb2 apply one recipe across a thousand languages, leav…

领域：cs.CL作者：Marek Šuppa、Ivan Vykopal、Jaroslav Kopčan

#cs.CL

📎 arXiv🕒 10-23 01:52🔗 arxiv.org

论文精选 83

Bilingual Language Models Capture Cross-Linguistic Influence in Second-Language Reading· 双语模型捕捉二语阅读的跨语言影响

While monolingual language model (LM) surprisal is widely established as a predictor of re…

领域：cs.CL作者：Abderrahmane Issam、Irene E. Winther、Stefan L. Frank

#cs.CL

📎 arXiv🕒 10-23 01:42🔗 arxiv.org

论文精选 82

How Ground Truth Shapes Retrieval Scores: Evidence-Aware Evaluation of a Retrieval-Augmented Generation System for Breast Lesions· 乳腺病变管理文献中的检索系统评估

Purpose: To describe an evidence-aware retrieval evaluation protocol for a clinical retrie…

领域：cs.CL作者：Miguel Mochizuki Silva、Samuel Ismael、Mateus C. Dias

#cs.CL

📎 arXiv🕒 10-23 01:36🔗 arxiv.org

论文精选 80

Would an LLM Match into a Residency? Uncertainty, Reasoning and Image Use of Open Language Models on the Spanish MIR Examination· 大语言模型能通过医学考试吗？

Accuracy on medical examinations is common evidence offered for the clinical use of langua…

领域：cs.CL作者：Miguel Fernandez-de-Retana、Unai Zulaika、Ruben Sanchez-Corcuera

#cs.CL

📎 arXiv🕒 10-23 01:23🔗 arxiv.org

论文精选 83

LongZero: Enhancing Long-Context Reasoning in LLMs from Zero Data· LongZero: 无数据增强长文本推理

Reinforcement learning (RL) is an effective way to strengthen the long-context reasoning o…

领域：cs.CL作者：Zhennan Wan、Jianfei Chen

#cs.CL

📎 arXiv🕒 10-22 01:59🔗 arxiv.org

论文精选 85

Environment Duels: A Microcosm of LLM Co-Evolution· 环境对决：大模型共进化的小宇宙

Through a tangle of real-world interactions, from synthetic training pipelines to data sha…

领域：cs.CL作者：Shibo Jin、Zhiqiu Xu、Yiling Zou

#cs.CL

📎 arXiv🕒 10-22 01:55🔗 arxiv.org

论文精选 85

Quant-Arena: Benchmarking LLM Agents for Quantitative Trading Research· 量化交易大模型基准测试

Quantitative research is an iterative loop: a researcher proposes a market hypothesis, imp…

领域：cs.CL作者：Houde Qian、Zheng Jiang、Wenyi Li

#cs.CL

📎 arXiv🕒 10-22 01:54🔗 arxiv.org

论文精选 85

Did the Problem Actually Get Harder? Testing LLMs on Paired P and NP-Complete Problems Hardened by Adversarial Multi-Agent Rewriting· 问题真的变得更难了吗？

Many NP-complete (NPC) problems have a polynomial-time (P) twin that differs in a single s…

领域：cs.CL作者：Nurul Labib Sayeedi、Md. Faiyaz Abdullah Sayeedi、Zabir Al Nazi

#cs.CL

📎 arXiv🕒 10-22 01:50🔗 arxiv.org

论文精选 82

GoiBERT: Compact Vietnamese Language Models via Two-Stage Knowledge Distillation· GoiBERT：通过两阶段知识蒸馏实现的紧凑越南语模型

Vietnamese pre-trained encoders have improved rapidly, but the strongest ones are large, m…

领域：cs.CL作者：Quang Anh Kiet Tran、Nam Anh Tran、Manh Hung Trinh

#cs.CL

📎 arXiv🕒 10-22 01:24🔗 arxiv.org

论文精选 83

NarraSpace: Multimodal Scene-Grounded Evidence for Picture-Description Dementia Classification· NarraSpace：痴呆症图像描述分类的多模态证据

Picture-description dementia classification uses a shared visual scene, but a transcript-l…

领域：cs.CL作者：Qingming Li、Ruiyan Sun、Satoshi Nakamura

#cs.CL

📎 arXiv🕒 10-21 01:55🔗 arxiv.org

论文精选 85

Benchmarking Arabic Dialects for Math Problems· 阿拉伯方言数学问题基准测试

Mathematical reasoning remains underexplored for low-resource languages, including Dialect…

领域：cs.CL作者：Noor Abo Mokh、Teresa Lynn、Nizar Habash

#cs.CL

📎 arXiv🕒 10-21 01:50🔗 arxiv.org

论文精选 83

Capable Teacher, Myopic Scorer: Reward-Free Teacher Repair for Agentic On-Policy Distillation· 有能力的教师，短视的评分者：无奖励教师修复

Agentic on-policy distillation (OPD) trains a student agent on its own trajectories while …

领域：cs.CL作者：Shu Zhao、Tan Yu

#cs.CL

📎 arXiv🕒 10-21 01:40🔗 arxiv.org

论文精选 85

Cache Engineering for Agents· 代理的缓存工程

Large Language Model (LLM) agents can handle increasingly complex tasks, navigating long t…

领域：cs.CL作者：Tianyu Fu、Ruoxi Wang、Zihan Min

#cs.CL

📎 arXiv🕒 10-21 01:32🔗 arxiv.org

论文精选 82

Large Flow Language Models· 大型流语言模型

Continuous flow language models offer a path to parallel generation with tools for control…

领域：cs.CL作者：Arnav Chavan、Chanhyuk Lee、Udbhav Bamba

#cs.CL

📎 arXiv🕒 10-21 01:20🔗 arxiv.org

论文精选 83

OpenVLHarness: Real World Visual Agentic Tool-Use with Multimodal Memory· OpenVLHarness: 多模态记忆的视觉代理工具使用

Multimodal large language models (MLLMs) have made rapid progress on visual reasoning, yet…

领域：cs.CL作者：Yu Zhou、Cheng-Fu Yang、Rui Sun

#cs.CL

📎 arXiv🕒 10-20 01:59🔗 arxiv.org

论文精选 82

LatComm: Reusable Context Compression for Latent Communication between Large Language Models· LatComm：多语言模型间隐式通信的可复用上下文压缩

Systems that combine several large language models (LLMs) often use one context for many q…

领域：cs.CL作者：Huawei Lin、Jinting Liu、Tan Yu

#cs.CL

📎 arXiv🕒 10-20 01:58🔗 arxiv.org

论文精选 83

Competing Learning in Hybrid Architectures· 混合架构中的竞争学习

Many modern language models have adopted hybrid architectures that incorporate both linear…

领域：cs.CL作者：Xingyu Shen、Yingfa Chen、Ziqing Qiao

#cs.CL

📎 arXiv🕒 10-20 01:55🔗 arxiv.org

论文精选 83

KEEP-Bench: Evaluating Evidence Preservation in Context Compression· KEEP-Bench：评估上下文压缩中的证据保留

Context compression improves long-context efficiency but may introduce faithfulness issues…

领域：cs.CL作者：Ziqing Man、Di Wu、Jia-Chen Gu

#cs.CL

📎 arXiv🕒 10-20 01:55🔗 arxiv.org

论文精选 82

RefCraft: Benchmarking Visual-Reference-Conditioned Agentic Generation through Web Development· RefCraft：通过网页开发评估视觉参考条件下的代理生成

Beyond basic tool use and coding skills, effectively incorporating visual references into …

领域：cs.CL作者：Yuchi Wang、Rongyao Fang、Weihao Yu

#cs.CL

📎 arXiv🕒 10-20 01:55🔗 arxiv.org

论文精选 83

Behavioral Couplings: From Formation to Behavioral Emergence in Language Models· 行为耦合：从形成到行为涌现

Fine-tuning large language models (LLMs) can induce behaviors beyond those explicitly targ…

领域：cs.CL作者：Dongyoon Hahm、Woogyeol Jin、Dennis Wei

#cs.CL

📎 arXiv🕒 10-17 01:58🔗 arxiv.org

论文精选 82

DiscoveryRSI: Recursively Learning to Discover Valuable Hypotheses· DiscoveryRSI: 递归学习发现有价值假设

Scientific discovery requires developing worthwhile hypotheses into practical investigatio…

领域：cs.CL作者：Jianyuan Zhong、Chenwei Wu、Zhaochen Yu

#cs.CL

📎 arXiv🕒 10-17 01:56🔗 arxiv.org

论文精选 83

Beyond the Answer: Student-Conditioned Trace Construction for Reasoning Distillation· 超越答案：面向学生的推理蒸馏

A reasoning trace can accompany a correct answer without providing effective supervision f…

领域：cs.CL作者：Shaoang Li、Yanhang Shi、Yufei Li

#cs.CL

📎 arXiv🕒 10-17 01:52🔗 arxiv.org

论文精选 83

Acceptance Is Not Speed: Length Inflation in Training-Free Lossy Speculative Decoding· 接受并非速度：无训练损失性投机解码中的长度膨胀

Speculative decoding typically characterizes speedup by draft tokens accepted per verifica…

领域：cs.CL作者：Chiatzen Wang、Bill Xu、Haochen Gu

#cs.CL

📎 arXiv🕒 10-17 01:41🔗 arxiv.org

论文精选 82

Steering beyond Language: Improving Multilingual Training-Free Text Embeddings via Paired-Residual Subspace Removal· 超越语言：通过配对残差子空间移除改善多语言无训练文本嵌入

While deriving text embeddings directly from frozen LLMs has emerged as a promising direct…

领域：cs.CL作者：Ailiang Lin、Zhuoyun Li、Yusong Wang

#cs.CL

📎 arXiv🕒 10-17 01:30🔗 arxiv.org

论文精选 82

OmniVerifier-C1: Decomposed Graded Verifier for Rule-Governed Generative World Modeling· OmniVerifier-C1: 多规则生成世界建模的分解评分验证器

Generative visual verifiers play a crucial role in advancing multimodal intelligence. To e…

领域：cs.CL作者：Bowei Liu、Xinchen Zhang、Xuhuan Li

#cs.CL

📎 arXiv🕒 10-16 01:58🔗 arxiv.org

论文精选 82

Training Reasoning Models to Generate Diverse Scientific Ideas· 训练推理模型生成多样化的科研想法

One of the first steps of a research pipeline is generating ideas. Large Language Models (…

领域：cs.CL作者：Manya Wadhwa、Pranav Narayanan Venkit、Hiroaki Hayashi

#cs.CL

📎 arXiv🕒 10-16 01:52🔗 arxiv.org

论文精选 82

MindEval2: Clinically Grounded Patients and Rubrics for Evaluating Language Models in Mental Health Support· MindEval2：评估语言模型的心理健康支持

With people increasingly turning to AI chatbots for mental health support, the need to eva…

领域：cs.CL作者：John Mendonça、Daniel Tschernutter、José Pombal

#cs.CL

📎 arXiv🕒 10-16 01:51🔗 arxiv.org

论文精选 85

Grid-Bench: Can LLMs Predict How Discrete Worlds Change?· Grid-Bench：测试 LLM 预测离散世界变化的能力

Spatial instructions describe how to move objects, transform regions, or modify their surr…

领域：cs.CL作者：Zixiang Xu、Zirui Song、Wanrong Zheng

#cs.CL

📎 arXiv🕒 10-16 01:50🔗 arxiv.org

论文精选 83

Knowledge That Is Linear Before Fine-Tuning Generalizes After It· 微调前的线性知识泛化

Not all facts a language model can recall support further learning equally well. We distin…

领域：cs.CL作者：Faridoun Mehri、Hinrich Schütze

#cs.CL

📎 arXiv🕒 10-16 01:47🔗 arxiv.org

论文精选 82

Latent Compaction with Test-Time Training· 测试时训练的潜在压缩

Long contexts enlarge key--value (KV) caches, increasing memory use and attention costs du…

领域：cs.CL作者：Yingfa Chen、Xingyu Shen、Chaojun Xiao

#cs.CL

📎 arXiv🕒 10-15 01:57🔗 arxiv.org

论文精选 81

Manacá-1B-Instruct-2: Reproducible Safety Alignment for a Native Brazilian-Portuguese LLM. Negative DPO Results, a Controlled Composition Scan, and a...· Manacá-1B-Instruct-2: 巴西葡萄牙语大模型的安全对齐

Brazilian Portuguese remains underserved by open-source language models, and the few that …

领域：cs.CL作者：Bruno Leonardo Santos Menezes、Carlos Leonardo Souza Cardoso、Fabio André Machado Porto

#cs.CL

📎 arXiv🕒 10-15 01:53🔗 arxiv.org

论文精选 83

PerSo: Investigating Facets of LLMs' Social Intelligence in Real-World Persuasion Scenarios· 大模型社交智能研究

Understanding the social intelligence of large language models (LLMs) is increasingly impo…

领域：cs.CL作者：Ali Zamani、Soroush Esfahanian、Erfan Shahabi

#cs.CL

📎 arXiv🕒 10-15 01:39🔗 arxiv.org

论文精选 82

PrismaCoder: Pushing the Boundaries of Open Multimodal Code Intelligence through Decomposed Rubric Synthesis· 普里玛码农：通过分解评分合成推动开放多模态代码智能的边界

Multimodal code intelligence connects the symbolic world of programs with the visual world…

领域：cs.CL作者：Qiushi Sun、Jingyang Gong、Yang Liu

#cs.CL

📎 arXiv🕒 10-15 01:32🔗 arxiv.org

论文精选 85

VASPO: Self-Distillation with Privileged Value Estimation· VASPO: 增强价值自蒸馏

On-policy self-distillation provides dense supervision for reasoning models via a privileg…

领域：cs.CL作者：Fei Wu、Zhenrong Zhang、Jianshu Zhang

#cs.CL

📎 arXiv🕒 10-15 01:17🔗 arxiv.org

论文精选 82

Tev: Scaling Long-Horizon Agentic Environments with Dense Progress Rewards· Tev: 使用密集进度奖励扩展长时段代理环境

Long-horizon agents have drawn growing attention for their ability to work autonomously --…

领域：cs.CL作者：Wenbo Hu、Shraman Pramanick、Krishna Kumar Singh

#cs.CL

📎 arXiv🕒 10-14 01:59🔗 arxiv.org

论文精选 82

DoubtLess: Training-Free Contrastive Decoding for Efficient Reasoning· DoubtLess: 无训练对比解码方法

Reasoning models achieve strong performance through long chains of thought (CoT), but ofte…

领域：cs.CL作者：Minghao Luo、Liang Chen

#cs.CL

📎 arXiv🕒 10-14 01:59🔗 arxiv.org

论文精选 83

Hybrid Reasoning Agents· 混合推理代理

How can we build an agent that thinks beyond human language, yet still acts through it in …

领域：cs.CL作者：Jiaru Zou、Yunzhe Qi、Sirui Chen

#cs.CL

📎 arXiv🕒 10-14 01:59🔗 arxiv.org

论文精选 82

Back-Translation Still Matters: Leveraging Monolingual Data for Low-Resource Machine Translation with LLMs· 回译仍重要：利用单语数据改善低资源语言机翻

Pre-trained Large Language Model (LLMs) are effective on many tasks in zero-shot and few-s…

领域：cs.CL作者：Divyanshu Bathla、Parag Singla、Jitin Singla

#cs.CL

📎 arXiv🕒 10-14 01:50🔗 arxiv.org

论文精选 80

Where Do LLM Ensemble Gains Come From? Decomposing Candidate Value, Diversity, and Aggregation· LLM 集合效应来源的研究

LLM ensemble gains are often measured against a single baseline, but the value attributed …

领域：cs.CL作者：Donghwan Kim

#cs.CL

📎 arXiv🕒 10-14 01:36🔗 arxiv.org

论文精选 83

RSI-Anything: Unlocking Recursive Self-Improvement with World Models· RSI-Anything：解锁递归自我改进的世界模型

Recursive self-improvement (RSI) aims to transform interaction experience into persistent …

领域：cs.CL作者：Taoyong Cui、Pinxian Zeng、Wanghan Xu

#cs.CL

📎 arXiv🕒 10-10 01:50🔗 arxiv.org

论文精选 82

When VLMs Destroy What They Know: Spatial Representations Fall Below Lexical Priors· 当大模型破坏已知：空间表示逊于词汇先验

Vision-Language Models (VLMs) increasingly analyze and reason about images, yet it remains…

领域：cs.CL作者：Shubh Garg、Chen Shani

#cs.CL

📎 arXiv🕒 10-10 01:50🔗 arxiv.org

论文精选 83

Structural Automata for Neural Semantic Parsing· 神经语义解析的结构自动机

Neural semantic parsers generate meaning representations token by token, yet well-formedne…

领域：cs.CL作者：Xiao Zhang

#cs.CL

📎 arXiv🕒 10-10 01:42🔗 arxiv.org

论文精选 82

Task Context Shapes Component Roles in Language Models· 任务上下文如何塑造语言模型组件角色

Large language models reuse the same attention heads across tasks, but identifying these s…

领域：cs.CL作者：Sovesh Mohapatra、Dani S Bassett

#cs.CL

📎 arXiv🕒 10-10 01:40🔗 arxiv.org

论文精选 82

Can Language Models Learn When to Abstain? Policy Learning and Calibrated Answer Selection· 语言模型的选择性回答

A useful language model should know when to answer and when to say "I don't know." Can we …

领域：cs.CL作者：Juan Terven、Diana-Margarita Córdova-Esparza、Ruben Gomez

#cs.CL

📎 arXiv🕒 10-10 00:33🔗 arxiv.org

论文精选 82

Predicting Alignment Generalization with Value Representations· 用价值表示预测对齐泛化

LLM developers post-train their models to exhibit prosocial values and behavioral traits, …

领域：cs.CL作者：Andy Liu、Mehar Bhatia、Karolina Stanczak

#cs.CL

📎 arXiv🕒 10-09 01:47🔗 arxiv.org

论文精选 82

Latent Core Tokenizer: Compress, but Meaningfully· 潜在核心分词器：有意义的压缩

Tokenizers are commonly optimized for compression, but a compact vocabulary does not neces…

领域：cs.CL作者：Felermino D. M. A. Ali、Millicent Ochieng、Ogbemi Ekwejunor-Etchie

#cs.CL

📎 arXiv🕒 10-09 01:32🔗 arxiv.org

论文精选 83

Which Skill to Distill? SGUID: Selecting a Compact Skill Bank for Model-Skill Co-Evolution· 选择模型技能共演化所需的紧凑技能库

Skills, reusable procedural guidance added at inference, can substantially improve LLM dow…

领域：cs.CL作者：Yuhan Liu、Xiyao Ma、Zhongkai Sun

#cs.CL

📎 arXiv🕒 10-09 01:27🔗 arxiv.org

论文精选 83

VFold: Symmetry-Aware Cross-Layer Value Cache Compression· VFold: 跨层值缓存压缩

While caching key-value (KV) states accelerates Large Language Model (LLM) decoding, this …

领域：cs.CL作者：Neha Verma、Sungwon Kim、Kenton Murray

#cs.CL

📎 arXiv🕒 10-09 01:14🔗 arxiv.org

论文精选 85

HarnessSQL: Harness-Native Training for SQL Agents in Realistic Database Environments· HarnessSQL: SQL 代理的本机训练框架

Text-to-SQL models are commonly trained to map questions directly to static queries, where…

领域：cs.CL作者：Haolin Yang、Jipeng Zhang、Jian Xie

#cs.CL

📎 arXiv🕒 10-09 00:36🔗 arxiv.org

论文精选 82

EngramEdit: Decoupled Knowledge Updates in LLMs through Conditional Memory· EngramEdit：通过条件记忆实现 LLM 知识更新

Conditional memory architectures such as DeepSeek Engram use input n-grams to look up lear…

领域：cs.CL作者：Hongru Cai、Ran Wei、Wenjie Wang

#cs.CL

📎 arXiv🕒 10-08 01:58🔗 arxiv.org

论文精选 83

Your Prompt Should Do More: Effects of Retrieval Instructions in Embedding Models

Prompted embedding models have recently received increasing attention, particularly for re…

领域：cs.CL作者：Amanda Myntti、Jenna Kanerva、Veronika Laippala

#cs.CL

📎 arXiv🕒 10-08 01:51🔗 arxiv.org

论文精选 82

PHRBench: A Behavioral Evaluation of Post-Hallucination Reasoning in LLMs· 幻觉后推理的基准测试

Hallucinated information can propagate through multi-stage LLM systems and become part of …

领域：cs.CL作者：Linghao Meng、Feng He、Xuan Yang

#cs.CL

📎 arXiv🕒 10-08 01:25🔗 arxiv.org

论文精选 83

CoTrace: Data Recipes for Training Terminal Agents with Harness-Model Co-Evolution· 终端智能体用共演化训练的数据配方

Terminal-agent capability depends jointly on model weights and the runtime harness that fo…

领域：cs.CL作者：Jixuan Chen、Jiaxin Zhang、Qinyuan Ye

#cs.CL

📎 arXiv🕒 10-08 01:06🔗 arxiv.org

论文精选 83

Document-Level Text Simplification in Estonian Using Large Language Models· 爱沙尼亚语文档级文本简化

Document-level text simplification involves transformations that go beyond sentence-intern…

领域：cs.CL作者：Meeri-Ly Muru、Eduard Barbu

#cs.CL

📎 arXiv🕒 10-08 00:38🔗 arxiv.org

论文精选 85

IdeaAnchor: Teaching LLMs to Turn Literature into Research Ideas· IdeaAnchor: 从文献生成科研想法

Scientific research often begins by synthesizing ideas from a set of related papers to ide…

领域：cs.CL作者：Ziyu Chen、Yilun Zhao、Jiashuo Sun

#cs.CL

📎 arXiv🕒 10-07 01:59🔗 arxiv.org

论文精选 85

AdvSim2Real : Training Web Agents Against Adaptive Prompt Injection in a Web World Model· AdvSim2Real：在 Web 世界模型中对抗适应性提示注入

Web agents complete user requests by reading and acting on pages that third parties write,…

领域：cs.CL作者：Sarim Hashmi、Mukul Ranjan、Kshitij Mishra

#cs.CL

📎 arXiv🕒 10-07 01:56🔗 arxiv.org

论文精选 82

The Missing Minimal Pair: Stereotype Evaluation in LLMs· 缺失的最小对立对：大模型中的刻板印象评估

A common approach to measuring bias in Large Language Models is to compare the log-likelih…

领域：cs.CL作者：Nataliya Stepanova、Ivan Titov、Emily Allaway

#cs.CL

📎 arXiv🕒 10-07 01:41🔗 arxiv.org

论文精选 85

Denoising Hierarchical Representations: Joint Continuous Diffusion for Language Modeling· 分层连续扩散语言模型

Diffusion Language Models (DLMs) hold the promise of order-agnostic, parallel text generat…

领域：cs.CL作者：Mathias Ollu、Nikos Komodakis

#cs.CL

📎 arXiv🕒 10-07 01:38🔗 arxiv.org

论文精选 75

Holdout Best-of-N: Unbiased Evaluation and Its Cost

Reusing the scores that select a Best-of-$N$ winner can overstate its expected reward. We …

领域：cs.CL作者：Shrey Shah、Yinheng Li

#cs.CL

📎 arXiv🕒 10-07 01:26🔗 arxiv.org