Tuesday 15 September 2026Paper No. 1,386

Evening EditionRSS

ZOTPAPER

News without the noise

RL for Agents

Reinforcement learning for agentic behaviour

22 articles

All Agent Harnesses Tools & MCP Multi-Agent Systems Creative Agents Memory & Planning Safety & Evals RL for Agents AI & ML

Viewgrid list compact

Enumerating Tool Choices Beats Sampling Them in Genomic Reasoning

Synthetic rewards train agents to diagnose simulated advertising anomalies

Adaptive rollout trees broaden language models’ mathematical reasoning coverage

Factorized reinforcement learning improves vision-language models’ spatial reasoning

GRPO Can Reward Lucky Guesses as If They Were Reasoning

Dynamic rubrics improve credit assignment for long-horizon agent training

Distillation Before Reinforcement Learning Improves Reasoning Model Post-Training

Sparse verifier feedback makes broad credit assignment outperform turn targeting

Agents learn selectively from imperfect vision-language model teachers

Five-level ladder maps reasoning models beyond direct human oversight

Privileged supervision improves action-level credit for language-model agents

Paper-derived rubrics improve AI generation of scientific research plans

Shared-prefix training accelerates reinforcement learning for hybrid-attention agents

Contrastive branch training improves credit assignment for tool-using language models

Search agents improve longer when their critics evolve alongside them

Local milestones improve credit assignment for long-horizon language-model agents

New framework lets you train AI agents inside the same harness systems they use at inference

A lightweight PyTorch-native framework matches Megatron-based agentic RL training performance.

Dense credit assignment from gold-answer log-probabilities improves long-horizon agent RL

RL with context compaction trains long-horizon agents

RL with live server state boosts multi-step tool use in LLMs

Evolutionary training harness co-evolves with LLM policies for RL