Single Transformer layer training can match full-parameter RL, paper says
RisingSayak · x · 2026-07-21
A new paper argues that reinforcement learning gains in Transformers may come disproportionately from only a few layers.
- The authors test whether training a single Transformer layer can match full-parameter RL training.
- Across seven models, they find a recurring pattern: layers around 40%–60% of depth contribute the most.
- In the figure shown, selectively training the most useful layers can match or surpass full RL training on math benchmarks, sometimes with fewer trainable parameters and better performance than full-parameter RL.
More from Research
- Anthropic masterclass spotlights how to build and observe AI agents — _jaydeepkarale · 2026-07-21
- NeurIPS 2026 workshop calls papers on on-device intelligence — YiMaTweets · 2026-07-21
- AI Security Institute says every tested model tried to cheat in cyber evaluations — connoraxiotes · 2026-07-21
- AI companies are buying old books to avoid training on AI-generated slop — CackleRooster · 2026-07-21
- Sakana says multiple diffusion models plus MCTS beat test-time scaling on coding and math — SakanaAILabs · 2026-07-21
- Soofi S 30B-A3B releases a full pretraining report and claims open-model leads in English and German — abursuc · 2026-07-21