acceptodds
Under review as a conference paper at ICLR 2027

Hesitation Has a Geometry: Entropy-Trained Hyperbolic Probes for Sparse Activation Steering

13 views

Abstract

When a large language model solves a mathematical problem, its reasoning is largely hierarchical, and the solution often branches at a few tokens where the next-token entropy is high. Such tree-like structure embeds in hyperbolic space with far lower distortion than in Euclidean space. Activation steering, however, usually edits the hidden states of a pretrained model by adding one fixed Euclidean vector at every token, even though most tokens of a solution are already determined by the context. We propose Hyperbolic Entropy Steering (HEST), which embeds the hidden states in the Poincaré ball with a lightweight probe whose only label is the model's own next-token entropy. Where this entropy exceeds a threshold, HEST moves the embedded state along the geodesic of steepest descent of a readout of the probe and maps the change back to the hidden state. For the Busemann readout of a learned ideal point, we prove that a step of fixed length lowers it by the same amount at every state. On three instruction-tuned models from the Qwen2.5-Math and Llama-3.1 families, HEST with the Busemann readout improves greedy accuracy on MATH-500 and GSM8K in five of six settings, by up to 1.8 points, whereas a contrastive steering vector added at every token lowers accuracy. With a Euclidean probe trained in the same way, this gain disappears on Qwen2.5-Math-1.5B-Instruct. The gains are largest on problems where the model hesitates often, and accuracy on the remaining problems is almost unchanged.

Primary area:
interpretability and explainable AI
Keywords:
activation steering, large language models, hyperbolic geometry, next-token entropy, mathematical reasoning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_pVhsOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper argues that activation steering should act only at the few high-entropy tokens where a math solution branches, instead of adding one fixed vector at every token. It proposes Hyperbolic Entropy Steering (HEST): a probe trained without correctness labels maps hidden states into a Poincare ball so that a readout orders them by the model's own next-token entropy, and at these "hesitation points" HEST moves the state along the geodesic that lowers the readout and decodes the next token from the edited state.

On three models, HEST changes greedy accuracy on MATH-500 and GSM8K by -0.30 to +1.80 points while steering a median of 0.8-3.0% of the tokens per solution, with gains in five of six settings. CAA added at every token lowers accuracy in all six settings. Per-setting sign tests give p = 0.09-0.64. On the development model, a Euclidean probe lowers accuracy, which the authors take as evidence that the gain depends on the hyperbolic geometry.

Critical strengths
  1. Choosing where to steer by entropy is effective, and picking these positions needs no correctness labels. On the development model, edits of the same size give 76.00 (+1.80) on MATH-500 at hesitation points, but applied at every token or at step boundaries they fall below greedy decoding, to 74.00 (-0.20) and 73.80 (-0.40). In all six settings, within the group with 0-10 hesitation points, which holds 71-81% of the problems, HEST turns at most 3% of the correctly solved problems into wrong ones. Unlike the common practice of adding the same vector at every token, this shows that where to intervene matters.

  2. The ablations and reporting are thorough. The paper varies where to steer, the direction (probe, fixed, reverse) and the geometry (Euclidean probe), and checks that the edited states realize the step of Proposition 1 to within 10^-4 on average.

Critical weaknesses
  1. The gains are small and not statistically significant. Per-setting sign tests give p = 0.09-0.64, and the pooled p = 0.050 treats six settings that share problems as independent. Summed over the four settings other than the development model, whose step length was chosen on the test sets, HEST fixes 49 problems and breaks 40, almost the same number. Each setting uses a single probe, so the variation across probe training runs is unknown.

  2. The direction controls do not show that the gain comes from the probe direction. Under the same gate and edit size, the fixed direction, which uses the CAA direction as is, also reaches 76.00 (+1.80) on MATH-500 (13/4 fixed/broken, against 16/7 for HEST), and the reverse direction also gains +0.40 and +0.15 points on MATH-500 and GSM8K. The fixed direction gains less on GSM8K (+0.08 vs. +0.53), but over both benchmarks only the probe direction is right on 39 problems and only the fixed direction on 33. No random-direction control at hesitation points is reported.

  3. The only external baseline is CAA, which differs from HEST in edit size and position. CAA uses multiplier 1, and on the development model its vector is about five times the size of a HEST edit and is added at every token. HEST is not compared with HARP [1], which also acts at high-entropy tokens, or with simple decoding alternatives at hesitation points.

[1] R. Storai and S. Hwang. HARP: Hesitation-aware reframing in transformer inference pass. NAACL, 2025.

Questions for authors
  1. Statistical robustness. Please report the mean +/- std of the accuracy change over 3-5 probes per setting trained with different seeds. Please also select the step length on held-out training problems rather than the test sets, and evaluate on a larger pool, such as the remaining MATH test problems not included in MATH-500. If the gains hold under these conditions, weakness 1 would be largely addressed.

  2. Random-direction control. Please run a random direction (several seeds) and the fixed direction with the same gate, KV rollback and edit size in all six settings. The fixed direction is currently reported only for the development model. Weakness 2 would be addressed if HEST clearly beats both.

  3. Mechanism at the steered token. At what fraction of steered positions does the edit change the argmax token, and does the model's actual next-token entropy at that position decrease? How do these quantities differ between the probe, fixed and reverse directions?

10 comments
AuthorsOpenReview ↗

Response to Reviewer pVhs (1/5): W1 and Q1, part 1

We sincerely thank the reviewer for the detailed and constructive review, and in particular for stating precisely which experiments would address each concern. We have run all of them, and we believe the paper is considerably stronger as a result. All new results are marked in blue in the revised PDF, and the line, table and appendix numbers below refer to it.

W1. The gains are small and not statistically significant. Per-setting sign tests give p = 0.09-0.64, and the pooled p = 0.050 treats six settings that share problems as independent. Summed over the four settings other than the development model, whose step length was chosen on the test sets, HEST fixes 49 problems and breaks 40, almost the same number. Each setting uses a single probe, so the variation across probe training runs is unknown.

Q1. Statistical robustness. Please report the mean +/- std of the accuracy change over 3-5 probes per setting trained with different seeds. Please also select the step length on held-out training problems rather than the test sets [...]

Answer: We fully agree with this concern. In the submission, the step length of the development model was chosen on the test sets, and every setting used a single probe, so the variation across probe training runs was unknown. We have rerun the evaluation exactly as the reviewer suggests.

1. The step length is selected on held-out training problems (lines 287 to 291, Appendix G.2). For every model and dataset, we draw 500 problems from the training split that are disjoint from the 100 problems of the probe and from the test sets. The MATH validation set is stratified by difficulty level to match MATH-500. We run HEST on these problems with each step length in {0.1, 0.25, 0.5, 1} and select the one with the highest accuracy. The thresholds are unchanged and are calibrated without labels, so no test label enters any hyperparameter. Table 1 of the revised paper reports this selection, and its numbers of correctly solved validation problems are as follows.

ModelDatasett = 0.1t = 0.25t = 0.5t = 1Selected t
Qwen2.5-Math-1.5BMATH4134164144100.25
Qwen2.5-Math-1.5BGSM8K4784804814821
Llama-3.1-8BMATH3473403483480.5
Llama-3.1-8BGSM8K4644604624630.1
Qwen2.5-Math-7BMATH4144144094151
Qwen2.5-Math-7BGSM8K4794814804821

Within a setting, the four step lengths differ by only 3 to 8 of the 500 problems, so HEST is robust to this choice.

2. Mean and standard deviation over three probes (Table 2, and Table 9 in Appendix H.1). We train three probes per setting with different random seeds and evaluate each of them on the test set at the selected step length. The accuracy of every probe, together with the mean and the standard deviation, is shown below.

ModelDatasettGreedyHEST, each of the three probesHEST, mean ±\pm std
Qwen2.5-Math-1.5BMATH-5000.2574.2075.00 / 74.60 / 74.4074.67 ±\pm 0.31
Qwen2.5-Math-1.5BGSM8K184.8485.29 / 85.06 / 84.6985.01 ±\pm 0.31
Llama-3.1-8BMATH-5000.550.8052.00 / 51.80 / 51.8051.87 ±\pm 0.12
Llama-3.1-8BGSM8K0.184.6185.06 / 84.23 / 84.7684.69 ±\pm 0.42
Qwen2.5-Math-7BMATH-500184.0084.60 / 85.40 / 85.2085.07 ±\pm 0.42
Qwen2.5-Math-7BGSM8K195.3895.68 / 95.38 / 95.6095.55 ±\pm 0.16

HEST improves the mean accuracy of greedy decoding in all six settings, with a standard deviation of 0.12 to 0.42 points across probes, and 16 of the 18 individual runs are at or above greedy decoding (lines 337 to 343).

The evaluation on a larger pool of problems, which is the third request of Q1, is reported in the next comment.

AuthorsOpenReview ↗

Response to Reviewer pVhs (2/5): W1 and Q1, part 2

Q1. [...] and evaluate on a larger pool, such as the remaining MATH test problems not included in MATH-500. If the gains hold under these conditions, weakness 1 would be largely addressed.

Answer: We thank the reviewer for this suggestion, which has given the paper its strongest result. We run the development model on the full MATH test set of 5,000 problems, with the probe and the validation-selected step length t = 0.25 unchanged (Section 4.3, Figure 2, Appendix H.2). The revised paper has a new section and a new figure for this experiment.

Overall result and the remaining problems. The results on all 5,000 problems, and on the 4,500 problems that are not in MATH-500, are as follows.

ProblemsGreedyHESTFixed / brokenExact two-sided sign test
Full MATH test set, 5,000 problems75.8276.78134 / 86p = 0.0015
Remaining problems not in MATH-500, 4,500 problems76.0076.98125 / 81p = 0.0026

The improvement is statistically significant on the full test set, and it is equally large and also significant on the remaining problems that are not included in MATH-500. The gain is therefore not specific to MATH-500.

Breakdown of the test set. Table 10 of the revised paper divides the 5,000 problems by difficulty level, by subject and by the number of hesitation points in the greedy solution, and Figure 2 shows the same numbers. They are as follows.

Part of the MATH test setProblemsGreedyHESTFixed / broken
Level 143794.7496.348 / 1
Level 289489.0489.8211 / 4
Level 3113184.8886.0324 / 11
Level 4121473.8974.6332 / 23
Level 5132454.6855.5959 / 47
Algebra118791.6692.2515 / 8
Prealgebra87185.7686.4518 / 12
Number theory54079.2680.3715 / 9
Counting and probability47469.8371.5219 / 11
Geometry47964.5165.9713 / 6
Precalculus54662.4564.1018 / 9
Intermediate algebra90360.5861.1336 / 31
0 to 10 hesitation points364089.7390.2548 / 29
11 to 24 hesitation points91047.1449.4557 / 36
At least 25 hesitation points45021.3323.1129 / 21
All problems500075.8276.78134 / 86

Accuracy increases in each of these 15 parts, at every difficulty level and in every subject.

Where the gain comes from (lines 400 to 410, Figure 3, Table 11). The last three rows divide the problems by the number of hesitation points in their greedy solution, which is fixed before any intervention. The group with 0 to 10 hesitation points contains 73% of the problems, and greedy decoding already solves 89.73% of them. HEST raises the accuracy of the two groups that hesitate more from 47.14% to 49.45% and from 21.33% to 23.11%. The same pattern holds in the six settings of Table 2: summed over all settings and probes, HEST fixes 81 problems and breaks 33 in the group with at least 25 hesitation points. HEST thus leaves the problems that greedy decoding already solves almost unchanged and improves those on which the model hesitates often.

With the step length selected on held-out problems, three probes per setting and the larger pool, the gains hold under all three conditions that the reviewer names in Q1.

AuthorsOpenReview ↗

Response to Reviewer pVhs (3/5): W2 and Q2

W2. The direction controls do not show that the gain comes from the probe direction. Under the same gate and edit size, the fixed direction, which uses the CAA direction as is, also reaches 76.00 (+1.80) on MATH-500 (13/4 fixed/broken, against 16/7 for HEST), and the reverse direction also gains +0.40 and +0.15 points on MATH-500 and GSM8K. The fixed direction gains less on GSM8K (+0.08 vs. +0.53), but over both benchmarks only the probe direction is right on 39 problems and only the fixed direction on 33. No random-direction control at hesitation points is reported.

Q2. Random-direction control. Please run a random direction (several seeds) and the fixed direction with the same gate, KV rollback and edit size in all six settings. The fixed direction is currently reported only for the development model. Weakness 2 would be addressed if HEST clearly beats both.

Answer: We thank the reviewer for this request. The submission reported the fixed direction for the development model only and had no random direction, so it could not show that the gain comes from the direction of the probe. We have run the controls exactly as the reviewer specifies (Section 5, Table 3, Appendix J.1).

Setup (lines 1350 to 1361). Every control keeps the layer, the threshold, the gate, the KV rollback and the fallback of HEST and changes only the edit. The fixed direction is the unit CAA direction. A random direction is a vector with independent standard normal entries in the hidden space, normalized to unit length, and we draw three of them with different random seeds. The reverse direction is HEST with the step taken away from the ideal point. In each setting, the fixed and the random directions have the median relative edit size of HEST at its validation-selected step length. We run all controls in all six settings.

Results. Table 3 of the revised paper reports the test accuracy of every variant, and Table 15 in Appendix J.1 gives every control run with its numbers of problems fixed and broken. HEST is run here with the probe that selects the step length, which is the probe that the controls use. The random column is the mean and standard deviation over the three random directions. The accuracies are as follows.

ModelDatasetGreedyHESTFixedEuclideanReverseRandom, mean ±\pm std
Qwen2.5-Math-1.5BMATH-50074.2075.0074.8073.8074.2074.20 ±\pm 0.35
Qwen2.5-Math-1.5BGSM8K84.8485.2984.7684.6984.4684.56 ±\pm 0.63
Llama-3.1-8BMATH-50050.8052.0049.4050.4050.2050.20 ±\pm 0.92
Llama-3.1-8BGSM8K84.6185.0684.2384.0884.3184.53 ±\pm 0.23
Qwen2.5-Math-7BMATH-50084.0084.6083.8084.2082.2084.33 ±\pm 0.12
Qwen2.5-Math-7BGSM8K95.3895.6895.5395.3895.3095.58 ±\pm 0.36

HEST against the fixed and the random directions. HEST has the highest accuracy in all six settings (lines 475 to 483). It is the only variant whose mean accuracy over the six settings is above that of greedy decoding, and the only one that fixes more problems than it breaks. The fixed direction is below greedy decoding in four of the six settings. The mean accuracy of the random directions over the six settings is 78.90%, which is close to the 78.97% of greedy decoding, so a small edit at hesitation points is close to neutral in a random direction.

The reverse direction no longer gains. The reviewer notes that in the submission the reverse direction also gained on the development model. With the step length selected on held-out problems, the reverse direction does not exceed greedy decoding in any of the six settings, and it fixes 69 problems and breaks 91. Stepping along the direction of the probe improves accuracy, and stepping against it lowers it.

Conclusion. With the same gate, the same KV rollback and the same edit size, neither a fixed direction nor random directions give the gain of HEST. The direction of the probe is the one that gives a consistent gain. We have added this conclusion to Section 5 and to the conclusion of the paper.

Reviewer_Lqs2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Reasoning of large language models is hierarchical when solving a mathematical problem, the solution branching from a few tokens where next-token entropy is high. The authors argue that this tree-like structure embeds better in hyperbolic space than in Euclidean space. With this idea, they propose a hyperbolic entropy steering, which embeds the hidden states in the poincare ball labeling them by next-token entropy and moves the states above a certain entropy threshold along the geodesic of steepes descent of a readout of the probe, mapping the change back to the hidden state. They show their method on two models and two datasets, showing that the improvement is largest on problems where the model hesitates often.

Critical strengths
  1. The model’s own uncertainty decides where to intervene, and it also trains the probe that decides how. The method needs no correctness labels, contrastive data or fine-tuning. This matters because, if the effect holds, the recipe is cheap and widely applicable.

  2. The paper includes negative results (OlympiadBench) and counts of problems fixed and broken. It also reports significance tests and openly states its limitations. Stating negative results is valuable for the community so I see it as a strength.

Critical weaknesses
  1. The main improvements are not statistically established and overall look fairly small. I think it should be possible to estimate the variability of the results, varying the training run of the probe and the 100 problems used to train it. And it should be cheap to run. Without these estimates, we can't say whether there is a significant improvement with HEST, which is a claim of the paper.

2.If I see correctly from table 3b, a plain fixed direction applied to the same tokens with same sample size matches HEST. This makes it unclear whether the hyperbolic space is justified, which is another claim of the paper.

Questions for authors
  1. Can you compute a significance of the results from Table 2?

  2. Can you discuss the results of table 3b where the ablation shows comparable ressults to HEST?

  3. Which readout should a user take? It should be possible to choose the Bussman or the radius variant before running the expeirment, not based on the result.

  1. Can you justify the choice of layers? I couldn't find any justification for choosing layer 19 in the main text. It might help to do a layer sweep to corroborate the choice.
10 comments
AuthorsOpenReview ↗

Response to Reviewer Lqs2 (1/3): W1 and Q1

We sincerely thank the reviewer for the thoughtful comments and for the concrete suggestions on how to strengthen the evidence. We have followed them, and we believe the paper is considerably stronger as a result. All new results are marked in blue in the revised PDF, and the line, table and appendix numbers below refer to it.

W1. The main improvements are not statistically established and overall look fairly small. I think it should be possible to estimate the variability of the results, varying the training run of the probe and the 100 problems used to train it. And it should be cheap to run. Without these estimates, we can't say whether there is a significant improvement with HEST, which is a claim of the paper.

Q1. Can you compute a significance of the results from Table 2?

Answer: We fully agree that the submission did not estimate the variability of the results, and we thank the reviewer for pointing out that this is cheap to do. We have rerun the evaluation in the following way.

1. Variability across probes (Table 2, and Table 9 in Appendix H.1). We now train three probes per setting with different random seeds and evaluate each of them on the test set. In addition, the step length is no longer chosen on the test sets: for every model and dataset it is selected on 500 held-out training problems that are disjoint from the 100 problems of the probe (lines 287 to 291, Table 1). The accuracy of every probe, together with the mean and the standard deviation, is shown below.

ModelDatasettGreedyHEST, each of the three probesHEST, mean ±\pm std
Qwen2.5-Math-1.5BMATH-5000.2574.2075.00 / 74.60 / 74.4074.67 ±\pm 0.31
Qwen2.5-Math-1.5BGSM8K184.8485.29 / 85.06 / 84.6985.01 ±\pm 0.31
Llama-3.1-8BMATH-5000.550.8052.00 / 51.80 / 51.8051.87 ±\pm 0.12
Llama-3.1-8BGSM8K0.184.6185.06 / 84.23 / 84.7684.69 ±\pm 0.42
Qwen2.5-Math-7BMATH-500184.0084.60 / 85.40 / 85.2085.07 ±\pm 0.42
Qwen2.5-Math-7BGSM8K195.3895.68 / 95.38 / 95.6095.55 ±\pm 0.16

HEST improves the mean accuracy of greedy decoding in all six settings. The standard deviation across probes is 0.12 to 0.42 points, and 16 of the 18 individual runs are at or above greedy decoding (lines 337 to 343).

2. Significance (Section 4.3, Figure 2, Table 10). On MATH-500, a change of one point corresponds to only five problems, so we test significance on a much larger set. We run the development model on the full MATH test set of 5,000 problems, with the probe and the step length unchanged. HEST raises accuracy from 75.82% to 76.78%. An exact two-sided sign test on the problems whose outcome changes gives p = 0.0015. The improvement is therefore statistically significant. Table 10 of the revised paper divides the test set by difficulty level and by subject, and the results are as follows.

Part of the MATH test setProblemsGreedyHESTFixed / broken
Level 143794.7496.348 / 1
Level 289489.0489.8211 / 4
Level 3113184.8886.0324 / 11
Level 4121473.8974.6332 / 23
Level 5132454.6855.5959 / 47
Algebra118791.6692.2515 / 8
Prealgebra87185.7686.4518 / 12
Number theory54079.2680.3715 / 9
Counting and probability47469.8371.5219 / 11
Geometry47964.5165.9713 / 6
Precalculus54662.4564.1018 / 9
Intermediate algebra90360.5861.1336 / 31
All problems500075.8276.78134 / 86

Accuracy increases at each of the five difficulty levels and in each of the seven subjects, so the gain does not come from one part of the benchmark.

We hope that these two additions establish the improvement that the reviewer asks about. The comparison with the fixed direction of Table 3b is answered in the next comment.

AuthorsOpenReview ↗

Response to Reviewer Lqs2 (2/3): W2 and Q2

W2. If I see correctly from table 3b, a plain fixed direction applied to the same tokens with same sample size matches HEST. This makes it unclear whether the hyperbolic space is justified, which is another claim of the paper.

Q2. Can you discuss the results of table 3b where the ablation shows comparable ressults to HEST?

Answer: We thank the reviewer for raising this point, which is central to the claim of the paper. The reviewer reads Table 3b of the submission correctly: on the development model on MATH-500, a fixed direction applied at the same tokens with the same edit size reached the same accuracy as HEST. That table covered one model and one benchmark. We now run this comparison in all six settings, and we add a Euclidean probe, a reverse direction and three random directions (Section 5, Table 3, Appendix J).

Setup (lines 1350 to 1361). Every variant keeps the layer, the threshold, the gate and the fallback of HEST and changes only the edit. The fixed direction is the unit CAA direction. The Euclidean probe is trained in the same way as the probe of HEST and differs only in the curvature. The size of the edit is matched to that of HEST in every setting, so the direction is the only difference.

Results. Table 3 of the revised paper reports the test accuracy of every variant, with the step length selected on held-out training problems. HEST is run here with the probe that selects the step length, which is the probe that the controls use. The accuracies are as follows.

ModelDatasetGreedyHESTFixedEuclideanReverseRandom, mean ±\pm std
Qwen2.5-Math-1.5BMATH-50074.2075.0074.8073.8074.2074.20 ±\pm 0.35
Qwen2.5-Math-1.5BGSM8K84.8485.2984.7684.6984.4684.56 ±\pm 0.63
Llama-3.1-8BMATH-50050.8052.0049.4050.4050.2050.20 ±\pm 0.92
Llama-3.1-8BGSM8K84.6185.0684.2384.0884.3184.53 ±\pm 0.23
Qwen2.5-Math-7BMATH-50084.0084.6083.8084.2082.2084.33 ±\pm 0.12
Qwen2.5-Math-7BGSM8K95.3895.6895.5395.3895.3095.58 ±\pm 0.36

Discussion of the fixed direction. The agreement in Table 3b of the submission does not extend beyond that one setting. HEST has the highest accuracy in all six settings (lines 475 to 483). The fixed direction is below greedy decoding in four of the six settings, and over the six settings it breaks more problems than it fixes, 89 against 80. For Llama-3.1-8B-Instruct on MATH-500, HEST reaches 52.00% and the fixed direction 49.40%, against 50.80% for greedy decoding. A fixed direction thus does not match HEST once more than one setting is considered.

Why the hyperbolic space is justified. The Euclidean probe is the direct test of this claim, because it uses the same training data, the same loss and the same readout, and only the curvature changes (lines 484 to 488, Appendix J.2). It is below HEST in all six settings, and its mean accuracy over the six settings is 78.76%, below the 78.97% of greedy decoding and the 79.61% of HEST. This agrees with the geometric argument of Section 3.2: in the hyperbolic probe space the Busemann direction depends on the state, whereas its Euclidean counterpart is the same direction at every state.

Taken together, a fixed direction, a Euclidean probe, a reverse direction and random directions all fail to give the gain of HEST under the same gate and the same edit size. We have added this conclusion to Section 5 and to the conclusion of the paper.

AuthorsOpenReview ↗

Response to Reviewer Lqs2 (3/3): Q3 and Q4

Q3. Which readout should a user take? It should be possible to choose the Bussman or the radius variant before running the expeirment, not based on the result.

Answer: We thank the reviewer for this question, and we agree that the choice must be made before the experiment. The answer is the Busemann readout, and the revised paper now states this explicitly in Section 3.5 (lines 291 to 293).

The reason is Proposition 1 and does not depend on any result. The decrease of the Busemann readout holds at every state and for every step length, whereas that of the radius readout requires that the step is not longer than the distance of the state to the origin. A user can therefore apply the Busemann readout with any step length, without checking the states.

Following this choice, all main results of the revised paper use the Busemann readout only. This includes the selection of the step length in Table 1, the test accuracy in Table 2, the full MATH test set in Section 4.3 and the controls in Table 3. The Busemann readout is also robust to the step length. On the development model on MATH-500, it is above greedy decoding at each of the four step lengths that we consider, as the table below shows (Table 20 in Appendix L.2).

Step length t0.10.250.51
Busemann readout, accuracy on MATH-500 (%), greedy 74.2074.8075.0076.0075.40

Q4. Can you justify the choice of layers? I couldn't find any justification for choosing layer 19 in the main text. It might help to do a layer sweep to corroborate the choice.

Answer: We thank the reviewer for pointing out that the main text gave no justification, and we have added it to Section 3.5 (lines 278 to 283).

We take the layer from prior work and do not tune it. For the two Qwen models, we use layer 19, which is the layer that SEAL uses for DeepSeek-R1-Distill-Qwen-1.5B and 7B, two models that are fine-tuned from Qwen2.5-Math. For Llama-3.1-8B-Instruct, we use layer 13, which is the layer that CAA and ODESteer use.

The layer sweep that the reviewer suggests supports this choice. On the development model on MATH-500, we compare layer 19 with an earlier and a later layer, with every other setting unchanged (Appendix L.1, Table 19). The accuracies are as follows.

Layer919, used by HEST25
Accuracy on MATH-500 (%), greedy 74.2074.8076.0074.40

All three layers are above greedy decoding, and layer 19 is the best of them. The main text now points to this sweep.

We thank the reviewer again for these questions, which have helped us to make the paper clearer, and we are glad to provide any further detail.

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

The paper proposes Hyperbolic Entropy Steering (HEST), an inference-time intervention for mathematical reasoning. A nonlinear probe maps hidden states into a Poincaré ball and is trained to rank the model's own next-token entropy, with an additional trajectory-local triplet objective. At high-entropy positions, HEST moves the embedding along a descent geodesic of either a Busemann or radial readout, then approximately realizes this change in the hidden state using damped Gauss–Newton updates. Proposition 1 establishes an exact decrease of the probe readout under the geometric step. Experiments cover three instruction-tuned models on MATH-500 and GSM8K, plus OlympiadBench on the development model. The selected Busemann configuration improves accuracy in five of six main settings, by up to 1.8 percentage points, with small median intervention rates. Ablations examine intervention positions, directions, and Euclidean versus hyperbolic probes.

Critical strengths
  1. An actionable intervention with modest supervision requirements. The probe uses model-generated entropy rather than correctness annotations, and the same uncertainty signal identifies intervention positions. The implementation specifies how the edited state is mapped back and how the cache is recomputed (Sections 3.1–3.4, pp. 4–5). This makes the proposal concrete and potentially useful where labeled steering data are unavailable. This advantage concerns probe training; step-length selection does use benchmark accuracy.

  2. The method is accompanied by meaningful diagnostics and transparent negative evidence. The exact readout-decrease result gives the geometric step a clear interpretation, and Appendix C checks its approximate realization in hidden space. The paper also reports fixes versus breaks, Euclidean and gated fixed-direction controls, failed transfer to OlympiadBench, and limitations of its statistical tests (Appendices H, M, and N). These disclosures substantially improve the assessability of a method whose gains are small.

Critical weaknesses
  1. The accuracy improvements are not yet supported by a convincing confirmatory evaluation. The Busemann step length is selected using MATH-500 and GSM8K accuracy on the development model, and those same results enter the headline evaluation (Section 3.5, p. 6). The largest gain, 1.8 points, corresponds to 16 fixes and 7 breaks on 500 problems. Appendix H reports individual two-sided paired-test p-values of 0.09–0.64 across the six settings; its pooled result is approximately 0.050 and explicitly assumes independence despite shared benchmark problems across models. Each setting also uses one trained probe. The transfer settings provide useful evidence, but they do not establish that the small gains are robust to problem sampling and probe training. These limitations matter because improved mathematical accuracy is the principal empirical contribution; they do not establish that the true effect is zero.

  2. The evidence does not sufficiently isolate the benefit of hyperbolic, state-dependent steering over a simpler matched intervention. The headline CAA baseline intervenes at every token with a substantially larger edit, whereas the gated, size-matched fixed direction already achieves the same 76.00% MATH-500 accuracy as HEST (Table 3, p. 9). Across the two development benchmarks, Appendix M reports 39 problems favoring HEST and 33 favoring that fixed direction; an exact two-sided sign test on these counts gives approximately 0.556. The Euclidean comparison is more favorable to HEST and should be credited, but it is restricted to one model and one trained probe per setting, with HEST's step selected on the test sets while the Euclidean step only matches edit size (Appendix N, p. 26). Thus neither superiority to the simpler gated direction nor a robust geometry-specific advantage is established. Moreover, decreasing an entropy-correlated probe readout does not itself establish a decrease in actual next-token entropy after intervention.

Questions for authors
  1. Assessment-critical validation: With step lengths selected on disjoint development problems, how stable are the paired accuracy gains across independently trained probes? In particular, does the improvement remain over both greedy decoding and the gated, size-matched fixed-direction control? Please separate the already reported exploratory results from any new confirmatory evidence.

  2. Assessment-critical comparison: What happens when the gated fixed-direction control and the Euclidean probe are evaluated on the other two models under the same validation and tuning budget as HEST? The current strongest evidence for geometry comes from the development model, while the simpler direction matches its largest reported gain.

  3. Mechanistic clarification: At the same prefix and edited position, how often does HEST reduce the actual next-token entropy, and how does that change relate to fixes and breaks? Appendix E's entropy decrease at the following position in unsteered greedy solutions does not answer this intervention-level question. This distinction affects whether the proposed explanation is supported beyond the probe-space guarantee.

3 comments
AuthorsOpenReview ↗

Response to Reviewer tYBj (1/3): W1 and Q1

We sincerely thank the reviewer for the careful, precise and constructive review. The comments have led us to rerun the whole evaluation, and we believe the paper is considerably stronger as a result. All new results are marked in blue in the revised PDF, and the line, table and appendix numbers below refer to it.

W1. The accuracy improvements are not yet supported by a convincing confirmatory evaluation. The Busemann step length is selected using MATH-500 and GSM8K accuracy on the development model, and those same results enter the headline evaluation (Section 3.5, p. 6). The largest gain, 1.8 points, corresponds to 16 fixes and 7 breaks on 500 problems. Appendix H reports individual two-sided paired-test p-values of 0.09–0.64 across the six settings; its pooled result is approximately 0.050 and explicitly assumes independence despite shared benchmark problems across models. Each setting also uses one trained probe. The transfer settings provide useful evidence, but they do not establish that the small gains are robust to problem sampling and probe training. These limitations matter because improved mathematical accuracy is the principal empirical contribution; they do not establish that the true effect is zero.

Q1. Assessment-critical validation: With step lengths selected on disjoint development problems, how stable are the paired accuracy gains across independently trained probes? [...] Please separate the already reported exploratory results from any new confirmatory evidence.

Answer: We fully agree with this concern. In the submission, the step length of the development model was selected with test accuracy, and every setting used one probe. We have replaced this evaluation with the protocol that the reviewer asks for.

1. The step length is now selected on disjoint held-out problems (lines 287 to 291, Appendix G.2). For every model and dataset, we draw 500 problems from the training split that are disjoint from the 100 problems of the probe and from the test sets. The MATH validation set is stratified by difficulty level to match MATH-500. We run HEST on these 500 problems with each step length in {0.1, 0.25, 0.5, 1} and select the one with the highest accuracy. The thresholds are unchanged from the submission and are calibrated without labels, so no test label enters any hyperparameter. Table 1 of the revised paper reports the number of correctly solved validation problems for every step length. HEST is robust to this choice: within a setting, the four step lengths differ by only 3 to 8 of the 500 problems.

2. Every setting is now run with three probes (Table 2, and Table 9 in Appendix H.1). We train three probes per setting with different random seeds and evaluate each of them on the test set at the selected step length. The accuracy of every probe, together with the mean and the standard deviation, is shown below.

ModelDatasettGreedyHEST, each of the three probesHEST, mean ±\pm std
Qwen2.5-Math-1.5BMATH-5000.2574.2075.00 / 74.60 / 74.4074.67 ±\pm 0.31
Qwen2.5-Math-1.5BGSM8K184.8485.29 / 85.06 / 84.6985.01 ±\pm 0.31
Llama-3.1-8BMATH-5000.550.8052.00 / 51.80 / 51.8051.87 ±\pm 0.12
Llama-3.1-8BGSM8K0.184.6185.06 / 84.23 / 84.7684.69 ±\pm 0.42
Qwen2.5-Math-7BMATH-500184.0084.60 / 85.40 / 85.2085.07 ±\pm 0.42
Qwen2.5-Math-7BGSM8K195.3895.68 / 95.38 / 95.6095.55 ±\pm 0.16

HEST improves the mean accuracy of greedy decoding in all six settings. The standard deviation across probes is 0.12 to 0.42 points, and 16 of the 18 individual runs are at or above greedy decoding (lines 337 to 343).

3. A larger pool of problems (Section 4.3, Figure 2, Table 10). To address problem sampling, we also run the development model on the full MATH test set of 5,000 problems, with the probe and the validation-selected step length unchanged. HEST raises accuracy from 75.82% to 76.78%. An exact two-sided sign test on the problems whose outcome changes gives p = 0.0015. Accuracy increases at each of the five difficulty levels and in each of the seven subjects.

Exploratory and confirmatory results. All results of HEST, of the controls and of HARP in Tables 1 to 3 and in Figures 2 and 3 are new and follow this protocol.

The comparison with the gated fixed direction, which is the second part of Q1, is answered together with W2 and Q2 in the next comment.

AuthorsOpenReview ↗

Response to Reviewer tYBj (2/3): W2 and Q2

W2. The evidence does not sufficiently isolate the benefit of hyperbolic, state-dependent steering over a simpler matched intervention. The headline CAA baseline intervenes at every token with a substantially larger edit, whereas the gated, size-matched fixed direction already achieves the same 76.00% MATH-500 accuracy as HEST (Table 3, p. 9). Across the two development benchmarks, Appendix M reports 39 problems favoring HEST and 33 favoring that fixed direction; an exact two-sided sign test on these counts gives approximately 0.556. The Euclidean comparison is more favorable to HEST and should be credited, but it is restricted to one model and one trained probe per setting, with HEST's step selected on the test sets while the Euclidean step only matches edit size (Appendix N, p. 26). Thus neither superiority to the simpler gated direction nor a robust geometry-specific advantage is established.

Q1. [...] In particular, does the improvement remain over both greedy decoding and the gated, size-matched fixed-direction control?

Q2. Assessment-critical comparison: What happens when the gated fixed-direction control and the Euclidean probe are evaluated on the other two models under the same validation and tuning budget as HEST? The current strongest evidence for geometry comes from the development model, while the simpler direction matches its largest reported gain.

Answer: We thank the reviewer for this important point. In the submission, both controls were evaluated on the development model only. We now run both of them, together with a reverse direction and three random directions, in all six settings (Section 5, Table 3, Appendix J).

Setup (lines 1350 to 1361). Every control keeps the layer, the threshold, the gate, the cache recomputation and the fallback of HEST and changes only the edit. The fixed direction is the unit CAA direction, and the Euclidean probe is trained in the same way as the probe of HEST and differs only in the curvature. In each setting, the size of the edit is matched to the median relative edit size of HEST at its validation-selected step length. The direction is therefore the only difference, and no test label enters these choices.

Results. Table 3 of the revised paper reports the test accuracy of every variant. HEST is run here with the probe that selects the step length, which is the probe that the controls use. The random column is the mean and standard deviation over three random directions. The accuracies are as follows.

ModelDatasetGreedyHESTFixedEuclideanReverseRandom, mean ±\pm std
Qwen2.5-Math-1.5BMATH-50074.2075.0074.8073.8074.2074.20 ±\pm 0.35
Qwen2.5-Math-1.5BGSM8K84.8485.2984.7684.6984.4684.56 ±\pm 0.63
Llama-3.1-8BMATH-50050.8052.0049.4050.4050.2050.20 ±\pm 0.92
Llama-3.1-8BGSM8K84.6185.0684.2384.0884.3184.53 ±\pm 0.23
Qwen2.5-Math-7BMATH-50084.0084.6083.8084.2082.2084.33 ±\pm 0.12
Qwen2.5-Math-7BGSM8K95.3895.6895.5395.3895.3095.58 ±\pm 0.36

The gated fixed direction. HEST has the highest accuracy in all six settings, and it is the only variant whose mean accuracy over the six settings is above that of greedy decoding (lines 475 to 483). The fixed direction is below greedy decoding in four of the six settings and breaks more problems than it fixes. The difference is clearest for Llama-3.1-8B-Instruct on MATH-500, where HEST reaches 52.00% and the fixed direction 49.40%, against 50.80% for greedy decoding. The agreement between the two methods that the reviewer points out in the submission therefore does not extend beyond that one setting.

The Euclidean probe (lines 484 to 488, Appendix J.2). The Euclidean probe is the direct test of the geometry, because only the curvature changes. It is below HEST in all six settings, and over the six settings it breaks more problems than it fixes, 85 against 73. The comparison is thus no longer restricted to one model.

Validation and tuning budget. The step length has no common scale across a hyperbolic probe, a Euclidean probe and a raw direction in the hidden space. We therefore give every control the edit size that HEST has at its validation-selected step length, so that all variants edit the hidden state by the same amount and only the direction varies (lines 1395 to 1398). Table 15 in Appendix J.1 lists the matched edit size of every setting.

The last sentence of W2, on the actual next-token entropy, is answered with Q3 in the next comment.

AuthorsOpenReview ↗

Response to Reviewer tYBj (3/3): Q3

W2. [...] Moreover, decreasing an entropy-correlated probe readout does not itself establish a decrease in actual next-token entropy after intervention.

Q3. Mechanistic clarification: At the same prefix and edited position, how often does HEST reduce the actual next-token entropy, and how does that change relate to fixes and breaks? Appendix E's entropy decrease at the following position in unsteered greedy solutions does not answer this intervention-level question. This distinction affects whether the proposed explanation is supported beyond the probe-space guarantee.

Answer: We thank the reviewer for this clarification, and we agree with both points. Proposition 1 is a statement about the readout of the probe, and the entropy at the following position of unsteered solutions does not measure the effect of the intervention. We have added the measurement that the reviewer asks for as the new Appendix K (lines 1423 to 1438). At every steered position of the test decodes, we compare the next-token distribution before the edit with the distribution after it, for the same prefix and the same position.

How often the edit lowers the actual entropy. Table 18 of the revised paper reports, for every direction, the percentage of edits that change the most likely token and the percentage that lower the actual next-token entropy. Each cell below gives these two percentages in this order.

ModelDatasetHESTFixedReverseRandom
Qwen2.5-Math-1.5BMATH-5004.3 / 52.94.3 / 66.24.3 / 51.23.8 / 50.2
Qwen2.5-Math-1.5BGSM8K10.7 / 55.612.2 / 65.010.2 / 46.78.0 / 51.5
Llama-3.1-8BMATH-5005.2 / 52.07.3 / 79.35.8 / 49.54.5 / 53.2
Llama-3.1-8BGSM8K2.8 / 50.73.1 / 54.12.6 / 51.32.4 / 51.7
Qwen2.5-Math-7BMATH-5003.3 / 64.33.2 / 97.62.5 / 28.72.0 / 51.5
Qwen2.5-Math-7BGSM8K8.2 / 52.110.6 / 65.66.1 / 49.45.4 / 55.8

HEST lowers the actual next-token entropy at 51 to 64% of its edits, and it changes the most likely token at 2.8 to 10.7% of them, more often than a random direction of the same size in all six settings. HEST therefore acts through a small number of token choices at hesitation points.

Relation to fixes and breaks. We group the edits of HEST by the outcome of their problem relative to greedy decoding. The numbers below are pooled over the six settings and the three probes.

Outcome of the problemNumber of editsEntropy decreases (%)Most likely token changes (%)
Fixed by HEST5,39655.37.1
Broken by HEST5,26253.87.3
Unchanged135,76655.75.8

The entropy decreases about equally often in the three groups, so the change of entropy at the edited position does not by itself separate fixes from breaks. The edits of problems whose outcome changes alter the most likely token more often than the others.

The effect grows with the step length. Table 17 of the revised paper extends the sweep of the development model on the MATH validation set. Both quantities increase smoothly with the step length, while the accuracy remains stable.

Step length tValidation accuracy (%)Most likely token changes (%)Entropy decreases (%)
0.182.62.548.1
0.2583.24.951.1
0.582.87.852.4
182.013.355.7
281.225.759.5
580.439.161.0

What this means for the explanation. The fixed direction lowers the entropy most often, at 54 to 98% of its edits, and yet its accuracy is below that of HEST in all six settings (Table 3). We have therefore made the explanation in the paper more precise. The main text now states that Proposition 1 describes the readout of the probe (lines 244 to 245), and Appendix K states that the gain of HEST comes from which token the edit selects, and not from a lower entropy alone.

We thank the reviewer again for these questions, which have made both the evaluation and the explanation of the paper more precise. We are glad to provide any further detail.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers: Summary of the Revision

We sincerely thank the three reviewers for their careful reading and their constructive, detailed suggestions. We are encouraged that the reviewers find the method concrete and useful, in particular that the model's own entropy decides where to steer and trains the probe without correctness labels (Reviewers tYBj, Lqs2 and pVhs), and that the paper is supported by Proposition 1 and by thorough ablations and diagnostics (Reviewers tYBj and pVhs). We have taken every comment seriously and have rerun the whole evaluation. All new results are marked in blue in the revised PDF, and the main text now has 10 pages. Below we summarize how the revision addresses the concerns raised by more than one reviewer. The individual responses give the details.

1. Statistical robustness of the gains (Reviewers tYBj, Lqs2 and pVhs)

  • The step length is now selected on 500 held-out training problems per setting, and the test labels are used only to report accuracy (Table 1).
  • Every setting is now run with three probes trained with different random seeds, and we report the mean and the standard deviation (Table 2, Table 9).
  • We evaluate on the full MATH test set of 5,000 problems. HEST raises accuracy from 75.82% to 76.78%, and an exact two-sided sign test on the problems whose outcome changes gives p = 0.0015. Accuracy increases at every difficulty level and in every subject (Section 4.3, Figure 2, Table 10).

2. External baselines (Reviewer pVhs)

  • We have added HARP, with its official code and three runs per setting. The test accuracies of Table 2 of the revised paper are as follows.
ModelDatasetGreedyCAAHARP, mean ±\pm stdHEST, mean ±\pm std
Qwen2.5-Math-1.5BMATH-50074.2072.4074.13 ±\pm 0.7674.67 ±\pm 0.31
Qwen2.5-Math-1.5BGSM8K84.8484.3884.94 ±\pm 0.5085.01 ±\pm 0.31
Llama-3.1-8BMATH-50050.8044.2050.27 ±\pm 1.3351.87 ±\pm 0.12
Llama-3.1-8BGSM8K84.6180.2184.46 ±\pm 1.0584.69 ±\pm 0.42
Qwen2.5-Math-7BMATH-50084.0083.2084.87 ±\pm 0.5085.07 ±\pm 0.42
Qwen2.5-Math-7BGSM8K95.3895.0095.53 ±\pm 0.1395.55 ±\pm 0.16

HEST has the highest mean accuracy in all six settings, and it is the only method that improves on greedy decoding in all six. It needs 1 to 14% more token forward passes than greedy decoding, against 5.5 to 30 times for HARP on the median problem (Table 14).

3. Whether the gain comes from the direction of the hyperbolic probe (Reviewers tYBj, Lqs2 and pVhs)

  • A fixed direction, a Euclidean probe, a reverse direction and three random directions are now run in all six settings, with the gate, the cache recomputation and the edit size of HEST (Table 3, Appendix J).
  • HEST has the highest accuracy in all six settings. Its mean accuracy over the six settings is 79.61%, against 78.97% for greedy decoding, 78.75% for the fixed direction, 78.76% for the Euclidean probe, 78.44% for the reverse direction and 78.90% for the random directions.

4. What the edit changes at the steered token (Reviewers tYBj and pVhs)

  • The new Appendix K compares the next-token distribution before and after the edit, at the same prefix and position. HEST changes the most likely token at 2.8 to 10.7% of the steered positions and lowers the actual next-token entropy at 51 to 64% of them (Table 18).

5. Where the gain comes from

  • The gain of HEST is largest on the problems on which the model hesitates often. Summed over all settings and probes, HEST fixes 81 problems and breaks 33 in the group with at least 25 hesitation points, and it leaves the problems that greedy decoding already solves almost unchanged (Figure 3, Table 11).

Other changes. Section 3.5 now states that the Busemann readout is the default and why, and gives the justification of the layer (Reviewer Lqs2). Figures 2 and 3 are new, and Figure 1 is redrawn.

We thank the reviewers again for their time and effort, which have made the paper considerably stronger. We hope these revisions address the reviewers' concerns, and we would be grateful if they could be taken into account in the final assessment. We are happy to answer any further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.