LoCo: local preference learning cuts reasoning tokens by up to 45% without accuracy loss
pmttyji · reddit · 2026-11-04
The paper proposes LoCo, targeting overthinking in large reasoning models that keep deliberating after sufficient information is gathered. Unlike trajectory-level compression methods, LoCo models the local decision of whether to continue:
- Detects candidate overthinking states from the model's own traces
- Freezes reasoning prefixes and regenerates shorter verified continuations from the same states
- Trains on state-conditioned preference pairs with DPO, teaching when to continue vs conclude rather than just shortening trajectories
- Training data built entirely from self-generated math traces
Results (5 math + 3 code benchmarks):
- Completion length cut 41.7%–45.4% on math and 23.4%–36.3% on code, with accuracy maintained or improved
- Best AUC on all code benchmarks; math-only training transfers effectively to code generation
Resources: arXiv 2611.04814; code and LoCo-7B weights on GitHub/HuggingFace.
Related event: Tencent's LoCo Teaches Reasoning Models When to Stop Thinking(2 posts)→
More from Research
- TLAPS-Bench and agentic AI surface a read-consistency bug in SlateDB's WAL spec — tianyin_xu · 2026-11-05
- CellType claims 27B biology model beats all 12 frontier models at predicting cell behavior — alifcoder · 2026-11-05
- No-Prerequisite Information Geometry Tutorial Released With Video and PDF Notes — FrnkNlsn · 2026-11-05
- structRFM: structure-guided RNA foundation model trained on 21M pairs lands in Nature — EricTopol · 2026-11-05
- Yandex open-sources Yambda-5B, a 4.79B-interaction music recommendation dataset — teortaxesTex · 2026-11-05
- Adversarial speculative decoding boosts long-horizon agent success with 0.06% params — kastnerkyle · 2026-11-05