acceptodds
Under review as a conference paper at ICLR 2027

Stability Makes Confidence Reliable: Test-Time Adaptation for LLM 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.

Primary area:
foundation or frontier models, including LLMs
Keywords:
TestTime Adaptation, Large Language Models, LLM Reasoning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_4to7OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Test-Time Adaptation via Stability-Aware Confidence Optimization (TASCO) for LLM reasoning. Motivated by an observed association between local confidence stability and answer correctness, TASCO optimizes a continuous prefix shared across unlabeled queries from a target task while keeping the backbone frozen. It introduces two alternative strategies: Random Perturbation (RP) combines entropy minimization with a penalty on the variance of trajectory-average token log-probabilities across generations under Gaussian-perturbed prefixes; Sharpness-Aware Perturbation (SAP) minimizes predictive entropy at a gradient-guided perturbed prefix while holding the generated trajectory fixed. The learned prefix is held fixed during final answer generation. Experiments on five LLM variants across reasoning benchmarks report improved average accuracy and shorter final generations relative to zero-shot CoT, alongside comparisons with existing adaptation methods. The paper supports its approach with ablations, behavioral analyses, and local sensitivity analysis.

Critical strengths
  1. The stability-aware perspective is supported by complementary diagnostic and behavioral analyses. Appendix B reports significant confidence-controlled associations between local stability and correctness in four of the five examined models, providing empirical motivation for the optimization objective. Behavioral analyses compare TASCO with confidence-only prefix optimization through perturbation robustness (Table 3), early decoding distributions (Figure 6), and an external LLM judge's assessment of step correctness, informativeness, redundancy, and premature commitment (Table 9). The qualitative example in Table 10 illustrates a different solution path that preserves a valid root missed by the confidence-only baseline. Together, these analyses connect the optimization design to both probabilistic and semantic aspects of generation, helping interpret the reported accuracy gains.

  2. The reported accuracy gains are supported by informative controls and transfer experiments. SAP improves average accuracy over TTSV by 2.9, 4.8, and 4.4 percentage points on the three backbones in Table 1. The unoptimized-prefix control and RP's same-rollout Perturbed Confidence control directly examine the contributions of prefix optimization and the stability objective (Figure 4). Furthermore, prefixes transferred between mathematical benchmarks improve over zero-shot CoT on all 12 evaluated source-target pairs without target-side re-optimization on Qwen2.5-Math-7B (Figure 8), supporting generalization to mathematical queries beyond the source adaptation set.

Critical weaknesses
  1. The efficiency evaluation could more fully account for adaptation cost and task-startup latency. The reported token savings concern final answer generation, while obtaining the prefix also requires confidence-only warm-up and iterative optimization. RP introduces additional perturbed rollouts, and SAP requires two gradient computations per update; distinct perturbed prefixes also limit exact prompt-KV reuse across branches. Reporting the time and compute needed to obtain a usable prefix, together with the number of queries required to amortize this cost, would clarify the practical resource trade-off. This concern applies to prefix adaptation, while final answer generation uses the fixed learned prefix.

  2. Continual adaptation in deployment could be discussed more explicitly. Real-world deployment can involve a continuing stream of queries whose task distribution evolves over time. The authors' cross-benchmark transfer experiments provide useful evidence that learned prefixes generalize across the evaluated mathematical benchmarks (Figure 8). Yet these experiments apply a fixed prefix to each target benchmark, leaving open how the prefix should be maintained as deployment queries evolve. It would be better if the authors discussed a continual adaptation strategy, clarifying when to reuse an existing prefix, update it incrementally using incoming unlabeled queries, or optimize a new prefix.

Questions for authors
  1. Adaptation cost and deployment readiness. On the same hardware, how much wall-clock time is required for confidence-only warm-up and subsequent prefix optimization for each TASCO variant? Please report the initial delay before the adapted prefix is ready and the amortized end-to-end cost, including final generation, across different query counts relative to zero-shot CoT and confidence-only adaptation.

  2. Transfer baseline. In Figure 8, how does a source-trained TASCO prefix compare with a confidence-only prefix trained on the same source queries and transferred without target-side adaptation? This comparison would help clarify the contribution of stability optimization to the observed transfer gains.

  3. Adaptation data requirements. How sensitive is TASCO to the number of unlabeled queries used for prefix adaptation? Can a prefix learned from a small subset retain its benefits on held-out queries from the same task?

4 comments

Response to Reviewer 4to7

We are grateful to reviewer 4to7 for the positive assessment of our stability-aware perspective, empirical findings, and generalization results. We also appreciate the constructive feedback and have conducted additional experiments to address the reviewer's concerns.

W1 & Q1: The efficiency evaluation could more fully account for adaptation cost and task-startup latency. The reported token savings concern final answer generation, while obtaining the prefix also requires confidence-only warm-up and iterative optimization. RP introduces additional perturbed rollouts, and SAP requires two gradient computations per update; distinct perturbed prefixes also limit exact prompt-KV reuse across branches. Reporting the time and compute needed to obtain a usable prefix, together with the number of queries required to amortize this cost, would clarify the practical resource trade-off. This concern applies to prefix adaptation, while final answer generation uses the fixed learned prefix.
Adaptation cost and deployment readiness. On the same hardware, how much wall-clock time is required for confidence-only warm-up and subsequent prefix optimization for each TASCO variant? Please report the initial delay before the adapted prefix is ready and the amortized end-to-end cost, including final generation, across different query counts relative to zero-shot CoT and confidence-only adaptation.

Response to weakness 1 and question 1: We thank the reviewer for raising concerns about adaptation costs and task-startup latency. To address these concerns, we provide three analyses: comparing per-question prefix optimization and inference times; breaking down the two optimization stages to quantify startup latency; and evaluating amortized end-to-end costs across query volumes, accounting for prefix reuse.

  • Optimization and Inference Efficiency. First, we have added an Optimization and Inference Efficiency subsection reporting prefix optimization and inference times on MATH500 and AIME24 using Qwen2.5-Math-7B on a single NVIDIA A800 GPU. Both TASCO variants undergo five epochs of confidence-only warm-up followed by five epochs of stability-aware optimization. As shown in the table below, TASCO-RP incurs 2.3–2.7 times the optimization cost of TTSV due to repeated perturbed rollouts, whereas TASCO-SAP reuses the generated trajectory with only one additional forward and backward pass, increasing optimization time by less than 5%.

    MethodMATH500 Opt. (s)MATH500 Infer. (s)AIME24 Opt. (s)AIME24 Infer.(s)
    Zero-Shot CoT--24.7--36.9
    TTSV6.4s17.68.922.1
    TASCO-RP17.113.120.115.9
    TASCO-SAP6.614.49.115.2
  • Task-Startup Latency. To further clarify task-startup latency, we report the time breakdown of the two optimization stages on MATH500. As shown below, both variants require 26.7 minutes for confidence-only warm-up, followed by 116.0 minutes of stability optimization for TASCO-RP and 28.4 minutes for TASCO-SAP. The total time to obtain an adapted prefix is therefore 142.7 minutes for TASCO-RP and 55.1 minutes for TASCO-SAP.

    MethodWarm-up (min)Stability Opt. (min)Total (min)
    TTSV——53.4
    TASCO-RP26.7116.0142.7
    TASCO-SAP26.728.455.1
  • Amortized End-to-End Efficiency. Since the learned prefix can be reused, we further examine how adaptation costs are amortized across different query volumes. On MATH500, TASCO-SAP requires approximately 55.1 minutes for initial adaptation but saves 10.31 seconds per query relative to Zero-Shot CoT, requiring about 321 subsequent queries to offset this cost. Relative to TTSV, only 32 queries are needed to recover SAP's additional adaptation cost. The table below reports the amortized total time per query, including initial adaptation and final generation, across different query volumes. As query volume increases, both TASCO variants progressively amortize their upfront costs. TASCO-RP requires more queries due to its higher optimization cost but benefits from faster inference at larger scales, whereas TASCO-SAP achieves lower overall costs with fewer queries.

    Method1003005001,0002,000
    Zero-Shot CoT24.7224.7224.7224.7224.72
    TTSV49.6528.2824.0120.8119.20
    TASCO-RP98.6641.5930.1821.6217.34
    TASCO-SAP47.4625.4321.0217.7216.06

Revised PDF: Section 4.6 ("OPTIMIZATION AND INFERENCE EFFICIENCY").

W2: Continual adaptation in deployment could be discussed more explicitly. Real-world deployment can involve a continuing stream of queries whose task distribution evolves over time. The authors' cross-benchmark transfer experiments provide useful evidence that learned prefixes generalize across the evaluated mathematical benchmarks (Figure 8). Yet these experiments apply a fixed prefix to each target benchmark, leaving open how the prefix should be maintained as deployment queries evolve. It would be better if the authors discussed a continual adaptation strategy, clarifying when to reuse an existing prefix, update it incrementally using incoming unlabeled queries, or optimize a new prefix.

Response to weakness 2: We thank the reviewer for recognizing the generalization of TASCO and for raising the question of how prefixes should be maintained as deployment queries evolve. We agree that this is an interesting direction and conduct a small prototype experiment to explore possible maintenance strategies.

We construct a query stream consisting of MATH500, Minerva Math, and GPQA, with 100 queries sampled from each dataset. Starting from a prefix optimized on MATH500, we use the first 50 queries in each stage to observe the incoming distribution and perform adaptation, and the remaining 50 for evaluation. We consider a semantic similarity-based strategy that chooses between prefix reuse, incremental update, and re-optimization according to the similarity between incoming queries and the existing adaptation data.

Query StreamFixed ReuseIncremental UpdateRe-optimization
MATH50078.074.068.0
Minerva Math32.040.030.0
GPQA30.034.038.0

The results show that reusing the existing prefix works well when the query distribution remains similar, while further updates may overfit the small set of new queries. Incremental updates are more effective for related but different mathematical problems, whereas re-optimization performs best after the shift to a more distinct domain. These findings provide preliminary evidence for selecting a prefix maintenance strategy based on changes in the query distribution.

Q2: Transfer baseline. In Figure 8, how does a source-trained TASCO prefix compare with a confidence-only prefix trained on the same source queries and transferred without target-side adaptation? This comparison would help clarify the contribution of stability optimization to the observed transfer gains.

Rseponse to question 2: We thank the reviewer for the suggestion. To evaluate the contribution of stability optimization to cross-benchmark transfer, we compare TASCO with the confidence-only baseline TTSV by directly transferring their learned prefixes to other benchmarks for evaluation. As shown below, TASCO consistently outperforms TTSV, further demonstrating the benefits of stability-aware optimization for cross-benchmark transfer.

SourceTargetTTSVTASCOΔ (pp)
MATH500AMC2356.164.4+8.3
Minerva Math29.433.4+4.0
AIME248.921.2+12.3
Minerva MathMATH50066.271.2+5.0
AMC2356.761.2+4.5
AIME2413.519.1+5.6
AIME24MATH50068.474.0+5.6
AMC2359.163.1+4.0
Minerva Math28.734.2+5.5
AMC23MATH50064.470.4+6.0
Minerva Math31.635.3+3.7
AIME2413.325.4+12.1
Overall mean41.3647.74+6.38

Q3: Adaptation data requirements. How sensitive is TASCO to the number of unlabeled queries used for prefix adaptation? Can a prefix learned from a small subset retain its benefits on held-out queries from the same task?

Rseponse to question 3: We thank the reviewer for raising this question. To evaluate the sensitivity of TASCO-SAP to adaptation data size, we sample 25, 50, 100, 200, 300, and 500 unlabeled problems from the MATH training split to optimize separate prefixes, and evaluate each prefix on the held-out MATH500 benchmark across three backbone models.

As shown in the table, performance generally improves with more adaptation queries. Notably, using only 300 unlabeled queries, all three models achieve accuracy comparable to or higher than that obtained with 500 queries. These results suggest that TASCO-SAP does not require a large adaptation set to achieve competitive performance on unseen queries from the same task.

Modeln = 25n = 50n = 100n = 200n = 300n = 500
Qwen2.5-Math-1.5B59.061.463.669.670.672.4
Qwen2.5-Math-7B45.658.864.668.275.074.0
LLaMA-3.1-8B-Instruct41.244.246.645.848.849.8

Thank you for the detailed clarfications and additional experiments. The adaptation-time breakdown and amortization analysis substantially address my efficiency concerns, while the transfer baseline and held-out evaluation strengthen the evidence for generalization. Overall, these additions clarify the method’s practical trade-offs, so I will maintain my positive score.

Dear Reviewer 4to7,

Thank you for your positive feedback and for taking the time to consider our response. Your suggestions helped us clarify the method’s practical trade-offs and strengthen our evaluation of generalization. We also appreciate your continued support of our work.

We will incorporate the additional analyses and experimental results from our discussion into the final manuscript.

Authors 35531

Reviewer_4SZfOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work proposes TASCO, a test-time soft prefix optimization normalization method, which adds both random perturbations and adversarial perturbations during prefix updating. The work is motivated by the fact that high-confidence prefixes do not always lead to correct results, but those with flatter nearby regions are more likely to be reliable. The paper conducts experiments on several benchmarks, including MATH500, AMC23, Minerva, AIME24, and GPQA, with various backbones and compares the method to representative baselines, and the gain is considerable.

Critical strengths

The work is well motivated. It points out the limitation of confidence-driven test-time optimization, which is indeed a key question for this series of works.

The experiments are solid, including several baseline methods and detailed ablation studies.

The writing is clear and easy to follow.

Critical weaknesses

The work lacks a clear comparison of learning costs with baselines, including the inference required during training and the training process. The adversarial training step may introduce additional optimization steps.

The source of the performance gain is still not clear. The Appendix mentions that baselines are implemented according to their original works, meaning that hyperparameters in this work may be different from those of the baselines. Thus, it is not clear whether the gain is from the method itself or from the tuned hyperparameters.

Questions for authors

See weaknesses 1 and 2.

As the overall idea of this work is normalization during prefix tuning, it is a natural question whether other simpler normalization methods can yield the same result, because adversarial methods are computationally costly. A natural idea is normalization through a broader data distribution: By involving more test queries, can we naturally prevent falling into unstable local regions and achieve prefixes that are more robust and generalize well, without additional normalization methods?

3 comments

Response to Reviewer 4SZf

We are grateful that the reviewer found our work well motivated and the experiments solid. We also appreciate the constructive feedback and address the points raised with additional experiments, analyses, and clarifications.

W1: The work lacks a clear comparison of learning costs with baselines, including the inference required during training and the training process. The adversarial training step may introduce additional optimization steps.

Response to weakness 1: We thank the reviewer for requesting a computational cost comparison. We have added an Optimization and Inference Efficiency subsection comparing different methods' test-time prefix optimization and final answer generation times on MATH500 and AIME24 using Qwen2.5-Math-7B. We report the average optimization and inference times per question.

TASCO-RP requires 2.3–2.7 times the optimization time of TTSV due to additional rollouts under perturbed prefixes. In contrast, TASCO-SAP reuses the original rollout, requiring just one additional forward and backward pass for the prefix update. This limits its optimization overhead to less than 5% relative to TTSV on both benchmarks. Including amortized prefix optimization and final inference, TASCO-SAP reduces the total time per question by 12.5% on MATH500 and 21.6% on AIME24, while achieving higher accuracy on both benchmarks.

MethodMATH500 Opt. (s)MATH500 Infer. (s)AIME24 Opt. (s)AIME24 Infer. (s)
Zero-Shot CoT--24.7--36.9
TTSV6.417.68.922.1
TASCO-RP17.113.120.115.9
TASCO-SAP6.614.49.115.2

W2: The source of the performance gain is still not clear. The Appendix mentions that baselines are implemented according to their original works, meaning that hyperparameters in this work may be different from those of the baselines. Thus, it is not clear whether the gain is from the method itself or from the tuned hyperparameters.

Response to weakness 2: We thank the reviewer for this important question. To address concerns about the source of TASCO's performance gains and hyperparameter fairness, we explain three aspects: first, the rationale for baseline hyperparameter configurations; second, the contribution of stability optimization through controlled ablation studies; and third, the consistency of performance gains across perturbation scales.

  • Fairness of Baseline Configurations. First, different test-time adaptation methods involve distinct optimization objectives and hyperparameter configurations. To ensure a fair comparison, we follow the experimental protocols and hyperparameter settings reported in each baseline's original study and rerun the optimization and evaluation procedures. Related studies have adopted this practice [1, 2]. We use the reported configurations to reproduce each baseline’s performance as faithfully as possible and support a fair comparison.

  • Contribution of Stability Optimization. More importantly, our controlled ablations further clarify the contribution of stability optimization. Figure 4 shows that TASCO-RP outperforms Perturbed Confidence under the same perturbation rollout budget, indicating that the gains cannot be attributed solely to additional perturbation sampling. The comparison with Stability Only further supports jointly optimizing confidence and stability.

  • Perturbation Scale Sensitivity. The perturbation-scale sensitivity analysis in Figure 5, with numerical results provided below, further supports this conclusion. At zero perturbation, the objective reduces to confidence-only optimization (TTSV), while both TASCO variants achieve higher accuracy at every tested nonzero perturbation scale across all five benchmarks. These findings indicate that the benefits of stability optimization persist across different perturbation scales rather than depending on a specific hyperparameter choice, further supporting the contribution of the proposed method.

    Table 1. Sensitivity to Gaussian Scale σ\sigma (TASCO-RP).

    σ=0\sigma=0σ=0.0025\sigma=0.0025σ=0.005\sigma=0.005σ=0.01\sigma=0.01σ=0.02\sigma=0.02
    MATH50068.669.870.470.870.2
    AMC2346.848.249.450.650.2
    AIME248.810.29.610.49.6
    Minerva20.622.121.722.421.7
    GPQA24.825.727.327.827.2

    Table 2. Sensitivity to Perturbation Radius ρ\rho (TASCO-SAP).

    ρ=0\rho=0ρ=0.2\rho=0.2ρ=0.4\rho=0.4ρ=0.8\rho=0.8ρ=1.0\rho=1.0
    MATH50068.671.270.671.869.8
    AMC2346.849.650.852.251.4
    AIME248.810.812.012.111.4
    Minerva20.621.021.721.321.0
    GPQA24.826.327.226.726.3

Revised PDF: Section 4.6 ("OPTIMIZATION AND INFERENCE EFFICIENCY").

Q1: As the overall idea of this work is normalization during prefix tuning, it is a natural question whether other simpler normalization methods can yield the same result, because adversarial methods are computationally costly. A natural idea is normalization through a broader data distribution: By involving more test queries, can we naturally prevent falling into unstable local regions and achieve prefixes that are more robust and generalize well, without additional normalization methods?

Response to question 1: We appreciate the reviewer's insightful suggestion and designed an experiment to test whether increasing the number and diversity of unlabeled adaptation queries could replace explicit stability optimization. We pooled problems from MATH, GSM8K, OlympiadBench, and College Math, clustered them using semantic embeddings, and sampled 1,000 problems across clusters. For each backbone, we optimized a shared prefix using only the confidence objective and evaluated it on five benchmarks. The original TTSV and TASCO-SAP instead followed the main setup, optimizing a separate prefix for each benchmark.

  • Broader-Distribution Prefix Optimization. As Table 3 shows, the pooled-query approach achieves lower average accuracy than the original TTSV across all three backbones, despite using a broader adaptation distribution. TASCO-SAP outperforms this approach by 3.2, 7.9, and 5.0 percentage points, respectively. Although broader adaptation data may help prevent overfitting to a small, narrowly distributed set of queries, optimizing a shared prefix across diverse tasks may hinder full adaptation to individual target distributions. Thus, under the evaluated settings, broader adaptation data does not achieve the same accuracy gains as explicit stability optimization.

    ModelMethodMATH500AMC23MinervaAIME24GPQAAvg.
    Qwen2.5-Math-1.5BTTSV68.646.820.68.824.833.9
    TTSV (More Queries)66.045.620.29.226.933.6
    TASCO-SAP71.852.221.312.126.836.8
    Qwen2.5-Math-7BTTSV71.059.133.517.936.443.6
    TTSV (More Queries)71.457.832.411.829.340.5
    TASCO-SAP73.466.239.124.238.948.4
    LLaMA-3.1-8B-InstructTTSV47.822.822.16.227.825.3
    TTSV (More Queries)42.824.720.47.528.224.7
    TASCO-SAP50.228.426.511.731.829.7
  • Cross-Task Transfer Evaluation. We further investigate whether more diverse adaptation queries can improve cross-task generalization. We compare the pooled-query prefix with a TTSV prefix optimized solely on MATH500 and directly transferred to other benchmarks. Table 4 shows that on Qwen2.5-Math-7B, the pooled-query approach improves average transfer accuracy across four target benchmarks from 31.6% to 32.8%. These results suggest that broader adaptation data may modestly benefit cross-task transfer.

    TTSV (MATH500)TTSV (More Queries)
    AMC2356.157.8
    Minerva29.432.4
    AIME248.911.8
    GPQA31.829.3
    Avg.31.632.8

References

[1] Sheng, L., Liang, J., He, R., Wang, Z., & Tan, T. (2025). The Illusion of Progress? A Critical Look at Test-Time Adaptation for Vision-Language Models. (NeurIPS).
[2] Lee, K., Jang, H., Lee, D., Alistarh, D., & Lee, N. (2026). The Unseen Frontier: Pushing the Limits of LLM Sparsity with Surrogate-Free ADMM. (ICLR).

Dear Reviewer 4SZf,

Thank you for your positive evaluation and constructive feedback on our work. We have addressed your concerns in our rebuttal and would appreciate any further suggestions that we could incorporate into the final version of our paper.

Reviewer_yLynOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies test-time adaptation of frozen LLMs for reasoning, where a continuous prefix is optimized on the unlabeled test questions using the model's own entropy as a label-free signal. The authors observe that, after confidence-only optimization, questions whose trajectory confidence stays stable under small Gaussian noise on the prefix are more likely to be answered correctly. Based on this observation, they propose TASCO, which adds a stability term to entropy minimization. Random Perturbation penalizes the variance of trajectory confidence across re-decoded noisy prefixes, and Sharpness-Aware Perturbation minimizes the entropy at a gradient-guided worst-case perturbation of the prefix. The method is evaluated on various mathematical and scientific reasoning benchmarks with both general and reasoning-enhanced models, and reports higher accuracy than confidence-only adaptation and shorter generations than standard prompting, together with ablations and cross-benchmark transfer.

Critical strengths
  • The paper addresses a timely problem, and the proposed stability term is simple to add on top of existing entropy-based test-time adaptation.
  • The evaluation covers several model families including reasoning-enhanced models.
Critical weaknesses
  • One of my main concerns is the lack of comparison to prior work. To my understanding, Sharpness-Aware Perturbation is the standard SAM objective (Foret et al., 2021) applied to the prefix, but SAM is not cited at all. The authors should clearly state what is new relative to SAM.
  • I am also not sure whether the baselines are sufficient. The headline gains in Table 1 mostly come from entropy minimization, which the existing baseline TTSV already provides, and the net gain of the stability term is a few points on average. Label-free methods that update the model parameters with the same entropy signal are cited in Section 2.2 but not compared. For the reasoning-enhanced models, the text says that TTSV and LTPO reduce accuracy, but Table 2 does not report them, and TASCO-RP is also missing from Table 2. Please correct me if I overlooked these results.
  • The comparison may not be fair in terms of tuning. The baselines use the configurations of their original authors (Appendix A.1), whereas the perturbation scales of TASCO are chosen on the evaluation benchmarks (Figure 5). Given the modest margin over TTSV, this matters.
  • The title claims that stability makes confidence reliable, but the paper only shows that stability correlates with correctness under the confidence-only prefix (Figure 1, Appendix B). It does not show that the confidence of the adapted model is a better indicator of correctness, nor whether the accuracy gain really comes from stability rather than from the stability term acting as a generic regularizer.
  • Some details are missing. I could not find the value of λrand\lambda_{\mathrm{rand}}, the number of optimization steps, or the convergence criterion. Please let me know if I have overlooked them.
Questions for authors
  • Refer to the above.
7 comments

Response to Reviewer yLyn

We are grateful that the reviewer provided this constructive feedback. We have conducted additional experiments and provided further clarifications to address the reviewer's concerns.

W1: One of my main concerns is the lack of comparison to prior work. To my understanding, Sharpness-Aware Perturbation is the standard SAM objective (Foret et al., 2021) applied to the prefix, but SAM is not cited at all. The authors should clearly state what is new relative to SAM.

Response to weakness 1: We thank the reviewer for highlighting the important related work on SAM [1]. During our literature review, we considered SAM and related studies, including ASAM [2] and SAR [3]. These studies primarily focus on generalization in conventional neural network training or the robustness of vision models under distribution shifts. Since our work focuses on generative reasoning adaptation for LLMs, we prioritized studies directly relevant to this setting in the Related Work section and did not sufficiently cover the broader literature on sharpness-aware optimization. We acknowledge the technical relevance of SAM to TASCO and will add the corresponding citations and discussions to the Related Work and Method sections of the revised manuscript to clarify their relationship.

Technical Relationship and Differences Between TASCO-SAP and SAM. SAM addresses the problem that low training loss does not necessarily ensure good generalization in supervised learning. It seeks flatter solutions by minimizing the worst-case loss within a neighborhood of model parameters. In contrast, TASCO-SAP addresses the problem that high confidence does not necessarily indicate correct reasoning in label-free test-time adaptation. Motivated by the empirical association between local confidence stability and reasoning correctness, TASCO-SAP optimizes a task-shared continuous prefix by minimizing predictive entropy under gradient-guided local prefix perturbations while keeping the LLM backbone frozen.

Revised PDF: Section 2 (Section "TEST-TIME ADAPTATION").

W2: I am also not sure whether the baselines are sufficient. The headline gains in Table 1 mostly come from entropy minimization, which the existing baseline TTSV already provides, and the net gain of the stability term is a few points on average. Label-free methods that update the model parameters with the same entropy signal are cited in Section 2.2 but not compared. For the reasoning-enhanced models, the text says that TTSV and LTPO reduce accuracy, but Table 2 does not report them, and TASCO-RP is also missing from Table 2. Please correct me if I overlooked these results.

Response to weakness 2: We thank the reviewer for raising concerns about the sufficiency of baseline comparisons and the sources of TASCO’s performance gains. Recent LLM reasoning work reports accuracy gains of a few percentage points over strong baselines [4–7]. TASCO’s gains over TTSV are comparable in magnitude and represent a meaningful improvement over entropy minimization alone. To assess gain consistency and broaden our baseline comparisons, we provide results across three independent random seeds, a comparison with parameter-updating entropy minimization, and additional results for reasoning-enhanced models.

  • 1. Additional Benefits of Stability Optimization. We repeated the comparison with TTSV using three independent random seeds. Across three models and five benchmarks, both TASCO variants outperform TTSV in mean accuracy, confirming additional gains from stability optimization.

    ModelMethodMATH500AMC23MinervaAIME24GPQAAvg.
    Qwen2.5-Math-1.5BZero-Shot CoT40.2 ± 1.834.1 ± 2.49.4 ± 1.46.2 ± 3.519.4 ± 2.121.9 ± 0.5
    TTSV69.2 ± 1.645.4 ± 2.019.2 ± 1.46.2 ± 2.624.2 ± 0.632.9 ± 1.4
    TASCO-RP70.8 ± 0.449.9 ± 0.922.1 ± 1.39.9 ± 1.026.7 ± 1.135.9 ± 0.5
    TASCO-SAP71.6 ± 0.949.4 ± 1.420.1 ± 2.211.8 ± 1.626.2 ± 0.635.9 ± 0.7
    Qwen2.5-Math-7BZero-Shot CoT51.8 ± 1.642.9 ± 2.014.1 ± 2.913.3 ± 2.330.7 ± 1.330.6 ± 1.8
    TTSV70.0 ± 2.257.8 ± 3.732.8 ± 2.517.0 ± 3.436.0 ± 2.542.7 ± 2.8
    TASCO-RP74.5 ± 1.266.7 ± 2.038.1 ± 2.120.4 ± 2.437.7 ± 2.647.4 ± 2.1
    TASCO-SAP72.6 ± 0.865.2 ± 2.039.0 ± 2.423.4 ± 2.438.5 ± 2.847.8 ± 1.8
    LLaMA3.1-8B-InstructZero-Shot CoT47.0 ± 0.825.1 ± 0.921.7 ± 0.77.9 ± 1.228.7 ± 0.726.1 ± 0.8
    TTSV47.1 ± 0.722.5 ± 2.819.6 ± 2.75.2 ± 3.127.2 ± 1.724.3 ± 1.6
    TASCO-RP47.2 ± 1.228.6 ± 0.825.3 ± 1.510.9 ± 1.232.7 ± 0.729.0 ± 1.1
    TASCO-SAP49.7 ± 0.826.4 ± 2.026.0 ± 0.712.2 ± 1.232.2 ± 0.829.2 ± 0.8
  • 2. Comparison with Parameter-Updating Entropy Minimization. We compare TASCO with EM, which updates model parameters by minimizing entropy. TASCO achieves higher average accuracy across all three models.

    ModelMethodMATH500AMC23MinervaAIME24GPQA
    Qwen2.5-Math-1.5BEM64.416.547.56.722.3
    TASCO-RP70.850.622.410.427.8
    TASCO-SAP71.852.221.312.126.8
    Qwen2.5-Math-7BEM75.262.535.720.037.1
    TASCO-RP75.868.837.121.237.4
    TASCO-SAP73.466.239.124.238.9
    LLaMA-3.1-8B-InstructEM48.422.823.910.024.9
    TASCO-RP46.829.125.010.433.3
    TASCO-SAP50.228.426.511.731.8
  • 3. Additional Baselines for Reasoning-Enhanced Models. We add TTSV, LTPO, and TASCO-RP results to Table 2. Across both DeepSeek-R1-Distill-Qwen models, TTSV and LTPO lower average accuracy than Zero-Shot CoT, whereas both TASCO variants improve it.

    ModelMethodAIME24 Acc.AIME24 Tok.AIME25 Acc.AIME25 Tok.MATH500 Acc.MATH500 Tok.AMC23 Acc.AMC23 Tok.
    DeepSeek-R1-Distill-Qwen-1.5BZero-Shot CoT22.8758022.5767478.6370059.46242
    TTSV23.3775116.4788677.8386862.25014
    LTPO14.3700113.3689972.8280550.44836
    TASCO-RP28.7699729.4681280.4372368.84329
    TASCO-SAP30.4675129.2683881.4361669.44282
    DeepSeek-R1-Distill-Qwen-7BZero-Shot CoT44.2689925.0709989.4339979.44672
    TTSV36.2753722.4771584.4345176.45126
    LTPO27.4661722.4681680.8322768.04712
    TASCO-RP45.3634731.4678389.4321189.34114
    TASCO-SAP49.2618834.2665290.4304188.14287

Revised PDF: Section 4.2 Appendix E (Section "ADDITIONAL RESULTS ON REASONING BENCHMARKS").

W3:The comparison may not be fair in terms of tuning. The baselines use the configurations of their original authors (Appendix A.1), whereas the perturbation scales of TASCO are chosen on the evaluation benchmarks (Figure 5). Given the modest margin over TTSV, this matters.

Response to weakness 3: We thank the reviewer for raising concerns about hyperparameter tuning and the fairness of the comparison. To address these concerns, we discuss the rationale for baseline configurations, examine perturbation-scale sensitivity, and present cross-model results to further assess the contribution of stability optimization.

  • Fairness of Baseline Configurations. First, test-time adaptation methods have distinct optimization objectives and hyperparameter settings. To ensure a fair comparison, we follow each baseline's original experimental protocols and hyperparameter configurations and rerun the corresponding optimization and evaluation procedures in a unified experimental environment. Recent studies also adopt this practice [8, 9]. Moreover, selecting hyperparameters using independent validation data is an established practice in related LLM test-time optimization research [6, 10, 11]. Therefore, following the established configurations of different methods allows each baseline to achieve its intended performance, rather than imposing identical hyperparameter settings across methods, thereby ensuring a fairer comparison.

  • Perturbation Scale Sensitivity. Second, Figure 5 presents a sensitivity analysis rather than benchmark-specific hyperparameter selection. In our main experiments, we use fixed perturbation scales of σ=0.01\sigma=0.01 for TASCO-RP and ρ=0.8\rho=0.8 for TASCO-SAP across all benchmarks. These settings are not always optimal. For instance, TASCO-SAP achieves higher accuracy with ρ=0.4\rho=0.4 on Minerva and GPQA, yet we use ρ=0.8\rho=0.8 in the main experiments. Across all five benchmarks, both TASCO variants outperform Confidence Only at every tested nonzero perturbation scale. These results show that gains from local confidence stability persist across perturbation scales without benchmark-specific tuning, supporting its contribution beyond confidence optimization alone.

    Table A. Sensitivity to Gaussian Scale σ\sigma (TASCO-RP). Accuracy (%) on Qwen2.5-Math-1.5B.

    Benchmarkσ=0\sigma=0σ=0.0025\sigma=0.0025σ=0.005\sigma=0.005σ=0.01\sigma=0.01σ=0.02\sigma=0.02
    MATH50068.669.870.470.870.2
    AMC2346.848.249.450.650.2
    AIME248.810.29.610.49.6
    Minerva20.622.121.722.421.7
    GPQA24.825.727.327.827.2

    Table B. Sensitivity to Perturbation Radius ρ\rho (TASCO-SAP). Accuracy (%) on Qwen2.5-Math-1.5B.

    Benchmarkρ=0\rho=0ρ=0.2\rho=0.2ρ=0.4\rho=0.4ρ=0.8\rho=0.8ρ=1.0\rho=1.0
    MATH50068.671.270.671.869.8
    AMC2346.849.650.852.251.4
    AIME248.810.812.012.111.4
    Minerva20.621.021.721.321.0
    GPQA24.826.327.226.726.3
  • Consistent Gains Beyond Confidence Optimization. TASCO's gains extend across models and tasks. On LLaMA-3.1-8B-Instruct, TTSV achieves 25.3% average accuracy, below Zero-Shot CoT's 26.0%, whereas TASCO-SAP reaches 29.7%, a 4.4 gain over TTSV. This further supports the additional benefits of local confidence stability beyond confidence-only optimization.

Comments on the paper

1 comment

Dear Area Chair and Reviewers,

As the discussion period ends, we briefly summarize our rebuttal.

First, we sincerely thank all three reviewers (yLyn, 4SZf, and 4to7) for their careful evaluations and constructive feedback. We appreciate that all three reviewers gave our work positive scores.

During the rebuttal, to address the reviewers' main concerns, we conducted additional experiments to further validate the relationship between local stability and reasoning reliability, as well as the additional benefits of stability-aware optimization. We also strengthened our empirical evidence through more comprehensive comparisons and ablation studies, along with evaluations of cross-task generalization, adaptation efficiency, and continual deployment. All new results and discussions have been incorporated into the revised manuscript, with substantive changes highlighted in blue.

We emphasize that simply increasing confidence does not guarantee correct reasoning, and improving reasoning reliability without external supervision remains an important challenge in LLM test-time adaptation. By examining the model's own confidence signals, we observed an interesting phenomenon: high-confidence reasoning that remains stable under local perturbations tends to be more reliable. Inspired by this observation, we propose TASCO, which incorporates local stability into confidence optimization, translating this empirical insight into an effective method for label-free test-time adaptation.

We hope that the insights from our work can inspire the community to further explore LLM reasoning reliability and ways to understand and effectively utilize model-internal signals. We also hope the Area Chair will give thoughtful consideration to the contributions and research value of our work.

We sincerely thank the Area Chair and all reviewers again for their time and effort. We wish everyone continued success in their research and hope that every excellent research contribution receives the recognition and acceptance it deserves!

Best regards,

Authors of Submission 35531

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.