EAU: co-evolving environment, agent and user lifts multi-turn RL, +36.7% on ColBench
xiye_nlp · x · 2026-10-22
University of Alberta researchers introduce EAU, a meta-agent framework for environment–agent–user co-evolution. Instead of treating the user simulator and environment as fixed, a meta-agent monitors trajectories, tool traces and rewards, diagnoses failures, and hot-swaps validated changes to the simulator, environment, reward design and RL config mid-run, with improvements persisting across runs.
- On three interactive benchmarks and two backbone models, EAU consistently beats prior RL pipelines: overall score up 14.7–15.0% over the strongest baseline, and up to 36.7% higher successful task completion on ColBench.
- The user/environment dynamics use an editable symbolic representation while the agent remains a neural LLM optimized with RL.
More from coding & agent
- Replacing LLM routing calls with 15ms CPU encoders: workshop on deterministic routing — camerongreen95 · 2026-10-22
- Cloudflare's security-audit-skill hits 35.7k stars, turns coding agents into auditors — Pytlicek · 2026-10-22
- Running Cursor Agents from a Shared Notion Board: Welcome to the Software Factory — nbaschez · 2026-10-22
- Multiplayer AI Manifesto team scores Claude Tag, Dust, QM, Viktor and Superconductor across 18 weighted criteria — sergeykarayev · 2026-10-22
- 7 prompts to make Claude build LBO models like a $180k investment banking analyst — alex_verem · 2026-10-22
- Oblivion Remastered mod built with Claude AI boosts FPS by 37% and slashes stuttering — PikaPikaDude · 2026-10-22