acceptodds
Under review as a conference paper at ICLR 2027

HD-Fuse: Single-Pass Hallucination Detection via Confidence Reshaping with Conformal Likelihood-Ratio Fusion under Logprob-Only APIs

Abstract

Hallucination detection under API logprob-only contracts is practically important but constrained: production LLM APIs expose only final-token logprob distributions, while hidden states, attention weights, and repeated sampling are unavailable or expensive. We present a systematic empirical study of single-pass, logprob-only hallucination detection for short-form QA ( tokens) and identify a validated recipe that outperforms individual signals: extract a 5-dimensional confidence vector, model the test as a Gaussian likelihood ratio (LDA/QDA), fuse complementary likelihood-ratio members (including thermodynamic trajectory statistics of the top- distribution) via a calibrated cross-validated rank ensemble, and attach a split-conformal -value with a finite-sample false-positive-rate guarantee, at zero extra passes. On real public benchmarks the recipe attains the best cross-cell mean AUC at one pass versus five; we confirm the signal and false-positive control on a production API (DeepSeek), so the logprob-only claim does not rest on local quantised weights. Ablation attributes the gain to signal decorrelation. Scope and prerequisites (short answers; a per-(model, benchmark) calibration fold) are stated up front, so the contribution is a validated deployment recipe plus an interpretable diagnostic framework; longer-form generation is future work.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Hallucination detection, Large language models, Log‑probability, Conformal prediction, Likelihood‑ratio fusion

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 4 reviews, above 27% of the others.
Reviewer_Yd2rOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces a new method called HD-Fuse, that detects when an AI model might be hallucinating. It uses token log probabilities from one generated response without requiring access to internal model states or multiple generations. It combines five confidence measures and features that track changes in token probabilities through an ensemble of statistical detectors (a technique called a likelihood-ratio test (LRT)). A split-conformal calibration layer is introduced to provide statistical false-positive-rate guarantees. Tests on HaluEval QA and TruthfulQA across four open-weight models and report the best average AUC among the compared methods (0.835). The paper also includes a DeepSeek API experiment, component studies, cross-validation, and cost measurements. The method requires labelled calibration data for each model and benchmark.

Critical strengths
  • The method proposed in the paper requires only a single generated response and token-level output log probabilities and It does not require access to any internal model hidden states or multiple generations. This makes it useful for logprob only API settings, which are studied in the paper.
  • The paper evaluates HD-Fuse across two public benchmarks and four open-weight models. It also includes a production-API experiment and component ablation studies. Results show improved average hallucination detection performance over several existing baselines.
Critical weaknesses
  • In Table 3, the authors report performance for the thermodynamic trajectory components individually and for the full HD-Fuse ensemble. However, they do not show how the full ensemble performs when the trajectory-based components are removed.
Questions for authors

It would be nice if it is clarified how the 250 calibration prompts were used? Were the examples used to calibrate p-values kept separate from those used to fit the detectors and choose ensemble weights?

Reviewer_yNAJOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HD-Fuse, the method to detect the hallucination in LLM response, in the setting that only API log-probabilities are available. From the top K log-probability of the first few generated tokens, it extract a 5 dimensional features and model the answers with a Gaussian (LDA). The thermodynamic trajectory statistic is introduced, and another 5 scored is constructed based on these statistics. HD-Fuse fuses these metrics with LDA as the metric detector. On HaluEval and TruthfulQA across four open-weight LLMs, HD-Fuse achieves the best mean AUC using one forward pass versus five for sampling-based baselines, and the approach is additionally validated on the DeepSeek API.

Critical strengths
  1. This method only needs the log probabilities returned by LLM, which matches the current API exposes. It requires only one greedy forward pass instead of several pass in sampling-based detectors.

  2. Beyond headline AUC, the paper compares against single-pass and sampling-based baselines, also the ablation of the ensemble components, sensitivity of the number of tokens are conducted.

Critical weaknesses
  1. In Table 3, the improvement of HD-Fused is small (0.01) compared to logistic fusion, this gain is only 0.006 in Table 4. An ablation study needs to be conducted to analyze why this number reduced with more calibrated samples (250 vs 400)?

  2. The results reported in Tab 4 are worse than Tab 3, while Tab 4 conduct the running on different held-out validation set, with different calibration values. Therefore, the 250 calibration might be accidentally chosen to favor the performance of the method.

  3. In Equation 1, there are two features, which is context and response, is measured by the log of length. There is no connection to model these two features by a Gaussian. An ablation study or analysis needs to show the choice of the paper.

  4. Lots of information in this paper is not well defined, for example, logistic fusion in Section 3.4, and the connection between Section 3.3. and Section 3.4 (ThermoPath-KDE, ThermoPath-LDA, KDE-LRT, DD-LRT) needs to be cited or explicit explain the computation, which will make this paper easier to follow.

Questions for authors

Please see the weaknesses section

Reviewer_FExWOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper addresses hallucination detection in a constrained API setting where only output token log-probabilities are available, precluding access to hidden states and repeated sampling. The evaluation is restricted to short-form question answering, with responses limited to 16 tokens. The proposed method combines five features extracted from the first three generated tokens, including token negative log-probability, top-(K) entropy, the top-1/top-2 log-probability gap, and prompt and response lengths. These signals are modelled using class-conditional Gaussian distributions to construct a likelihood-ratio detector, with linear discriminant analysis (LDA) as the default configuration.

Critical strengths

The setting is practically relevant, and the scope is stated up front. Short answers, logprob-only access, a calibration fold per model and benchmark.

The paper is honest about negative results. Long-form window-level localization is at chance (0.55). Template benchmarks have a construction artifact, flagged explicitly. Calibration doesn’t transfer across model families.

The production-API note is useful practical knowledge. DeepSeek masks alternative logprobs at temperature 0, so the pass has to run at a nonzero temperature with the argmax logprob reconstructed from the top-K list.

There is a cost comparison with wall-clock times, not just pass counts.

Critical weaknesses

The results suggest that the majority of the performance gain is attributable to supervised learning rather than the proposed ensemble. A standard logistic regression model achieves nearly identical performance to the full method (0.824 vs. 0.835; 0.817 vs. 0.823 under cross-validation), indicating that the rank ensemble contributes only a modest incremental improvement. Furthermore, the improvement over DD-LRT, a closely related method, appears to be relatively small (+0.016) and does not appear to reach conventional statistical significance.

The larger performance margins are primarily against unsupervised baselines, which are less informative for establishing the method's comparative advantage. Among the more relevant supervised comparisons, the method outperforms WEPR by 0.032 and HALT, a GRU trained on only 250 examples. However, the limited training data for HALT raises concerns about whether this comparison adequately reflects the potential of a supervised recurrent baseline.

Overall, the results demonstrate the effectiveness of supervised training for this task, but provide more limited evidence that the proposed ensemble offers a substantial advantage over simpler supervised alternatives.

Questions for authors

Are there any variants of the method that would be considered unsupervised?

Reviewer_qohaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The authors propose "HD-Fuse", a single-pass method for hallucination detection in LLMs using only output token probabilities. The motivation is that individual confidence signals may not be sufficient to distinguish correct and hallucinated responses, and combining different signals can improve detection performance. The proposed method consists of multiple parts. First, the authors extract five signals based on token confidence, entropy, probability gaps, and input and output lengths. Next, they introduce additional thermodynamic trajectory statistics to capture changes in token probability distributions during generation. They then train multiple likelihood-ratio-based detectors and combine their outputs using a cross-validation-weighted rank ensemble. The authors evaluate their method on HaluEval and TruthfulQA using four LLMs. They also evaluate the method on the DeepSeek API and compare it with existing single-pass and sampling-based hallucination detection methods.

Critical strengths

The paper addresses an important problem of hallucination detection using only output token probabilities, without requiring access to the model's internal representations or multiple generations. The proposed method is computationally efficient and achieves improvements over some of the existing baselines. The use of conformal calibration provides a practical advantage.

Critical weaknesses
  1. The paper can be significantly improved in terms of presentation. In the methodology section, many parts of the method and the proposed signals are not discussed adequately. The motivations behind many of these signals are unclear, and some seem to provide similar information. For example, in Section 3.3, the authors introduce spectral temperature without clearly explaining how lambda_t is estimated. Under their assumed exponential model, this parameter is equal to the top-1/top-2 logprob gap, meaning that spectral temperature would simply be the reciprocal of an existing signal. If another procedure for estimating the lambda_t is used, then it should be clearly mentioned and discussed in the paper. Also for the Girsanov-style drift is not clear how the drift is being computed. For a general reader the paper may be hard to follow as it mentions several concepts without explaining them adequately.

  2. The paper does not include qualitative examples or figures showing cases where the proposed signals help detect hallucinations, particularly when simpler signals fail. I think the space used for the existing figures could have been utilised more efficiently to provide such examples and better explain the method.

  3. The novelty of the proposed method is not clear to me. Previous works have already investigated training hallucination detectors based on output token probabilities. The proposed method seems to be a combination of existing ideas, and the authors do not clearly explain why the additional signals are needed or how they capture different aspects of hallucination. The ablation study does not isolate the influence of each signal. For example, the signals mentioned in Section 3.3 are not ablated separately. Given that these signals may provide similar information, ablation on each of them is needed.

  4. The scope of the proposed method seems limited. In many benchmarks and types of hallucination, the hallucination does not necessarily occur within the first few generated tokens. However, the method mainly relies on the first three tokens, so it seems to be more applicable to tasks where the answer appears at the beginning of the response or shortly afterwards.

Questions for authors

See Weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.