Beyond Confidence: Stability-Aware Test-Time Adaptation for LLM Reasoning

2609.11393 ↗ · 2026-09-10 · cs.AI

Bincheng Gu, Min Gao, Zongwei Wang, Yibing Bai, Yulan He, Junliang Yu

The finding

Headline claim

Optimizing a lightweight continuous prefix for confidence stability, not just low entropy, at test time improves frozen-LLM reasoning accuracy and token efficiency without training data or model updates.

The test

The authors evaluate TASCO (two variants: Random Perturbation and Sharpness-Aware Perturbation) against zero-shot CoT and four test-time adaptation baselines (SLOT, LatentSeek, LTPO, TTSV) on five reasoning benchmarks (MATH500, AMC23, Minerva, AIME24, GPQA) across LLMs from three families: Qwen2.5-Math-1.5B, Qwen2.5-Math-7B, and Llama3.1-8B-Instruct. The core empirical observation is that queries grouped by low trajectory-confidence variance have higher accuracy than high-variance groups, even controlling for mean confidence.

Effect size & scope

On Qwen2.5-Math-7B, TASCO-SAP reaches 48.4 average accuracy versus 31.2 for zero-shot CoT (+17.2 points) and 43.6 for the best baseline TTSV (+4.8); TASCO claims up to 17.2% accuracy gains and up to 28.1% fewer generated tokens. Gains appear consistently across both model scales and both perturbation variants, plus generalization to unseen datasets with the same optimized prefix. Numbers are shown fully only for the 1.5B and 7B models; the Llama3.1-8B table is truncated in the provided text.

Limitations

The authors themselves concede the theory characterizes robustness rather than correctness, relying on an empirical correlation between stability and accuracy. The method requires K perturbed rollouts per optimization step (Random Perturbation), adding inference-time compute that the abstract does not cost out against token savings. Optimization happens on the unlabeled test set itself, raising transduction concerns. The evaluation covers only small (1.5B to 8B) math/code-oriented models on five benchmarks; no frontier-scale model, no multi-step agentic tasks, and the accuracy claim of 'up to 17.2%' is a best-case, not an average.

Counterpoint

Baselines are competing test-time-optimization methods rather than simple strong alternatives: self-consistency with majority voting is not reported, and TTSV already gives +11.9 to +12.4, so the marginal gain of stability-awareness over good verifier-free optimization is roughly 3 to 5 points. Entropy minimization plus prefix tuning is prompt-engineering-adjacent; the learned prefix may exploit dataset-specific artifacts (the same test set is used for optimization), and 'cross-distribution generalization' needs scrutiny for contamination. The stability-accuracy correlation could partly reflect difficulty (easy queries are both more stable and more accurate).

Economy implication

Little or nothing yet at economy scale. The practical signal is that verifier-free, frozen-model adaptation methods keep getting cheaper and better, which would push inference cost per unit of reasoning capability down at the margin; but this is a small-model academic result, and the compute cost of perturbed rollouts during optimization is unquantified, so no near-term staffing or capex conclusion follows.

Who this matters for

Technical Capability: Researchers in test-time adaptation get a new, theoretically grounded label-free signal (local confidence stability) that beats entropy-only objectives on small reasoning models.

Market & Valuation: Weakly relevant: methods that improve frozen-model reasoning without post-training hint at lower adaptation costs, but the scale is too small to move any investment view.

Bears on (advisory · papers never write evidence)

[1.4] Inference-time compute yields gains that don't saturate quickly on hard reasoning.

Abstract

Test-time adaptation has emerged as a lightweight alternative to costly post-training for improving the reasoning capabilities of Large Language Models (LLMs) on downstream tasks. Predictive entropy provides a model-derived signal for such adaptation, guiding models toward higher-confidence reasoning states without external verifiers or reward models. However, higher confidence does not necessarily imply correctness, as LLMs may remain highly confident along incorrect reasoning trajectories. We observe that high-confidence reasoning is more likely to be correct when confidence remains stable under local perturbations. Based on this observation, we propose Test-Time Adaptation via Stability-Aware Confidence Optimization (TASCO), a framework that incorporates local stability into confidence-based test-time adaptation while keeping the LLM frozen. TASCO operationalizes local stability by optimizing a lightweight task-level prefix under two alternative perturbation strategies: Random Perturbation promotes distributional stability across trajectories induced by nearby perturbed prefixes, whereas Sharpness-Aware Perturbation targets worst-case local sensitivity. Experiments demonstrate that TASCO improves reasoning accuracy and token efficiency across diverse LLMs and reasoning benchmarks, while behavioral analyses show that it maintains stable confidence under local perturbations without prematurely concentrating the model's predictive distribution.