acceptodds
Withdrawn from ICLR 2027

HPSA: Hallucination Detection from a Single Hidden State, Without Training or Layer Search

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Large language models can produce fluent but unsupported statements, yet many detectors require labelled in-domain data, additional generations, or a trained probe. We study a simpler alternative: a single hidden state from the answer-generation pass, robust coordinate scaling, and a closed-form first-moment direction. The resulting detector, HPSA, uses no fitted classifier, extra sampling, or target labels to produce its score. Across four question-answering benchmarks and three dense decoder models, it reaches a mean out-of-fold ROC-AUC of over twelve model–benchmark cells and is the best detector without labels from the scored rows in seven cells. A geometric analysis finds a consistent location shift between the two label populations. Finally, a direction estimated from other benchmarks retains of the target's own label-free AUC on average. These results support a narrow conclusion: a fixed hidden-state statistic ranks answers usefully at the cost of one dot product.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
hallucination detection, large language models, internal representations, training-free methods, uncertainty estimation

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_6sX8OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper focuses on the task of hallucination detection for LLMs. The authors proposed single forward pass label-free method that is based only on the hidden states of the model. The method is motivated by the geometry of the hidden states, with the finding that hallucinated and correct answers differ by a location shift. Finally, the authors provide results on 4 datasets and 3 models for both in-domain and out-of-domain setups.

Critical strengths
  1. The scope of the paper is well defined in the text, and the research direction is important. The method seems quite novel.
  2. The proposed method shows competitive performance compared to more complicated methods and even outperforms them without any label tuning.
Critical weaknesses
  1. My main concern is the importance of the individual components of the method. In particular, the norm ∥z∥\|z\| is highly competitive and in some cases even outperforms the full method. This norm baseline (arguably the most important component) reduces to the distance between a test embedding and the median embedding of the training set. According to Table 7, the datasets contain only a small portion (<50% in most cases) of hallucinations, so the median should lie close to the mean of the correct-answer class, which dominates the sample. In other words, the method may largely measure how far an answer is from the typical correct one. This is supported by the AUC, which correlates strongly with the percentage of hallucinations across setups. Therefore, it is unclear whether the performance is a real advantage of the method or an artifact of the models, datasets, and choice of ROC-AUC as the metric. It is thus important to add a detailed ablation table covering all datasets and models, together with the average score, and to report other metrics, such as PR-AUC or PRR [1]. Additionally, since robust scaling seems to be so important, it would be useful to add further ablations: replacing ww with a random direction, an all-ones direction, or the first principal component, and using just the distance to the median as the final score.
  2. The experiment on transfer across benchmarks is questionable: it seems to require the source benchmark’s scale vector ss. If we can estimate ss from unlabelled states, we can estimate ww from the same states. I would suggest adding transfer results without any scaling. Also, following from weakness 1, I think it is important to add a transfer experiment from a dataset in which hallucinations are a minority to a dataset in which they are a majority.
  3. While the method is label-free, it still requires data. So the method still needs hundreds of in-distribution hidden states from the same model according to Table 13.
  4. The paper overlooks several important baselines: single-forward-pass methods, e.g., seq.prob [1], RAUQ [2], which are directly comparable to the proposed method, and more advanced sampling-based methods [3,4].
  5. Minor issues:
    1. I don't think it is fair to flip the sign of the score based on the orientation of ROC-AUC for each dataset. The sign should be fixed across all datasets, so please report raw AUC for all methods.
    2. Notation issues: hi,jh_{i,j} is not defined in Section 2.2.
    3. Figure 1 is quite blurry and of low quality.
    4. The abbreviation HPSA is never explained.

[1] Benchmarking Uncertainty Quantification Methods for Large Language Models with LM-Polygraph (Vashurin et al., TACL 2025)
[2] Efficient Hallucination Detection for LLMs Using Uncertainty-Aware Attention Heads. (A. Vazhentsev et al., ICML 2026)
[3] Semantic Density: Uncertainty Quantification for Large Language Models through Confidence Measurement in Semantic Space. (Qiu et al., NeurIPS 2024)
[4] Uncertainty Quantification for LLMs through Minimum Bayes Risk: Bridging Confidence and Consistency (Vashurin et al., NeurIPS 2025)

Questions for authors

See weaknesses

Reviewer_aWSkOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HPSA, a training-free hallucination detection for question answering. It reads the last-layer residual state at the end-of-turn token, robust scales each coordinate using the median and IQR of unlabelled states, and scores the answer by projecting onto the normalized mean of those scaled states. This method does not use labels, probe, extra generation or layer search. The method is evaluated on MATH, CoQA, TruthfulQA, and TriviaQA on three models.

Critical strengths

The method is a training-free and white-box detection technique, which would help explainability.

Critical weaknesses
  1. The paper is not easy to follow.
  2. The scaler ablation is degenerate. Section 2.2 says the scaler and direction are fit on the same rows, then if so, mean/SD scaling makes bar(z) equals to zero, which makes the direction numerical noise.
  3. The statistic does not identify what it detects, including which end of its axis means “hallucination”.
Questions for authors
  1. For each MATH cell, how many rows are graded by Sympy, and how many by the string fallback?
  2. What is the diagnostic split, and why does Qwen/CoQA score 0.549 there and 0.747 in table 8?
Reviewer_gidKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HPSA, a simple method for detecting hallucinations from one final-layer hidden state per answer. It standardizes each coordinate using the median and interquartile range of unlabeled calibration examples, then scores answers by projecting them onto the mean standardized vector. No classifier training or additional answer sampling is required. Across three models and four QA benchmarks, the method achieves a mean ROC-AUC of 0.729. The paper also examines calibration size, transfer across benchmarks, and sensitivity to answer length and activation scale.

Critical strengths
  1. The method is simple and inexpensive once the required hidden state is available. It uses a fixed layer and readout position, requires no labeled calibration examples, and avoids training a separate classifier.

  2. The experiments cover three models and four benchmarks, with useful analyses of calibration size and possible confounding factors. The authors also acknowledge settings where the score depends strongly on answer length or activation scale.

Critical weaknesses
  1. The main contribution is a simple readout and its empirical evaluation within an established research direction. Prior work has already used hidden states to assess truthfulness [1] and unlabeled generations to build hallucination detectors [2]. HPSA outperforms output entropy in seven of twelve settings but performs worse in five, with no significant difference in the reported paired test. These results suggest useful gains in some settings, but do not yet establish a consistent advantage over inexpensive alternatives.

  2. Two aspects of the experimental validation need clarification. On MATH, excluding references that SymPy cannot parse changes the positive rate from 23.9% to 0.2%, making independent validation of the text matching fallback particularly important. Separately, the mean-centering ablation makes the calibration mean zero, so the normalized direction in Equation (3) becomes undefined in exact arithmetic. Its performance drop therefore does not isolate the benefit of robust scaling.


[1] Amos Azaria and Tom Mitchell. 2023. The Internal State of an LLM Knows When It’s Lying. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 967–976, Singapore. Association for Computational Linguistics.

[2] Du, Xuefeng, Chaowei Xiao, and Yixuan Li. "Haloscope: Harnessing unlabeled llm generations for hallucination detection." Advances in Neural Information Processing Systems 37 (2024): 102948-102972.

Questions for authors
  1. What fraction of MATH examples use the text matching fallback, and how reliable are these labels under an independent mathematical grader or human audit? Does the advantage over entropy remain after correcting any labeling errors?

  2. How is the direction computed in the mean and standard deviation scaling ablation when the transformed calibration mean is zero? Can the authors retain median centering and change only the scaling factor to isolate its effect?

Reviewer_3BdNOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

HPSA reads the last-layer residual state at the final chat-template token & scales each coordinate by median and interquartile range, using unlabelled calibration states. It then scores each answer by projecting onto the normalised mean of the scaled states. It is valuidated using 3 models on 4 benchmarks. It is reported as the best label-free method in 4 of those 8 cells.

Critical strengths

Computational efficiency is its main strength. The results seem honest along with the limits . The length, norm and length-quartile controls are appreciated.by me. And the claims are appropriate. As a practical extra signal for serving systems that already expose hidden states, it has real utility.

Critical weaknesses

The mechanism is mislabelled and not understood. After centering on median, the first moment is the gap between mean and median, scaled by the IQR. That is a skewness direction pointing toward the heavier or more dispersed tail. It is not a “first-moment location” direction. The paper’s own geometry fits this as hallucinated answers have higher IQR . It also explains the ablations. With standard scaling the mean of the scaled data is zero by construction, so the direction is just noise. so this does not support the claim that robust scaling matters. So the detector works only when the hallucinated class happens to be the more dispersed tail. The paper doesnt test whether that holds under a shift in hallucination prevalence or on tasks where correct answers are the more variable class. The performance and baseline comparisons are also weak, compared to length-normalised entropy without efficiency gains. HIDE and HaloScope results reported are well below their published numbers, and the paper says it ported these baselines with AI help. That raises questions about either their reproducibility or the failthfulness of their reimplementation. Scope is also limited with limited number of models and datasets.

Questions for authors

Is the score equivalent to projecting onto (mean − median)/IQR? If so, please describe it that way and its implications.

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

Related papers

Discussion (0)

Sign in to comment.