The Morning
From the arXiv
Joint Optimization of Tool Creation and Use for Large Language Model Agents
his paper introduces SMITH, a reinforcement learning framework that jointly optimizes LLM agent tool creation and usage within a single policy. By training tool building and invocation together, and using independent reward signals for schema, code, and outcome correctness, SMITH enables agents to learn to create and effectively use tools. This approach significantly improves performance on procedural reasoning tasks compared to existing methods.
On-policy Distillation with Verifiable Reward
This paper proposes **OPDVR**, a novel method that seamlessly integrates on-policy distillation (OPD) and reinforcement learning with verifiable rewards (RLVR) for post-training large language models. The core innovation is reformulating OPD's implicit reward …
Parason: Revealing Subtask and Trial Parallelism in LLM Reasoning
This paper introduces Parason, a method that identifies and leverages two types of parallelism in LLM reasoning: Subtask Parallelism (breaking down a task) and Trial Parallelism (parallel exploration of hypotheses). The key contribution is demonstrating that T…
Recursive Experiential-Working Memory Evolution for Long-Horizon Agent Harnesses
This paper introduces Recuris, a novel architecture designed to overcome challenges in long-horizon tasks by employing a recursive Experiential-Working Memory. Recuris uses Working Memory to track task progress and guide skill selection from Experiential Memor…
The Invisible Editorial Layer: Formalizing Undisclosed Inference-Time Steering, Probability Placement, and the Attribution Problem in Deployed Language Models
This paper introduces the concept of an "invisible editorial layer" in deployed language models, where undisclosed interventions modify the model's probability distribution at inference time. This layer allows for systematic steering of generated text, creatin…
SkillForge: Evolving Verifiable Skills for Reinforcement Learning Agents
SkillForge addresses the limitation of episodic reinforcement learning agents by introducing a framework for continuous skill evolution. It allows skills to be verified and refined…
CAFE: Self-Improving Search Agents Need Co-Evolving Feedback
CAFE introduces a novel framework where a single model acts as both a search agent and a feedback critic, enabling self-improvement. The core method involves the agent learning to …
Confident at the moment of action: belief miscalibration in LLM play under hidden information
This paper investigates whether Large Language Models (LLMs) accurately reflect their knowledge when making decisions under uncertainty. The core method involves testing LLMs in a …
EviDx: Evidence-Aware Active Diagnosis with Scaffolded LLM Agents
EviDx is a novel framework for active clinical diagnosis using LLM agents. It addresses the limitations of static LLM diagnosis by creating interactive, evidence-seeking environmen…
Meta$^n$: Recursive Self-Improvement through Emergent Depth
Meta$^n$ achieves recursive self-improvement by repeatedly applying a fixed meta-operation ($Ω$) to its own outputs, rather than modifying the operation itself. This allows for eme…
The Town Square
Apple has unveiled its new M6 and M5 Ultra chips, promising significant advancements in performance and AI capabilities for its devices.
Workshops
This repository serves as a read-only mirror of a community plugin marketplace for Claude Cowork and Claude Code, allowing users to discover and access plugins for enhanced functionality.
This AI job search framework runs locally, using Claude to evaluate job postings, tailor CVs, write cover letters, and prepare for interviews.
