Xiaohongshu's HintTuning: 1K samples + one SFT cuts reasoning tokens 31.5%
小红书技术REDtech · wechat · 2026-10-22
Xiaohongshu's NeurIPS 2026 paper HintTuning tackles overthinking in reasoning models: on problems both variants solve, Thinking models output 5-8x more than same-family Instruct models; and feeding sliced CoT prefixes as hints lets Instruct models solve problems with only 27.5% of steps and 11.9% of tokens — most tasks hinge on the first few reasoning steps.
The method proposes a "Minimum Effective Hint": split CoTs at reflection words (wait, actually...), probe an Instruct model with prefixes to find the fewest segments needed to answer correctly, then bin 1,000 problems (s1K) into no-think/short-think/full-think states and run one standard SFT (2-10 hours on 8x H20). On Qwen3-4B-Thinking and DeepSeek-R1-Distill 7B/14B/32B, math benchmarks average 31.5% fewer reasoning tokens (24-66% per benchmark) at nearly equal accuracy; on AMO-Bench, Qwen3-4B rises from 10.3 to 20.5. Ablations show wrong trajectories still help training, and a model's own traces transfer better than a stronger model's. Paper, code, models and data are open-sourced.
More from Models
- ChatGPT web can actually listen to audio, but the same model can't via API or Codex — Possible_Statement84 · 2026-10-23
- MiniMax M3.1 Flash free for 3 days on 0G, with 1M context and 512K output — MiniMax_AI · 2026-10-23
- TLA+ expert corrects Grok: TLC only exhausts finite state spaces, not all behaviors — tianyin_xu · 2026-10-23
- Open LLM cost calculator shows Arabic tokenization and 6x price gaps across models — Moham-saed · 2026-10-23
- OpenAI blocks manual reset with 13% usage left: 'no reset needed right now' — AccomplishedSugar490 · 2026-10-23
- Meta's 816B Muse Spark Open scores 39, losing to 180B Qwen3.8-Flash-Next at 40 — BTA_Labs · 2026-10-23