acceptodds
Withdrawn from ICLR 2027

LLM Uncertainty Revealed in Behavioral Boundary Effects

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Large language models may produce fluent and confident answers even when those answers are unreliable. Existing black-box uncertainty estimators typically rely on repeated sampling from the same model or disagreement across an unordered collection of models. The former may fail when the model consistently reproduces the same error, while the latter may introduce unnecessary disagreement and inference cost due to differences in model architecture and training distribution. In this work, we investigate whether an ordered family of small models can serve as a more effective uncertainty probe, and propose **AnchorAUC**. Given an input, a target model first generates an anchor answer, after which a sequence of progressively smaller probe models independently answers the same input. Their semantic retention relative to the anchor forms a directional degradation trajectory along an ordered model-scale axis. AnchorAUC integrates this trajectory into a reliability score: answers that remain stable across a broader range of probe levels receive lower uncertainty, whereas answers that drift earlier exhibit a behavioral boundary associated with higher uncertainty. We evaluate AnchorAUC on five question-answering benchmarks, four target models, and two probe families. AnchorAUC achieves competitive AUROC and AUARC across diverse model–task configurations while requiring substantially lower inference cost than multi-anchor sampling. Further analyses of probe order, correctness transitions, and probe granularity show that the ordered trajectory contains information beyond generic cross-model disagreement.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Large Language Models, Uncertainty Quantification, Behavioral Boundary, Ordered Model Probing

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_ZE5TOpenReview ↗
Rating
1(−1)
Confidence
3
Informativeness
1(−1)
Presentation
2(−1)
Soundness
1(−1)
Read review
Paper summary

This paper asks how we can tell whether an LLM's answer is reliable without looking inside the model. The idea is intuitive: the target model answers first, then a few smaller models from the same family answer the same question, and we check how much they agree with the target model. The more the smaller models arrive at the same answer, the more reliable it is considered. The authors plot this agreement along model size and use the area under the curve as the score, which they call AnchorAUC. Experiments on five QA datasets, four target models and two sets of small models show that it outperforms several baselines at a much lower cost. But this work still miss some components and need to clarify something.

Critical strengths

The presentation of this paper is visually appealing, and the overall structure is complete. The main text is concise and easy to follow. The proposed method is simple and low-cost, the evaluation covers a broad range of settings, and the ablation studies are well designed and directly address the key claims.

Critical weaknesses
  1. While checking the tables, I noticed an unusual pattern. In Table 1, the SESDLG–SE difference is identical across all four anchor models (for example, 9.0 on TriviaQA), and in Table 2, TU's AUROC on the two closed-source anchors differs by exactly 0.5 on every dataset. I may be misunderstanding how the baselines were computed, and I would appreciate it if the authors could clarify how these numbers were obtained.

  2. As far as I can tell, the method can be viewed as a weighted average of three probes' agreement, and the gain over the plain mean is 2.3 AUROC points, so I have some reservations about the degree of novelty.

  3. I am a bit concerned that the score mostly reflects how hard a question is for small models. Cases where the anchor is correct but a probe is wrong account for 14%–21% of examples, and these correct answers would receive high uncertainty.

  4. A few case studies of failure cases would help readers understand when the method breaks down. Besides, the current tasks are all short-answer QA. I would be curious to see how the method performs on long-form generation or mathematical reasoning. Additionally, the correctness labels come mainly from an LLM judge. A small human-checked sample with agreement statistics would make the results more convincing.

Questions for authors

Hi, if you can add some experiments as mentioned in weakness and clarify some concerns, I'm willing to improve my score. Your presentation is very good.

5 comments
AuthorsOpenReview ↗

Responses to w1 and w2

We thank the reviewer for the positive feedback on the presentation of our work. We have further polished Figure 1 to enhance its visual appeal and make the overall presentation more intuitive and engaging. We address the reviewer's comments and questions point by point below.

1. Unusual Numerical Patterns in Tables 1 and 2

Thank you for pointing out these unusual numerical patterns. In an earlier version of our experiments, some results were still pending. We temporarily used placeholder values based on fixed differences for these entries. After the corresponding experiments were completed, we obtained the actual results, but inadvertently failed to update all the placeholder entries in the final LaTeX tables.

We have now replaced these placeholders with the actual experimental results, including the SE_SDLG results for Phi-4-14B, OLMo2-13B, and Mistral-Nemo-12B in Table 1, and the TU results for Claude Opus 5 and GPT-5.6 Sol in Table 2. All corresponding averages have also been recalculated. These discrepancies were confined to the reported tables and did not affect the metric calculations, or method implementations. We have cross-checked the corrected values against the original experimental records, and the main conclusions remain unchanged.

We apologize for this oversight and sincerely thank the reviewer for identifying it. The corrected tables are included in the revised manuscript.

2. On the Concern That AnchorAUC Is Essentially a Weighted Average

We do not deny the basic principle of weighted averaging itself, but it is worth emphasizing that, prior to our work, no study had organized models of different scales within the same model family into an ordered probe sequence, even using simple averaging. This organization is itself a new attempt, allowing us to systematically examine how model behavior changes with scale while controlling for differences across model families.

On this basis, we further recognize that simple averaging may overlook the differences in information carried by models of different scales, and find that incorporating model scale into the aggregation weights can further improve method performance. Furthermore, we systematically examine semantic retention, correctness changes, and behavioral transitions between neighboring model scales, revealing systematic associations between changes in model scale, answer semantics, and correctness. These findings not only provide empirical support for scale-aware aggregation strategies but also further demonstrate that models at different scales are not simply interchangeable independent voters; their relative scales and behavioral changes themselves contain valuable information about answer reliability.

Therefore, we believe the novelty of our work lies not merely in the weighted averaging itself, but more importantly in the scale-ordered probe organization and the systematic use of scale-related behavioral information to assess answer reliability.

AuthorsOpenReview ↗

Responses to w3

3. On the Concern That the Score Primarily Reflects Question Difficulty for Smaller Models

Thank you for raising this important concern. We would like to clarify that 14%–21% represents the average proportion of cases where the Anchor is correct but a Probe is incorrect, rather than the false-positive rate of the final score. An incorrect answer from an individual Probe does not necessarily lead to high uncertainty, as the final score depends on the consistency between multiple Probes and the Anchor,.

As you pointed out, our method does not completely separate question difficulty from answer reliability, since we believe question difficulty itself serves as a useful risk signal:

greater task difficulty is generally associated with a higher risk of incorrect answers.

Therefore, the key question is:

whether scale-related behavioral signals can still help distinguish correct and incorrect Anchor answers, even when smaller Probes have lower standalone accuracy.

To investigate this, we conducted additional experiments using Qwen3-14B as the Anchor and Qwen3-8B, Qwen3-4B, and Qwen3-1.7B as Probes on more challenging mathematical and reasoning tasks. The results show that our method can retain meaningful discriminative ability even when the smaller model has substantially lower accuracy than the Anchor:

Task TypeTask8B Accuracy (%)14B Accuracy (%)AUROC (%)
Mathematical ReasoningMATH-50089.1892.2790.97
Mathematical ReasoningBBH-Object Counting46.4056.4079.06
Logical ReasoningBBH-Five-Object Logical Reasoning58.8062.0073.71

For example, on BBH Object Counting, the 8B Probe achieves only 46.40% accuracy, compared with 56.40% for the 14B Anchor, yet our method achieves an AUROC of 79.06%. Similarly, on BBH Five-Object Logical Reasoning, the 8B Probe achieves 58.80% accuracy, while our method still achieves an AUROC of 73.71%. These results suggest that the usefulness of Probes does not simply depend on their standalone accuracy or their ability to answer every question correctly.

We nevertheless acknowledge that when a task substantially exceeds the capabilities of multiple Probes (e.g., when the accuracy of all probes approaches zero), their behavioral signals may become less informative. We have clarified this point in the revised manuscript.

AuthorsOpenReview ↗

Responses to w4

4.1 Failure Case Analyses

We identify two major types of failure cases.

Case 1: Shared model errors lead to low uncertainty for an incorrect answer.

Question: Alice and Bob each have a bag containing five balls, one in each of the following colors: blue, green, orange, red, and purple. Alice randomly transfers one ball to Bob's bag. Bob then randomly selects one ball from his bag and returns it to Alice. What is the probability that the two bags have identical compositions at the end?

Correct answer: 1/31/3.

Model responses:14B, 8B, and 4B: 1/61/6; 1.7B: 1/51/5

The anchor and the two stronger probes share the same incorrect answer. Consequently, the mathematical-equivalence matching variant assigns this incorrect anchor answer a low uncertainty score of U=0.0935U=0.0935.

Case 2: Errors introduced by the NLI evaluator itself.

Question: When was Episode 122 of Dragon Ball Super aired?

Correct answer: Japanese version: January 7, 2018. English version: August 3, 2019.

Model responses: 14B: "The main series has only 121 episodes, so there is no official release date for episode 122." 8B: "The original series ended with episode 122, which aired on July 1, 2015."

Neither response matches the reference answers, and the two responses contradict each other. However, the original full-text NLI evaluator assigns them a bidirectional average retention score as high as 0.9785, causing the incorrect anchor answer to fall into the low-uncertainty region. This case illustrates that the NLI evaluator itself can also be a source of errors in our framework.

4.2 Performance on Mathematical Reasoning and Long-Form Generation Tasks

Mathematical reasoning.

We added experiments on MATH-500 using Qwen3-14B as the anchor and Qwen3-8B, 4B, and 1.7B as probes.

Task8B Accuracy (%)14B Accuracy (%)AUROC (%)AUARC (%)
MATH-50089.1892.2790.9798.65

Long-form generation.

For long-form generation, we conducted closed-book generation experiments on 200 questions from ASQA. The primary bottleneck in this setting is the context-length limitation of the NLI evaluator. The nli-deberta-v3-base model used in our original experiments has a maximum input length of 512 tokens. We investigate two potential solutions:

  1. Sentence-level matching: Split the original answers into individual sentences, compute NLI scores at the sentence level, select the maximum matching score, and then average the resulting scores.
  2. Using a longer-context evaluator: Replace the original NLI evaluator with a longer-context discriminator, such as Kev-Qwen3-4B, a JEV-style open-source reproduction.
MethodAUROC (%)AUARC (%)
nli-deberta-v3-base, full-answer matching55.779.15
nli-deberta-v3-base, sentence-level matching61.8110.56
Kev-Qwen3-4B, full-answer matching78.2418.01

4.3 Human Validation of Correctness Labels

We randomly sampled 50 examples from each of the four datasets that use LLM-assisted correctness evaluation and manually verified their correctness labels. We then compared the human annotations with the automatically assigned labels on an example-by-example basis.

Among the 200 manually checked examples, 193 labels were consistent with the automatic evaluations, yielding an overall agreement rate of 96.5%. The results for individual datasets are reported below.

StrategyQA was excluded from this validation because its correctness labels are determined using deterministic rule-based evaluation rather than an LLM judge.

DatasetNumber of Verified ExamplesAgreement between Automatic and Human LabelsAgreement Rate
TriviaQA5050100%
Natural Questions504896%
HotpotQA504998%
CoQA504692%
Total20019396.5%
Reviewer_o9fUOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses black-box uncertainty estimation for LLMs, in particular cases where a model is stable yet incorrect. It proposes AnchorAUC, which runs an ordered sequence of progressively smaller probe models on the same input, scores each probe's answer against the target (anchor) answer with bidirectional NLI-based semantic retention, and takes the area under the resulting degradation curve on a size-normalized axis as the reliability score. Across five QA benchmarks, four anchor models and two probe families, the method matches or outperforms strong baselines, including cross-model disagreement, at substantially lower inference cost. Ablations (permutation, granularity, family matching, large anchors) indicate that the ordered trajectory carries information beyond generic model disagreement.

Critical strengths
  1. Ordered probes as a new signal. I believe turning unordered cross-model disagreement into a directional degradation trajectory is a simple and elegant idea, and the AUC summary is robust to probe spacing.
  2. Thorough ablations supporting the central claim. Permutation tests, granularity and position effects, and the boundary-alignment analyses provide, in my view, convincing evidence that ordering itself matters rather than mere disagreement.
  3. Favorable cost–accuracy trade-off. The concrete latency and parameter-equivalent analyses (e.g., a 39× latency reduction relative to TU) make the method practically attractive in black-box settings.
Critical weaknesses
  1. Dependence on an external NLI model. The retention score adds a further source of error and possible domain mismatch, and robustness to alternative NLI choices is not deeply examined.
  2. Evaluation limited to QA. It is unclear whether the ordering advantage holds for code generation, mathematical reasoning or long-form generation, where semantic comparison is harder.
  3. Heuristic notion of behavioral boundary. The boundary assumes capability tracks parameter count, which is not guaranteed across families or tasks, and the anchor placement for closed-source models (14B vs. 100B) feels somewhat ad hoc; clearer formalization and practitioner guidance might help.
  4. Evaluation and baseline gaps. Correctness for non-binary datasets relies on a single LLM judge, and kernel-based estimators such as KLE are not compared; human validation or multiple judges, and such a comparison, would strengthen the claims.
Questions for authors
  1. How sensitive are AUROC/AUARC and cost to the choice of NLI model (e.g., RoBERTa-large-MNLI, DeBERTa-large, or embedding-based entailment)?
  2. Were the sampling-based baselines re-sampled at tuned temperatures rather than temperature 0, and what were the per-method token budgets?
  3. For large anchors, would a small calibration set suffice to set the anchor's reference placement on the size axis, and how robust is performance to moderate misplacement?
  4. Could shared, correlated errors between anchor and probes become more frequent in specialized domains or with closer familial ties than the observed ≈1–2%?
2 comments
AuthorsOpenReview ↗

Response to o9fU (1/2)

We sincerely thank the reviewer for the thoughtful comments and constructive suggestions. Below, we address each concern and question in detail.

W1 & Q1. Dependence on an External NLI Model

We conducted a sensitivity analysis using different semantic comparators, with Qwen3-14B as the anchor and Qwen3-8B, Qwen3-4B, and Qwen3-1.7B as probes. The results are macro-averaged across five tasks.

As shown below, our method is robust to the choice of NLI model, while the additional computational overhead is negligible.

Semantic ComparatorAUROCAUARCCost (ms/sample) ↓
DeBERTa-v3-base78.3081.750.94
RoBERTa-large-MNLI78.6081.951.43
DeBERTa-large-MNLI78.6281.971.65
Embedding-based: MPNet-base-v276.8781.400.42

W2. Evaluation Limited to QA

We thank the reviewer for this valuable suggestion. We conducted additional experiments on mathematical reasoning and long-form generation to evaluate the applicability and limitations of our method beyond short-answer QA.

Mathematical reasoning. We added experiments on MATH-500 using Qwen3-14B as the anchor and Qwen3-8B, 4B, and 1.7B as probes.

Task8B Accuracy (%)14B Accuracy (%)AUROC (%)AUARC (%)
MATH-50089.1892.2790.9798.65

Long-Form Generation. We additionally conducted closed-book generation experiments on 200 questions from ASQA.

A key bottleneck for long-form generation is the context-length limitation of the NLI model. Specifically, the nli-deberta-v3-base model used in our paper supports a maximum input length of 512 tokens.

To address this limitation, we explored two approaches:

  1. Sentence-level matching: Split the original responses into sentences, compute sentence-level NLI scores, and average the maximum matching scores.
  2. Longer-context comparator: Replace the NLI model with a comparator supporting longer contexts, such as Kev-Qwen3-4B, an open-source reproduction of a Jev-style judge.

Both approaches improve performance on long-form generation, as shown below.

MethodAUROCAUARC
nli-deberta-v3-base, paragraph-level matching55.779.15
nli-deberta-v3-base, sentence-level matching61.8110.56
Kev-Qwen3-4B, paragraph-level matching78.2418.01

W3 & Q3. Heuristic Notion of Behavioral Boundary

We thank the reviewer for highlighting the need to clarify this concept. The behavioral boundary is defined based on the empirically observed retention sequence. Let the anchor retention be �1��1�, followed by probe retentions �2��2�. The boundary is defined as the first position satisfying:

$

bτ=min⁡{i:ri−1−ri≥τ},τ=0.2.b_\tau=\min\{i:r_{i-1}-r_i\geq\tau\},\qquad \tau=0.2.

$

If no such drop occurs, no boundary is detected in the sequence. Thus, the boundary represents an observable change in answer-retention behavior, rather than a precise boundary of the model's intrinsic capability.

For closed-source anchors with unknown parameter counts, 14B and 100B are predefined reference coordinates, not estimates of their actual parameter counts.

Beyond the two-endpoint comparison already included in the paper, we conducted an additional sensitivity analysis using GPT-5.6 Sol as the anchor across five reference positions.

Reference PositionAUROC (%)AUARC (%)
14B86.0096.20
20B85.9996.18
30B85.9896.16
50B86.1196.17
100B86.3096.20

These results indicate limited sensitivity to the choice of reference position under the evaluated settings.

AuthorsOpenReview ↗

Response to o9fU (2/2)

W4. Evaluation and Baseline Gaps

We included KLE as a baseline under the Qwen3-14B anchor setting. As shown below, AnchorAUC continues to outperform KLE across the evaluated benchmarks.

MethodTriviaQAStrategyQANQHotpotQACoQAAverage
KLE80.8 / 83.579.2 / 96.772.8 / 49.374.7 / 82.870.2 / 77.775.5 / 78.0
Probe-Qwen84.6 / 89.688.4 / 98.673.3 / 49.880.6 / 96.774.7 / 90.980.3 / 85.1
Probe-Llama93.1 / 92.579.8 / 97.279.6 / 52.776.3 / 94.981.6 / 91.582.1 / 85.8

We further conducted a manual verification of the automatically assigned correctness labels. Specifically, we randomly sampled 50 examples from each of the four datasets that use LLM-assisted correctness evaluation and compared the automatic labels against human annotations on an instance-by-instance basis.

Among the 200 verified examples, 193 labels agreed, yielding an overall agreement rate of 96.5%. The results for each dataset are presented below. StrategyQA was excluded from this verification because its correctness labels are determined using rule-based evaluation.

DatasetVerified SamplesAgreements Between Automatic and Human LabelsAgreement Rate
TriviaQA5050100%
Natural Questions504896%
HotpotQA504998%
CoQA504692%
Total20019396.5%

Q2. Were the Sampling-Based Baselines Re-sampled at Tuned Temperatures Rather Than Temperature 0, and What Were the Per-Method Token Budgets?

Yes. As specified in Appendix B of our paper, SAR, SE, SE_SDLG, and Self-Consistency use a sampling temperature of 1.0, with five additional generations per question.

The token budgets for each method are summarized below.

MethodAdditional Generations per QuestionTokens / Sample
AnchorAUC (Qwen 8B+4B+1.7B)31,581.0
AnchorAUC (Llama 8B+3B+1B)31,598.5
SAR / SE / SE_SDLG / Self-Consistency52,738.1
TU8 (2 per model across 4 models)4,380.9

Q4. Could Shared, Correlated Errors Between Anchor and Probes Become More Frequent in Specialized Domains or With Closer Familial Ties Than the Observed ≈1–2%?

We used Qwen3-14B as the Anchor and Qwen3-8B/4B/1.7B as Probes, and evaluated 500 sampled questions from MedQA (medical domain) and FinQA (financial domain). The results show that the average high-retention shared error rate remains comparable to the approximately 1–2% observed in common QA tasks.

TaskAll Models IncorrectHigh-Retention Shared Error Rate
MedQA26.0%0.4%
FinQA15.2%0.0%
Reviewer_ALf7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces a new way to estimate uncertainty for LLMs. Instead of using models of the same size but various families, the authors investigate whether the output of a suite of models of decreasing size degenerates and how quickly. This serves as support for the anchor's answer and a fast decay is associated with higher uncertainty. The authors show competitive performance.

Critical strengths

The idea seems novel and effective. The computational overhead is also lower compared to many other uncertainty estimates for LLMs. The authors run a couple of interesting experiments that provide more insights

Critical weaknesses

The method would not work when there is a task that are beyond the 8B capability (eg harder math, reasoning, ...). Then only the anchor might be the one solving the task and all others would fail. The introduction is therefore over-claiming a bit in my opinion. Either testing this (which is unlikely to work) or rephrasing this method toward QA / knowledge tasks would be important.

Questions for authors
  1. From the experiments it seems like a probe family is better than the other based on the task irrespective of the anchor model. Do I see this correctly? The fact that the anchor model does not have to be of the same model family seems to slightly go against the initial motivation of the paper. I think this should be adapted or investigated more.
  2. The ablation on how many probes to use, could be dependent on the task, right?
  3. Why do you not use the 2 probe setup when it works better on Llama?
  4. Did you test even more (e.g. on the Qwen family)?
6 comments
AuthorsOpenReview ↗

Response to Critical Weakness

We thank the reviewer for recognizing the novelty, efficiency, and diagnostic value of our approach.

Regarding the concern that our method may fail on tasks beyond the capabilities of the 8B probe, we agree that probe capability is an important factor determining the effectiveness of our approach. To investigate this concern, we conducted additional experiments using Qwen3-14B as the anchor and Qwen3-8B, Qwen3-4B, and Qwen3-1.7B as probes on more challenging mathematical and reasoning tasks, including MATH-500 and a subset of BBH (Big-Bench-Hard).

Task TypeTaskAll-Probe Error Rate (%)8B Accuracy (%)14B Accuracy (%)AUROC (%)
Mathematical ReasoningMATH-5007.2889.1892.2790.97
Mathematical ReasoningBBH-Object Counting40.8046.4056.4079.06
Logical ReasoningBBH-Five-Object Logical Reasoning21.2058.8062.0073.71

Either testing this (which is unlikely to work) or rephrasing this method toward QA / knowledge tasks would be important.

The results show that the applicability of our method is not restricted to knowledge-based QA tasks. It can also provide effective reliability signals on mathematical and logical reasoning tasks, provided that the selected Probes retain sufficient capability to produce informative behavioral signals.

The method would not work when there is a task that are beyond the 8B capability (eg harder math, reasoning, ...). Then only the anchor might be the one solving the task and all others would fail.

On BBH Object Counting, all three Probes give incorrect answers on 40.80% of samples, while the 8B Probe achieves only 46.40% accuracy, yet our method still achieves an AUROC of 79.06%. Similarly, on BBH Five-Object Logical Reasoning, 21.20% of samples are answered incorrectly by all three Probes, and our method achieves an AUROC of 73.71%, despite the 8B Probe achieving only 58.80% accuracy.

These results suggest that the usefulness of Probes does not simply depend on their standalone accuracy or their ability to answer every question correctly. Even when all Probes fail on a substantial proportion of individual samples, their collective behavioral signals can still provide useful discriminative information at the dataset level.

Nevertheless, we acknowledge that our method has a capability boundary. When a task substantially exceeds the capabilities of multiple Probes (e.g., when the accuracy of all probes approaches zero), their behavioral signals may become insufficiently informative, potentially leading to degraded discriminative performance. We have revised the Introduction and the discussion of limitations to clarify this capability-dependent scope and avoid overclaiming the general applicability of our method.

AuthorsOpenReview ↗

Response to Questions

Q1: The fact that the anchor model does not have to be of the same model family seems to slightly go against the initial motivation of the paper.

In terms of AUROC, Table 1 does show a consistent task-dependent advantage across different anchors. However, the magnitude of these advantages varies across anchors. For example, the AUROC advantage of Probe-Qwen on HotpotQA ranges from 4.3 to 13.0 percentage points across the four anchors. Furthermore, this directional consistency does not hold for all AUARC results. On CoQA, for instance, each probe family performs better under two of the four anchors. Therefore, the current results are insufficient to conclude that probe-family performance is entirely independent of the anchor model.

We agree that the original manuscript should more clearly distinguish between family consistency within the probe chain and family consistency between the anchor and the probes. The former is intended to reduce confounding variations introduced by differences in architecture, training procedures, and other family-specific characteristics within the probe sequence. The latter, however, is not a necessary condition for AnchorAUC. Cross-family differences may reduce shared errors between the anchor and probes, but they may also introduce additional disagreement with a correct anchor answer. We have revised the relevant discussion in the Introduction to clarify this distinction and better align the motivation with our experimental findings.

Q2: The ablation on how many probes to use could be dependent on the task, right?

Yes, we agree that the optimal number of probes may depend on the task. Our experimental results suggest different configurations for different types of tasks:

  • Open-domain factual QA: Two probes can serve as a cost-effective starting point, with the third probe added depending on validation results.

  • Context-grounded QA: The three-probe Probe-Qwen configuration currently provides the best performance among the tested configurations.

  • Difficult mathematical and compositional reasoning tasks: Increasing the capability ceiling of the probe family should take priority over simply increasing the number of probes.

Q3. Why do you not use the 2-probe setup when it works better on Llama?

The main experiments consistently use three probes primarily to maintain a unified analytical setting across different model configurations and to provide more observation points along the model-scale axis. This allows us to investigate the structure of the semantic-retention trajectory and its relationship with correctness transitions more systematically.

However, this analytical convenience does not imply that the three-probe configuration is always optimal in terms of uncertainty estimation performance. In practical deployment, the number of probes can and should be selected according to the specific task and the desired cost-performance trade-off. As discussed in Section 5.7, two-probe configurations already capture most of the benefits of a longer probe sequence and can provide a more efficient alternative in certain settings.

Q4. Did you test even more (e.g., on the Qwen family)?

We conducted additional experiments within the Qwen model family by introducing a stronger 14B probe and using a 32B model as the anchor, resulting in an expanded probe sequence of 14B, 8B, 4B, and 1.7B.

On the five QA benchmarks, this expanded configuration improves macro-averaged AUROC and AUARC by 3.46 and 1.47 percentage points, respectively.

TaskAnchor: 14B, Probes: 8B+4B+1.7B (AUROC/AUARC)Anchor: 32B, Probes: 14B+8B+4B+1.7B (AUROC/AUARC)
TriviaQA84.63 / 89.6289.66 / 91.73
NQ73.33 / 49.7577.96 / 52.63
HotpotQA80.55 / 96.7286.54 / 98.37
CoQA74.74 / 90.9076.59 / 92.33
StrategyQA88.42 / 98.5688.21 / 97.83
Macro average80.33 / 85.1183.79 / 86.58

On BBH Object Counting, all three Probes give incorrect answers on 40.80% of samples, while the 8B Probe achieves only 46.40% accuracy, yet our method still achieves an AUROC of 79.06%. Similarly, on BBH Five-Object Logical Reasoning, 21.20% of samples are answered incorrectly by all three Probes, and our method achieves an AUROC of 73.71%, despite the 8B Probe achieving only 58.80% accuracy.

Can you tell me more about why it is that way? E.g. by binning the answers based on capability or showcasing a few examples that explain it.

Nevertheless, we acknowledge that our method has a capability boundary. When a task substantially exceeds the capabilities of multiple Probes (e.g., when the accuracy of all probes approaches zero), their behavioral signals may become insufficiently informative, potentially leading to degraded discriminative performance. We have revised the Introduction and the discussion of limitations to clarify this capability-dependent scope and avoid overclaiming the general applicability of our method.

I could see some follow-up where the smaller probes process parts of the traces from the anchor as additional information. I could imagine that they are still helpful with the additional information somehow. Just as a side note. Not necessary for this paper.

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

Related papers

Discussion (0)

Sign in to comment.