AI
Mitigating the Length-Scaling Tax with Online Distillation
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 29/09/2026, 20:00:00 UTC
Briefing created: 01/10/2026, 23:22:24 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 01/10/2026, 14:18:24 UTC.
Briefings
Researchers developed Length Self-Distillation (LSD) to mitigate the Length-Scaling Tax (LST) in reinforcement learning post-training. The method routes solved prompts to on-policy distillation while keeping the original RL objective for unsolved tasks. Experiments show LSD achieves comparable or better performance than RL across various variants, significantly reducing response length growth on easy queries. On single-turn reasoning, LST dropped from 19.0% to -3.7%, and on multi-turn agentic tasks, it decreased from 31.4% to 13.7%. The approach uses an exponential moving average of the online policy as its teacher, requiring no external model.
Key points
- Length Self-Distillation (LSD) routes solved prompts to on-policy distillation while retaining the original RL objective for unsolved prompts.
- LSD reduces LST from 19.0% to -3.7% on single-turn reasoning and from 31.4% to 13.7% on multi-turn agentic tasks.
Context and limitations
The study covers single-turn and multi-turn reasoning tasks but does not specify limitations on other problem types or contexts.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 01/10/2026, 23:22:24 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeieqghlaxa2wifahbkz3hlwaie4tjdiwa7p7jdllptwrmglouca2pi