Zhejiang University and Tencent’s STEER targets entropy collapse in RL training
jiqizhixin · x · 2026-07-24
Zhejiang University and Tencent propose STEER to fix entropy collapse in RL for LLMs
Researchers from Zhejiang University and Tencent introduce STEER, a method aimed at a common failure mode in reasoning training: entropy collapse.
Instead of using a heuristic to blindly adjust entropy, STEER estimates how entropy changes and then reweights tokens adaptively. The authors claim this targets a key flaw in RL for LLMs more directly than prior interventions.
According to the post, the method outperforms state-of-the-art baselines on:
- 6 math-reasoning benchmarks
- 3 coding benchmarks
The paper is titled Rethinking Entropy Interventions in RLVR: An Entropy Change Perspective and the code is open-sourced.
More from Research
- MissionReport benchmark: only 3 of 540 AI-generated government reports pass all criteria — Exp_Mark · 2026-10-27
- Finch 1: a geospatial foundation model trained on imagery from 50+ sensors — Scobleizer · 2026-10-27
- Skyfall AI's SKY-001 world model turns sketches into CAD builds in 3.9s, vs 199s for GPT-6 Astra — rohanpaul_ai · 2026-10-27
- Physicists Let AI Pick Its Own Problem—It Wrote a Publishable Paper in 2 Hours — burny_tech · 2026-10-27
- Researcher flags unresolved data transparency issues in compositionality evals — najoungkim · 2026-10-27
- Separate Forms from Estimators: Why Steer by Difference of Means When You Can Use SGD? — johnhewtt · 2026-10-27