Title: Mixture of Debaters: Learn to Debate at Architectural Level in Multi-Agent Reasoning

Source: http://arxiv.org/abs/2606.29425v1

Summary:

This paper introduces Mixture of Debaters (MoD), a novel framework that embeds multi-agent dialectical reasoning directly into a single model's Mixture-of-Experts architecture. By replacing multi-model communication with lightweight expert routing and dynamic self-debate, it establishes a highly efficient and latency-optimized primitive for agentic reasoning.

Title: LLM-Guided Planning for Multi-hop Reasoning over Multimodal Nuclear Regulatory Documents

Source: http://arxiv.org/abs/2606.29399v1

Summary:

This work proposes a novel agentic reasoning framework centered on LLM-guided planning for complex, multi-hop reasoning tasks over extensive document sets, demonstrating a significant breakthrough in agent capabilities. The emphasis on planning as the dominant performance factor (+38.0pp gain) establishes a crucial paradigm for how agents can tackle real-world problems requiring sophisticated, structured inference and knowledge accumulation.

Title: Coverage-Driven KV Cache Eviction for Efficient and Improved Inference of LLM

Source: http://arxiv.org/abs/2606.29563v1

Summary:

This paper introduces K-VEC, a coverage-aware KV-cache eviction strategy that resolves performance drops in long-context tasks by maximizing token retention across attention heads and layers. By theoretically linking token coverage to mutual information, it provides a foundational efficiency breakthrough for deploying long-context LLMs under resource constraints.

Title: Flow Matching in Feature Space for Stochastic World Modeling

Source: http://arxiv.org/abs/2606.29059v1

Summary:

This work introduces FlowWM, a novel stochastic world modeling architecture that successfully performs generative flow matching directly in high-dimensional pretrained feature spaces. By incorporating a differentiable one-step projection, it establishes a new primitive for planning agents to forecast diverse, temporally consistent futures.

Title: Agent Safety Is Action Alignment

Source: http://arxiv.org/abs/2606.28739v1

Summary:

This paper redefines agent safety by identifying the category error of using chatbot refusal training for action-taking LLM agents. It establishes 'action alignment' enforced outside model weights via least privilege as the necessary paradigm to prevent the reasoning collapse of multi-step agents.

Title: Grounded Iterative Language Planning: How Parameterized World Models Reduce Hallucination Propagation in LLM Agents

Source: http://arxiv.org/abs/2606.27806v1

Summary:

The paper introduces Grounded Iterative Language Planning (GILP), a novel agentic framework that addresses LLM planning hallucinations by pairing API-based LLMs with small parameterized world models. By implementing a consistency gate that prompts the LLM to revise plans when its imagined states conflict with the world model's predictions, GILP provides a new agentic reasoning loop that significantly reduces error propagation.

Title: Drop-Then-Recovery: How Redundant Are Vision-Language-Action Models?

Source: http://arxiv.org/abs/2606.27755v1

Summary:

This paper presents an architectural efficiency breakthrough for Vision-Language-Action (VLA) models by systematically analyzing block redundancy and block sensitivity via a novel 'GateProbe' metric. It reveals that VLA language backbones are highly redundant, demonstrating that drastically reducing language block capacity can maintain or even improve downstream robotic control performance under identical fine-tuning budgets.

Title: Joint Learning of Experiential Rules and Policies for Large Language Model Agents

Source: http://arxiv.org/abs/2606.27136v1

Summary:

This work introduces JERP, a novel agentic framework that simultaneously updates an external pool of natural-language rules and the model's parametric policy from the same interaction trajectories. By keeping prompt-based rules synchronized with the evolving policy, it establishes a unified paradigm for optimizing both interpretive guidance and internal capabilities in multi-step interactive environments.

Title: Joint Learning of Experiential Rules and Policies for Large Language Model Agents

Source: http://arxiv.org/abs/2606.27136v1

Summary:

This work introduces JERP, a novel agentic framework that simultaneously updates an external pool of natural-language rules and the model's parametric policy from the same interaction trajectories. By keeping prompt-based rules synchronized with the evolving policy, it establishes a unified paradigm for optimizing both interpretive guidance and internal capabilities in multi-step interactive environments.

Title: Einstein World Models

Source: http://arxiv.org/abs/2606.26969v1

Summary:

This paper proposes a blueprint for LLM-based reasoning systems that integrates visual-temporal rollouts directly into the reasoning trace as inspectable hypotheses. By extending tool-calling into the domain of visual thought experiments, it introduces a novel framework for grounding complex physical and counterfactual reasoning beyond text alone.