- Who it helps
- Researchers studying efficient reasoning and model training.
- Availability
- November 3 paper and public implementation are available. The repository links a model checkpoint; its contents were not tested.
The training idea
Local preference learning compares continuations from matched reasoning states. The training uses self-generated mathematical reasoning and evaluates transfer to coding.
What the reported savings mean
The authors report shorter completions while maintaining or improving accuracy on their tests. Fewer generated tokens do not directly establish lower total training or deployment costs. The README specifies Python 3.12 and CUDA 12.8 and describes development on H20 hardware; that is not a verified minimum hardware requirement.
LoCo trains reasoning on preferred local continuations ↗LoCo implementation ↗
What you can do
- Read the training requirements and evaluation conditions before planning a reproduction. Open reference ↗
Where the evidence stops
Paper abstract and repository documentation read November 4. Training, model loading and benchmark reproduction were not executed.
Sources read. Primary sources read November 4, 2026 UTC. No installation, model inference or complete workflow executed.
Sources & further reading
Paper · paper
LoCo trains reasoning on preferred local continuations ↗
Identifies the subject, documented capability or specific verification limit.
Muyao Yuan and coauthors · page checked 2026-11-04
GitHub · official
LoCo implementation ↗
Identifies the subject, documented capability or specific verification limit.
Muyao Yuan and collaborators · page checked 2026-11-04
Coverage note: Independent attention and end-to-end execution remain unverified.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
LoCo aims to shorten reasoning by learning which continuation to prefer at an intermediate reasoning state.