Multi-Agent Coordination, Evidence Verification, and Adaptive Embodied Control

Today’s selection of 8 noteworthy AI/ML papers from arXiv, covering advances in LLM alignment, model merging, fact-checking, multi-agent coordination, embodied control, and financial AI.

1. A Geometric Perspective on Stabilizing Value Conflict Resolution

Authors: Saket Reddy, Andy Liu | Categories: cs.LG, cs.AI Link: arxiv.org/abs/2607.17946v1

The authors show geometrically that chain-of-thought (CoT) reasoning smooths the loss landscape in LLMs during RLHF training, helping resolve optimization instability from scalar rewards. They demonstrate that value conflict-focused CoT generalizes to different kinds of moral reasoning, and introduce a new CoT design that further improves performance on complex value conflicts.

Takeaway: This paper offers a principled geometric explanation for why CoT improves moral reasoning, providing a concrete path toward more pluralistic and value-aware LLMs.

2. Rethinking Heterogeneous LLM Merging: A Weighted Model Averaging Perspective

Authors: Jiahe Fan, Yinghao Hou, Si Chen, Aiyuan Zhang, Hong Xie et al. | Categories: cs.AI Link: arxiv.org/abs/2607.18026v1

This work investigates whether billion-parameter LLMs with different architectures can be merged via simple weighted averaging without training or semantic alignment. Using dimensional adaptation and ratio-controlled interpolation, they find that small-ratio interpolation can improve performance by transferring complementary capabilities, though near-balanced interpolation collapses and a seesaw effect emerges across tasks.

Takeaway: A surprisingly strong baseline for heterogeneous model merging that challenges the necessity of complex fusion techniques—while also revealing fundamental limits of weighted averaging.

3. DeLIVeR: Decomposed Learning for Information-grounded Veracity Recognition via Reinforced Knowledge Graph Exploration

Authors: Cong Hoan Nguyen, Thomas Hoang, Hieu Minh Duong, Long Nguyen | Categories: cs.CL, cs.AI Link: arxiv.org/abs/2607.17935v1

DeLIVeR treats evidence retrieval for fact-checking as a reinforced strategic exploration task, using a Planner LLM to decompose claims into targeted questions for traversing Knowledge Graphs. Optimized via GRPO, it achieves 10-15% F1 improvement over HippoRAG2 on LIAR, FEVER, and PolitiFact using Qwen2.5-7B, bridging multi-hop reasoning gaps with an auditable evidence trail.

Takeaway: A clever shift from passive retrieval to active, reinforced question-planning that significantly boosts fact-checking accuracy while maintaining transparency.

4. Sparse Evidence Can Suffice: Agentic Evidence Seeking for Multimodal Video Misinformation Detection

Authors: Haochen Zhao, Yongxiu Xu, Xinkui Lin, Dong Xie, Jiarui Lu et al. | Categories: cs.CV, cs.AI Link: arxiv.org/abs/2607.18080v1

SIEVE reframes video misinformation detection by decoupling evidence acquisition from verification: an evidence-seeking agent actively identifies sparse, decision-relevant clues before judging veracity. Trained with supervised trajectories and reinforcement learning, it consistently outperforms baselines while producing compact, inspectable evidence packages.

Takeaway: Challenges the “more data is better” assumption in multimodal reasoning—sparse, targeted evidence can actually improve both accuracy and transparency.

5. Value-Aware Prediction for Robust Multi-Agent Coordination Under Communication Loss

Authors: Kemal Devrim Kafadar, Eren Özaltun, Mahmud Efnan Şanlı, Feyza Orak, Emirhan Gazi et al. | Categories: cs.MA, cs.LG, cs.RO Link: arxiv.org/abs/2607.17914v1

The authors extend multi-agent observation sharing under communication dropout (MARO) with value-aware weighting of the predictor’s loss function using advantage estimates. This focuses model capacity on high-return dynamics rather than stochastic noise, achieving over 20% improvement in mean returns and 64.7% reduction in performance variance under high communication attrition.

Takeaway: A principled solution to the practical problem of communication dropout in multi-agent systems—value-aware prediction prevents the performance collapse that plagues standard predictors.

6. Adaptive Adversaries: A Multi-Turn, Multi-LLM Benchmark for LLM Agent Security

Authors: Devina Jain, David Hartmann, Chuan Li | Categories: cs.CR, cs.AI, cs.LG Link: arxiv.org/abs/2607.18063v1

This benchmark introduces adaptive multi-round attacks against LLM defenders, showing that attack success rate jumps from 0-1% with a single turn to 5.4-14.0% with 15 adaptive rounds. Pooling three frontier attackers uncovers 1.4-2.2× more unique attacks than any single attacker, and the attack transcripts have low similarity to existing benchmarks.

Takeaway: A much-needed dynamic security benchmark that reveals how vulnerable current LLM agents are to multi-turn manipulation—and how defender rankings vary dramatically across scenarios.

7. Patch Policy: Efficient Embodied Control via Dense Visual Representations

Authors: Gaoyue Zhou, Zichen Jeff Cui, Ada Langford, Bowen Tan, Yann LeCun et al. | Categories: cs.RO, cs.LG Link: arxiv.org/abs/2607.18236v1

Patch Policy enables transformer-based robot policies to consume dense pre-trained ViT patch tokens directly via a block-causal attention mask, without the computational overhead of a full VLM. It achieves 40% relative improvement over global-pooled representations and surpasses fine-tuned OpenVLA-OFT by 18% while using only 0.7% of the parameters.

Takeaway: A lightweight yet powerful architecture that brings dense visual representations to real-time robot control—practical, efficient, and backed by strong empirical results across seven environment suites.

8. FinSAgent: Corpus-Aligned Multi-Agent RAG Framework for Evidence-Grounded SEC Filing Question Answering

Authors: Jijun Chi, Zhenghan Tai, Hanwei Wu, Tung Sum Thomas Kwok, Hailin He et al. | Categories: cs.IR, cs.CL, cs.MA Link: arxiv.org/abs/2607.18102v1

FinSAgent reframes SEC filing QA by injecting corpus-side conditioning to correct misalignment between model priors and filing structure. Using role-specialized agents anchored to 10-K item structure, database-aware query decomposition, and a learned feature-gated reranker, it improves retrieval coverage and answer correctness across five benchmarks—validated by 1,000 anonymous user ratings.

Takeaway: A rigorous approach to the “prior-corpus misalignment” problem in financial AI, demonstrating how domain structure can be baked into multi-agent retrieval for significant gains.

This content was generated with AI assistance. Paper information sourced from arXiv.