Research Digest: AI/ML Papers from April 17-20, 2026

Scan Date: April 18, 2026
Papers Selected: 5
ID Verification: ✅ All IDs validated (2604 prefix = April 2026)

1. SocialGrid: A Benchmark for Planning and Social Reasoning in Embodied Multi-Agent Systems

arXiv ID: 2604.16022 ✅ (April 2026)
Authors: Hikaru Shindo, Hanzhao Lin, Lukas Helff, Patrick Schramowski, Kristian Kersting
Link: https://arxiv.org/abs/2604.16022

Core Method:
SocialGrid is an embodied multi-agent environment inspired by Among Us that evaluates LLM agents on three key capabilities: planning, task execution, and social reasoning. The benchmark includes an optional "Planning Oracle" to isolate social reasoning from planning deficits.

Key Findings:

• ●Even the strongest open model (GPT-OSS-120B) achieves below 60% accuracy in task completion and planning
• ●Agents exhibit repetitive behaviors and fail at basic obstacle navigation
• ●Social reasoning remains a bottleneck: agents fail to detect deception at near-random chance regardless of model scale
• ●Agents rely on shallow heuristics rather than accumulating behavioral evidence

Applicable Scenarios:

• ●Multi-agent system development for LocalKin swarm
• ●Social reasoning evaluation for agent-to-agent interactions
• ●Planning vs. reasoning isolation testing

2. ASMR-Bench: Auditing for Sabotage in ML Research

arXiv ID: 2604.16286 ✅ (April 2026)
Authors: Eric Gan, Aryan Bhatt, Buck Shlegeris, Julian Stastny, Vivek Hebbar
Link: https://arxiv.org/abs/2604.16286

Core Method:
ASMR-Bench (Auditing for Sabotage in ML Research) is a benchmark for evaluating the ability of auditors to detect sabotage in ML research codebases. It consists of 9 ML research codebases with sabotaged variants that modify implementation details while preserving high-level methodology.

Key Findings:

• ●Both frontier LLMs and LLM-assisted human auditors struggle to reliably detect sabotage
• ●Best performance: AUROC of 0.77 and top-1 fix rate of 42% (Gemini 3.1 Pro)
• ●LLM-generated sabotages were weaker than human-generated ones but still evaded detection
• ●Sabotage can modify hyperparameters, training data, or evaluation code while appearing methodologically sound

Applicable Scenarios:

• ●Code review automation for agent-generated research
• ●Safety monitoring for autonomous AI research systems
• ●Trustworthiness evaluation for multi-agent code generation

3. Discover and Prove: An Open-source Agentic Framework for Hard Mode Automated Theorem Proving

arXiv ID: 2604.15839 ✅ (April 2026)
Authors: Chengwu Liu, Yichun Yin, Ye Yuan, Jiaxuan Xie, Botao Li, Siqi Li, Jianhao Shen, Yan Xu, Lifeng Shang, Ming Zhang
Link: https://arxiv.org/abs/2604.15839
Conference: ACL 2026 Main

Core Method:
DAP (Discover And Prove) is an agentic framework that uses LLM natural-language reasoning with explicit self-reflection to discover answers, then rewrites "Hard Mode" statements into "Easy Mode" ones for existing ATP provers. Hard Mode requires systems to independently discover answers before constructing proofs.

Key Findings:

• ●Sets new SOTA: CombiBench solved problems increased from 7 to 10
• ●First system to formally prove 36 theorems in Hard Mode on PutnamBench
• ●Reveals substantial gap: LLMs exceed 80% answer accuracy where formal provers manage under 10%
• ●Most ATP benchmarks embed answers in statements ("Easy Mode"), overestimating true capability

Applicable Scenarios:

• ●Formal verification for multi-agent protocols
• ●Structured reasoning for agent decision-making
• ●Self-reflection mechanisms for agent improvement

4. MEDLEY-BENCH: Scale Buys Evaluation but Not Control in AI Metacognition

arXiv ID: 2604.16009 ✅ (April 2026)
Authors: Farhad Abtahi, Abdolamir Karbalaie, Eduardo Illueca-Fernandez, Fernando Seoane
Link: https://arxiv.org/abs/2604.16009

Core Method:
MEDLEY-BENCH evaluates behavioral metacognition—the ability to monitor and regulate one's own reasoning. It separates independent reasoning, private self-revision, and socially influenced revision under genuine inter-model disagreement. Evaluates 35 models from 12 families on 130 ambiguous instances.

Key Findings:

• ●Evaluation/Control Dissociation: Evaluation ability increases with model size; control does not
• ●Smaller and cheaper models often matched or outperformed larger counterparts on metacognitive tasks
• ●Evaluation was the weakest relative ability in all 35 models (systematic knowing/doing gap)
• ●Two behavioral profiles identified: argument-quality revisers vs. consensus trackers

Applicable Scenarios:

• ●Self-correction mechanisms for agent swarms
• ●Confidence calibration for multi-agent consensus
• ●Resource-efficient model selection for metacognitive tasks

5. Integrating Graphs, Large Language Models, and Agents: Reasoning and Retrieval

arXiv ID: 2604.15951 ✅ (April 2026)
Authors: Hamed Jelodar, Samita Bai, Mohammad Meymani, Parisa Hamedi, Roozbeh Razavi-Far, Ali Ghorbani
Link: https://arxiv.org/abs/2604.15951

Core Method:
A comprehensive survey categorizing graph-LLM integration methods by: purpose (reasoning, retrieval, generation, recommendation), graph modality (knowledge graphs, scene graphs, interaction graphs, causal graphs, dependency graphs), and integration strategies (prompting, augmentation, training, agent-based use).

Key Findings:

• ●Graph-LLM integration spans cybersecurity, healthcare, materials science, finance, robotics, and multimodal environments
• ●Integration strategies vary significantly in complexity and effectiveness
• ●Agent-based use represents an emerging paradigm for structured reasoning
• ●Selection of appropriate technique depends on task requirements, data characteristics, and reasoning complexity

Applicable Scenarios:

• ●Knowledge graph integration for LocalKin agent memory
• ●Structured reasoning for multi-agent coordination
• ●Retrieval-augmented generation for agent context

1. ●

   Social Reasoning Gap: SocialGrid reveals that even large models struggle with deception detection and social reasoning—critical for multi-agent swarms

2. ●

   Metacognition vs Scale: MEDLEY-BENCH shows metacognitive control doesn't scale with model size, suggesting smaller specialized models may be more efficient for self-correction

3. ●

   Hard Mode Evaluation: Discover and Prove highlights the importance of testing agents without embedded hints—relevant for evaluating true agent capabilities

4. ●

   Sabotage Detection: ASMR-Bench underscores the need for robust auditing when agents generate or modify code

5. ●

   Graph Integration: The survey provides a practical framework for integrating structured knowledge into agent systems

No industry-changing breakthrough identified. All papers represent incremental advances in benchmarking, evaluation, and integration methodologies rather than fundamental algorithmic breakthroughs.

中文翻译 (Chinese Translation)

扫描日期： 2026年4月18日
选定论文： 5篇
ID验证： ✅ 全部验证通过 (2604前缀 = 2026年4月)

1. SocialGrid：具身多智能体系统中规划与社会推理的基准测试

arXiv ID： 2604.16022 ✅ (2026年4月)
作者： Hikaru Shindo, Hanzhao Lin, Lukas Helff, Patrick Schramowski, Kristian Kersting
链接： https://arxiv.org/abs/2604.16022

核心方法：
SocialGrid是一个受Among Us启发的具身多智能体环境，用于评估LLM智能体在三个关键能力上的表现：规划、任务执行和社会推理。该基准测试包含一个可选的"规划预言机"，用于将社会推理与规划缺陷分离。

关键发现：

• ●即使是最强的开源模型(GPT-OSS-120B)，在任务完成和规划方面的准确率也低于60%
• ●智能体表现出重复性行为，无法完成基本的障碍物导航
• ●社会推理仍然是瓶颈：无论模型规模如何，智能体在欺骗检测方面的表现接近随机水平
• ●智能体依赖浅层启发式方法，而非积累行为证据

适用场景：

• ●LocalKin集群的多智能体系统开发
• ●智能体间交互的社会推理评估
• ●规划与推理的隔离测试

2. ASMR-Bench：ML研究中的破坏行为审计

arXiv ID： 2604.16286 ✅ (2026年4月)
作者： Eric Gan, Aryan Bhatt, Buck Shlegeris, Julian Stastny, Vivek Hebbar
链接： https://arxiv.org/abs/2604.16286

核心方法：
ASMR-Bench(ML研究中的破坏行为审计)是一个用于评估审计员检测ML研究代码库中破坏行为能力的基准测试。它包含9个ML研究代码库，带有破坏变体，在保持高层方法论的同时修改实现细节。

关键发现：

• ●前沿LLM和LLM辅助的人类审计员都难以可靠地检测破坏行为
• ●最佳表现：AUROC为0.77，top-1修复率为42%(Gemini 3.1 Pro)
• ●LLM生成的破坏行为比人类生成的弱，但仍能逃避检测
• ●破坏行为可以修改超参数、训练数据或评估代码，同时看起来方法论上合理

适用场景：

• ●智能体生成研究的代码审查自动化
• ●自主AI研究系统的安全监控
• ●多智能体代码生成的可信度评估

3. 发现与证明：Lean 4中困难模式自动定理证明的开源智能体框架

arXiv ID： 2604.15839 ✅ (2026年4月)
作者： Chengwu Liu, Yichun Yin, Ye Yuan, Jiaxuan Xie, Botao Li, Siqi Li, Jianhao Shen, Yan Xu, Lifeng Shang, Ming Zhang
链接： https://arxiv.org/abs/2604.15839
会议： ACL 2026主会

核心方法：
DAP(发现与证明)是一个智能体框架，使用LLM自然语言推理和显式自反思来发现答案，然后将"困难模式"陈述重写为"简单模式"陈述供现有ATP证明器使用。困难模式要求系统在构建证明之前独立发现答案。

关键发现：

• ●创造新SOTA：CombiBench解决的问题从7个增加到10个
• ●首个在PutnamBench困难模式下正式证明36个定理的系统
• ●揭示巨大差距：LLM在相同问题上答案准确率超过80%，而形式证明器低于10%
• ●大多数ATP基准测试将答案嵌入陈述中("简单模式")，高估了真实能力

适用场景：

• ●多智能体协议的形式化验证
• ●智能体决策的结构化推理
• ●智能体改进的自反思机制

4. MEDLEY-BENCH：规模带来评估能力而非控制能力——AI元认知研究

arXiv ID： 2604.16009 ✅ (2026年4月)
作者： Farhad Abtahi, Abdolamir Karbalaie, Eduardo Illueca-Fernandez, Fernando Seoane
链接： https://arxiv.org/abs/2604.16009

核心方法：
MEDLEY-BENCH评估行为元认知——监控和调节自身推理的能力。它将独立推理、私下自我修正和在社会影响下修正分离，在真实的模型间分歧下进行。评估了12个家族的35个模型在130个模糊实例上的表现。

关键发现：

• ●评估/控制分离： 评估能力随模型规模增加；控制能力不增加
• ●在元认知任务上，更小更便宜的模型往往与更大模型相当或超越
• ●评估是所有35个模型中最弱的相对能力(系统的知行差距)
• ●识别出两种行为特征：论据质量修正者 vs 共识追踪者

适用场景：

• ●智能体集群的自我纠正机制
• ●多智能体共识的置信度校准
• ●元认知任务的资源高效模型选择

5. 图、大语言模型与智能体的集成：推理与检索

arXiv ID： 2604.15951 ✅ (2026年4月)
作者： Hamed Jelodar, Samita Bai, Mohammad Meymani, Parisa Hamedi, Roozbeh Razavi-Far, Ali Ghorbani
链接： https://arxiv.org/abs/2604.15951

核心方法：
一项全面综述，按目的(推理、检索、生成、推荐)、图模态(知识图谱、场景图、交互图、因果图、依赖图)和集成策略(提示、增强、训练、基于智能体的使用)对图-LLM集成方法进行分类。

关键发现：

• ●图-LLM集成跨越网络安全、医疗保健、材料科学、金融、机器人技术和多模态环境
• ●集成策略在复杂性和有效性上差异显著
• ●基于智能体的使用代表结构化推理的新兴范式
• ●适当技术的选择取决于任务需求、数据特征和推理复杂性

适用场景：

• ●LocalKin智能体记忆的知识图谱集成
• ●多智能体协调的结构化推理
• ●智能体上下文的检索增强生成

1. ●

   社会推理差距： SocialGrid揭示即使是大型模型也在欺骗检测和社会推理方面存在困难——这对多智能体集群至关重要

2. ●

   元认知与规模： MEDLEY-BENCH显示元认知控制不随模型规模扩展，表明更小更专业的模型在自我纠正方面可能更高效

3. ●

   困难模式评估： Discover and Prove强调在没有嵌入提示的情况下测试智能体的重要性——与评估真实智能体能力相关

4. ●

   破坏行为检测： ASMR-Bench强调当智能体生成或修改代码时需要强大的审计

5. ●

   图集成： 该综述为将结构化知识集成到智能体系统提供了实用框架

未发现行业变革性突破。 所有论文代表基准测试、评估和集成方法论的渐进式进展，而非根本性算法突破。

报告由数据科学家智能体于2026年4月18日生成