acceptodds
Withdrawn from ICLR 2027

GRA: Grounding Ratio Analysis for Prompt-Anchored Update Trajectories in Hallucination Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Large language models (LLMs) can generate fluent but unsupported or factually incorrect responses, undermining their reliability in real-world deployment. Existing white-box hallucination detectors often focus on hidden-state snapshots or aggregate internal representations, leaving the generation-time update process underexplored. In this paper, we shift the focus from static hidden-state analysis to update anchoring and propose Grounding Ratio Analysis (GRA), a white-box hallucination detector that measures prompt anchoring in generation-time residual updates. For each generated token and Transformer layer, GRA measures whether a residual-stream update remains anchored to the prompt region or shifts toward the generated-prefix region, forming a token-layer grounding trajectory. We summarize this trajectory with mean-trend pooling and use a lightweight probe for response-level reliability estimation. Experiments on five open-weight LLMs and five question-answering and hallucination detection datasets show that GRA achieves the highest average AUROC among evaluated methods on every evaluated model. Further ablations show that temporal trajectory information is important, prompt-anchoring signals appear across different update sources, and raw GRA trajectory statistics are already directly separable. These results suggest that prompt-anchored update dynamics provide an effective and interpretable signal for hallucination detection.

Primary area:
interpretability and explainable AI
Keywords:
hallucination detection, large language models, residual stream, prompt grounding, white-box detection

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_dwdNOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes GRA, a white-box method that detects whether a given answer is a hallucination by checking whether the model relies more on the prompt or on the generated tokens. Specifically, each time the model generates a token, GRA measures how much of the model's internal update is similar to the prompt tokens rather than to the generated tokens, and calls this proportion the grounding ratio. It then summarizes how the grounding ratio changes over the answer into features for a simple probe to detect hallucinations.

The paper claims three main contributions.

  1. A detection framework that measures whether the model relies on the prompt or on the generated tokens during generation.
  2. It shows that the mean and trend of the grounding ratio over the answer are an effective representation for hallucination detection.
  3. An evaluation across five open-weight LLMs and five QA datasets.

For validation, GRA reports the highest average AUROC among the compared methods on all five LLMs.

Critical strengths

None.

Critical weaknesses
  1. The performance advantage of the proposed method may still come from indirectly using the length of the LLM's answer. This is a big concern, since prior work shows that a simple detector using the length of response as feature alone can match complex hallucination detectors [1]. Although the paper groups the answers by length to show that GRA does not rely on length, two observations suggest that GRA uses length:
  • The grounding ratio tends to decrease as the answer gets longer, because Eq. 3 and 4 use sums instead of averages over the prompt tokens and over the generated tokens. Lookback Lens [2] shares a similar idea with the paper but is not cited. It computes a similar ratio from attention weights, but it averages the attention weights over the prompt tokens and over the generated tokens, and thus avoids this dependence on the answer length.
  • Table 5 in the supplementary splits the test answers into three length groups (fewer than 6, 6 to 15, and more than 15 tokens) and reports the AUROC within each group. For 17 of the 25 model–dataset pairs, the AUROC on all answers in Table 1 is higher than the AUROC within any length group. For example, for Qwen3-14B on HaluEval, the AUROC is 94.01 on all answers, but only 71.17, 83.02, and 63.15 within the three groups. This suggests that GRA is much less accurate among answers of similar length.
  1. The results in Table 1 are not reliable for two reasons:
  • The test set is not held-out. The paper describes the 20% test set as held-out, but the released code get_test4.py shows it was also used to select the checkpoint and to adjust the learning rate.
  • Table 1 uses a single seed. The five-seed average of GRA in Table 2 of the supplementary differs from its result in Table 1, and in 10 of the 25 model–dataset pairs this difference is larger than the gain of GRA over ICR Probe. So it is unclear whether these gains are stable.

[1] Janiak, Denis, Jakub Binkowski, Albert Sawczyn, Bogdan Gabrys, Ravid Shwartz-Ziv, and Tomasz Jan Kajdanowicz. "The illusion of progress: Re-evaluating hallucination detection in llms." In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing.

[2] Chuang, Yung-Sung, Linlu Qiu, Cheng-Yu Hsieh, Ranjay Krishna, Yoon Kim, and James Glass. "Lookback lens: Detecting and mitigating contextual hallucinations in large language models using only attention maps." In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing.

Questions for authors
  1. How well does a classifier detect hallucinations if it uses only the prompt and answer lengths as input? Please compare it with GRA on the same test set. This needs no new computation, since get_test.py already saves the number of tokens in each answer.
  2. What are the AUROCs of GRA and ICR Probe when the checkpoint is selected on a separate validation set? Please report the mean and standard deviation over the five seeds.
Reviewer_iApiOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a hallucination detection method by measuring the ratio of how each layer's update aligns with residuals in prompt vs. in generated-prefix. The proposed method, GRA, computes this ratio for each generated token and layer, then aggregates them into a feature vector. The authors then train a linear model on that feature to predict hallucination. The authors evaluate GRA on 5 datasets with 5 LLMs and compare it against 5 baselines. The results show that GRA outperforms other baselines on most settings. Ablation shows that without the trained linear model, the feature itself already achieves non-trivial performance, suggesting that the proposed method captures hallucination-related features.

Critical strengths
  1. GRA is simple yet effective to detect hallucinated outputs.
  2. The authors conduct extensive experiments and ablations to justify the design choices.
  3. The paper is generally well written and easy to read.
Critical weaknesses
  1. The idea of GRA is similar to Lookback Lens [1], which is not discussed in the related work and baselines. I'd like to see a comparison and discussion of why GRA is better at capturing the grounding ratio.
  2. The paper evaluates GRA across response lengths but did not evaluate it across prompt lengths, whereas both are natural potential confounds. It is therefore necessary to report the performance across prompt lengths as well. It would further strengthen the claim if the authors could evaluate GRA on hallucinations in a RAG scenario (e.g., RAGTruth [2]).
  3. In Section 3.1, the paper chooses to use FFN sublayer update as Δhi(ℓ)\Delta h_i^{(\ell)} without justification. Although it later provides ablation in Section 5.2, its conclusion, "We therefore use FFN-GRA as the default instantiation of GRA, not because it is uniformly the strongest variant, but because the FFN update is a direct sublayer update written to the residual stream and provides a clear default update source for analysis," (Lines 431-358) is not convincing, as Attn is also a direct sublayer update. Since both sublayers directly update the residual, it is unclear to me why the grounding ratio can be determined by one update.
  4. Generations grounded in context don't mean it is hallucination-free, especially when the context is noisy or contains misinformation [3]. The authors should discuss this issue in the limitations. Also, it would be interesting to see how robust GRA is under knowledge conflict.

[1]: Chuang et al. 2024. Lookback Lens: Detecting and Mitigating Contextual Hallucinations in Large Language Models Using Only Attention Maps. [2]: Niu et al. 2024. RAGTruth: A Hallucination Corpus for Developing Trustworthy Retrieval-Augmented Language Models. [3]: Xu et al. 2024. Knowledge Conflicts for LLMs: A Survey.

Questions for authors
  1. Lines 366-371: What are the statistics for the short, medium, and long length data (e.g., number of samples and average length)?
  2. What dataset is used for experiments on response length and label robustness?
  3. The GRA is a supervised method. What is its OOD performance (e.g., train on one dataset and test on another)?
Reviewer_EwN2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Grounding Ratio Analysis, a white-box hallucination detection method that measures the alignment of generation-time residual updates with prompt and generated-prefix representations. The resulting token-layer trajectories are summarized using mean-trend pooling and classified with a lightweight probe. Experiments on five LLMs and five datasets show improved average AUROC over existing baselines.

Critical strengths
  1. The paper presents an interesting perspective on hallucination detection by analyzing generation-time residual updates rather than static hidden states.
  2. The evaluation covers five LLMs and five datasets, with consistent average AUROC improvements over strong baselines. The temporal pooling and update-source ablations provide additional support for the proposed features.
Critical weaknesses
  1. Insufficient justification of the grounding ratio. GRA sums absolute alignment scores over prompt and generated-prefix tokens. Consequently, the ratio may be strongly influenced by the relative lengths of these regions rather than genuine grounding. The current response-length analysis does not adequately isolate this confounding factor.
  2. Limited evidence for the proposed interpretation. High alignment with prompt representations does not necessarily imply factual correctness, especially when the prompt contains only a question without supporting evidence. Additional controlled experiments are needed to establish whether GRA captures factual grounding rather than general representation similarity.
Questions for authors
  1. How does GRA perform when the alignment scores are normalized by the number of tokens in each region?
  2. Can the authors provide controlled experiments varying prompt length and generated-prefix length while preserving the same factual content?
  3. Can the authors demonstrate that GRA captures factual grounding rather than general representation similarity, for example by comparing supported and unsupported responses under matched prompts?

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

Related papers

Discussion (0)

Sign in to comment.