Tencent paper: continuously harder task environments beat co-evolution, +8.6pp on Terminal-Bench

rohanpaul_ai · x · 2026-09-09

A new Tencent paper argues that as agents improve, training tasks cannot stay still: continuously harder environments outperform co-evolution.

Key insight: synthetic terminal tasks eventually become too easy; once solved reliably, they stop providing useful RL signal.
Method: instead of building tasks around model failures, evolve the tasks themselves—less familiar setups, rarer required skills, more steps—verifying each harder version before introducing it.
Results: on Terminal-Bench 2.1, Qwen3.6-27B hit 71.5% vs 62.9% with co-evolution; Qwen3.6-35B-A3B hit 64.9% vs 55.1%. Tested with Hy4 preview, Claude Opus 5, and GPT-5.6 Sol.

More from coding & agent