AutoThink cuts reasoning traces up to 65% while boosting decision accuracy by 11.5–18.4 points
jiqizhixin · x · 2026-11-23
Researchers from StartLux, the University of Hong Kong and SUSTech present StartLux-Decision in an arXiv preprint, introducing AutoThink: estimate decision probabilities first, invoke explicit reasoning only for uncertain cases, then combine initial and post-reasoning predictions.
- Reasoning traces up to 65% shorter while largely preserving decision accuracy
- On DecisionPivot, AutoThink improves pair accuracy over Instant variants by 11.5–18.4 percentage points
- Trained on Qwen3.5 (0.6B–27B): full-parameter SFT, then adapters with reasoning supervision + RLCD²; the RL objective rewards calibrated probabilities via proper scoring rules and softly penalizes overlong traces
- Inference uses Read-or-Think: Instant gives one-pass predictions, AutoThink reasons only when uncertain
More from Research
- How computer enumeration became a legitimate proof aid, from 1959 to 1976 — moarbugs · 2026-11-23
- Full MICCAI talk: AI and data scaling laws for mapping human brain diseases — PTenigma · 2026-11-23
- ENIGMA Q&A: Why Open Datasets Made Alzheimer's AI Classification Nearly Solved — PTenigma · 2026-11-23
- 25 Ways to Make VLA Robots Faster, Mapped From Model to Server — TheTuringPost · 2026-11-23
- A website shows a coordinated AI agent fleet attacking P vs NP live — kimmonismus · 2026-11-23
- AI finds formal flaws in top economics papers, echoing long-standing critiques — elonmusk · 2026-11-23