rl

This paper shows that language agents fail to reliably connect past actions with their outcomes when using interaction history, and proposes a simple outcome-labeling annotation plus a learned calibrator that selectively records experience, improving task success and reducing action repetition.

The tweet highlights the benefits of working in a small AI lab (under 20 people) where there are no divisions over pretraining, SFT, RL, agents, or evaluation, and no internal politics, with everyone being responsible and aware.

The article mentions that Jerry Tworek, former VP of RL at OpenAI, suggested user data could influence the solution of Navier-Stokes problems.

The tweet highlights the value of understanding different AI fine-tuning methods like SFT and RL, their impact on model behavior, and recommends using mathematical analysis and visual aids such as ASCII art to grasp these concepts.

The River API was tested and outperformed Tinker on reinforcement learning runs with identical training code, featuring improvements in details like routing replay and being open-sourced with ablations.

Prime RL now supports expressing and training multi-agent systems, enabling use cases like agentic judge, self-play, user simulation, and complex agent collaboration.

KAT-Coder-V2.5-Dev is an open-weight MoE coding model with 35B total parameters (3B active), achieving state-of-the-art results on agentic coding benchmarks through SFT and RL training.

An open-source end-to-end post-training recipe covering SFT, RL, and test-time scaling with models, datasets, and code.

This paper analyzes evaluation and design paradigms in deep reinforcement learning, demonstrating that canonical paradigms can lead to incorrect conclusions and providing insights into scaling, capacity, and complexity.

This article explains vLLM's weight syncing API for reinforcement learning, covering how it facilitates weight updates and KV cache recompute in RL training, with a focus on reducing complexity for training frameworks.

Microsoft released FastContext-1.0, a lightweight repository-exploration subagent for LLM coding agents that reduces main-agent token consumption by up to 60% while improving resolution rates by up to 5.5%.

Kyutai Labs released a new paper on using reinforcement learning to post-train speech models (Moshi and PersonaPlex) for more human-like interaction, including when to respond, wait, or give listening cues.

A compilation of reinforcement learning algorithm interview questions curated by @sheriyuo, shared by @arjunkocher.

This paper introduces DRPO, which replaces the hard mask in DPPO with a smooth advantage-weighted quadratic regularizer to improve stability and efficiency in LLM reinforcement learning by providing continuous gradient corrections beyond trust-region boundaries.

Introduces MosaicLeaks, a benchmark of 1,001 multi-hop deep research tasks that chain private enterprise documents with public web queries to evaluate privacy leakage. Finds that models leak sensitive information at multiple levels, and proposes PA-DR, a reinforcement learning framework that reduces leakage while improving task accuracy.

John Schulman speculates that inoculation prompting during RL training could inadvertently make models better at sandbox escapes and hacking.

This week, 9 new records were added to the autoresearch ecosystem, bringing the total to 383, covering multiple open-source tools and projects such as the AutoResearch-RL reinforcement learning framework, lance-autoresearch database kernel optimization, and Clio prediction market backtesting framework.

Poolside is hosting a 2-day model research hackathon in London to push an open-weight agent model further using RL and fine-tuning on Laguna XS.2, with partners including NVIDIA, Prime Intellect, and Hugging Face, and a prize of an NVIDIA DGX Spark.

Teknium observes that the Hermes agent initially behaves inefficiently but gains large efficiency boosts after solving a task once, likening it to "linearized RL."