acceptodds
Under review as a conference paper at ICLR 2027

NuRA: Numerical Manifolds in Pretrained LLMs Guide Test-Time Adaptation

10 views

Abstract

Despite advances in mathematical reasoning, reliable numerical prediction remains a challenge for large language models (LLMs). We find that pretrained numeric output representations exhibit angular relationships associated with magnitude and digit patterns, which we describe as the Numerical Manifold. The patterns extend across tokenizers and model families, including vision-language and math-specialized models. In this work, we propose Numerical Representation Adaptation (NuRA), which turns pretrained numerical geometry into a label-free learning signal for test-time adaptation. NuRA adapts hidden representations by minimizing angular dispersion, encouraging probability to concentrate on aligned numerical directions without labeled test data or additional offline training. We analyze how pretrained numerical relationships shape probability redistribution during adaptation. Experiments across numerical tasks and model families show broad accuracy gains with updates to approximately 0.01% of MLP parameters.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language models, numerical prediction, numerical representations, test-time adaptation, representation geometry, entropy minimization

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_8VDeOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces a test-time adaption method for improved numerical predictions. The authors first present the numerical manifold hypothesis which describes the angular relationship between numeric output directions, and empirically validate it across multiple models. Based this hypothesis, the authors propose to use the angular dispersion of the numerical prediction as a test-time objective for improving numerical predictions. An intuitive explanation for the gradient update is provided through the relation between change of output token probabilities and change of angular dispersion. The authors compare their method to alternative decoding and test-time adaptation strategies across both VLMs and LMs on tasks focussing on number understanding and mathematical reasoning and show its empirical performance.

Critical strengths
  • The paper introduces an interesting test-time adaptation method that goes beyond entropy signals by incorporating the geometric properties of number representations instead. The proposed method thus also goes beyond simple distance-based priors often enforced in previous work, which allows to capture a richer structure in the representations.
  • Empirical evaluation is shown across a variety of numerical, mathematical reasoning, and visual tasks and the proposed method seems to consistently improve over most baselines.
  • The authors critically discuss the method's limitations in the form of potential miscorrections (Figure 7).
Critical weaknesses
  • The method is not guaranteed to improve prediction accuracy, as it may also enforce overconfidence for some experiments. While the method seems to improve overall prediction quality as empirically shown, it would be insightful to also understand to which degree the overall predictions change, e.g. by quantifying how often the described failure modes occur.
  • The need for multiple forward and backward passes during the weight updates appears to be the main computational bottleneck of the proposed method. Consequently, the latency of the method does not change much even for very sparse weight updates. While the latency is lower than full chain-of-thought reasoning, it still remains behind alternative inference methods and it remains unclear if it can be improved further.
Questions for authors
  • The paper discusses the possibility of miscorrections as a potential failure mode, where correct predictions are changed to incorrect ones. How often does this failure mode occur in the presented setups?
  • The statement that numerical structure is not only found in input embeddings, as often described in previous work, but also contained in numeric output directions appears of particular importance for models using weight sharing between input embeddings and output weights. Does the method show performance differences between models using weight-sharing and other models?
  • The method's performance seems to depend highly on the temperature parameter. What is a practical way of choosing this hyperparameter?
  • To clearly highlight the difference between entropy minimisation and the proposed angular dispersion objective, it would be helpful to apply the entropy minimisation method only on the numerical tokens in a further ablation.
4 comments
AuthorsOpenReview ↗

Response to Reviewer 8VDe (Part 1/2)

We sincerely appreciate the time and thought you have devoted to reviewing our work. Your recognition of the numerical geometry and empirical results is encouraging, and your questions have helped us examine the method’s behavior and practical use more closely. We address each point below with additional experiments and clarification.

W1 / Q1 The paper discusses the possibility of miscorrections as a potential failure mode, where correct predictions are changed to incorrect ones. How often does this failure mode occur in the presented setups?

We welcome the opportunity to discuss this practical question in more detail. We compared unadapted greedy decoding with NuRA on the same questions from Tables 1–3. Across the three task groups, 77 of 4,050 initially correct answers became incorrect (1.90%), while 293 of 5,436 initially incorrect answers became correct (5.39%). The table reports both directions of change, with each rate calculated over the corresponding initially correct or incorrect predictions. Net gain is the number of corrections minus miscorrections; the token row covers aligned final-answer token positions for direct numerical prediction only.

TaskLevelPaired unitsMiscorrection rateCorrection rateNet gain
Direct numerical predictionAnswer4,02027/1,404 (1.92%)121/2,616 (4.63%)+94
Direct numerical predictionToken30,633312/15,094 (2.07%)680/15,539 (4.38%)+368
Mathematical reasoningAnswer3,46615/1,841 (0.81%)43/1,625 (2.65%)+28
Visual numerical predictionAnswer2,00035/805 (4.35%)129/1,195 (10.79%)+94
OverallAnswer9,48677/4,050 (1.90%)293/5,436 (5.39%)+216

Your suggestion helped us assess the gains alongside the risk of changing correct answers. NuRA corrected 5.39% of initially incorrect answers while changing 1.90% of initially correct answers to incorrect ones, with corrections outnumbering miscorrections in every task group. These rates show how adaptation recovers errors while largely preserving correct predictions. We have added this paired analysis to Appendix D.4.

Q2 The statement that numerical structure is not only found in input embeddings, as often described in previous work, but also contained in numeric output directions appears of particular importance for models using weight sharing between input embeddings and output weights. Does the method show performance differences between models using weight-sharing and other models?

Thank you for raising this distinction and suggesting an interesting direction to explore. We compared tied and untied models within the Qwen3 and Llama families, and the table reports their gains over greedy decoding in average answer accuracy (percentage points) from Table 1 and Appendices D.1–D.2.

ModelShared weightsDirect numericalMathematical reasoning
Llama-3.2-1BYes+1.71+4.08
Llama-3.2-3BYes+2.49+2.88
Llama-3.1-8BNo+2.96+2.13
Qwen3-4BYes+2.04+0.48
Qwen3-8BNo+1.99+1.61

The similar gains on Qwen3-4B and Qwen3-8B in direct numerical prediction (+2.04 and +1.99 points) suggest that NuRA's benefits extend across both tied and untied models.

We also explored input embeddings as geometric priors in the two untied models. The table reports Ans. Acc. / Token Acc. (%), averaged across the six direct numerical prediction datasets, with the adaptation columns indicating the source of the geometric prior.

ModelGreedyInput embedding weightsOutput head weights
Qwen3-8B37.68 / 55.4539.50 / 58.9539.67 / 59.15
Llama-3.1-8B24.64 / 37.6627.32 / 42.8827.60 / 43.45

Both sources improve numerical prediction, with output head weights giving slightly higher answer and token accuracy in both models. This suggests that useful numerical geometry is present on both sides, while output head weights provide a prior directly connected to the logits used for prediction. We have added these comparisons to Appendix F.4.

Our response continues in Part 2.

AuthorsOpenReview ↗

Response to Reviewer 8VDe (Part 2/2)

Q3 The method's performance seems to depend highly on the temperature parameter. What is a practical way of choosing this hyperparameter?

We appreciate your attention to this practical aspect of the method and agree that guidance on choosing the temperature would make NuRA easier to use across models and tasks. NuRA uses the same temperature transformation as softmax sampling, with γ = 1/T, but applies it to the numerical distribution used for adaptation without rescaling the final decoding logits. We believe one practical guide is to consider how concentrated this distribution is before adaptation, using a smaller γ when one numerical token dominates and a larger γ when probability is spread too evenly across numerical tokens.

When the distribution is very sharp, other numerical tokens receive little weight, so their geometric relationships have limited influence on the objective. A smaller γ gives these alternatives more weight, allowing adaptation to make greater use of the numerical geometry. Too much smoothing, however, weakens the model's preferences among numerical tokens, and a very small γ also reduces how strongly the adaptation weights respond to changes in the logits. We find it helpful to view the choice of γ as a balance between preserving these preferences and allowing plausible alternatives to contribute to the update.

This perspective is consistent with our observations across models and tasks. Numerical predictions in mathematical reasoning were often more concentrated, and we use γ = 0.3 for Qwen3 on the reasoning tasks, compared with 0.6–0.8 on direct numerical prediction. We also observed sharper answer distributions in DeepSeek-R1-Distill-Qwen-1.5B, for which we use γ = 0.125 for reasoning and 0.25 for direct numerical prediction. These observations suggest that models within the same family and related tasks can share patterns in prediction concentration, which may provide a useful starting point when choosing γ for a new setting. We have added this practical guidance and the complete configurations to Appendix C, together with sensitivity results under the main adaptation protocol in Appendix F.5.

Q4 To clearly highlight the difference between entropy minimisation and the proposed angular dispersion objective, it would be helpful to apply the entropy minimisation method only on the numerical tokens in a further ablation.

Thank you for this helpful suggestion. We compared the two objectives on the same numerical tokens with the adaptation settings fixed, and NuRA achieved 0.75–2.71 percentage points higher answer accuracy across all four settings. We have added this ablation to Appendix C of the revised manuscript.

W2 The need for multiple forward and backward passes during the weight updates appears to be the main computational bottleneck of the proposed method. Consequently, the latency of the method does not change much even for very sparse weight updates. While the latency is lower than full chain-of-thought reasoning, it still remains behind alternative inference methods and it remains unclear if it can be improved further.

Thank you for highlighting this limitation. We agree that sparse updates provide only modest latency savings because forward and backward computation is still required. Our sparsity experiments examine how few parameters need to change to improve numerical prediction, and we find that updating a small, gradient-selected subset retains most of the token-accuracy gain from full updates, suggesting that existing numerical representations can be adjusted while leaving almost all pretrained weights unchanged.

We appreciate your emphasis on efficiency and have explored several alternatives, including updating only the final MLP layers, directly optimizing the final hidden state, and updating normalization parameters following the approach of TENT (Wang et al., 2021). Table 4 reports the final-four-layer comparison, and Appendix F.1 reports the normalization-only results. Although these alternatives did not match the accuracy gains of our selected MLP updates, improving efficiency remains an important goal. We believe that stopping adaptation once the numerical prediction distribution stabilizes is a promising direction, and we will explore the resulting accuracy–latency trade-off in future work. We have expanded the discussion of computational cost in Appendix H.

References

Wang, D., Shelhamer, E., Liu, S., Olshausen, B., and Darrell, T. (2021). Tent: Fully Test-Time Adaptation by Entropy Minimization. ICLR.

Thank you for your elaborate rebuttal and clarifications, particularly regarding the amount of miscorrections as a potential failure mode. I believe this adds a valuable analysis of the method to the paper.

AuthorsOpenReview ↗

Thank you for your feedback

Thanks again for taking another look and for your support of our work. We're really glad the miscorrection analysis was useful, and your feedback helped us strengthen the paper.

Reviewer_bJd8OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how numerical relationships learned during pretraining can guide test-time adaptation in LLMs. The authors observe that numerical tokens exhibit structured angular relationships in the pretrained unembedding matrix and propose NuRA, a label-free adaptation method that minimizes angular dispersion among numerical token representations by updating a small fraction of MLP parameters. The method achieves performance improvements across various numerical tasks and model families, including conventional LLMs, math-specialized models, and vision-language models. The paper further provides theoretical insights into the proposed objective and analyses of the adaptation dynamics, parameter sparsity, and hyperparameter sensitivity.

Critical strengths

The paper’s key strengths are listed below:

  • Interesting and novel approach for test-time adaptation. The paper proposes an interesting and novel approach that uses numerical relationships learned during pretraining in the model’s unembedding matrix to construct a label-free test-time adaptation objective. Notably, updating only a very small fraction of the model's parameters (e.g., 0.01%) achieves measurable performance improvements across different numerical tasks and model families, including conventional LLMs, math-specialized models, and vision-language models.
  • Thorough analysis and clear presentation. Beyond the performance evaluations, the paper provides extensive analyses of the proposed method, including sparse parameter updates, anchor trajectories, and hyperparameter sensitivity (Section 4.3). Moreover, the paper is well presented, with clear and informative figures (e.g., Figures 2 and 3) that help explain the underlying intuition and methodology.
Critical weaknesses

The paper’s main weaknesses are listed below:

  • Dependence on pretrained numerical relationships and generalizability. It seems that NuRA relies on numerical relationships learned during pretraining to guide how probabilities are redistributed. However, these relationships are fixed, while the relationships relevant to a prediction can depend on the input problem. For example, for 242^4, the model might assign probability to 8, 16, and 32 (with slightly higher mass on the incorrect 32), but the unembedding geometry might encode stronger relationships between nearby numbers such as 31, 32, and 33 rather than powers of two. In such cases, NuRA might reinforce a geometrically coherent but incorrect group of predictions, potentially reducing angular dispersion while decreasing the probability of the correct answer. This raises concerns about generalization to tasks whose relevant numerical relationships are not well captured by the pretrained geometry or where two conflicting relationships interfere, as in my example. One possible suggestion could be to evaluate NuRA on tasks with different numerical structures, such as exponentiation or modular arithmetic, and analyze whether adaptation improves or worsens initially incorrect predictions, particularly when the relevant numerical relationships interfere or differ from those encoded in the unembedding matrix.
  • Interaction with chain-of-thought reasoning. While the paper evaluates NuRA on mathematical reasoning benchmarks (e.g., GSM8K and MATH500), where adaptation is triggered at numerical decoding steps, it does not explicitly analyze its effect on intermediate computations during CoT reasoning. In particular, NuRA could improve intermediate arithmetic calculations, but might also reinforce incorrect predictions and propagate errors through subsequent reasoning steps. It also remains unclear whether NuRA should be applied at every numerical step or only selectively. One possible suggestion would be to compare CoT alone with CoT + NuRA applied at all numerical steps, only at uncertain steps, or only at the final answer, measuring both intermediate and final-answer accuracy.
Questions for authors

A list of questions is provided below:

  • Missing results in Table 5. Why are some entries missing for the program-aided and CoT baselines? Were these settings not evaluated?
  • Evaluation setup in Figure 4. Could the authors explicitly state in the figure caption which dataset is used for evaluating Llama-3.2-3B?
  • Definition of numerical tokens. How exactly is the numerical token set N\mathcal{N} constructed? Does it only contain numerical strings (e.g., "1", "2", "3"), or also verbalized numbers (e.g., "one", "two", "three")?
  • CoT accuracy in Table 5. Table 5 reports generation latency and length for CoT prompting, but not answer accuracy. How does its accuracy compare to direct inference and NuRA?
  • Sensitivity to parameter selection. Appendix K compares fixed and dynamic parameter selection on Addition-50. Do these findings generalize to other tasks and models? How does reselecting the binary mask ss at each adaptation step affect performance?
5 comments
AuthorsOpenReview ↗

Response to Reviewer bJd8 (Part 1/2)

We appreciate your careful reading and the thought behind your questions. Your suggestions have motivated additional comparisons on adaptation during reasoning and parameter selection. We address the evaluation details first, then discuss your broader question about the generality of pretrained numerical relationships.

Q1 — Missing results in Table 5. Why are some entries missing for the program-aided and CoT baselines? Were these settings not evaluated?

We appreciate to clarify the design of Table 5. We evaluated program-aided inference on Addition and MATH401 and CoT prompting on GSM8K and ChartQA. On mathematical and visual questions, producing an executable solution requires the model to interpret the problem, determine the reasoning steps, and translate them into code, so changes in accuracy can reflect both reasoning and numerical computation. Our adaptation comparison keeps the original generation workflow and introduces updates at numerical prediction positions, providing a more direct assessment of numerical adaptation. For Addition and MATH401, executable computation provides a natural reference, while CoT adds reasoning steps and generation time without necessarily improving accuracy. The dashes mark combinations outside this timing evaluation, whose scope we have clarified in Appendix H.

To further assess how numerical adaptation interacts with reasoning, we evaluated NuRA under both question-only and CoT prompting on GSM8K, with results for different adaptation positions reported under W2/Q4.

Q2 — Evaluation setup in Figure 4. Could the authors explicitly state in the figure caption which dataset is used for evaluating Llama-3.2-3B?

Thank you for pointing this out. Figure 4 evaluates Llama-3.2-3B on Addition-50 using reference answer prefixes, as described in Appendix F.2. We have added the dataset name to the figure caption.

Q3 — Definition of numerical tokens. How exactly is the numerical token set constructed? Does it only contain numerical strings (e.g., "1", "2", "3"), or also verbalized numbers (e.g., "one", "two", "three")?

Thank you for asking us to clarify this definition. We include only tokens whose individually decoded strings consist entirely of ASCII digits, such as “1”, “23”, or “001” when available in the tokenizer. Verbalized numbers such as “one” and “two” are excluded, as are tokens containing whitespace, signs, or decimal points.

In our evaluation, the mean total probability assigned to these tokens at answer positions before adaptation exceeds 99% in each setting, as shown below. Under the numerical-answer format, little probability remains outside this set, including on verbalized numbers. We have added the construction rule and examples to Appendix C.1, and report the probability measurements in Appendix C.3.

ModelNumerical-token probability (%)
Llama-3.1-8B99.895
Qwen3-4B99.999
DeepSeek-Math-7B99.997
Qwen2.5-VL-3B99.093

W2 — Interaction with chain-of-thought reasoning. One possible suggestion would be to compare CoT alone with CoT + NuRA applied at all numerical steps, only at uncertain steps, or only at the final answer, measuring both intermediate and final-answer accuracy.

Q4 — CoT accuracy in Table 5. Table 5 reports generation latency and length for CoT prompting, but not answer accuracy. How does its accuracy compare to direct inference and NuRA?

We appreciate your suggestion. Table 5 uses a small subset to measure generation costs (Appendix H), so we believe accuracy is better assessed on a larger set of questions. We evaluated Llama-3.1-8B on GSM8K under question-only and CoT prompting. CoT primarily supports reasoning (although numerical prediction may also improve), so we compare NuRA within each prompting setting and follow your suggested adaptation positions to assess its effect on intermediate calculations and final answers.

PromptingAdaptation positionsAnswer Acc. (%)Intermediate calculation Acc. (%)
Question-onlyNone84.4698.59
Question-onlyAll numerical positions87.4998.67
CoTNone85.9098.37
CoTAll numerical positions87.6498.63
CoTUncertain numerical positions87.2698.55
CoTFinal answer only85.7498.37

Uncertain positions have a numerical top-1 probability below 50%.

Our response continues in Part 2.

AuthorsOpenReview ↗

Response to Reviewer bJd8 (Part 2/2)

Your suggested comparison helps clarify where adaptation is useful, and we find improvements in both intermediate calculation and answer accuracy under both prompting settings. Adapting only uncertain positions retains most of the answer-accuracy gain under CoT, while adapting only the final answer provides no improvement. One possible explanation is that the final answer already follows the preceding reasoning, so changing its numerical prediction alone may not correct an error established earlier. These results suggest that adapting during reasoning can be more useful than intervening only at its conclusion.

Q5 — Sensitivity to parameter selection. Appendix K compares fixed and dynamic parameter selection on Addition-50. Do these findings generalize to other tasks and models? How does reselecting the binary mask at each adaptation step affect performance?

We appreciate your suggestion to extend this comparison. We evaluated fixed and dynamic selection on three datasets with Llama-3.2-3B, using a 0.01% parameter budget and five adaptation steps.

DatasetQuestionsFixed maskReselected each step
GSM8K1,319145/1426 (10.17%)143/1426 (10.03%)
MATH401401349/986 (35.40%)345/986 (34.99%)
Addition-5050105/418 (25.12%)108/418 (25.84%)

Entries report correct numerical tokens / evaluated numerical tokens, followed by accuracy. Reselecting the mask slightly improves Addition-50 but slightly reduces accuracy on GSM8K and MATH401, showing no consistent advantage across these tasks. Fixed selection achieves similar performance without repeated mask selection.

Our main results show gains with fixed selection across multiple model families, supporting its practical use beyond Llama-3.2-3B. We agree that comparing it with dynamic selection on these models would help clarify when reselecting parameters is useful, and we would be happy to extend this comparison as computational resources allow.

W1 — Dependence on pretrained numerical relationships and generalizability. However, these relationships are fixed, while the relationships relevant to a prediction can depend on the input problem. In such cases, NuRA might reinforce a geometrically coherent but incorrect group of predictions, potentially reducing angular dispersion while decreasing the probability of the correct answer.

Your example raises an interesting question, and we appreciate the thought behind it. For 242^4, recovering 16 from an initially preferred 32 depends on both the current probabilities and the learned geometry. Although 32 has slightly higher individual probability, numerical tokens geometrically related to 16 may together provide stronger support for its direction. Low-probability tokens such as 31 and 33 contribute little to the anchor, so their proximity to 32 alone cannot determine the update.

In particular, information spread across several numerical tokens could help 16 overtake 32 during adaptation (for example, when these tokens reflect an approximate estimate of the answer’s magnitude), provided the learned relationships and current probabilities favor that change. We believe this helps explain how a fixed representation can remain useful across different problems, as the probabilities in each prediction give different weight to these relationships. Those relationships can extend beyond numerical adjacency, as suggested by the magnitude and digit patterns in Figure 1 and prior findings on Fourier-like features and numerical heuristics (Zhou et al., 2026; Nikankin et al., 2025). These observations suggest that the learned structure may be useful beyond tasks based on numerical proximity.

We agree that adaptation depends on the available numerical structure (gains are larger on Llama-8B than on Llama-1B/3B), and we are glad to pursue your suggested direction in our future work, using exponentiation and modular arithmetic to study when the learned geometry helps correct errors or reinforces them.

References

Zhou, T., Fu, D., Soltanolkotabi, M., Jia, R., and Sharan, V. (2026). FoNE: Precise Single-Token Number Embeddings via Fourier Features. arXiv:2502.09741, revised April 2026.

Nikankin, Y., Reusch, A., Mueller, A., and Belinkov, Y. (2025). Arithmetic Without Algorithms: Language Models Solve Math With a Bag of Heuristics. ICLR.

Reviewer_bJd8OpenReview ↗

Response to Authors

I thank the authors for their responses to my questions and concerns. Regarding my initial questions:

  • Q1: I suggest reformatting Table 5 or making the explanation more explicit in the main text, rather than only in Appendix H. As it is currently presented, it is not immediately clear why program-aided inference is not evaluated for GSM8K and ChartQA, while CoT is not evaluated for Addition and MATH401.
  • Q3: Great, thanks for this analysis. For future studies, it could be interesting to look into how the behavior changes when using problems that incentivize the use of verbalized numbers.
  • Q4: It seems that a lot of important details (e.g., see above, or the fact that Table 5 uses only a small subset of questions) are hidden in the appendix. It would be good to make these explicit in the main text.
  • Q5: Alright, thanks for the analysis.

Regarding my concerns (in the order of your response):

  • W2: Thanks for the analysis, great to see some numbers on this. However, I am a bit skeptical about the setup: ignoring NuRA for now, there seems to be almost no increase in performance from question-only to CoT prompting. Do you use zero-shot CoT? Also, do you let the model reason freely before committing to a final answer in the question-only setup? This might produce largely identical CoT-like responses in both conditions and could explain the moderate accuracy increase. The analysis would be much more interesting for cases (model + dataset combinations) where question-only prompting vs. CoT prompting actually results in a meaningful difference. Could you maybe redo this analysis? Sorry, I know time is short.

  • W1: My example might be a bit oversimplified. What I’m trying to get at is that, if I understood NuRA correctly, it only takes the current problem into account implicitly, through the model’s predicted probabilities. But these probabilities could of course be wrong. For instance, the correct answer might be the second most likely prediction, while the incorrect answer happens to lie in a region of the unembedding that is geometrically more consistent with other predictions. At the same time, plausible alternatives around the correct answer (e.g., natural “off-by-one” errors) might not be assigned much probability by the model, so their geometric relationships barely contribute to the objective.

    In that case, NuRA could end up reinforcing a region of the pretrained geometry that is consistent with the model’s incorrect prediction, rather than moving toward the correct answer. And since it has no explicit notion of which numerical relationships are actually relevant to the problem being solved, there’s nothing inherently preventing this, right? I could imagine this being particularly problematic when the model has learned strong numerical patterns during pretraining that don’t necessarily apply to the current task.

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

NuRA adapts numerical predictions by minimizing angular dispersion among pretrained numeric output directions. It updates a small set of gradient-selected MLP parameters and restores the original parameters after each prediction. The paper evaluates calculation, mathematical reasoning, visual numerical prediction, and uncertainty estimation across several model families.

Critical strengths

The paper gives numerical geometry a concrete use. Permuting token directions reduces performance, and Appendix I shows better error-detection AUROC than numerical entropy on all eight reported model-dataset pairs.

Gains extend to complete generated answers. Llama-3.2-3B GSM8K accuracy improves from 76.88% to 83.24%, and Qwen2.5-VL-3B average visual-task accuracy improves from 38.30% to 43.20%. The paper also reports failed adaptations and distinguishes reference-prefix analyses from free generation.

Critical weaknesses

The closest aggregate adaptation control is missing. Appendix E.1 uses full-vocabulary entropy for the entropy-minimization baseline, whereas NuRA restricts its objective to numeric tokens. The comparison therefore changes both geometry and support. Numerical entropy is tested for error detection and in three selected adaptation examples, but there is no aggregate free-generation comparison under matched adaptation settings. The permutation control does not resolve this issue because permuted geometry can itself produce harmful updates.

The supporting experiments do not consistently evaluate the main method. Section 4.3 cites Appendix Q for sensitivity, but its Llama sweeps use normalized input embeddings, dense MLP/normalization updates, and reference prefixes. These differ from the main output-head geometry and sparse free-generation method. The principal accuracy tables also lack a consolidated configuration and selection protocol. Separately, the reproducibility statement promises proofs comparing with token-entropy adaptation in Appendix E, but that appendix contains tasks, models, baselines, and Algorithm 1.

Questions for authors

How does NuRA compare with numerical-entropy minimization in aggregate free-generation accuracy when support, temperature, trigger, update budget, and parameter pool are matched?

4 comments
AuthorsOpenReview ↗

Response to Reviewer hPPt (Part 1/2)

We appreciate the time and care you have devoted to reviewing our work. Your comments identify important comparisons and experimental details to clarify, and we have added experiments and revised the corresponding descriptions. We respond to each point below.

W1 / Q1 How does NuRA compare with numerical-entropy minimization in aggregate free-generation accuracy when support, temperature, trigger, update budget, and parameter pool are matched?

We appreciate your suggestion to compare the two objectives on the same set of numerical tokens (the support of the adaptation distribution). In our original evaluation, we used full-vocabulary Shannon entropy as a general baseline to align our evaluation with established entropy-minimization methods for test-time adaptation and LLM reasoning (Wang et al., 2021; Agarwal et al., 2025). We agree that minimizing Shannon entropy over the same numerical tokens is also a useful and closer comparison, which directly isolates the gains from the learned geometry.

We found that numerical tokens already received more than 99% of the probability mass on average at answer positions before adaptation, suggesting that entropy minimization over the two token sets may yield similar results in this setting. We used the same numerical token set, inverse temperature, trigger rule, update budget, and eligible parameters for both objectives, and report free-generation answer accuracy below alongside the full-vocabulary entropy baseline.

ModelNumerical-token probability (%)Full-vocabulary entropy Acc. (%)Entropy over numerical tokens Acc. (%)NuRA Acc. (%)
Llama-3.1-8B99.89524.7024.8927.60
Qwen3-4B99.99937.2837.3539.43
DeepSeek-Math-7B99.99754.8954.9755.72
Qwen2.5-VL-3B99.09341.6741.0343.20

We report mean numerical-token probability before adaptation and average answer accuracy across datasets within each task group. NuRA outperformed entropy minimization over numerical tokens by 0.75–2.71 percentage points across the four settings, while the two entropy baselines differed by at most 0.64 points. These results support the benefit of the geometric objective beyond the choice of numerical token set.

The model had largely narrowed its prediction to numerical tokens, so computing entropy over this set leaves out many tokens but very little probability mass. We believe this helps explain the similar accuracy of the two entropy baselines. Our error-detection experiments show a similar pattern, with the two entropy scores differing by less than 0.01 in AUROC in six of eight settings (Appendix E, Table 27). Our full-vocabulary angular-dispersion ablation also retains clear gains, with further improvements from focusing on numerical tokens (Table 4). We expect this focus to be useful when other tokens receive more probability, as it keeps the objective centered on numerical predictions and the geometry studied in Section 3.

W1 The permutation control does not resolve this issue because permuted geometry can itself produce harmful updates.

The permutation control tests the contribution of the learned token–direction correspondence, and the additional entropy comparison measures the benefit of angular dispersion on the same numerical token set.

References

Wang, D., Shelhamer, E., Liu, S., Olshausen, B., and Darrell, T. (2021). Tent: Fully Test-Time Adaptation by Entropy Minimization. ICLR.

Agarwal, S., Zhang, Z., Yuan, L., Han, J., and Peng, H. (2025). The Unreasonable Effectiveness of Entropy Minimization in LLM Reasoning. NeurIPS 38.

Our response continues in Part 2.

AuthorsOpenReview ↗

Response to Reviewer hPPt (Part 2/2)

W2 Section 4.3 cites Appendix Q for sensitivity, but its Llama sweeps use normalized input embeddings, dense MLP/normalization updates, and reference prefixes.

We conducted the analysis in Appendix F.5 with the dense-update version to understand how adaptation responds to hyperparameters across tasks. We agree that evaluating the main configuration makes this analysis more directly relevant and have repeated the experiments using its numerical token set, output-head geometry, sparse MLP updates, and generation protocol. The table summarizes token-accuracy ranges (%) for Llama-3.2-3B on the three text tasks and Qwen2.5-VL-3B on ChartQA-100.

SweepAddition-50MATH401GSM8KChartQA-100
Learning rate14.62–24.0834.23–37.467.60–11.0570.55–75.09
Inverse temperature γ4.42–24.502.54–37.051.45–10.2070.55–73.36
Numerical token set22.50–24.9634.79–35.299.45–11.2772.31–73.52
Update steps11.28–24.5532.31–35.508.19–11.0571.27–73.09

We varied the learning rate (0.001–0.05), γ (0.1–2), the numerical token set, and update steps (0–10). We have added these results and the corresponding experimental settings to Appendix F.5.

W2 The principal accuracy tables also lack a consolidated configuration and selection protocol.

We summarize the configurations for all main and supplementary accuracy results below, with each entry giving the learning rate / γ. For each input, we select the 0.01% of MLP parameters with the largest absolute initial gradients and keep this set fixed through adaptation for up to five updates. We have added these settings and the parameter-selection procedure to Appendix C.

We use γ = 1/T to adjust the numerical distribution used for adaptation, with smaller values softening concentrated predictions and larger values sharpening flatter predictions. This provides a practical basis for choosing the temperature, a question also raised by Reviewer 8VDe. Because γ affects both the probability weights and the gradient, its effect depends partly on the learning rate, which controls the update size. We report both parameters together with the sensitivity results to clarify these choices.

Numerical prediction

ModelAddition-50NUPABB-ArithNB-MixedNB-ArithMATH401
Llama-3.2-1B/3B, Llama-3.1-8B0.005 / 0.50.0025 / 0.50.0025 / 0.50.0025 / 0.50.0025 / 0.50.0025 / 0.5
Qwen3-4B/8B0.025 / 0.80.025 / 0.80.0125 / 0.60.0125 / 0.60.025 / 0.60.0125 / 0.6
DeepSeek-Math-7B0.01 / 10.005 / 10.005 / 10.01 / 10.005 / 10.01 / 1
DeepSeek-R1-Distill-Qwen-1.5B0.02 / 0.250.04 / 0.250.02 / 0.250.04 / 0.250.02 / 0.250.02 / 0.25

Mathematical reasoning

ModelGSM8KMATH500TabMWPFinQA
Llama-3.2-1B/3B, Llama-3.1-8B0.005 / 0.250.0025 / 0.50.005 / 0.250.0025 / 0.5
Qwen3-4B/8B0.025 / 0.30.025 / 0.30.025 / 0.30.025 / 0.3
DeepSeek-Math-7B0.01 / 0.50.01 / 0.50.01 / 0.50.01 / 0.5
DeepSeek-R1-Distill-Qwen-1.5B0.02 / 0.1250.02 / 0.1250.02 / 0.1250.02 / 0.125

Visual numerical prediction

ModelChartQAMathVistaTallyQA
Qwen2.5-VL-3B0.01 / 10.005 / 0.80.01 / 1
InternVL3.5-4B0.005 / 0.80.01 / 10.01 / 1

W2 Separately, the reproducibility statement promises proofs comparing with token-entropy adaptation in Appendix E, but that appendix contains tasks, models, baselines, and Algorithm 1.

Thank you for pointing this out. We have corrected the sentence to “Appendix C documents the evaluation tasks, models, baseline definitions, and adaptation algorithm.”

Thank you for your thoughtful rebuttal and extra results. They resolved most of my concerns and I’ll increase the rating to 3.

AuthorsOpenReview ↗

Thank you for your feedback

Thank you very much for re-evaluating our work and for your support. We are glad our response resolved your concerns and appreciate your helpful feedback throughout the review process.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.