Multi-Agent RL, Self-Taught LLM Evolution, GUI Grounding, and Audio Anti-Fraud

Today’s selection of 8 noteworthy AI/ML papers from arXiv, covering cooperative multi-agent reinforcement learning for emergency relief logistics, self-improving LLM training frameworks, evolutionary meta-path discovery for fake news detection, efficient GUI grounding, frozen-weight skill optimization for audio anti-fraud, audio-augmented robot manipulation, ReLU network Jacobian theory, and trusted back-translation for matching decompilation.

1. CoRe-MARL: Cooperative Redistribution Under Unknown Dynamics Using Recurrent Multi-Agent Reinforcement Learning

Authors: Naimur Rahman Chowdhury, Shatabdi Sen Prapti, Md. Salehin Seyam, Limon Bin Hossain | Categories: cs.LG, cs.AI Link: arxiv.org/abs/2609.18639

This work develops CoRe-MARL, a cooperative multi-agent reinforcement learning framework for emergency relief redistribution among decentralized local centers facing uncertain demand and supply. Formulated as a Dec-POMDP with recurrent networks to capture evolving dynamics and MAPPO for centralized training with decentralized execution, the framework aims to improve service in the worst-case region and reduce inter-regional service gaps. Experiments show MAPPO reduces the service gap and enhances worst-served center performance compared to recurrent IPPO and heuristic baselines, while maintaining competitive network-wide service across diverse trajectory patterns.

Takeaway: A compelling application of multi-agent RL to equitable humanitarian logistics, demonstrating that recurrent MAPPO can adapt to dynamics that agents never directly observe—a setting that closely mirrors real-world decentralized response networks.

2. STRETCH the Boundaries: A Unified Self-Taught Framework for Progressive LLM Evolution

Authors: Yajie Yu, Mark Lee, Yue Feng | Categories: cs.CL Link: arxiv.org/abs/2609.18642

STRETCH addresses capability stagnation in self-improvement training by introducing a dynamic Stretch Zone mechanism that continuously aligns question difficulty with the model’s evolving solving ability. Within a single parameter space, the model alternates between a Scaffolder generating adaptive challenges and a Learner optimizing solving trajectories via reinforcement learning, stabilizing training while mitigating reward hacking. Experiments on negotiation and operations research benchmarks show consistent gains over strong prompting and domain-specific baselines, with analysis confirming dynamic difficulty alignment is critical for sustained improvement.

Takeaway: The dual-loop Scaffolder-Learner design offers a principled answer to the “curriculum problem” in self-play—keeping challenges perpetually at the edge of competence is a simple idea with apparently strong empirical payoff.

3. Reasoning through Evolution: Automatic Meta-path Discovery for LLM-based Fake News Detection

Authors: Ziyi Zhou, Xiaoming Zhang, Hui Pang, Yuting Zhang, Tiesunlong Shen et al. | Categories: cs.AI, cs.LG Link: arxiv.org/abs/2609.18597

MAGER is a multi-agent genetic evolution framework that automatically discovers meta-paths optimized for LLM reasoning, compressing complex propagation graphs into informative subgraphs for fake news detection. This alleviates information overload and modality mismatch, enabling frozen LLMs to perform structure-aware veracity reasoning in zero-shot and few-shot settings. A graph in-context learning strategy retrieves semantically and structurally similar demonstrations, and extensive experiments show substantial improvements over frozen LLM baselines in data-efficient settings.

Takeaway: Evolving graph structures specifically tailored to LLM reasoning is a clever reframing of the propagation-network paradigm—it sidesteps GNN supervision while retaining structural evidence.

4. RankGround: Efficient High-Resolution GUI Grounding via Lightweight Reranker-Guided Crop Selection

Authors: Liyang Fan, Xinping Bi, Yitai Li, Shuaimin Li, Hui Li et al. | Categories: cs.CV, cs.CL, cs.HC Link: arxiv.org/abs/2609.18690

RankGround is a two-stage framework achieving accurate GUI grounding with a single VLM call per query, using GroundRanker—a lightweight multimodal reranker that selects the most promising crop from a dense candidate set. Ranking supervision data is constructed from existing grounding datasets using strict containment criteria and boundary-aware augmentation, with a two-stage curriculum (pointwise then listwise) refining crop selection. Results show 1.4× faster inference and 5.5% average localization accuracy improvement over the second-best method across backbones and screen scales.

Takeaway: Solving the accuracy-efficiency trade-off in GUI agents via learned crop ranking is a practically important contribution—single-VLM-call grounding could meaningfully reduce the cost of building multimodal interface agents.

5. FRAUDSkill: Structured Frozen-Weight Skill Optimization for Audio Anti-Fraud Detection

Authors: Chengxian Hu, Zhiming Ma, Mingjun Pan, Yifan Wang, Shun Zhang et al. | Categories: cs.SD, cs.CL Link: arxiv.org/abs/2609.18766

FRAUDSkill is a structured frozen-weight adaptation framework that keeps the underlying audio-language model unchanged while optimizing an external layer of skill programs, route-specific policies, and decision rules for anti-fraud detection. Structured output control combined with validation-guided multi-path inference ensures protocol-compliant predictions across service-scenario identification, fraud detection, and fraud-type classification. On the TeleAntiFraud benchmark, it achieves 73.50% Macro-F1—outperforming the frozen-model baseline by 31.96%—while reducing invalid outputs to just 1.94%.

Takeaway: A strong demonstration that external skill programs can outperform fine-tuning for evolving, policy-laden deployment scenarios, where fraud patterns and labeling rules change faster than weights can be retrained.

6. Dreaming the Sound of Contact: Leveraging Video and Audio Generation for Zero-Shot Force-Aware Manipulation and Data Generation

Authors: Guanhua Ji, Tianyu Li, Dayoon Suh, Yuqian Zhang, Boyan Zhang et al. | Categories: cs.RO, cs.AI Link: arxiv.org/abs/2609.19137

This work augments generated video with generated audio to shape bounded, time-varying desired-force profiles using the loudness of contact sounds, addressing the missing force information in purely kinematic video-generated trajectories. A pipeline derives motion trajectories and force profiles from structured natural-language prompts, executed on a Franka Panda robot with a closed-loop force regulator. The approach succeeds on contact-rich tasks where kinematic-only baselines fail, and also serves as a data generation engine for training closed-loop policies.

Takeaway: Using generated contact sound as a proxy for force is an inventive cross-modal trick—it unlocks contact-rich manipulation from video generation models that were never designed to reason about physical interaction.

7. Rank and computation of the pathlifting Jacobian of a DAG ReLU network

Authors: Manon Verbockhaven | Categories: stat.ML, cs.LG Link: arxiv.org/abs/2609.18682

This paper provides a self-contained proof of the rank of the pathlifting Jacobian of a DAG ReLU network via induction on the number of hidden nodes, using the network’s skeleton matrix—a sparse encoding of network paths—as the key tool. The intermediate propositions link pathlifting, its Jacobian, network parameters, and the skeleton matrix, yielding a computation method without backpropagation that is reported to be super efficient in practice. A Python module implements the propositions for feedforward networks and experimentally quantifies the computational gains.

Takeaway: Theoretical results on ReLU network Jacobians are rare and valuable, and a backpropagation-free computation with practical speedups makes this worth attention from both theory and systems communities.

8. Echo: Learning-based Matching Decompilation using Trusted Back Translation

Authors: Jun Bi, Xiangxin Fang, Aarsh Chaube, José Wesley De Souza Magalhães, Rodrigo C. O. Rocha et al. | Categories: cs.CR, cs.AI Link: arxiv.org/abs/2609.18706

Echo is a matching decompilation system that uses compilation not only for verification but as trusted feedback to guide iterative search, addressing trust issues in neural decompiler outputs. It generates candidate programs and compilation configurations with a domain-specific model, measures assembly-level similarity, synthesizes promising code-configuration pairs, and repairs remaining mismatches via rule-based, neural, and reasoning-based refinement. Echo produces 2.43× more exact matches than the strongest baseline on function-level benchmarks and matches 2.75× and 7.4× as many functions as GPT-5.6 and Codex respectively on the Mirai malware binary.

Takeaway: The insight that compilation can serve as a trusted feedback signal—not just a final check—is what makes Echo notable, and the margins over frontier code models on real malware are striking.

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