acceptodds
Under review as a conference paper at ICLR 2027

Jailbreaks for Black-Box Uncertainty Quantification in Large Reasoning Models

Abstract

While Large Reasoning Models (LRMs) excel at complex reasoning, alignment through reinforcement learning often induces systemic overconfidence. In production environments, where logits may be unavailable, robust black-box uncertainty quantification (UQ) is essential for trustworthiness and safety. Focusing on question-answering for LRMs, we show that existing black-box methods, such as paraphrase-based self-consistency and confidence verbalization, offer little to no improvement over simple repeated sampling, suggesting that alignment suppresses useful output variability. We introduce prompt-level relaxation operators that broaden the model's effective output distribution by approximating the effect of an optimal policy obtained with a stronger KL-regularization parameter, hence closer to the reference model. Theoretically, we demonstrate that relaxation improves calibration. We propose Jailbreak for Uncertainty (J4U), a jailbreak-derived technique for UQ that empirically reproduces the behavioral signatures predicted by our relaxation theory. Across 3 datasets and 4 LRMs, including a closed-source production model, J4U's improvement over repeated sampling achieves statistical significance in up to 6 more LRM–dataset-metric settings than the strongest black-box UQ state-of-the-art baseline we evaluate, with average ECE reductions up to 5 larger. These results provide a practical tool for UQ in black-box LRM deployment.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Uncertainty Quantification, Confidence, Jailbreak, Black-box, Large Reasoning Models

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_6iZHOpenReview ↗
Rating
1
Confidence
3(+1)
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes Jailbreak for Uncertainty (J4U), which applies jailbreak-inspired prompt transformations to benign questions and estimates confidence from the frequency of the majority answer across multiple samples. The transformations include random suffixes, ASCII-art substitution, and text decomposition and reconstruction. The paper interprets these transformations as relaxing RL-induced constraints, motivated by a KL-regularized policy model. Experiments on four models and three main datasets report improvements in calibration metrics over repeated sampling and two prompting baselines. Additional analyses examine answer entropy, reasoning-trace similarity, sampling budget, and performance on GSM8K.

Critical strengths

The method has a simple deployment interface (a prompt-engineering method). It requires only textual queries and responses, without model modification or access to logits.

The empirical calibration results merit investigation. The paper reports improvements beyond ECE, including Brier score and NLL, across several model–dataset combinations.

Critical weaknesses

I am confused when reading the paper perhapse I didn't understand the full picture.

The abstract moves from

"Focusing on question-answering for LRMs, we show that existing black-box methods, such as paraphrase-based self-consistency and confidence verbalization, offer little to no improvement over simple repeated sampling"

to

", suggesting that alignment suppresses useful output variability."

I find it hard to understand the logic here. How was this conclusion reached? The paper didn't perform SFT or RL, nor did the authors compare a base model against a chat model. The actual comparison was against various perturbation methods, yet the conclusion shifted to alignment affecting model capabilities. I’m not really sold on this explanation. The paper mentions "RL" 95 times, yet the actual experiment doesnt seems to be related to RL. How do the authors prove that "jailbreak-style" prompts don't introduce confounding effects related to RL alignment? Can the prompts that authors used actually trigger a jailbreak in the first place? I don't see any experiments showing whether the base models exhibited the same response to the perturbations they tested. Sec 3 defines Pi_ref as the base model, and Pi_RL as the RL model but all the experiments are done only on Pi_RL, Pi_ref only exists in the writing.

The experiments are done on small language models (the largest one here may be luna). The method is basically a string level method (requires no adapter or training), and I believe its easy to scale.

The writing is poor. Sec 3 is quite long but the content seems to be basic definitions (e.g. ECE) . Also, the paper lacks a diagram illustrating the methodology; Figures in the appendix are with tiny font sizes

Questions for authors

The main issue is that I don't see what purpose J4U serves beyond the explanation provided in the paper. Looking at the metrics, the accuracy remains roughly the same under the prompt-based setup, though the ECE drops. If the main conclusion is that RL reduces output diversity, then the experiments and the conclusion don't align.

5 comments

Thank you for your detailed review and for recognizing the practical interface and empirical calibration results. J4U addresses a limitation of repeated-sampling confidence: a model can repeatedly give the same incorrect answer, so strong agreement need not imply correctness. J4U applies jailbreak-inspired transformations to benign questions and uses the frequency of the majority answer as its confidence score. The experiments test whether this score better reflects the probability that an answer is correct while maintaining comparable accuracy. A change in the majority answer is not required for this improvement.

We revised the abstract and Section 5.4 to clarify this distinction, added practical motivation and Figure 1 to illustrate the mehtodology and its objectives, and expanded the comparisons with REPHRASE, temperature-adjusted repeated sampling, and a logit-based baseline. All additions and changes are visible in blue in the revised paper. Please find below our responses to your concerns.

Critical Weaknesses:

1. The abstract suggests that alignment suppresses useful variability. We revised the abstract to distinguish the hypothesis motivating J4U from the experimental findings. RL-induced sharpening is motivated by prior work and a hypothesis that drives the design of J4U; our empirical finding is that J4U improves confidence calibration on the evaluated LRMs. The scope of the theoretical explanation is detailed in point 3.

2. What comparisons with perturbation methods establish. Comparing against existing UQ methods, including prompt perturbations, directly tests J4U’s practical contribution. These comparisons establish the relative effectiveness of the UQ methods; they do not identify the effect of alignment. We strengthened this evaluation with a paired comparison against REPHRASE (Appendix D.5, Table 15). Averaged over the three datasets, J4U-ART and J4U-PROG improve ECE, Brier score, and NLL relative to REPHRASE on each of the four models. The improvements are statistically significant in all 12 model–metric comparisons for ART and 10 of 12 for PROG.

We also added a simple logit-based baseline on Qwen3-4B and gpt-oss-20b (Appendix D.4). J4U achieves lower ECE and Brier score on every evaluated model–dataset pair, without requiring access to token probabilities. These comparisons strengthen the empirical contribution independently of the proposed explanation involving alignment.

3. RL appears throughout the paper, but the experiments do not involve RL training. RL defines the policy model analyzed in Section 4. Under its assumptions, Proposition 4.2 establishes movement toward the reference policy, and Theorem 4.3 shows how calibration can improve while the most probable answer remains unchanged: a moderate relaxation reduces excessive confidence, with strict calibration improvement guaranteed in the large-sample limit. This result motivates the search for prompt transformations with this behavior. These guarantees apply to operators satisfying Definition 4.1. We have not established that J4U satisfies this definition. The experiments directly evaluate its calibration performance at a finite sampling budget on released models (i.e. πRL\pi_{RL}). They support the practical benefit of J4U, while identifying relaxation as its mechanism requires further evidence. We clarified this distinction in the abstract and Section 5.4.

4. Could the improvements arise from effects unrelated to alignment? The new experiments test a concrete alternative explanation: whether simply increasing answer diversity reproduces J4U’s performance. We evaluated temperature-adjusted RS on Qwen3-4B and two datasets, with 100 inputs per dataset and ten samples per input (Appendices D.6–D.7).

On SuperGPQA, J4U-PROG and RS at T=2.7T=2.7 have similar mean answer entropies, 0.85 and 0.82. PROG achieves lower ECE, 0.22 versus 0.29, and higher accuracy, 0.46 versus 0.35. For ART, the closest RS run matches its ECE of 0.29.

On SynthPAI, the closest RS run reaches entropy 0.35 versus 0.36 for ART and PROG. Its ECE is 0.30, compared with 0.23 and 0.24, respectively. Only 34.3% of its outputs contain a valid label, versus 100% for J4U. Generation limits were 16,384 tokens for RS and 4,096 for J4U.

Thus, similar mean answer entropy does not necessarily produce similar calibration or accuracy. Increasing temperature does not consistently reproduce the performance of J4U-ART and J4U-PROG in these experiments.

5. Can these prompts actually trigger a jailbreak? ART and PROG are adapted from published jailbreak attacks (Jiang et al., 2024; Kang et al., 2024), while SUFFIX uses a simplified random suffix. “Jailbreak-inspired” describes the origin of these transformations. Here, the transformations are applied to benign questions and evaluated for confidence estimation. Their usefulness for UQ is measured directly by the calibration results, rather than inferred from attack success.

6. The reference policy is not evaluated, and base models might respond similarly. We have not compared the released models with their corresponding pre-alignment reference policies, so we cannot determine whether the base models would respond similarly. Section 5.4 now explicitly states that answer entropy and trace similarity characterize J4U’s effects but do not establish movement toward πref\pi_{ref}.

7. The experiments use small language models. The evaluation includes three open-weight models from 4B to 32B parameters and GPT-5.6 Luna, accessed through a production API. The evidence therefore extends beyond the small open-weight models. On GPT-5.6 Luna and SynthPAI, ART reduces ECE from 0.32 to 0.26 and NLL from 9.86 to 7.45, while accuracy remains 0.61 (Table 7). J4U requires no model modification in either setting and is applicable to any model. These results demonstrate applicability across the evaluated models, although they do not isolate the effect of model size.

8. Section 3 is long, the method lacks a diagram, and appendix figures are difficult to read. We shortened Section 3 to approximately one page by defining the theoretical mode, accuracy indicator, and ECE inline and removing the separate definition of perfect calibration. Figure 1 now illustrates the complete procedure: prompt transformation, repeated sampling, majority prediction, and confidence estimation. We also enlarged the axis labels, tick labels, and legends in Figures 2–4. The new introduction paragraph explains why confidence calibration matters in diagnostic support and text anonymization.

Questions for Authors:

1. What purpose does J4U serve when accuracy stays roughly the same but ECE decreases? The objective of J4U is to improve calibration without degrading accuracy. Improving accuracy is not required for an uncertainty estimator to be useful. Accuracy measures how often the model is correct; calibration measures whether its confidence agrees with its probability of correctness. Similar accuracy can therefore coexist with substantially more reliable confidence scores.

For example, on Qwen3-4B and SynthPAI, RS achieves 56% accuracy and ECE 0.34. With the same 10 samples per question, J4U-PROG achieves 58% accuracy and ECE 0.23; Brier score and NLL also improve (Table 4). A user needs to know both what answer the model gives and how much confidence to place in it. Similar accuracy does not make these confidence estimates equally useful.

2. If the main conclusion is that RL reduces output diversity, the experiments and conclusion do not align. The main empirical conclusion is that J4U improves black-box confidence estimation on the evaluated LRMs. Across the twelve model–dataset pairs, PROG and ART reduce ECE by 15.8% and 13.8% on average relative to RS, compared with 3.1% for REPHRASE (Table 1). The revised abstract makes this practical contribution explicit: better-calibrated confidence estimates obtained through text queries and responses alone.

Despite the lengthy response, posted on the final day of rebuttal, my main concern remains unresolved. The paper’s theoretical explanation centers on RL alignment, and I had expected experiments directly testing this explanation. The authors now explicitly acknowledge that they have not established that J4U satisfies their definition of a relaxation operator. Presenting this connection as a hypothesis leaves the link between the theory and the actual method unsupported. Similarly, citing jailbreak attacks from 2024 does not demonstrate that these adapted prompts successfully jailbreak the models evaluated in late 2026, given how quickly attacks and defenses evolve. Calling them “jailbreak-inspired” explains their origin, but supplies no experimental evidence that they bypass alignment on these models.

In light of these, I will maintain my rating and increase my confidence

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

This paper investigates black-box uncertainty quantification for large reasoning models (LRMs), motivated by the observation that reinforcement learning-based alignment can induce overconfidence and suppress useful output variability. The authors introduce the concept of relaxation operators, which theoretically improve confidence calibration by moving the model's output distribution closer to its pre-alignment policy. Building on this idea, they propose Jailbreak for Uncertainty (J4U), a method that applies three jailbreak-inspired prompt transformations to benign questions to increase response diversity and improve confidence estimation through repeated sampling. The method is evaluated across four LRMs and three datasets, against repeated sampling, confidence verbalization, and rephrasing-based approaches. The strongest J4U variants achieve up to 15.8% average reduction in expected calibration error compared with repeated sampling.

Critical strengths

The paper shows strong empirical improvements in black-box uncertainty quantification. J4U achieves substantial calibration improvements across four LRMs and three datasets without requiring model internals or additional training. It reduces average ECE by up to 15.8%, compared with 3.1% for the strongest prior baseline, and achieves statistically significant improvements in up to 18 of 36 evaluation settings without significant accuracy degradation. These results demonstrate the practical value of J4U, and this is particularly important for proprietary models with limited or no access.

Critical weaknesses

The paper's theoretical guarantees rely on relaxation operators that are equivalent to increasing the KL-regularization parameter of an RL-optimal policy. However, the authors do not establish that the proposed J4U transformations satisfy this condition. Instead, they rely on increased output entropy and reasoning-trace diversity as indirect evidence, neither of which necessarily implies that the resulting policy is closer to the pre-alignment reference. Furthermore, J4U independently samples randomized prompt transformations, inducing a mixture of output distributions that need not correspond to any single relaxation parameter. Theorem 4.3 also guarantees strict calibration improvement only asymptotically, while providing no strict improvement guarantee for the finite sampling setting used in the experiments. This disconnect is consequential because the relaxation framework is presented as a central technical contribution, yet its applicability to J4U remains unverified. Although the experiments demonstrate meaningful calibration improvements, they do not establish the proposed theoretical mechanism or the conditions under which it generalizes.

Questions for authors

How does J4U compare against a simpler baseline that increases output diversity, such as temperature-adjusted repeated sampling with matched output entropy?

1 comment

Thank you for recognizing the practical value of the method. Following your comment, we added a control with temperature-adjusted repeated sampling (App. D.6). All additions and editions are visible in blue in the revised paper. Please find below our answers to your concerns.

Critical weaknesses:

1. Scope of theoretical guarantee. Proposition 4.2 shows that a relaxation operator moves the policy closer to πref\pi_{ref}. Theorem 4.3 gives sufficient conditions under which a moderate relaxation improves calibration. For J4U, we observe experimentally that J4U improves calibration. Whether the transformations achieve these gains by moving the policy toward πref\pi_{ref} remains open. We revised the abstract and Section 5.4 to clarify what the entropy and trace-diversity measurements establish.

2a. Random transformation and mixture of output distributions. We clarified in Section 5.1 that each J4U variant independently samples a prompt transformation at each call. The resulting mixture need not correspond to a single relaxation parameter, as Definition 4.1 requires. Even in this case, the confidence score has a precise interpretation: as KK grows, it converges to the largest answer probability under the mixture.

2b. Finite sampling and scope of the guarantee. Under the assumptions of Theorem 4.3, a moderate relaxation preserves the most probable answer and strictly improves calibration in the large-sample limit. The finite-sample bound does not guarantee a strict ECE reduction at K=10K=10. At this sampling budget, J4U-PROG yields lower ECE than repeated sampling in all 12 model–dataset pairs, and J4U-ART in 11 of 12. The reductions are statistically significant in 6 pairs for J4U-PROG and 7 for J4U-ART (Tables 4–7).

Questions For Authors: Temperature-adjusted repeated sampling. We ran the control you suggested (new Appendix D.6). With Qwen3-4B on SuperGPQA and SyntPAI (100 questions, 10 sampled answers per question), we run repeated sampling at temperatures from 0.7 to 10 and compare J4U with the runs whose mean answer entropy is closest. J4U-ART and J4U-PROG are better than entropy-matched repeated sampling in all 4 settings in terms of accuracy, ECE, Brier score and NLL, with a single tie (ECE of J4U-ART on SuperGPQA, 0.29 for both). The gain of J4U is therefore not explained by increase of output diversity alone. We recall that tuning the temperature lies outside our fully black-box setting and is e.g. impossible on GPT Luna.

Finally, REPHRASE can also be used as a comparison because paraphrasing can increase answer diversity without using a jailbreak-inspired transformation. On gpt-oss-20b and HLE, and considering Tables 5 and 9: repeated sampling yields a mean answer entropy of 0.30 and an ECE of 0.72; REPHRASE raises entropy to 0.48, while ECE remains at 0.72; J4U-ART raises entropy only to 0.32, yet lowers ECE to 0.58. In this setting, REPHRASE produces the largest increase in answer entropy and no observable gain in calibration.

Reviewer_o4WfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

To address the systemic overconfidence of Large Reasoning Models (LRMs), this paper proposes a novel Jailbreak for Uncertainty (J4U) for UQ that reproduces the behavioral signatures predicted by using the relaxation theory. The authors evaluate their methods on four LRMs and three datasets. The results show that the methods improve higher performance than repeated sampling and two black-box UQ methods.

Critical strengths

1.The paper explores an interesting idea by using jailbreak prompts on uncertainty quantification. The work finds an interesting way to improve calibration in LRMs. 2.The paper has good writing and hardly has grammar errors. The organization of motivation, related works, methods, and experiments is clear. 3. The proposed method is simple, practical, and effective. And it does not require additional training, making it suitable for black-box models.

Critical weaknesses

1.The comparison on different datasets and selected comparison methods are insufficient. To be specific, the paper compares four LRMs model (three for open-weight models and one for closed-source model). And the selected comparison methods are limited, containing repeated sampling, verbalized confidence, and rephrase. So it is better to extand experiments can not show proposed methods effective. 2.The paper needs more theoretical analysis. Although authors introduce relaxtion operators to demonstrate J4U improves confidence calibration, the experiments can not show J4U make the data distribution of model output closer to reference policy. 3.The novelty should further to be demonstrated. The paper doesn't clarify the difference between existing prompt perturbation and their methods. Although the paper is the first to use jailbreaks, they need add more evaluation and comparison methods to show the fairness and validity. 4.The authors should add experiments to show the improvements if come from jailbreak transformations or simply from increased output diversity. 5. The overall writing is good, just some error. It is better to correct existing some grammar errors. For example, "Let x a relaxation operator" lacks "be" in page 4. 6. The proposed three different prompt transformations should further be justified. Specially, J4U-Suffix shows limited improvements compared with J4U-Art. The authors are better to explain why these transformations are selected and why performance differences between them.

Questions for authors

1.Can you provide stronger illustrations or experimental analysis to clarify J4U output closer to reference policy? 2.Can you add more comprehensive experiments of other related comparison methods on different datasets? 3.What are the important differences between J4U and prompt perturbation methods? It is better to provide figures of motivation and framework.

2 comments

Thank you for recognizing the practical value of the method. Following your comment, we added: a new figure to illustrate the method (Fig. 1), an experiment exploring the impact of temperature related increases of output diversity (App D.7) and three new controls test whether the gains come from generic perturbation (a direct comparison with REPHRASE, App. D.5; temperature-adjusted repeated sampling, App. D.6; a grey-box logit baseline, App. D.4). All additions and editions are visible in blue in the revised paper. Please find below our answers to your concerns.

Critical weaknesses:

1. Datasets and baselines. We added temperature-adjusted repeated sampling (RS-adj-temp) using Qwen3-4B on SuperGPQA and SyntPAI (new Appendix D.6) to test if increasing the temperature of the model can reproduce J4U-ART calibration gains at comparable entropy. We report the results under point 4. In addition to purely black-box baselines, we added a simple grey-box baseline based on logits in Appendix D.4 for gpt-oss-20b and Qwen3-4B (GPT-5.6 Luna does not expose logits, and we could not store them for DeepSeek-R1-32B because of memory constraint). Overall, J4U matches the accuracy of this logit-based baseline and is better calibrated in ECE and Brier score on every LRM-dataset pairs, without necessitating access to token probabilities. In term of datasets, in addition to the three main datasets, Appendix B.2 tests the "open-ended" subpart of the HLE dataset. Appendix D.3 examines the GSM8K dataset, where repeated sampling is already well calibrated.

2. Theoretical analysis. Proposition 4.2 shows that a relaxation operator moves the policy closer to πref\pi_{ref} and Theorem 4.3 gives sufficient conditions for improved calibration. Our experiments establish that J4U-ART and J4U-PROG improve calibration on average. They leave open whether either method achieves this gain by moving the policy closer to πref\pi_{ref}. We revised the abstract and Section 5.4 second paragraph to make clear what the entropy and trace-diversity measurements establish.

3. Novelty and comparison versus prompt perturbations. J4U repurposes three families of jailbreak-inspired transformations for confidence estimation on benign questions. REPHRASE varies the wording of the input. J4U instead appends a random suffix (J4U-SUFFIX), asks the model to reconstruct the input from fragments (J4U-PROG), or represents a selected word in ASCII art (J4U-ART). The new Fig. 1 illustrates J4U-ART. We provide in a new appendix (D.5) a direct comparison between all three J4U variants and REPHRASE. Averaged over all dataset, both J4U-PROG and J4U-ART achieved better calibration than REPHRASE on every LRM and for all three calibration metrics. The improvement is statistically significant in all 12 cases for J4U-Art and in 10 out of 12 for J4U-Prog. Accuracy is overall comparable or higher, with a single significant decrease. We also added temperature-adjusted repeated sampling to test whether a simpler way of increasing answer diversity can reproduce the calibration gain (see point 4).

4. Add experiments to explore increased output diversity. We added a temperature-adjusted repeated-sampling baseline to test whether increasing answer diversity can reproduce J4U’s calibration gain (new Appendix D.6). With Qwen3-4B on SuperGPQA and SyntPAI (100 questions, 10 sampled answers per question), we ran repeated sampling at temperatures from 0.7 to 10 and compared J4U with the runs whose mean answer entropy is closest. J4U-Art and J4U-Prog are better than entropy-matched repeated sampling in all 4 settings in terms of accuracy, ECE, Brier score and NLL, with a single tie (ECE of J4U-ART on SuperGPQA, 0.29 for both). The gain of J4U is therefore not explained by diversity alone. We recall that tuning the temperature lies outside our fully black-box setting and is e.g. impossible on GPT Luna. We additionally analyse the effect of temperature on RS (new Appendix D.7). On SuperGPQA, raising the temperature first improves calibration moderately, up to T = 2, then degrades it as valid answers and accuracy drop; on SynthPAI, it has no clear effect on calibration. At high temperatures, most responses no longer contain a valid answer. Even with the best temperature chosen a posteriori, RS does not reach the ECE of J4U on SynthPAI, nor that of J4U-PROG on SuperGPQA.

Finally, REPHRASE can be used as a comparison because paraphrasing can increase answer diversity without using a jailbreak-inspired transformation. On gpt-oss-20b and HLE, and considering Tables 5 and 9: repeated sampling yields a mean answer entropy of 0.30 and an ECE of 0.72; REPHRASE raises entropy to 0.48, while ECE remains at 0.72; J4U-ART raises entropy only to 0.32, yet lowers ECE to 0.58. In this setting, REPHRASE produces the largest increase in answer entropy and no observable gain in calibration.

5. Grammar. Thank you for this positive assessment. We corrected this mistake and thoroughly proofread the paper.

6. Prompt transformations. We selected one representative per major class of one-shot black-box jailbreaks in the taxonomies of Shen et al. (2025) and Li et al. (2024): suffix-based, semantic obfuscation, and template-based restructuring. The aim was to test whether the effect holds across attack families rather than for a single pattern. J4U-SUFFIX is a deliberately simplified version of optimized suffix attacks (Zou et al., 2023): it uses random rather than gradient-optimized characters, since gradients are unavailable in the black-box setting, and it alters the prompt only marginally. Consistently, it produces the smallest increase in entropy (+11.8% vs. +64.6% and +47.7%, Table 2), the smallest increase in trace diversity (Table 3), and the smallest calibration gain (Appendix B.1).

Questions For Authors:

Q1. Can you provide stronger illustrations or experimental analysis to clarify J4U output closer to reference policy? Proposition 4.2 establishes a decrease in KL divergence to the reference policy for relaxation operators. For J4U, we observe improved calibration, but answer entropy and trace diversity do not measure distance to the pre-alignment reference. Whether J4U moves the deployed policy closer to that reference therefore remains open.

Q2. Can you add more comprehensive experiments of other related comparison methods on different datasets? We added two baselines. First, we evaluated temperature-adjusted repeated sampling on Qwen3-4B and SuperGPQA to test whether it reproduces J4U’s calibration gain at comparable answer entropy (see point w4 above). Second, we added a simple grey-box logit-based baseline (see point 1 above) on all 4 LRMs and the 3 datasets used in main experiments. Results show that J4U-Prog and J4U-Art consistently improve calibration w.r.t. this new baseline.

Q3.What are the important differences between J4U and prompt perturbation methods? It is better to provide figures of motivation and framework. We added Fig. 1. It shows how J4U-ART randomly selects a word from the input and presents it in ASCII art for each model call. In the illustrated example, the majority answer stays the same, but its frequency falls from 9/10 to 7/10. This frequency is the confidence score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.