- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
This paper proposes RaTI (Reliability-aware Truth Inference), a multi-model question answering framework that combines candidate-level hallucination estimation with reliability-weighted Minimum Bayes Risk (MBR) decoding. The central motivation is that agreement among multiple language models does not necessarily imply factual correctness, especially when models produce correlated errors.
RaTI first obtains one candidate response from each of five heterogeneous lightweight language models. It then uses a pretrained hidden-state-based hallucination detector to estimate the factual risk of each response. These risk scores are converted into normalized inverse-risk credibility weights, which are incorporated into an MBR objective based on cosine distances between sentence embeddings. The final prediction is selected from the original candidate set by minimizing the weighted semantic disagreement.
The authors evaluate RaTI on six question answering benchmarks and compare it against individual models, collaborative inference baselines, and alternative truth-inference strategies. The paper also includes analyses of reliability estimation, model complementarity, robustness, computational overhead, and several additional experiments in the appendix. The reported results suggest that instance-level reliability information can improve answer selection in heterogeneous multi-model settings.
Critical strengths
-
Well-motivated problem formulation. The paper addresses an important limitation of consensus-based multi-model inference: multiple models may agree on an incorrect answer. Distinguishing candidate reliability from inter-model agreement is a meaningful perspective for improving collaborative question answering.
-
Simple and interpretable framework. The proposed method combines hallucination-risk estimation with weighted semantic consensus in a straightforward manner. The individual components are relatively easy to understand, and the final decision does not require an additional autoregressive aggregation model.
-
Broad empirical evaluation. The experiments cover six QA benchmarks and compare the method with several lightweight models and collaborative inference baselines. The reported improvements are generally consistent across F1, Exact Match, and BLEU.
-
Useful component and reliability analyses. The comparison between perplexity-based and hidden-state-based scoring provides evidence that the choice of reliability estimator matters. The reported ROC-AUC improvement from 0.667 to 0.728 supports the usefulness of the hidden-state signal.
-
Additional experimental evidence. The appendix provides comparisons with uniform and fixed model-level weighting, sensitivity analyses for sentence encoders and weighting functions, ensemble-size and heterogeneity studies, paired bootstrap confidence intervals, and candidate-oracle analysis. These experiments help characterize the behavior and limitations of the proposed framework.
-
Practical potential. The method avoids iterative inter-model communication and additional autoregressive answer synthesis. This makes it potentially useful in settings where multiple small language models are already available and an inexpensive answer-selection mechanism is desired.
Critical weaknesses
- Limited methodological novelty.
The main algorithm is a relatively direct combination of an existing pretrained hallucination detector, inverse-risk weighting, and embedding-based MBR decoding. The hallucination detector is not newly developed or trained by the authors, and the semantic disagreement function is the standard cosine distance between sentence embeddings. Although the integration is practically reasonable, the paper does not clearly establish a substantial algorithmic contribution beyond combining these established components. The relationship to prior confidence-weighted consensus and risk-aware reranking methods could be discussed more carefully.
- Insufficient validation of the central reliability assumption.
The method assumes that hallucination scores produced by a pretrained detector provide meaningful relative reliability estimates across different candidate models and answer types. However, the detector is trained on external context-grounded hallucination datasets, while the evaluated tasks include extractive QA, multiple-choice reasoning, relation extraction, and numerical reasoning. The paper reports an overall ROC-AUC of 0.728, but this moderate discrimination performance does not fully establish that the scores are consistently reliable across datasets, model families, or different types of errors.
Moreover, Equation (6) uses inverse-risk weighting, which may greatly amplify differences between small hallucination scores. Since the authors explicitly acknowledge that these scores are not calibrated probabilities, additional analysis of score distributions, failure cases, and the stability of resulting weights would strengthen the reliability argument. The existing weight-mapping sensitivity experiments are useful but do not completely address this issue.
- Limited evidence isolating the contribution of reliability-weighted MBR from simpler selection rules.
Appendix D.1 provides a useful comparison with uniform, global model-level, and dataset-level MBR weighting. Nevertheless, the paper does not fully establish whether the combination of reliability scoring and semantic MBR is substantially better than simply selecting the candidate with the lowest predicted hallucination risk. Such a detector-only argmin baseline is particularly important because the proposed method uses a strong pretrained detector and operates on only five candidate answers.
Similarly, comparing against a semantic-consensus-only method with the same candidate set, and examining the interaction between detector quality and MBR selection, would provide clearer evidence about the necessity of the proposed joint decision rule.
- Incomplete end-to-end computational efficiency evaluation.
The efficiency analysis primarily reports the number of autoregressively generated tokens. However, RaTI additionally evaluates a Qwen2.5-3B-based hallucination detector on every candidate and computes sentence embeddings for semantic comparison. These operations introduce nontrivial computation and memory requirements, even though they do not generate new output tokens.
Consequently, equal generated-token consumption does not imply equal inference latency, FLOPs, energy consumption, or memory usage. End-to-end measurements, including detector execution, embedding computation, and aggregation, are necessary to substantiate the practical efficiency advantages, particularly when comparing RaTI against a larger single model or conventional voting.
- Potential limitations of semantic distance for factual answer selection.
The proposed MBR objective uses cosine distances between sentence embeddings as a proxy for disagreement between candidate answers. However, semantic similarity does not necessarily reflect factual agreement. For example, two responses containing different numerical values, dates, or named entities may be close in embedding space despite expressing contradictory facts.
This issue is particularly relevant for DROP, relation extraction, and mathematical reasoning. The current evaluation does not sufficiently isolate cases where semantic similarity conflicts with factual correctness. More targeted experiments on numerical contradictions, entity substitutions, and semantically similar but factually incompatible answers would help validate the choice of loss function.
- Remaining concerns about evaluation scope and reproducibility.
The main evaluation uses context-grounded QA benchmarks, while the broader motivation concerns factual reliability in multi-model question answering. It remains unclear how well the method generalizes to closed-book factual QA, settings without supporting passages, or questions requiring external verification. The additional GSM8K experiment is useful, but its improvement over the strongest collaborative baseline is relatively small.
Furthermore, several benchmarks are evaluated on subsets rather than their full original datasets. The paper reports subset sizes but does not sufficiently explain their selection procedure and the sensitivity of results to the chosen evaluation samples. More detailed reporting of dataset selection, preprocessing, answer normalization, and inference configuration would improve reproducibility.
Questions for authors
How does RaTI compare against a detector-only selection baseline that simply chooses the candidate with the lowest predicted hallucination score? Can the authors report this comparison on all six benchmarks using exactly the same candidate pool?
Can the authors provide per-dataset and per-model-family evaluations of hallucination detection quality, including ROC-AUC and score calibration? In particular, how reliable is the detector for numerical reasoning and multiple-choice QA?
8 comments
Response to Weakness 2 and Question 1
We thank the reviewer for these valuable comments. We provide additional analyses as follows.
-
Reliability Evaluation. The HS estimator achieves an overall ROC-AUC of 0.728, indicating moderate but meaningful discrimination capability. Although EM-based correctness labels may introduce noise by treating semantically equivalent answers as incorrect, RaTI consistently outperforms HS-lower, which directly selects the lowest-risk candidate, demonstrating the complementary benefit of semantic consensus.
Method DROP RE RACE SQuAD NewsQA TYDIQA HS-lower 48.0 75.8 40.7 70.1 40.0 52.6 RaTI 54.5 79.7 43.9 73.8 41.2 53.7 -
Robustness of Reliability-to-Credibility Mapping. We adopt normalized inverse mapping to assign higher weights to lower-risk candidates, with normalization ensuring that weights sum to one and smoothing preventing numerical instability. Although inverse mapping may amplify small risk differences, sensitivity analysis (Appendix D.3, Figure 9) shows similar F1 performance across Uniform, Linear, Exponential, and Inverse mappings (69.6, 70.0, 69.9, and 70.5, respectively). Inverse mapping achieves the best overall EM (57.8), F1 (70.5), and BLEU (65.5), supporting its effectiveness.
Response to Weakness 3
We thank the reviewer for this comment. To clarify the contributions of the two main components, Figure 3 reports an ablation study comparing hallucination-based reliability estimation (HS), semantic MBR, and their combination (HS-MBR).
The results show that HS-MBR achieves the best performance across all three metrics, with average F1, EM, and BLEU scores of 70.5, 57.8, and 65.5, respectively. In comparison, HS achieves 67.9, 54.5, and 63.1, while MBR achieves 69.6, 56.5, and 64.2. These results demonstrate that reliability estimation and semantic consensus provide complementary benefits: HS reduces the influence of unreliable candidates, while MBR further exploits semantic agreement. Their combination consistently outperforms either component alone, validating the effectiveness of the proposed framework.
Response to Weakness 1
We thank the reviewer for this insightful comment. We clarify the methodological novelty and algorithmic contributions of RaTI as follows.
Methodological Novelty. The novelty of RaTI lies in shifting multi-model aggregation from consensus-based selection to reliability-aware consensus inference. Rather than relying solely on agreement among candidate responses, RaTI jointly considers their instance-level reliability and semantic relationships. This is particularly important because agreement among models does not necessarily guarantee factual correctness.
Algorithmic Contribution. The key algorithmic contribution is to integrate candidate-level reliability into the semantic MBR objective. Specifically, hallucination-derived credibility weights are used to measure the expected semantic disagreement:
\sum_{j=1}^{n}W_j L(R_i,R_j),$$ where ($W_j$) represents the credibility of candidate ($R_j$), and ($L(R_i,R_j)$) measures semantic disagreement. Unlike conventional MBR, which treats all candidates equally, RaTI assigns greater influence to more reliable candidates when evaluating the consensus risk.Relationship to Existing Methods. RaTI differs from existing confidence-weighted consensus methods in that it estimates the reliability of each candidate response for each individual question, rather than relying primarily on model-level confidence or generation probabilities. It also differs from risk-aware reranking methods, which typically use risk scores to filter, rank, or independently select candidates. Instead, RaTI incorporates reliability weights directly into the MBR objective, allowing reliability and semantic consensus to jointly determine the final answer.