Large language models (LLMs) often generate infeasible or incorrect solutions in long-horizon planning tasks. This paper proposes a symbolic feedback-driven iterative self-refinement framework that uses natural language prompting, a symbolic verifier, and a plan recognizer to significantly improve the feasibility and correctness of LLM planning, enhancing system robustness and reliability.

This paper proposes Tree of Evidence (ToE), a hierarchical evidence reasoning framework for automated fact-checking that models each claim as a dynamically expanding argument tree. ToE integrates a reinforcement learning-driven multi-source retrieval agent, an evidence evaluation agent, and an argument tree aggregation algorithm to iteratively decompose, retrieve, and verify claims through an explainable evidence chain. Theoretical analysis guarantees convergence of the retrieval policy, and experiments show improvements of 4 to 24 percentage points over competitive baselines, especially on adversarially poisoned inputs.

Explicit reasoning does not necessarily improve multimodal emotion recognition accuracy but makes predictions more interpretable. Fast thinking (direct answers) improves recall, while slow thinking (deliberative reasoning) favors precision. MER-R1 is a reinforcement learning framework that jointly optimizes both through dual-objective disentanglement and confidence calibration, achieving state-of-the-art performance.

DysLexLens is an end-to-end, evidence-traceable low-resource LLM framework that analyzes dyslexic learners' experiences with AI tools by mining Reddit discussions. It employs dictionary-driven filtering, knowledge-graph reasoning, quantitative metrics, and qualitative validation to extract meaningful insights from noisy social media data.

The paper introduces ODYSSEY, a categorical framework for building verifiable, local truth-preserving foundation models by composing 'foundries'—modular architectural components that specify local contexts, representation families, restriction maps, gluing rules, obstruction policies, update obligations, and human-facing views. Universal Foundry Learning formalizes construction via left and right Kan extensions, while Foundry SQL provides a query surface using TICKET certification. The system is fully implemented and tested across diverse foundries, and will be presented as a 2.5-hour tutorial at ICML 2026.

This paper proposes a unified three-stage training paradigm to enable large language model (LLM) agents to internalize a world model for foresight planning. By addressing the format-capability gap through World Model Agentic Mid-Training, Format-Eliciting SFT, and Foresight-Conditioned Reinforcement Learning, the approach outperforms baselines on search and mathematical reasoning tasks.

A new study investigates how manipulating personality traits like agreeableness in LLM agents affects team performance across different tasks. Results show that while low agreeableness causes adversarial communication, its impact on objective outcomes depends on task structure: negligible in coding but detrimental in open-ended collaboration and bargaining.

Inspired by the development of the Internet, this paper proposes the concept of a world-wide AI-model network (AI-ModelNet) to enable interconnection, capability sharing, and collaborative reasoning among AI models, addressing the high training costs and deployment complexities of large models and the bottleneck of heterogeneous model collaboration. The paper reviews current single-model and multi-model research, presents the system vision and hierarchical architecture, validates feasibility through a prototype system and diverse applications, and discusses future research directions.

Gemma 4 is now significantly faster in Ollama 0.31 on Apple Silicon via multi-token prediction (MTP), powered by MLX. Performance improves up to 90% on coding-agent benchmarks.

Simon Willison shares Jon Udell’s critique of the phrase “human in the loop,” which Udell argues cedes authority to machines. Instead, Udell proposes “agent in the loop,” where agents are invited into human-led processes, not the reverse.