New RL Trick: Teaching Models to Save Reasoning Tokens
andrey_kurenkov · x · 2026-07-17
Thinking Machines Inkling proposed a practical Reinforcement Learning (RL) trick: incorporating a penalty term into the optimization objective, specifically Reward = Task Reward − λ × (number of reasoning tokens), alongside task success rate.
By adjusting the λ value across different rollouts and pairing it with varying effort instructions, the model learns to treat "reasoning" as a resource to be conserved, rather than endlessly maximizing its thinking length.
More from Research
- Deep Learning Course Adds Mechanistic Interpretability Lecture This Winter — CSProfKGD · 2026-10-26
- INCLUDE-2.0 benchmark evaluates regional knowledge across 91 languages and 86 countries — ABosselut · 2026-10-26
- CoRL paper: flow-matching VLAs can run 1-step Euler inference with no retraining — kaixhin · 2026-10-26
- CoVer at EMNLP 2026: Training Code LLMs to Generate and Self-Verify Simultaneously — algo_diver · 2026-10-26
- Robotic dexterity should be judged on speed, force control and tool use, not copying humans — DJiafei · 2026-10-26
- Turing-complete malware can run in PyTorch's backward pass via a zero-valued loss — retr0jirachi · 2026-10-26