AI Paper Daily

AI 论文日报 · 2026-07-02

覆盖北京时间日期 2026-07-01–2026-07-02。面向 Agent、LLM reasoning/planning/tool use/memory、coding agents、RAG、评测与 RL for LLMs 的每日筛选。

candidate count389

new included count366

selected count15

Top 3

RepoRescue: An Empirical Study of LLM Agents on Whole-Repository Compatibility Rescue
对 coding agent 的任务分解、验证或仓库级能力给出新基准/方法，可转化为 FJ 的代码代理评测素材。
Can Agents Generalize to the Open World? Unveiling the Fragility of Static Training in Tool Use
对 coding agent 的任务分解、验证或仓库级能力给出新基准/方法，可转化为 FJ 的代码代理评测素材。
Leveraging LLM-Based Agentic Systems to Generate Quantum Applications for Test Optimization
对 coding agent 的任务分解、验证或仓库级能力给出新基准/方法，可转化为 FJ 的代码代理评测素材。

Top Picks

01

RepoRescue: An Empirical Study of LLM Agents on Whole-Repository Compatibility Rescue

Zhihao Lin, Mingyi Zhou, Zhensu Sun, Yizhuo Yang, Renyu Yang, David Lo, Li Li
published BJT 2026-07-02 · updated BJT 2026-07-02 · cs.SE

对 coding agent 的任务分解、验证或仓库级能力给出新基准/方法，可转化为 FJ 的代码代理评测素材。

方法要点: Open-source libraries and tools are widely reused, but compatibility maintenance is expensive. Once maintainers leave, useful repositories can stop working as runtimes and dependencies evolve.
为什么重要: 命中偏好信号：agent, agents, tool, tools, reasoning, benchmark；适合 Lucian 的 agent/coding/personalization 研究候选池。
局限/风险: 日报阶段未完整阅读 PDF；需核对公开代码、数据、baseline 与 ablation。
Lucian 下一步: 阅读 abstract + intro + experiment table；符合方向则加入 autoresearch backlog。

Relevance 5/5Novelty 4.0/5Substance 3.5/5Evidence 3.0/5Actionability 3.6/5Total 19.1/25

02

Can Agents Generalize to the Open World? Unveiling the Fragility of Static Training in Tool Use

Song-Lin Lv, Weiming Wu, Rui Zhu, Zi-Jian Cheng, Lan-Zhe Guo
published BJT 2026-07-01 · updated BJT 2026-07-01 · cs.AI

对 coding agent 的任务分解、验证或仓库级能力给出新基准/方法，可转化为 FJ 的代码代理评测素材。

方法要点: While Large Language Model (LLM) agents demonstrate proficiency in static benchmarks, their deployment in real-world scenarios is hindered by the dynamic nature of user queries, tool sets, and interaction dynamics. To address this generalization gap, we formalize OpenAgent (Tool-Use Agent in Open-World), a problem setting characterized by distributional shifts across query, action, observation, and domain dimensions.
为什么重要: 命中偏好信号：agent, agents, tool, reasoning, rag, benchmark；适合 Lucian 的 agent/coding/personalization 研究候选池。
局限/风险: 日报阶段未完整阅读 PDF；需核对公开代码、数据、baseline 与 ablation。
Lucian 下一步: 阅读 abstract + intro + experiment table；符合方向则加入 autoresearch backlog。

Relevance 5/5Novelty 4.0/5Substance 3.5/5Evidence 3.0/5Actionability 3.6/5Total 19.1/25

03

Leveraging LLM-Based Agentic Systems to Generate Quantum Applications for Test Optimization

Ming Tao, Yuechen Li, Tao Yue, Man Zhang, Aitor Arrieta Marcos
published BJT 2026-07-01 · updated BJT 2026-07-01 · cs.SE, quant-ph

对 coding agent 的任务分解、验证或仓库级能力给出新基准/方法，可转化为 FJ 的代码代理评测素材。

方法要点: Quantum computing is increasingly explored for software engineering (SE) optimization, but translating natural-language (NL) task-level requirements into executable quantum applications still demands substantial quantum and programming expertise. We present QPipe, a large language model (LLM)-based multi-agent architecture that autonomously turns NL requirements into traceable quantum-application workflows through specialized agents for requirement parsing, formulation, code generation, review, execution, and verif
为什么重要: 命中偏好信号：agent, agents, multi-agent, feedback, rag, benchmark；适合 Lucian 的 agent/coding/personalization 研究候选池。
局限/风险: 日报阶段未完整阅读 PDF；需核对公开代码、数据、baseline 与 ablation。
Lucian 下一步: 阅读 abstract + intro + experiment table；符合方向则加入 autoresearch backlog。

Relevance 5/5Novelty 4.0/5Substance 3.5/5Evidence 3.0/5Actionability 3.6/5Total 19.1/25

04

Beyond Document Grounding: Span-Level Hallucination Detection over Code, Tool Output, and Documents

Ádám Kovács, Bowei He, Xue Liu, István Boros, Szilveszter Tóth, Gábor Recski
published BJT 2026-07-01 · updated BJT 2026-07-01 · cs.CL

对 coding agent 的任务分解、验证或仓库级能力给出新基准/方法，可转化为 FJ 的代码代理评测素材。

方法要点: Hallucination detection for retrieval-augmented generation (RAG) is usually evaluated on natural-language document evidence. However, grounded generation systems increasingly rely on structured inputs: source code, developer-tool output, markdown documents, tables, and repository metadata.
为什么重要: 命中偏好信号：agent, tool, rag, retrieval, benchmark, eval；适合 Lucian 的 agent/coding/personalization 研究候选池。
局限/风险: 日报阶段未完整阅读 PDF；需核对公开代码、数据、baseline 与 ablation。
Lucian 下一步: 阅读 abstract + intro + experiment table；符合方向则加入 autoresearch backlog。

Relevance 5/5Novelty 4.0/5Substance 3.5/5Evidence 3.0/5Actionability 3.6/5Total 19.1/25

05

AGI Maze as a Benchmark Framework for World-Modeling Agents

Alexey Potapov
published BJT 2026-07-01 · updated BJT 2026-07-01 · cs.AI

把记忆从被动存储推进到可训练/可评估的 agent 能力，值得直接映射到长期用户画像与情节一致性系统。

方法要点: Large language models (LLMs) are powerful pattern-completion systems, but their default operating mode - predicting the next token from a static context - does not reliably produce persistent, manipulable representations of an external world. Many tasks that look like "reasoning" in text become substantially harder once the environment is partially observable, stateful, and requires memory and structured hypotheses about hidden state.
为什么重要: 命中偏好信号：agent, agents, memory, reasoning, benchmark, evaluation；适合 Lucian 的 agent/coding/personalization 研究候选池。
局限/风险: 日报阶段未完整阅读 PDF；需核对公开代码、数据、baseline 与 ablation。
Lucian 下一步: 阅读 abstract + intro + experiment table；符合方向则加入 autoresearch backlog。

Relevance 5/5Novelty 4.0/5Substance 3.5/5Evidence 3.0/5Actionability 3.6/5Total 19.1/25

06

When Classic Cache Policies Fail: Learning-Augmented Replacement for Semantic Retrieval Buffers

Yushi Sun, Bowen Cao, Wai Lam
published BJT 2026-07-01 · updated BJT 2026-07-01 · cs.DB, cs.CL

把记忆从被动存储推进到可训练/可评估的 agent 能力，值得直接映射到长期用户画像与情节一致性系统。

方法要点: LLM agents increasingly rely on retrieval buffers to store and reuse past experience, yet the cache management policies governing these buffers remain largely ad-hoc. We formalize this as an online semantic cache replacement problem with switching costs, where items are matched by embedding similarity and hit quality is continuous rather than binary.
为什么重要: 命中偏好信号：agent, agents, memory, feedback, rag, retrieval；适合 Lucian 的 agent/coding/personalization 研究候选池。
局限/风险: 日报阶段未完整阅读 PDF；需核对公开代码、数据、baseline 与 ablation。
Lucian 下一步: 阅读 abstract + intro + experiment table；符合方向则加入 autoresearch backlog。

Relevance 5/5Novelty 4.0/5Substance 3.5/5Evidence 3.0/5Actionability 3.6/5Total 19.1/25

07

TRACE: State-Aware Query Processing over Temporal Evidence Graphs for Conversational Data

Maolin Wang, Yu Wang, Zichun Liu, Baiyuan Qiu, Chenbin Zhang, Jiguang Shen, Haoran Yang, Hao Miao
published BJT 2026-07-01 · updated BJT 2026-07-01 · cs.CL

把记忆从被动存储推进到可训练/可评估的 agent 能力，值得直接映射到长期用户画像与情节一致性系统。

方法要点: Conversational data is increasingly used as a persistent source of user state for long-running assistants and AI agents. However, querying this data remains challenging because conversations naturally evolve: plans are revised, preferences change, and later messages frequently supersede or contradict earlier information.
为什么重要: 命中偏好信号：agent, agents, memory, reasoning, retrieval, benchmark；适合 Lucian 的 agent/coding/personalization 研究候选池。
局限/风险: 日报阶段未完整阅读 PDF；需核对公开代码、数据、baseline 与 ablation。
Lucian 下一步: 阅读 abstract + intro + experiment table；符合方向则加入 autoresearch backlog。

Relevance 5/5Novelty 4.0/5Substance 3.5/5Evidence 3.0/5Actionability 3.6/5Total 19.1/25

08

The Illusion of Safety: Multi-Tier Verification of AI vs. Human C++ Code

Saif Mahmud, Fadul Sikder, Yuede Ji, Zhang Haotian, Jeff, Lei
published BJT 2026-07-01 · updated BJT 2026-07-01 · cs.SE

把记忆从被动存储推进到可训练/可评估的 agent 能力，值得直接映射到长期用户画像与情节一致性系统。

方法要点: Large language models increasingly generate C++, a memory-unsafe language where a single overlooked violation can become an exploitable bug. Yet most security evaluations of AI-generated code rely on static analysis alone, which flags warnings without confirming runtime violations or reasoning about untested paths.
为什么重要: 命中偏好信号：tool, tools, memory, reasoning, benchmark, evaluation；适合 Lucian 的 agent/coding/personalization 研究候选池。
局限/风险: 日报阶段未完整阅读 PDF；需核对公开代码、数据、baseline 与 ablation。
Lucian 下一步: 阅读 abstract + intro + experiment table；符合方向则加入 autoresearch backlog。

Relevance 5/5Novelty 4.0/5Substance 3.5/5Evidence 3.0/5Actionability 3.6/5Total 19.1/25

09

MECoBench: A Systematic Study of Multimodal Agent Collaboration in Embodied Environments

Qingyun Liu, Jiwen Zhang, Jingyi Hu, Siyuan Wang, Zhongyu Wei
published BJT 2026-07-01 · updated BJT 2026-07-01 · cs.MA, cs.AI, cs.CL, cs.CV

对 coding agent 的任务分解、验证或仓库级能力给出新基准/方法，可转化为 FJ 的代码代理评测素材。

方法要点: Recent multimodal large language models (MLLMs) have strong potential as embodied agents, but their ability to collaborate in visually grounded environments remains underexplored. To address this gap, we introduce MECoBench, a multimodal embodied cooperation benchmark with an evaluation platform spanning diverse real-world tasks, two cooperation structures, and three collaboration modes.
为什么重要: 命中偏好信号：agent, agents, benchmark, evaluation, eval, code；适合 Lucian 的 agent/coding/personalization 研究候选池。
局限/风险: 日报阶段未完整阅读 PDF；需核对公开代码、数据、baseline 与 ablation。
Lucian 下一步: 阅读 abstract + intro + experiment table；符合方向则加入 autoresearch backlog。

Relevance 5/5Novelty 4.0/5Substance 3.5/5Evidence 3.0/5Actionability 3.6/5Total 19.1/25

10

Theory of Mind and Persuasion Beyond Conversation: Assessing the Capacity of LLMs to Induce Belief States via Planning and Action

Ben Slater, Matteo G. Mecattaf, Lucy G. Cheke, John Burden, Winnie Street
published BJT 2026-07-01 · updated BJT 2026-07-01 · cs.CL

围绕推理/规划状态的可验证性或学习机制，适合用于构建更稳的长程 agent loop。

方法要点: Theory of Mind (ToM) benchmarks for Large Language Models (LLMs) typically rely on passive question-answering formats, but the deployment of LLMs in increasingly agentic and autonomous forms demands new evaluations. In this paper we evaluate an agent's ability to induce specific belief states in other agents by taking actions rather than using conversational persuasion, a capability we call Non-Conversational Planning ToM (NCP-ToM).
为什么重要: 命中偏好信号：agent, agents, reasoning, planning, benchmark, evaluation；适合 Lucian 的 agent/coding/personalization 研究候选池。
局限/风险: 日报阶段未完整阅读 PDF；需核对公开代码、数据、baseline 与 ablation。
Lucian 下一步: 阅读 abstract + intro + experiment table；符合方向则加入 autoresearch backlog。

Relevance 5/5Novelty 4.0/5Substance 3.5/5Evidence 3.0/5Actionability 3.6/5Total 19.1/25

11

Retrieved Images as Visual Thought: Training-Free Multimodal In-Context Learning for the Open-vs-Closed Gap

Bingchen Huang, Zhiling Wang, Yifu Chen, Yuanchao Du
published BJT 2026-07-01 · updated BJT 2026-07-01 · cs.CV

对 coding agent 的任务分解、验证或仓库级能力给出新基准/方法，可转化为 FJ 的代码代理评测素材。

方法要点: Recent work on Thinking with Images makes vision a dynamic part of reasoning, but does so through generation: the model invokes external tools, synthesizes code, or imagines new imagery, each at the cost of a tool protocol, brittle code, or an expensive training pipeline. A fourth route makes vision dynamic without generating anything, by retrieving labeled exemplar images and reasoning over them, yet it remains underexplored despite being train-free.
为什么重要: 命中偏好信号：tool, tools, reasoning, retrieval, benchmark, eval；适合 Lucian 的 agent/coding/personalization 研究候选池。
局限/风险: 日报阶段未完整阅读 PDF；需核对公开代码、数据、baseline 与 ablation。
Lucian 下一步: 阅读 abstract + intro + experiment table；符合方向则加入 autoresearch backlog。

Relevance 4.9/5Novelty 4.0/5Substance 3.5/5Evidence 3.0/5Actionability 3.6/5Total 19.0/25

12

Adversarial Pragmatics for AI Safety Evaluation: A Benchmark for Instruction Conflict, Embedded Commands, and Policy Ambiguity

Brett Reynolds
published BJT 2026-07-02 · updated BJT 2026-07-02 · cs.AI, cs.CL, cs.SE

提供新的评测视角，可帮助把 agent 系统从演示推进到可复现实验。

方法要点: Safety evaluations for language models increasingly depend on judgments about ambiguous natural-language behaviour: whether a model has followed an instruction, refused appropriately, complied with a policy, resisted an embedded command, or misreported progress in an agentic task. Existing benchmarks often compress these distinctions into pass/fail labels, obscuring whether failures arise from capability limits, policy ambiguity, instruction conflict, scaffold failure, or unstable evaluator judgments.
为什么重要: 命中偏好信号：agent, tool, rag, benchmark, evaluation, eval；适合 Lucian 的 agent/coding/personalization 研究候选池。
局限/风险: 日报阶段未完整阅读 PDF；需核对公开代码、数据、baseline 与 ablation。
Lucian 下一步: 阅读 abstract + intro + experiment table；符合方向则加入 autoresearch backlog。

Relevance 4.8/5Novelty 4.0/5Substance 3.5/5Evidence 3.0/5Actionability 3.6/5Total 18.9/25

13

Rise From The Ashes: LLM-based Static Analysis for Deep Learning Framework Bugs

Shaoyu Yang, Haifeng Lin, Chunrong Fang, Xiang Chen, Wei Cheng, Jiawei Liu, Yiyu Zhang, Hongyu Liu et al.
published BJT 2026-07-01 · updated BJT 2026-07-01 · cs.SE

对 coding agent 的任务分解、验证或仓库级能力给出新基准/方法，可转化为 FJ 的代码代理评测素材。

方法要点: Deep learning (DL) frameworks are critical AI infrastructures that often hide bugs with serious security implications. While dynamic approaches such as fuzzing are effective in uncovering these bugs, they require real test execution and incur high computational costs.
为什么重要: 命中偏好信号：agent, multi-agent, rag, retrieval, evaluation, eval；适合 Lucian 的 agent/coding/personalization 研究候选池。
局限/风险: 日报阶段未完整阅读 PDF；需核对公开代码、数据、baseline 与 ablation。
Lucian 下一步: 阅读 abstract + intro + experiment table；符合方向则加入 autoresearch backlog。

Relevance 4.8/5Novelty 4.0/5Substance 3.5/5Evidence 3.0/5Actionability 3.6/5Total 18.9/25

14

A Task-State Representation for Long-Horizon Mobile GUI Agents

Yujie Zheng, Zikang Liu, Xin Zhao, Ji-Rong Wen
published BJT 2026-07-01 · updated BJT 2026-07-01 · cs.CL

把记忆从被动存储推进到可训练/可评估的 agent 能力，值得直接映射到长期用户画像与情节一致性系统。

方法要点: While long-horizon mobile GUI agents typically rely on thought-action-observation loops, they struggle to separate persistent task states from transient screen observations. As execution histories grow, this entanglement imposes a severe context burden, causing agents to forget initial requirements, hallucinate progress, or repeatedly interact with stale interfaces.
为什么重要: 命中偏好信号：agent, agents, memory, reasoning, long-horizon, benchmark；适合 Lucian 的 agent/coding/personalization 研究候选池。
局限/风险: 日报阶段未完整阅读 PDF；需核对公开代码、数据、baseline 与 ablation。
Lucian 下一步: 阅读 abstract + intro + experiment table；符合方向则加入 autoresearch backlog。

Relevance 4.8/5Novelty 4.0/5Substance 3.5/5Evidence 3.0/5Actionability 3.6/5Total 18.9/25

15

Identifying and Resolving Pitfalls of Knowledge-Based VQA Benchmarks: Auditing, Repairing, and Augmenting

Qian Ma, S M Rayeed, Charles V. Stewart, Qiong Wu, Yao Ma
published BJT 2026-07-01 · updated BJT 2026-07-01 · cs.CL, cs.CV, cs.IR, cs.MM

围绕推理/规划状态的可验证性或学习机制，适合用于构建更稳的长程 agent loop。

方法要点: Knowledge-Based Visual Question Answering (KB-VQA) aims to evaluate whether Visual Language Models (VLMs) can retrieve, ground, and reason over external structured knowledge beyond visual evidence. In practice, answer accuracy is widely adopted as the primary evaluation metric, implicitly treating correctness as a proxy for knowledge-grounded reasoning.
为什么重要: 命中偏好信号：reasoning, retrieval, benchmark, evaluation, eval, rl；适合 Lucian 的 agent/coding/personalization 研究候选池。
局限/风险: 日报阶段未完整阅读 PDF；需核对公开代码、数据、baseline 与 ablation。
Lucian 下一步: 阅读 abstract + intro + experiment table；符合方向则加入 autoresearch backlog。

Relevance 4.8/5Novelty 4.0/5Substance 3.5/5Evidence 3.0/5Actionability 3.6/5Total 18.9/25

版本更新提醒

未发现需要单独提醒的重大版本更新。

数据源失败或不确定性说明

Hugging Face Daily Papers 2026-07-01 failed: URLError: <urlopen error [Errno 101] Network is unreachable>
Hugging Face Daily Papers 2026-07-02 failed: URLError: <urlopen error [Errno 101] Network is unreachable>

检索类别：cs.AI, cs.CL, cs.CV, cs.LG, stat.ML, cs.SE, cs.IR。去重状态：历史 seen IDs 185，窗口候选 389，新论文 366，精选 15。