Test‑time adaptation offers a lightweight alternative to costly post‑training for enhancing Large Language Models (LLMs) on downstream reasoning tasks. Existing methods rely on predictive entropy as an internal signal, steering the model toward higher‑confidence reasoning states without external verifiers or reward models.
Higher confidence, however, does not guarantee correctness; LLMs can remain overly confident while following erroneous reasoning paths. We observe that confidence is more likely to be correct when it stays stable under local perturbations. Building on this insight, we introduce Test‑Time Adaptation via Stability‑Aware Confidence Optimization (TASCO), a framework that incorporates local stability into confidence‑driven adaptation while keeping the LLM frozen.
TASCO operationalizes local stability by optimizing a lightweight task‑level prefix under two perturbation strategies:
- Random Perturbation: Generates multiple reasoning trajectories from nearby perturbed prefixes and encourages distributional consistency across them.
- Sharpness‑Aware Perturbation: Targets worst‑case local sensitivity, reducing the model’s over‑reaction to tiny changes.
Experiments across diverse LLMs and reasoning benchmarks demonstrate that TASCO improves both reasoning accuracy and token efficiency. Behavioral analyses reveal that TASCO maintains stable confidence under local perturbations without prematurely collapsing the predictive distribution.
Review: By embedding local stability into confidence optimization, TASCO boosts LLM reasoning reliability without modifying model weights, presenting a promising direction for lightweight test‑time adaptation.