acceptodds
Under review as a conference paper at ICLR 2027

U-Space: Uncovering When and Why Uncertainty Arises in Language Models

Abstract

Large language models are informing decisions with ever-higher stakes. As the consequences of their errors grow, a central question becomes harder to ignore: how much can we trust an individual answer? When a wrong answer can cause serious harm, the cost of error may far exceed that of deferring to an expert. Yet recognizing when to defer remains difficult because language models can present incorrect conclusions with fluent explanations and an authoritative tone. Uncertainty quantification seeks to address this disconnect by estimating the reliability of individual predictions. However, many existing methods require repeated generations or separately trained components, and their scalar estimates do not reveal where uncertainty arises or how it evolves during reasoning. Recent work has also shown that generation length can be strongly associated with uncertainty estimates and correctness, raising the question of how much of an estimator's predictive power comes from uncertainty-specific information rather than output length alone. Mechanistic interpretability offers a way to address these limitations by connecting human-interpretable concepts to intermediate model states. Building on this capability, we introduce the U-Space, a low-dimensional subspace that makes a model's evolving uncertainty measurable and interpretable. We identify semantic anchors for doubt and certainty, map their unembedding directions back into the residual space, and combine their contrasts into an orthogonal basis. The U-Lens projects each token state onto these basis vectors, yielding an interpretable token-level uncertainty map that can be inspected directly or aggregated into a scalar uncertainty score. Our approach requires no correctness labels, repeated generations, or training. Across reasoning benchmarks, its confidence score outperforms established baselines under both standard and length-controlled evaluation and transfers more reliably than supervised estimators.

Primary area:
interpretability and explainable AI
Keywords:
uncertainty quantification, mechanistic interpretability

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_hW8xOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2(+1)
Read review
Paper summary

This paper proposes U-Space/U-Lens, aiming to move beyond the limitation of traditional uncertainty quantification methods that produce only a single scalar score. It constructs an interpretable latent subspace based on four predefined semantic dimensions—ambiguity, incompleteness, conflicting evidence, and general uncertainty—and reads out token-level uncertainty along the reasoning trajectory. The semantic signal is further combined with predictive entropy to form the final trace-level uncertainty score. Its main contribution is an interpretable semantic-direction construction method that requires neither correctness labels nor additional training, together with an internal-representation analysis of when and why uncertainty arises.

Critical strengths
  1. The motivation of this paper is very good. Attributing uncertainty along different dimensions is a highly worthwhile direction to explore.
  2. The method has a certain degree of interpretability. The technical approach of constructing label-free interpretable latent directions through semantic anchors is itself insightful.
  3. The experimental coverage is relatively broad, and the exploration of the method’s generality is fairly proactive. Although the length-related experiments are debatable, the experiments themselves are well designed. The authors recognize that merely showing correlational activation heatmaps is insufficient and therefore further attempt causal interventions, which is commendable.
  4. Computational efficiency and deployment friendliness are clear advantages.
  5. The presentation quality of the paper is good. The figures and tables are all well designed.
Critical weaknesses
  1. The methodological novelty and the primary evaluation objective are not truly aligned. The paper claims that U-Space decomposes uncertainty into four interpretable dimensions and emphasizes its goal of answering when and why uncertainty arises. However, in the main experiments, the authors ultimately aggregate the four dimensions into a single score and use conventional correct/incorrect discrimination metrics as the primary evaluation criteria. As a result, the paper’s most distinctive contribution, namely “structured uncertainty attribution,” is not directly validated. The current validation of the category semantics mainly relies on only 12 manually constructed matched prompt pairs per category and several qualitative examples. A more appropriate experimental design would treat uncertainty attribution itself as the primary task and correctness prediction as a downstream application.

  2. The authors divide uncertainty into four dimensions, but the orthogonality of these dimensions is imposed by the algorithm rather than being an intrinsic property of the dimensions themselves. For example, in Appendix C, the authors obtain different uncertainty directions using matched-prompt activation differences, and the pairwise cosine similarities among several directions are 0.51, 0.60, and 0.65, indicating substantial correlations. Moreover, after enforcing orthogonality algorithmically, do the resulting category directions still preserve their original semantic meanings?

  3. The paper motivates the necessity of length-controlled evaluation by observing that incorrect answers are substantially longer than correct answers. Prior work cited by the paper has shown that different correctness evaluation methods may exhibit substantial length bias [1]. String-matching-based methods are highly sensitive to response length, whereas LLM-based evaluation criteria are much less sensitive to length. I would like to know how much of the observed length–correctness association remains if the same set of samples is relabeled using an LLM-based evaluator.

  4. The current experimental results are not sufficient to establish that the advantage of U-Lens over existing methods has been evaluated fairly and reliably. Many baselines perform substantially worse in this paper than the levels commonly reported in their original papers or other standard evaluations. For example, Semantic Entropy achieves an average AUROC of 0.790 across 30 model–task combinations in the original paper, whereas it achieves only 0.646 in this paper; similarly, P(True) decreases from approximately 0.698 in the original paper to 0.567 here [2].

  5. There is an inconsistency between the code and the paper that may directly affect the reported results. The paper defines the U-Space category direction by separately normalizing the centroids of the positive and negative anchors, i.e., μc±=norm(mean(norm(ga))),\mu_c^\pm=\mathrm{norm}(\mathrm{mean}(\mathrm{norm}(g_a))), However, in src/uspace.py, the code normalizes only the individual anchor directions; after computing the positive and negative centroids, it does not normalize the two centroids separately before taking their difference. This means that the U-Space actually constructed by the code is different from the formulation described in the paper, which may in turn affect Acone, U-Lens, token-level attribution, steering, and all other results that depend on this basis.

[1] Santilli, A., Golinski, A., Kirchhof, M., Danieli, F., Blaas, A., Xiong, M., Zappella, L., & Williamson, S. (2025). Revisiting Uncertainty Quantification Evaluation in Language Models: Spurious Interactions with Response Length Bias Results. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), 743–759.

[2] Farquhar, S., Kossen, J., Kuhn, L., & Gal, Y. (2024). Detecting Hallucinations in Large Language Models Using Semantic Entropy. Nature, 630(8017), 625–630.

Questions for authors
  1. Could the authors provide evidence that the main conclusions are robust to the correctness evaluator? I would like to know whether the length–correctness association and the relative advantage of U-Lens persist when the same outputs are relabeled using an LLM-based evaluator or another less length-sensitive correctness criterion.
  2. Could the authors clarify whether the four orthogonalized U-Space axes still preserve category-specific semantics after orthogonalization?
  3. Could the authors explain why several baselines perform substantially worse than the results reported in their original papers?
4 comments

Thank you for the thoughtful review. We appreciate that you find the attribution of uncertainty to distinct dimensions worthwhile and are especially grateful for raising the inconsistency between the code and the paper.

W1: Attribution vs. aggregated evaluation. This touches on a key difficulty of uncertainty quantification. Uncertainty itself has no ground truth, so it is almost always evaluated as a proxy for answer correctness, a label that exists only at the output (Beigi et al., 2024). It is therefore not obvious how to test whether a model is uncertain at a given point of its reasoning, or why. We use correctness because it is the established protocol and allows a fair comparison with existing estimators, but we agree that it does not test the attribution itself. A first step toward such a test is to use inputs whose source of uncertainty is fixed by construction. We ran a test, specified before any trace was generated, on the Known-Unknown Questions dataset (Amayuelas et al., 2024), whose unanswerable questions are annotated as ambiguous, unsolved, about the future, or controversial. The matching category is identified with an AUROC of 62 to 65 across the three models, against about 50 for random bases and 55 to 59 for counting the anchor words, with both margins significant on every model. Notably, the source becomes visible over the course of the reasoning rather than in its final state, which is what the token-level readout is designed for. We added the test to Appendix E: Interpretability of Contrastive Activation Differences. We see this as an encouraging but partial answer to an open question and would be glad to discuss how attribution during reasoning could be evaluated more directly.

W2 and Q2: Decodability after orthogonalization. Each orthogonalized axis still decodes to the vocabulary of its category. We decoded every column of UℓU_\ell through the J-Lens over the full vocabulary and added the result to the new Appendix D: Axis and Orthogonalization Ablation. On all three models, each axis keeps the anchors of its own category at the 98th to 100th percentile of the vocabulary, while the anchors of the other categories fall by 15 to 22 percentile points. Orthogonalization thus makes each axis more specific to its category.

W3 and Q1: Length-confound of correctness evaluator. We relabeled every response of seed 42 with an LLM judge, Qwen3-32B, a model we do not evaluate. The judge sees the question, the reference answer and the final answer of each response, but never the reasoning trace, so the length of the trace cannot influence its verdict. It agrees with our graders on 93.9 to 99.7 percent of the responses per model and benchmark. The association between length and correctness does not weaken. The AUROC of generation length for predicting an error is as high or higher under judge labels in 11 of 12 model and benchmark combinations. The lead of U-Lens over the strongest baseline under length control also persists and grows slightly, from 1.3, 1.2 and 4.2 to 2.0, 1.8 and 4.8 AUROC points. We added this analysis to Appendix H: Extended Results.

W4 and Q3: Baseline performance delta. Thank you for raising this. Part of the gap concerns the sampling temperature. The submission sampled Semantic Entropy (SE) at 0.5, the setting recommended by Kuhn et al. (2023), and the temperature sweep in our response to Reviewer UVBE shows that this is not the strongest setting for our models. The revised paper therefore samples at each model's own temperature. Another difference lies in the setting. The 0.790 of Farquhar et al. averages short-form QA with sentence-length answers, whereas our responses are reasoning traces with a median length of 3.8k to 14.9k tokens, mostly on hard multi-step benchmarks. SE was not designed for such traces. Thus, we adapted it further in its favor. We first identified where each model states its final answer and adapted SE so that its entailment model, which can only read a limited number of tokens, receives the portion of each answer where that statement appears. This raises Semantic Entropy from 64.6 to 67.8 AUROC on average, and on TriviaQA, the benchmark closest to the setting of Farquhar et al., it reaches 79.5 to 84.1. U-Lens still leads it by 3.2 AUROC points on average and by 3.3 under length control, both significant. The same difference in setting applies to P(True), which here judges answers to hard multi-step problems reached after long reasoning rather than short factual answers.

W5: Equation and code. Thank you for checking the code this closely and catching this. You are right, the code averages the normalized anchor directions of each pole and normalizes only the difference of the two centroids, whereas Eq. 1 additionally normalized each centroid. The error was in the equation. All results in the paper were produced with the released implementation and are therefore unaffected. We corrected Eq. 1 in the revised version.

Thank you again for your time and thoughtful feedback on our submission.

As the discussion period is approaching its end on November 18 (AoE), we wanted to kindly follow up to see if our previous responses and the revised manuscript have adequately addressed your concerns. If any points remain unclear or if you have any further questions, please let us know. We would be more than happy to clarify them promptly before the deadline.

Best regards, Authors

The authors have addressed most of my concerns, particularly regarding reproducibility, and I have increased the Soundness score to 2. As a side note, I am also working on uncertainty attribution research. I believe the problem with current uncertainty quantification methods is that they struggle to identify what specific risks are present and what interventions should be taken, which greatly limits the usefulness of uncertainty in supporting practical decision-making. I had hoped to find some inspiration in the authors' work, but they still follow the traditional UQ evaluation paradigm, which is somewhat disappointing. Nevertheless, overall, this is a good paper.

Thank you for revisiting our work and increasing the Soundness score! We share your concern regarding current uncertainty estimates and their limited practical value, which is precisely why we proposed our interpretable U-Space approach. We debated evaluation approaches extensively but chose not to introduce a new paradigm, as our length-controlled evaluation already highlights another deficit in current evaluations (the potential length shortcut) affecting many UQ methods. Comparability with existing methods was another factor, though we remain keen to explore alternatives that better capture practical usefulness. Intrinsic ground truth for model uncertainty remains hard to capture, but we did consider controlled task difficulty as a proxy, especially in domains where difficulty can be increased gradually. Overall, we see our readouts through verbalizable concepts as a promising step toward more detailed uncertainty attribution. Our steering results suggest that the identified directions causally influence model outputs, opening a promising avenue for intervention, while leaving open which interventions are appropriate for each uncertainty type. We hope our work offers something useful for your research and would welcome your suggestions for evaluating these aspects more directly. If our clarifications have addressed your remaining concerns and you deem our work useful for the scientific community, we would appreciate it if that could be reflected in the overall recommendation as well.

Reviewer_rSGZOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper propose U-space of residual stream for uncertainty estimation for reasoning LLM. It first build a 4 axis subspace by Jacobian lens, taking diff of means doubt and certainty words and orthonormalise the result. the U-Lens then reads the space by token map and trace score. It is based on Jacobian lens and difference of means, and combine to propose new four-axis basis, the cone score and the map. It exp on 3 open source reasoning LLM, and report AUROC on length-controlled setting. it compaires with single pass, man-pass and supervised baselines.

Critical strengths
  1. The paper's method is well motivated, every design choice have a classical or popular motivation, like taxonomies of imperfect information (Parsons 1996) for 4 category; and use latest 2026 tech like Jacobian-lens from anthropic. And most of them are proved with evidence.
  2. The paper's content especially figure and table is well designed and impressive, like the novel colored points in table1 and the elegant figure1 above the intro.
Critical weaknesses
  1. The abstract says "its confidence score outperforms established baselines under both standard and length-controlled evaluation". But for uncertainty quantification results for each model (Table12), TokUR is higher on Gemma 4 (75.1 vs 73.2). And for Table 1: 71.1+4.7 against TokUR 69.6+4.9 unadjusted, a 1.5-point gap. The + is the spread across three models, there might be noise without 95% CI for each model.
  2. Some of design choice miss experiment evidence, like four categories v.s. score with one doubt-minus-certainty axis, and with each category alone; and Orthonormal basis.
Questions for authors
  1. How firm is the lead in the main table? for the table1, and table 12, can you provide 95% confidence interval on the margin . and can you provide adjusted uncertainty quantification results for each model?
  2. Does length control affect the evaluation? does leaving out truncated traces remove the longest errors? need finer bins or a length-regressed score; or results with truncated traces counted as errors.
  3. Can you do more test to solve weakness 2? like for Four category, score with one doubt-minus-certainty axis, and with each category alone. and for Orthonormal basis: score the trace with twenty random orthonormal four-dimensional bases at the same layer, the way the paper already does for its emotion readout;
2 comments

We thank you for a clear and constructive review. We appreciate that you see each design choice as well-motivated and that you valued the effort spent on figures and tables.

W1 and Q1: Missing confidence intervals. Thank you for this question. We computed 95% confidence intervals for the margin of U-Lens over every baseline, paired by item, with 1,000 bootstrap draws over the benchmark items applied to all three seeds, so that each margin matches the three-seed tables. For the cross-model averages in Table 1, U-Lens is ahead of every baseline with an interval that excludes zero in AUROC, AUPRC, and AURC, with and without length control. The 1.5-point gap to TokUR you mention is therefore not noise (+1.8 [+0.2,+3.3]+1.8\ [+0.2, +3.3] AUROC on the shared items). Per model under length control, U-Lens is ahead of all eleven baselines in every metric, and the interval excludes zero everywhere except against the computationally much more expensive Semantic Entropy in AUPRC and on Qwen in AUROC, where the margin is positive but not significant. (Semantic Entropy needs ten additional generations per question, whereas U-Lens reads the single generation it scores.) Without length control, the per-model picture is mixed, as you note, and TokUR's lead on Gemma is significant (−1.7 [−3.3,−0.2]-1.7\ [-3.3, -0.2]). Table 2 already reports the length-matched AUROC, AUPRC, and AURC of every method for each model. We added the interval table to Appendix H: Extended Results and now state the claim precisely in the abstract and introduction.

Q2: Truncation. Any samples that are not scoreable due to truncation or other factors are not included in the evaluation. We chose not to include these to prevent potentially correct samples from being counted as errors, biasing the evaluation. Due to the comparatively large generation budget, truncation only happens in 0.2 to 4.1% of samples. We have added per-model truncation rates to Table 14: Generation details in the appendix.

To test whether leaving them out removes the longest errors, we kept every truncated trace and counted it as an error. Such a trace usually stops before the end-of-thinking marker, so we read U-Lens at its last token and recomputed the single-pass baselines on the same items (on one seed). This makes generation length a stronger unadjusted predictor, from 66.6 to 70.3 AUROC on Qwen and from 68.4 to 71.6 on Magistral, which confirms that many of these traces are long errors. Under length control, however, U-Lens keeps its lead, with 71.6 vs. 69.9 on Qwen and 67.3 vs. 62.8 on Magistral. Gemma's eight truncated traces are too few to matter. Even if all of them were errors that U-Lens ranks as most confident, its length-matched AUROC would only move from 68.7 to 68.4. Finally, with 20 or 30 length bins instead of 10, the margin of U-Lens over the strongest baseline stays positive in 26 of 27 combinations of model, metric, and bin count. Both analyses are now in Appendix H: Extended Results.

W2 and Q3: Anchor construction ablation. Thank you for raising this point. We agree that these ablations are valuable additions. We added a new section to the appendix (Appendix D: Axis and Orthogonalization Ablation), where we first compare the paper's four-axis construction against a single doubt-minus-certainty axis and against random orthonormal bases. Additionally, we evaluate the impact of orthogonalization on the main construction and on each of the four axes individually. To summarize the findings, the correctness score is robust to a single-axis construction and to orthogonalization, while a random basis collapses. This aligns with the motivation for the construction, where the target of the four individual axes is the token-level readout that asks which kind of uncertainty a state expresses. Before orthogonalization, all four axes perform well as correctness indicators because they share a large common component that is also captured by the single pooled axis. This directly motivates orthogonalization, which removes this shared component and enables a differentiated readout of each axis without affecting the performance of the original combined basis used for scoring.

Thank you again for your time and thoughtful feedback on our submission.

As the discussion period is approaching its end on November 18 (AoE), we wanted to kindly follow up to see if our previous responses and the revised manuscript have adequately addressed your concerns. If any points remain unclear or if you have any further questions, please let us know. We would be more than happy to clarify them promptly before the deadline.

Best regards, Authors

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

This paper proposes U-LENS, a training-free, single-pass uncertainty estimator for reasoning LLMs. The core construction is U-SPACE: a low-dimensional subspace of the residual stream built from hand-picked uncertainty/certainty anchor words, pulled back into intermediate layers via the J-Lens mean Jacobian, contrasted (uncertainty centroid minus certainty centroid), and orthogonalized into four semantic directions (ambiguity, incompleteness, conflicting evidence, general uncertainty). At inference, the model's hidden state is projected onto this basis, yielding a token-by-category uncertainty map and a cone-alignment score A_cone. The final score multiplies A_cone at the end-of-thinking token by the mean predictive entropy over the reasoning trace. Experiments on three open-weight reasoning models and four benchmarks show that U-LENS outperforms established baselines under both standard and length-controlled evaluation, and activation steering along the U-SPACE directions causally induces hesitation.

Critical strengths
  1. The length-controlled evaluation is a genuine and timely contribution. The paper shows that generation length alone reaches 68.6% AUROC unadjusted, and that several multi-pass baselines collapse below single-pass predictive entropy once length is controlled. U-LENS drops by only 1.8 points and leads all nine model-metric comparisons. This directly addresses a known confound in the uncertainty quantification literature and makes the empirical case substantially more credible.

  2. The causal intervention provides evidence beyond correlation. Steering along the U-SPACE directions makes the model hesitate or lose confidence even on 2+2=4. This is a stronger form of validation than most interpretability papers provide, and it supports the claim that these directions capture representations that shape behavior rather than merely correlate with it.

Critical weaknesses
  1. The construction is heuristic, and the paper does not test its robustness. The four categories and the anchor word lists are hand-picked. App. C's diff-in-means decoding is post-hoc validation, not a construction-time guarantee. There is no sensitivity analysis showing what happens if anchors are swapped, perturbed, or drawn from a different source. Since the entire method rests on this choice, the lack of a robustness check is a material concern.

  2. The J-Lens pullback is a first-order approximation, but its error is never analyzed. The mean Jacobian assumes local linearity. The paper ablates readout depth and compares R-Lens vs. J-Lens, but never measures how linearization error propagates into the U-SPACE directions or the final score. If attention routing changes abruptly, the directions may not correspond to what the paper claims.

Questions for authors
  1. Can you randomly replace 20–50% of the anchor words and report how U-SPACE directions (cosine to the original basis) and U-LENS AUROC change? This would directly address the robustness concern.
  2. For the orthogonalized basis UℓU_\ell, does each column still decode to category-aligned vocabulary through the J-Lens? The paper only reports decoding for the pre-orthogonalization directions.
  3. How would U-LENS adapt for models without an explicit end-of-thinking token? Have you tried reading AconeA_{\text{cone}} at the last reasoning token or first answer token, and how much does performance drop?
  4. How were the steering λ\lambda values in Figs. 13–16 chosen? Is there a unified criterion, or were they tuned per category until the output changed?
  5. Semantic Entropy is compared at temperature 0.5. Would a temperature sweep change the comparison?
2 comments

Thank you for the positive assessment and the thoughtful review. We are glad that the steering experiments convinced you that the U-Space directions shape behavior rather than merely correlate with it.

W1 & Q1: Anchor construction robustness. We added an ablation to the end of Appendix B: Anchor Words in which we randomly replace or drop anchor terms in 10% increments and plot the impact on the AconeA_{\mathrm{cone}} and U-Lens scores as well as the cosine similarity to the unperturbed axes. Overall, the ablation shows that the construction is relatively robust. Replacing terms with random words is the harsher case. At 20%, the axes keep a cosine of at least 0.84 to the original basis, and the length-matched U-Lens AUROC drops by at most 0.6 points. Replacing 50% yields a cosine of 0.58 to 0.66 and costs at most 1.7 points. The AconeA_{\mathrm{cone}} readout reaches chance only when nearly all anchors are random.

Q2: Decodability after orthogonalization. Each orthogonalized axis still decodes to the vocabulary of its category. We decoded every column of UℓU_\ell through the J-Lens over the full vocabulary and added the result to the new Appendix D: Axis and Orthogonalization Ablation. On all three models, each axis keeps the anchors of its own category at the 98th to 100th percentile of the vocabulary, while the anchors of the other categories fall by 15 to 22 percentile points. Orthogonalization thus makes each axis more specific to its category.

W2: Jacobian pullback error impact. Thank you for raising this point. We tested this at our readout layer on 48 reasoning traces from Qwen. We shift the EOT state along each U-Space direction as in our steering experiments, let the full nonlinear model run on, and compare its response with the J-Lens prediction. The predicted change is only approximate in detail, with a cosine of 0.14 to 0.27 to the actual one, about three times that of a random direction. Its meaning, however, holds. In 99 to 100 percent of traces, the model raises the words of the steered category above those of the other three categories, against a chance level of 25 percent. The score itself reads the model's actual hidden state along these directions and never passes through the lens, so the approximation reaches it only through the directions tested here. In Appendix A, we added the paragraph Linearization error, where we provide the setup and the full results.

Q3: Models without an explicit end-of-thinking (EOT) token. To test this, we first read the state at alternative positions on our three models. The EOT marker remains the best place to read, and the last generated token is the natural fallback, with a length-matched U-Lens AUROC of 67.2 / 70.2 / 63.8 on Gemma / Qwen / Magistral against 68.9 / 71.8 / 68.0 at the marker (on a single seed). Second, we ran the full evaluation on Gemma-3-27B-it, which has no EOT marker. We used the paper's 1,000 MMLU-Pro items and the publicly released J-Lens. Without a marker, we read AconeA_{\mathrm{cone}} either at the last generated token or at the token just before the model writes its final answer.

Gemma-3-27B-it, MMLU-ProAUROCAUROC, length-matched
U-Lens, before the final answer74.471.7
U-Lens, last generated token73.570.8
Predictive Entropy71.070.0
DeepConf70.369.9
Self-Certainty73.469.3
Mean NLL70.068.7
Maximum Entropy73.468.4
TokUR72.267.5
SentenceSAR69.566.4
Semantic Entropy61.762.4
Feature-Gaps (supervised)67.361.9
P(True)56.754.5
MSP52.953.7

U-Lens leads every baseline at both read positions, before and after length control, so the readout does not depend on a dedicated EOT token.

Q4: Steering value selection. We did not select the doses systematically. The figures give qualitative evidence that adding a category direction changes the model's behavior in a way that matches the category. For each category, we tried a range of values by hand on the same prompt and show three at which the model still answers the question, so that the change can be read as induced uncertainty and not as a breakdown of generation. We now state this explicitly in Appendix J: Causal Intervention.

Q5: Semantic Entropy temperature. We chose 0.5 following Kuhn et al. (2023), who found this to work best. We have now swept it on Gemma over 0.25, 0.5, 0.75 and 1.0, on about 300 items each of TriviaQA and MMLU-Pro. Averaged over both, Semantic Entropy reaches 67.6 to 70.2 AUROC against 77.5 for U-Lens, and 66.9 to 71.9 against 72.3 after length matching, so the temperature does not change the comparison. The revised paper samples Semantic Entropy at each model's own temperature (1.0 for Gemma), its strongest setting after length matching. At these settings, averaged over all four benchmarks (Table 2), U-Lens still leads Semantic Entropy after length matching on every model, with 68.4 vs. 65.8 on Gemma, 71.0 vs. 68.6 on Qwen and 68.3 vs. 63.5 on Magistral.

Thank you again for your time and thoughtful feedback on our submission.

As the discussion period is approaching its end on November 18 (AoE), we wanted to kindly follow up to see if our previous responses and the revised manuscript have adequately addressed your concerns. If any points remain unclear or if you have any further questions, please let us know. We would be more than happy to clarify them promptly before the deadline.

Best regards, Authors

Comments on the paper

1 comment

We sincerely thank all reviewers for their thoughtful reviews and constructive questions!


We have uploaded a revised version of the paper in which everything we added in response to the reviews is marked in blue, so that it is easy to find. In the individual responses, we address the weaknesses you mentioned and the questions you raised and point to the corresponding changes.

If any point remains unaddressed or further questions arise, we are happy to provide more details and continue the discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.