GRIP: Parameter interpolation guided by fine-grained rewards

简体中文

Deterministic mechanism diagnostic, not the original benchmark.

Paper information

Field	Details
Paper	arXiv 2608.25583
Institution	Peking University (first-listed first-author affiliation).
First public release	2026-08-26 (arXiv v1)
Author code	The paper uses a public training framework, but no method-specific author repository was found as of 2026-08-28.
Local adapter or method key	grip
Local reproduction code	src/auto_research/post_training/latest_20260827.py

Paper method and evidence

Replace a global blend between thinking and instruction weights with reward-driven layer/parameter interpolation to keep reasoning accuracy while shortening output.

flowchart LR
  T[Thinking weights] --> I[Fine-grained interpolation]
  S[Instruction weights] --> I
  R[Rewards] --> I

Original paper figure

The original figure shows reward-driven interpolation. From the original paper; copyright remains with the authors.

Core equation

\[ \theta_{GRIP}=\alpha(r)\odot\theta_{think}+[1-\alpha(r)]\odot\theta_{inst}. \]

Reported results

The paper reports mean accuracy 76.5 across five reasoning benchmarks, matching or beating the thinking model at 73.0% of its token usage. No industrial A/B.

Local reproduction

auto-research post-train --algorithm grip --dataset arithmetic-smoke --steps 100 --seed 42 --offline

Arithmetic-smoke, 100 steps: 0.1250→0.2344 (+87.50%). Interpolate candidate policy probabilities, not LLM weights. Metrics, batch index. No original checkpoint, private data, external service, or full benchmark.