Reproduce “Beyond Confidence: Stability-Aware Test-Time Adaptation for LLM Reasoning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Bincheng Gu, Min Gao, Zongwei Wang, Yibing Bai, Yulan He, Junliang Yu
TASCO improves unlabeled test-time adaptation for LLM reasoning by optimizing a continuous prefix for both high confidence and local stability under prefix perturbations. Across mathematical and scientific reasoning benchmarks, it raises accuracy by up to 17.2 percentage points, reduces generated-token counts, and improves answer consistency without updating the underlying model.
Key insights
- Confidence and stability provide complementary signals: minimizing confidence alone can encourage brittle decisions, while reducing local confidence variation discourages fragile reasoning states.
- For small perturbations, the random-perturbation stability term is related to gradient norm and local curvature; SAP instead approximates worst-case confidence loss around the learned prefix.
- Stability correlates with correctness but is not a correctness guarantee. The method’s strongest gains occur on harder problems, and learned prefixes generally transfer across mathematical datasets.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.