acceptodds
Under review as a conference paper at ICLR 2027

Verbalized Confidence Calibration for Retrieval-Augmented Large Language Models via Multi-Source Evidence Fusion

Abstract

Large language models are increasingly deployed in retrieval-augmented generation (RAG) settings, yet they frequently verbalize high confidence when their answers are incorrect. Existing methods for calibrating verbalized confidence typically rely on token-level generation probabilities or heuristic confidence scores, and use the correctness of the predicted answer as supervision, without explicitly modeling the agreement and conflict across the model's parametric knowledge and the retrieved documents. We propose a two-stage framework for verbalized confidence calibration in retrieval-augmented multiple-choice question answering. In the first stage, information from the model's parametric knowledge and from each retrieved document is encoded into source-specific evidence vectors, each of which is used to parameterize a Dirichlet distribution over answer options. Using self-attention, a fusion module captures agreement and conflict among these sources and produces a fused Dirichlet distribution. The expected value of the probability assigned to the predicted answer under this distribution serves as a scalar confidence target. In the second stage, a lightweight confidence estimator takes the question-conditioned answer-option logits as input and learns to predict this scalar target without requiring additional document retrieval during confidence estimation. The predicted scalar confidence is then converted into a Gaussian-smoothed soft target distribution over verbalized confidence levels, and KL divergence regularization is employed to align the model's verbalized confidence distribution with this soft target. Experiments on multiple-choice question-answering benchmarks show that our method improves confidence calibration on the in-distribution dataset and on most out-of-distribution datasets, and, overall, yields more reliable uncertainty estimates than baseline methods.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large language models;Retrieval-augmented generation;Verbalized confidence calibration;Multi-source evidence fusion;Evidential deep learning

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_S3uyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a two-stage framework for calibrating verbalized confidence in large language models using multiple sources of evidence. In the first stage, the model's parametric knowledge and retrieved documents are converted into evidence representations. They are aggregated through an attention-based evidence fusion (AEF) module to parameterize a Dirichlet distribution over answer options. In the second stage, the trained AEF module generates soft confidence targets using only the model's logits, which are conditioned by the question, without requiring document retrieval as in the first stage. These targets are used to fine-tune the LLM through a loss function containing the Brier score, KL divergence, and answer prediction losses.

The method is evaluated on Llama-3.1-8B-Instruct and Qwen3-8B-Instruct, with training on MedQA and OOD evaluation on PubMedQA, TriviaQA, StrategyQA, and TruthfulQA. The reported results show improved average calibration and confidence discrimination compared with existing baselines, particularly on the in-distribution dataset. The paper also provides ablation studies of the training objectives, evidence fusion module, and several hyperparameters.

Critical strengths

S1. Transferring retrieval-informed confidence estimates into verbalized confidence.

The two-stage design provides a plausible framework for using external evidence during confidence estimator training while avoiding additional per-document LLM passes when producing confidence targets during Stage 2. In particular, masked-retrieval training connects the multi-source estimator to the setting without retrieval that supplies soft targets. This is meaningful because it attempts to retain the benefits of retrieval-informed supervision without requiring the full multi-source estimation procedure whenever confidence is expressed. The practical value of this separation is a substantive reason to consider the approach.

S2. Substantial in-distribution performance improvement.

On MedQA, the reported ECE decreases, and AUC improves substantially for Llama and Qwen relative to ConfTuner. Improvements in both calibration and discrimination would be practically meaningful, with the reliability diagram in Figure 2 providing qualitative support for better alignment between confidence and empirical accuracy. These in-distribution results constitute the strongest empirical evidence in the paper, subject to resolution of the evaluation-protocol concerns below.

Critical weaknesses

W1. Insufficient justification for the epistemic uncertainty interpretation.

The paper interprets u=K/Su=K/S as epistemic uncertainty, but the link between logit-derived evidence strength and the LLM's knowledge remains insufficiently validated. For the source-level mapping e=softplus⁡(z)e=\operatorname{softplus}(z), a common logit shift preserves normalized answer-option probabilities but changes the corresponding evidence strength and uncertainty. The epistemic meaning of this sensitivity needs justification. Moreover, the claimed uncertainty responses to source agreement and conflict are not directly validated by controlled experiments. Proposition 2 establishes variance reduction for fixed teacher, per-example optimization, rather than a calibration guarantee for the fine-tuned LLM. These gaps may weaken the paper's central theoretical motivation.

W2. Insufficient isolation of the retrieval-based evidence fusion contribution.

During Stage 2, all retrieval branches are masked, and the confidence targets depend only on no-retrieval logits. Although Appendix E.2 includes pmask=1p_{\mathrm{mask}}=1, it does not report the final verbalized confidence calibration results after Stage 2 under this setting, with all other training conditions kept unchanged. Thus, the contribution of retrieval-based evidence learning cannot be separated from general confidence distillation and KL regularization. Furthermore, the retrieval and no-retrieval evaluations in the supplementary implementation use potentially different predicted answers (204/1,268 cases in the supplied Llama test logits), making it difficult to isolate improvements in confidence estimation itself.

W3. Concerns about the evaluation protocol and reliability of reported calibration metrics.

The supplementary code uses different confidence definitions across evaluations. For MedQA, confidence is computed as the expectation of the confidence-token distribution, c^ID=∑j=09qj(j/9)\hat{c}_{\mathrm{ID}}=\sum_{j=0}^{9} q_j(j/9), whereas OOD evaluation uses the generated confidence value, c^OOD=jgen/9\hat{c}_{\mathrm{OOD}}=j_{\mathrm{gen}}/9, where jgenj_{\mathrm{gen}} is the generated confidence token. It is unclear which definitions were used for the reported results in Table 1 and whether the improvements hold under a consistent evaluation procedure. The PubMedQA prompt restricts answers to yes/no despite the supplied dataset containing "maybe" labels. These evaluation issues require clarification to establish the reliability of the reported results. Answer accuracy after fine-tuning is also not reported, making it difficult to assess whether calibration improvements preserve answer quality.

Questions for authors

Q1. Epistemic uncertainty interpretation (Sections 3- 4, Appendix C).

How can you justify the interpretation of u=K/Su=K/S as epistemic uncertainty despite its sensitivity to common logit shifts? Can the authors provide controlled evidence that the proposed uncertainty measure appropriately responds to supporting, conflicting, or irrelevant retrieved documents?

Q2. Retrieval contribution (Sections 4.2–4.3, Appendix E).

How does final verbalized confidence calibration change when AEF is trained without retrieval, while keeping the architecture and Stage 2 training identical? This may establish whether retrieval-based evidence learning provides benefits beyond general confidence distillation.

Q3. Retrieval versus no-retrieval comparison (Table 3).

Does Table 3 evaluate different predicted answers and correctness labels across retrieval conditions, as suggested by the supplementary implementation? If so, could the authors report results with the predicted answers and correctness labels held fixed to isolate the contribution of retrieval to confidence estimation?

Q4. Confidence extraction and metric computation (Table 1, supplementary code).

The supplementary code appears to use c^=∑j=09qj(j/9)\hat{c}=\sum_{j=0}^{9}q_j(j/9) for MedQA but the generated confidence value c^=jgen/9\hat{c}=j_{\mathrm{gen}}/9 for OOD datasets. Could the authors clarify which confidence definitions were used for Table 1 and whether the reported improvements hold under a consistent evaluation procedure?

Q5. PubMedQA evaluation protocol (Appendix G.2).

How were "maybe"-labeled PubMedQA examples handled given the yes/no-only evaluation prompt? Were these examples excluded, or was a different prompt used for the reported results? If neither, can the authors provide corrected results?

Q6. Answer prediction accuracy (Section 5, Tables 1–2).

Could the authors report answer accuracy before and after fine-tuning, including the loss ablations? This would clarify whether the calibration improvements preserve answer quality.

7 comments
AuthorsOpenReview ↗

Response to Reviewer Comment Point by Point

Thanks for your valuable and careful comment. Below, we answer your questions in detail and outline the corresponding revisions we will make in the final manuscript.

  1. Regarding question 1, we agree that our formulation does not guarantee invariance to common logit shifts and therefore does not fully justify interpreting u=K/Su=K/S as a direct measure of the LLM’s epistemic uncertainty. We have revised this interpretation to “evidence-based uncertainty” within the adopted Dirichlet parameterization. In our controlled experiment, mean uu is highest without retrieval and lower under all three document conditions, with supporting evidence yielding the lowest mean value. However, these results do not establish that uu reliably distinguishes evidence quality. We also explicitly acknowledge its sensitivity to common logit shifts as a limitation. The relevant content is shown on page 2-4, 10, 14 and 23 of the revised PDF, in the introduction, the related works, the preliminary, the section 4 , the conclusion, the Appendix B.2 and the Appendix K.

  2. Regarding question 2, training AEF without retrieval worsens average final verbalized confidence calibration for both backbones: average ECE and Brier score increase, while average AUC decreases. With the AEF architecture and Stage 2 training procedure held unchanged, these results support an additional contribution from retrieval-assisted AEF training beyond the confidence-distillation and regularization objectives. The relevant content is shown on page 9 of the revised PDF, in the Table 2 and the paragraph above Table 2.

  3. Regarding question 3, thank you for pointing out this confounding factor. Indeed, the original Table 3 allowed predicted answers and correctness labels to differ across retrieval conditions. We have added the requested comparison with predicted answers and correctness labels held fixed. Under this controlled setting, test-time retrieval provides no clear additional calibration benefit, while AEF trained with retrieval exhibits more stable calibration metrics across test conditions than AEF trained without retrieval. The relevant content is shown on page 16-17 of the revised PDF, in the Appendix E.1 and Table 4.

  4. Regarding question 4 and 5, we thank the reviewer for identifying these evaluation inconsistencies. For Question 4, the original evaluation used the expectation of the confidence-output distribution for MedQA but the generated confidence value for the OOD datasets. We have corrected this inconsistency by using generated confidence values throughout. Under a consistent evaluation procedure, our method retains the lowest average ECE and Brier score among the four main comparison methods for both backbones. For Question 5, “maybe”-labeled examples were retained in the original PubMedQA evaluation, although the prompt incorrectly allowed only “yes” and “no.” We have corrected the prompt to permit all three answers and reevaluated PubMedQA without excluding “maybe”-labeled examples. Although our method outperforms the compared methods on only some metrics on PubMedQA, it achieves the best average ECE, Brier score, and answer accuracy across all five datasets for both backbones. These results indicate that the overall benefits remain supported under the corrected evaluation procedure. Specifically, for methods that generate verbalized confidence, we extract the integer c∈{0,…,9}c\in\lbrace 0,\ldots,9\rbrace generated by the LLM and convert it to a confidence score in [0,1][0,1] using c/9c/9. The revised corrected results, and accompanying analysis appear on pages 7–9 and 22 of the revised manuscript.

  5. Regarding question 6, our method improves average answer accuracy after QLoRA fine-tuning, from 0.6042 to 0.6420 for Llama 8B and from 0.6089 to 0.6560 for Qwen 8B. Removing the language modeling loss produces the largest decrease in average accuracy among the loss ablations for both backbones. The detailed accuracy results are reported in Appendix G, on pages 19–20 of the revised manuscript.

AuthorsOpenReview ↗

The details and data for this comment

We have reorganized the detailed data in the response for the reviewers’ direct review.

  1. Regarding question 1, we agree that a common shift of the answer-option logits preserves their softmax probabilities but can change the evidence obtained through the softplus mapping. Consequently, our formulation does not guarantee shift invariance, and interpreting u=K/Su=K/S as a direct measure of the LLM’s epistemic uncertainty was too strong. In our implementation, AEF operates on the native logits of a fixed backbone under a specified prompting and extraction procedure. Common shifts are not explicitly introduced during normal inference. To examine responses to different types of evidence, we constructed a controlled synthetic-evidence experiment. Given each question, its answer options, and the ground-truth answer, we prompted DeepSeek to generate two supporting passages intended to substantiate the correct answer and two misleading passages intended to favor an incorrect alternative. The latter served as a conflicting-evidence condition. We additionally constructed a mismatched condition by assigning passages from other questions, providing a control for irrelevant evidence. Because the synthetic passages were generated with access to the ground-truth answer, this experiment evaluates controlled evidence interventions rather than the quality of a realistic retrieval pipeline. For this supplementary experiment, we trained AEF using supporting passages from the training split. The evaluated answer for each question was fixed to the LLM’s prediction made without retrieval. Predictions and correctness labels therefore remained identical across conditions, with 704 correct and 514 incorrect predictions. Confidence was computed as the Dirichlet mean assigned to that fixed answer.

Table 5. AEF Uncertainty and Confidence under Controlled Evidence Conditions.

MetricWithout RetrievalSupportingConflictingMismatch
Average Uncertainty (u)0.27810.23420.23630.2390
Confidence of Correct Answers0.59660.62730.57340.5612
Confidence of Incorrect Answers0.45100.41010.43940.4224

Supporting evidence increased mean confidence on correct predictions from 0.5966 to 0.6273 and decreased it on incorrect predictions from 0.4510 to 0.4101. In contrast, misleading and mismatched evidence reduced confidence in both groups. These results demonstrate confidence adjustments consistent with prediction correctness under this supporting-evidence training condition. The interpretation of uu requires a distinction between evidence strength and evidence quality. Within our Dirichlet parameterization, u=K/Su=K/S reflects the total concentration assigned across all answer options, rather than whether the evidence supports the ground-truth answer. Evidence assigned to an incorrect alternative can also increase total concentration and therefore reduce uu. Accordingly, a lower uu does not necessarily indicate more reliable evidence or a more accurate prediction. So, all three document conditions reduced mean uu relative to the no-retrieval baseline.

AuthorsOpenReview ↗

The details and data for this comment

  1. Regarding question 2, we added an ablation in which AEF is trained without retrieval, while keeping its architecture and the Stage 2 training procedure unchanged. The no-retrieval AEF is then used as the confidence teacher in Stage 2. The table reports final verbalized confidence calibration and answer accuracy, with averages computed across the five datasets. The results show that incorporating retrieval during AEF training improves the average confidence estimation performance of both models. For Llama-8B, average ECE decreases from 0.1552 to 0.0882, Brier score decreases from 0.2479 to 0.2028, and AUC increases from 0.6027 to 0.6990. Qwen-8B shows smaller but consistent improvements in these average metrics, although its average answer accuracy decreases slightly from 0.6613 to 0.6560.

Table 6. Effect of Retrieval during AEF Training on Final Verbalized Confidence Calibration and Answer Accuracy.

MetricModelMethodMed (ID)Pub (OOD)Tri (OOD)Str (OOD)Tru (OOD)Average
ECE ↓Llama 8Bw/o RE0.23130.04870.08750.11720.29140.1552
ECE ↓Llama 8BOurs0.05180.10080.05750.08310.14780.0882
ECE ↓Qwen 8Bw/o RE0.15840.04870.24360.12650.11160.1378
ECE ↓Qwen 8BOurs0.12750.03380.24170.14320.06620.1225
AUC ↑Llama 8Bw/o RE0.55960.58850.64570.62800.59150.6027
AUC ↑Llama 8BOurs0.71180.72340.69680.65350.70930.6990
AUC ↑Qwen 8Bw/o RE0.65950.75130.72130.59320.55870.6568
AUC ↑Qwen 8BOurs0.67870.76510.72390.59360.64030.6803
Brier ↓Llama 8Bw/o RE0.29300.18020.20540.23580.32510.2479
Brier ↓Llama 8BOurs0.21520.17000.18110.20800.23970.2028
Brier ↓Qwen 8Bw/o RE0.24630.16460.26680.24160.20140.2241
Brier ↓Qwen 8BOurs0.23130.15830.26280.24930.18530.2174
ACC ↑Llama 8Bw/o RE0.56720.75100.67030.67830.47740.6288
ACC ↑Llama 8BOurs0.57060.74870.71970.68610.48510.6420
ACC ↑Qwen 8Bw/o RE0.61010.72700.56140.65160.75640.6613
ACC ↑Qwen 8BOurs0.61040.72100.55570.65360.73930.6560
  1. Regarding question 3, specifically, we fixed each evaluated answer to the LLM’s prediction made without retrieval and kept its correctness label unchanged. We compared AEF trained with mixed retrieval exposure (pmask=0.5p_{\mathrm{mask}}=0.5) against AEF trained without retrieval (pmask=1p_{\mathrm{mask}}=1), evaluating each with and without retrieval at test time. For AEF trained with mixed retrieval exposure, adding retrieval at test time leaves the Llama 8B metrics unchanged at the reported precision. For Qwen 8B, ECE changes from 0.0755 to 0.0778, AUC from 0.7317 to 0.7331, and Brier remains 0.2138. In contrast, AEF trained without retrieval shows deterioration in all three metrics when retrieval is introduced at test time. For Qwen 8B, ECE increases from 0.0747 to 0.0973, and AUC decreases from 0.7405 to 0.7261. This comparison suggests that retrieval exposure during AEF training improves the stability of aggregate confidence-estimation metrics across the evaluated test conditions. We include this controlled comparison and revise Appendix E.1 accordingly. In particular, we remove the stronger attribution of the original AUC gains to improved confidence estimation through retrieval and replace it with the more limited interpretation supported by the fixed-answer results.

Table 7. AEF Confidence Estimation with Fixed Predicted Answers across Training and Test Retrieval Conditions.

Training ConditionTest ConditionLlama 8B ECE ↓Llama 8B AUC ↑Llama 8B Brier ↓Qwen 8B ECE ↓Qwen 8B AUC ↑Qwen 8B Brier ↓
With RetrievalWithout Retrieval0.04730.73810.20320.07550.73170.2138
With RetrievalWith Retrieval0.04730.73810.20320.07780.73310.2138
Without RetrievalWithout Retrieval0.04410.73860.20320.07470.74050.2116
Without RetrievalWith Retrieval0.04540.72710.20750.09730.72610.2191
Reviewer_ovzROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a two-stage framework to calibrate the verbalized confidence of LLMs in multiple-choice QA using multi-source evidence fusion. In Stage 1, an Attention-based Evidence Fusion (AEF) module maps no-retrieval and document-conditioned answer logits into Dirichlet distributions to model agreement and conflict across sources. In Stage 2, the frozen AEF module generates soft confidence targets (under a masked-retrieval setting) to fine-tune the LLM via a combined Brier score, KL distillation, and language modeling loss. The authors validate the method by training 4-bit quantized Llama-3.1-8B and Qwen3-8B on MedQA and evaluating on four OOD benchmarks, showing improved ECE, Brier scores, and AUC over baselines like ConfTuner.

Critical strengths

The idea of using evidential deep learning and Dirichlet distributions to explicitly model agreement and conflict between parametric knowledge and retrieved documents is well-motivated and yields a much richer supervision signal than binary correctness labels. Furthermore, the masked-retrieval strategy is a practical design choice that enables single-pass verbalized confidence estimation at inference time without needing costly multi-document forward passes, while still generalizing reasonably well across out-of-distribution datasets.

Critical weaknesses

There is a disconnect between the paper's framing as a calibration method for RAG and how the model actually operates in Stage 2 and at test time. From what I understand, during Stage 2 fine-tuning and OOD evaluation, retrieved documents are completely masked out and the model only uses no-retrieval logits, meaning it is not calibrating an active RAG generation pipeline at inference time. Additionally, the AEF formulation strictly relies on extracting logits over a fixed set of K answer options, which limits the method to multiple-choice setups. It is unclear how this was applied to TriviaQA—which is prompted for short-phrase answers in Appendix G.3—and there is a minor inconsistency in Table 3 where the text and table swap the results for AEFConf and AEFConf*.

Questions for authors
  1. How is the AEF module (which requires K option logits) applied to TriviaQA, given that Appendix G.3 shows a free-form short-phrase prompt rather than multiple-choice options?
  2. Since Stage 2 calibration and test-time inference run entirely in the masked (no-retrieval) setting, how does the verbalized confidence behave when retrieved documents are actually provided in the prompt at test time? If my understanding is incorrect, please elaborate.
  3. Could you clarify the discrepancy between the text in Appendix E.1 and Table 3 regarding AEFConf vs. AEFConf*? The text states the full-set model (AEFConf*) gets 0.0490 ECE and the 2,000-sample model (AEFConf) gets 0.1098, which is reversed in Table 3.
2 comments
AuthorsOpenReview ↗

Response to Reviewer Comment

Thanks for your valuable comment. Below, we answer your questions in detail and outline the corresponding revisions we make in the final manuscript.

  1. Regarding question 1, AEF cannot be directly applied to TriviaQA’s free-form answers because it requires logits over a predefined set of (K) answer options. We do not apply AEF to TriviaQA in our experiments. Instead, AEF provides confidence supervision during fine-tuning on MedQA, after that, the fine-tuned LLM directly generates answers and verbalized confidence scores without invoking AEF. TriviaQA is used for out-of-distribution evaluation of this fine-tuned LLM. We therefore neither construct answer options nor compute AEF targets for TriviaQA. Instead, we employ the free-form short-answer prompt in Appendix J.3(in the revised PDF) and evaluate the generated confidence against the correctness of the generated answer. This experiment assesses the transfer of verbalized confidence calibration learned on MedQA to an open-ended QA task, rather than the direct application of AEF to free-form answers.

  2. Regarding question 2, indeed, Stage 2 fine-tuning and the main evaluations use the masked, no-retrieval setting. And when retrieved documents are included in the prompt at test time, our method retains lower ECE and Brier scores than both the untuned baseline and ConfTuner, for both Llama 8B and Qwen 8B. Thus, our method remains effective for verbalized confidence calibration when retrieved documents are provided at inference time, despite Stage 2 training without retrieval. To investigate behavior when retrieved documents are provided at inference time, we additionally evaluated the Stage 2 models with retrieved documents included in the prompt. Model parameters were held fixed, and no additional fine-tuning was performed. Since the no-retrieval results have already been reported above, the following table presents only the results with retrieval. Under this setting, our method achieves lower ECE and Brier scores and higher answer accuracy than both the untuned baseline and ConfTuner for both backbones. It also achieves the highest AUC for Llama 8B, while its AUC for Qwen 8B is slightly lower than ConfTuner’s (0.7214 versus 0.7245). The relevant content is shown on page 21 of the revised PDF, in the Appendix I verbalized confidence calibration performance under RAG.

  3. Regarding question 3, thank you for pointing out this inconsistency. Appendix E.1 incorrectly reverses the model names and reports an incorrect ECE value for the full-set model. The correct correspondence is as follows: AEFConf uses the full training set and achieves an ECE of 0.0473, whereas AEFConf* uses 2,000 training samples and achieves an ECE of 0.1098, as reported in Table 3. The value 0.0490 in the text should therefore be corrected to 0.0473. We apologize for these reporting errors and will correct Appendix E.1 and ensure consistent model names and numerical values throughout the manuscript. The relevant content is shown on page 16 of the revised PDF, in the Appendix E.1 confidence estimation results.

AuthorsOpenReview ↗

The details and data for this comment

We have reorganized the detailed data in the response for the reviewers’ direct review.

  1. Regarding question 1, TriviaQA is used for out-of-distribution evaluation of this fine-tuned LLM. We therefore neither construct answer options nor compute AEF targets for TriviaQA. Instead, we employ the free-form short-answer prompt in Appendix J.3(in the revised PDF) and evaluate the generated confidence against the correctness of the generated answer. This experiment assesses the transfer of verbalized confidence calibration learned on MedQA to an open-ended QA task, rather than the direct application of AEF to free-form answers.

  2. Regarding question 2,

Table 4. Answer accuracy and verbalized confidence calibration on the MedQA test set with retrieved documents included in the prompt.

MethodLlama 8B ACCLlama 8B ECELlama 8B BrierLlama 8B AUCQwen 8B ACCQwen 8B ECEQwen 8B BrierQwen 8B AUC
Baseline0.55630.40130.40560.52880.52870.34960.36920.5391
ConfTuner0.55470.23880.26400.71100.57440.10350.21850.7245
Ours0.56960.07220.21730.71830.60900.08380.20590.7214
Reviewer_cVXEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a two-stage method for calibrating the confidence of multiple-choice LLMs. In the first stage, the method retrieves multiple documents with ToG-2, obtains answer-option logits from the question alone and from each question-document pair, maps these sources to Dirichlet evidence, and fuses them with an attention-based evidence fusion module. The expected probability of the predicted answer is used as a confidence target. In the second stage, the retrieval branches are masked and the target is distilled into the LLM with QLoRA using a Brier loss, a KL loss over smoothed confidence-token labels, and an answer language-modeling loss. Llama-3.1-8B-Instruct and Qwen3-8B-Instruct are trained on MedQA and evaluated on MedQA and four out-of-distribution QA datasets. The reported results improve average ECE, AUC, and Brier score over the listed baselines, with ablations supporting the use of the three losses.

Critical strengths
  1. The paper gives a clear mechanism for converting agreement and conflict across the model's parametric knowledge and several retrieved documents into a soft confidence target. This is more informative than supervising confidence only with a binary correct or incorrect label.

  2. The empirical study covers two model families, one in-distribution dataset and four out-of-distribution datasets. The loss ablations and the comparison between retrieval and no-retrieval confidence estimation are useful for understanding which parts of the method contribute to the reported calibration gains.

Critical weaknesses
  1. The practical cost and the actual target system are unclear. Constructing the first-stage signal uses ToG-2 retrieval and, with the chosen M = 4, five LLM forward passes per example: one without retrieval and one for each retrieved document. This is a substantial data-generation cost. More importantly, retrieval is removed during the second-stage calibration and final confidence evaluation. The method therefore appears to use RAG as an expensive teacher for calibrating a non-retrieval LLM, rather than calibrating the confidence of an end-to-end RAG system. This makes the title and some conclusions potentially misleading.

  2. The paper does not report QA accuracy or exact match for the baseline, ConfTuner, and the proposed method. This is a central omission. The language-modeling loss is explicitly motivated as preserving answer performance, and the ablation discussion states that removing it reduces the ability to predict correct answers, but no accuracy values are shown. ECE, AUC, and Brier score measure confidence quality; they do not show whether QLoRA changed the number of correct answers. Without accuracy, it is not possible to verify that the calibration gains were not obtained at the expense of task performance.

  3. The paper does not analyze how the quality of retrieved evidence affects confidence calibration. Since the method is motivated by agreement and conflict between the model's parametric knowledge and retrieved documents, it is important to distinguish correct, irrelevant, and misleading evidence. A controlled evaluation with correct and incorrect retrieved evidence would show whether the method becomes appropriately less confident under conflicting evidence or is simply influenced by whatever evidence is retrieved. Without this analysis, it is unclear whether the reported gains come from robust evidence fusion or mainly from the retrieval pipeline providing useful information.

Questions for authors
  1. What are the QA accuracy or exact-match results for the untuned baseline, ConfTuner, and the proposed method on every dataset? Please also report the change in accuracy before and after QLoRA, including the loss ablations if possible. This result would materially affect my assessment.

  2. What is the full cost of producing the first-stage targets, including the number of ToG-2 retrieval operations?How does the calibration gain change with a cheaper setting such as M = 1, and how does the cost compare with the ensemble and ConfTuner baselines?

  3. The proposed method does not appear to calibrate a RAG system itself; rather, it uses retrieval only to construct the confidence supervision signal and then calibrates the verbalized confidence of an LLM without retrieval. However, the title states that the method is for calibrating retrieval-augmented large language models. Could the authors clarify this discrepancy? Is the current title potentially misleading about the actual target of calibration?

2 comments
AuthorsOpenReview ↗

Response to Reviewer Comment Point by Point

Thank you for your valuable comments. Below, we answer your questions in detail and outline the corresponding revisions we make in the final manuscript.

  1. Regarding question 1, our method improves average answer accuracy after QLoRA fine-tuning, from 0.6042 to 0.6420 for Llama 8B and from 0.6089 to 0.6560 for Qwen 8B. Removing the language modeling loss produces the largest decrease in average accuracy among the loss ablations for both backbones. The detailed accuracy results are reported in Appendix G, on pages 19–20 of the revised manuscript. Table 5 reports answer accuracy for the untuned baseline, Ensemble, ConfTuner, and our method on every evaluated dataset. Table 6 presents the ablation results. These results support its role in preserving answer accuracy during confidence calibration.

  2. Regarding question 2, the full cost of preparing the first-stage inputs is approximately 31–32 seconds per sample, including entity extraction and linking, document retrieval and reranking, and LLM logit extraction, compared with approximately 3 seconds per sample for ConfTuner’s data preparation. The retrieval pipeline involves up to three graph-guided retrieval iterations per question, and one complete preprocessing pass over MedQA covers 12,723 questions across the training, validation, and test sets. Reducing M from 4 to 1 saves approximately 1 second per sample but yields worse calibration results. At inference, our method requires approximately 0.4 seconds per sample, comparable to ConfTuner and substantially lower than the ensemble baseline’s approximately 5.6 seconds. Thus, the main additional cost is offline preparation, and M=4 provides better results with a modest increase in preparation time. The relevant content is shown on page 20-21 of the revised PDF, in the Appendix H computational cost analysis.

  3. Regarding question 3, indeed, retrieval is used during Stage 1 to train the AEF confidence estimator, whereas Stage 2 fine-tuning and the main test-time evaluations operate without retrieved documents. The evaluated target is therefore the verbalized confidence of an LLM without retrieval at inference time, rather than that of an end-to-end RAG system. We clarify the scope of the paper accordingly and, replace “Verbalized Confidence Calibration for Retrieval-Augmented Large Language Models via Multi-Source Evidence Fusion” with “Retrieval-Assisted Verbalized Confidence Calibration for Large Language Models via Multi-Source Evidence Fusion” in the title to more accurately reflect the role of retrieval in our method. We also revise the corresponding statements in the abstract and introduction to describe the method as retrieval-assisted verbalized confidence calibration. The relevant content is shown on page 1-2 and 9 of the revised PDF, in the title, the abstract, the introduction, the title of the subsection 4.1 and the conclusion.

AuthorsOpenReview ↗

The details and data for this comment

We have reorganized the detailed data in the response for the reviewers’ direct review.

  1. Regarding question 1, Table 1 reports answer accuracy for the untuned baseline, Ensemble, ConfTuner, and our method on every evaluated dataset. Relative to the untuned baseline, our method improves MedQA accuracy from 0.5381 to 0.5706 for Llama and from 0.5527 to 0.6104 for Qwen. On the four OOD datasets, Llama improves consistently, while Qwen improves on three datasets but decreases on TruthfulQA, from 0.8017 to 0.7393. The Average column reports the average value across all five datasets. Table 2 presents the ablation results. Removing the language modeling loss decreases average accuracy from 0.6420 to 0.5823 for Llama and from 0.6560 to 0.6320 for Qwen. These results support its role in preserving answer accuracy during confidence calibration. However, the full objective does not maximize accuracy in every ablation comparison, indicating trade-offs between confidence calibration and answer accuracy.

Table 1. Accuracy of Comparing Methods.

ModelMethodMed (ID)Pub (OOD)Tri (OOD)Str (OOD)Tru (OOD)Average
Llama 8BBaseline0.53810.73300.65970.67390.41620.6042
Llama 8BEnsemble0.39340.55410.58440.52190.64420.5396
Llama 8BConfTuner0.56250.68400.64630.66230.45220.6015
Llama 8BOurs0.57060.74870.71970.68610.48510.6420
Qwen 8BBaseline0.55270.51000.54410.63610.80170.6089
Qwen 8BEnsemble0.54320.63300.51550.66810.69360.6107
Qwen 8BConfTuner0.59950.30820.54390.61800.78940.5718
Qwen 8BOurs0.61040.72100.55570.65360.73930.6560

Table 2. Accuracy for the Ablation Study.

ModelMethodMed (ID)Pub (OOD)Tri (OOD)Str (OOD)Tru (OOD)Average
Llama 8Bw/o LM0.47450.74800.66240.65210.37450.5823
Llama 8Bw/o Brier0.59790.74850.67620.69430.57570.6585
Llama 8Bw/o KL0.56400.75600.67390.67100.45040.6231
Llama 8Bw/o RE0.56720.75100.67030.67830.47740.6288
Llama 8BOurs0.57060.74870.71970.68610.48510.6420
Qwen 8Bw/o LM0.58290.60460.55030.63410.78820.6320
Qwen 8Bw/o Brier0.60090.63500.55840.66760.77110.6466
Qwen 8Bw/o KL0.60610.68900.55890.65310.75640.6527
Qwen 8Bw/o RE0.61010.72700.56140.65160.75640.6613
Qwen 8BOurs0.61040.72100.55570.65360.73930.6560
  1. Regarding question 2, in Table 3, we report the computational costs of data preparation, both training stages, and inference separately. For ConfTuner, preparation includes predicted-answer generation. For our method, it includes entity extraction and linking, ToG-2 document retrieval and reranking, and LLM logit extraction. Each question is processed by the ToG-2 retrieval pipeline once (involving up to three iterations of graph-guided document retrieval and reranking to produce a ranked candidate pool), and the retrieved passages are reused for the different document-count settings. Thus, reducing M from 4 to 1 only reduces the number of top-M passages fed into the LLM for forward-pass logit extraction, but does not reduce the number of retrieval-pipeline executions. Reducing the number of documents from M=4 to M=1 lowers the logit extraction cost, but only modestly reduces total preparation time because the retrieval pipeline remains unchanged. AEF training is lightweight, taking approximately 0.13 ms and 0.12 ms per sample per epoch for M=1 and M=4, respectively. On the MedQA test dataset, Increasing M from 1 to 4 improves ECE from 0.1324 to 0.0518, AUC from 0.6472 to 0.7118, and Brier score from 0.2429 to 0.2152. Although the M=1 variant improves ECE and Brier score over ConfTuner, its AUC is lower.

Table 3. Time per sample per epoch for three methods under Llama 8B and test metrics on the MedQA test set.

MethodPreparationStage 1 TrainingStage 2 TrainingInferenceECE ↓AUC ↑Brier ↓
Ensemble–––5.6 s0.25690.62980.3015
ConfTuner3 s–1 s0.4 s0.19310.67510.2636
Ours (M=1)31 s0.13 ms1 s0.4 s0.13240.64720.2429
Ours (M=4)32 s0.12 ms1 s0.4 s0.05180.71180.2152

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.