acceptodds
Under review as a conference paper at ICLR 2027

Geometry-Aware Hallucination Detection in Large Language Models

Abstract

Large language models (LLMs) frequently generate factually incorrect or unsupported content, commonly referred to as hallucinations. Prior work has explored decoding strategies, retrieval augmentation, and supervised fine-tuning for hallucination detection, while recent studies show that in-context learning (ICL) can substantially influence factual reliability. However, existing ICL demonstration selection methods often rely on surface-level similarity heuristics and exhibit limited robustness across tasks and models. We propose GA-ICL, a geometry-aware demonstration sampling framework for selecting in-context demonstrations that leverages latent representations extracted from frozen LLMs. By jointly modeling local manifold structure and class-aware prototype geometry, GA-ICL selects demonstrations based on their proximity to learned prototypes rather than lexical or embedding similarity alone. Across factual verification (FEVER) and hallucination detection (HaluEval) benchmarks, GA-ICL outperforms standard ICL selection baselines in the majority of evaluated settings, with particularly strong gains on dialogue and summarization tasks. The method remains robust under temperature perturbations and model variation, indicating improved stability compared to heuristic retrieval strategies. While lexical retrieval can remain competitive in certain question-answering regimes at smaller model scales, our results demonstrate that geometry-aware prototype selection provides a reliable and training-light approach for hallucination detection without modifying LLM parameters. Extended evaluations on Phi-14B and Qwen3-32B confirm that GA-ICL scales effectively to larger models, outperforming all compared baselines including on QA tasks where smaller models show boundary-condition limitations, offering a principled direction for improved ICL demonstration selection.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Metric learning

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_WTsaOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GA-ICL, a demonstration-selection method for hallucination detection using frozen language models. The method extracts hidden representations from labeled training examples and learns a lightweight projection network together with class prototypes. Its objective combines a modified proxy-anchor loss with a local manifold-based loss. Examples nearest to the learned momentum prototypes are selected as a fixed demonstration set and reused across test queries. Experiments on FEVER and three HaluEval tasks show improvements over several ICL selection baselines in many model–task configurations, particularly dialogue and summarization. The paper also provides component ablations, dimensionality analyses, and evaluations on larger models.

Critical strengths
  1. The problem is relevant and the approach is practically motivated. Improving demonstration selection while keeping the underlying LLM frozen is a useful direction, especially when updating the full model is impractical. The paper explicitly distinguishes training a lightweight selection module from a training-free approach.
Critical weaknesses
  1. The central equation appears inconsistent with its stated objective. Equation 1 seems to push positive pairs apart and pull negative pairs together. The authors should clarify the implemented loss.
  2. The necessity of manifold modeling is insufficiently established. Prototype learning and dimensionality reduction also contribute substantially. Comparisons with simpler supervised projections under matched dimensions and training budgets are needed.
  3. The statistical protocol is unclear. The paper describes deterministic selection and decoding but reports five-seed results. The sources of randomness and the statistical reliability of smaller gains require clarification.
Questions for authors
  1. Is Equation 1 a typographical error? What is the exact loss used in the implementation?
  2. How does GA-ICL compare with simpler supervised projection methods under matched dimensions and training budgets?
  3. What changes across the five seeds for each method, and are the gains statistically significant?
4 comments
AuthorsOpenReview ↗

Response to Reviewer WTsa(except W3/Q4)

We thank the reviewer for the careful reading. Below we address the three weaknesses and the four questions together.

W1 / Q1: Eq. (1) and the exact loss used

Eq. (1) as typeset is a transcription error and not a typo in a single symbol. The signs of the exponents were wrong. The implemented loss matches the textual description: same-class embeddings are pulled toward their proxy and different-class embeddings are pushed away. The error arose when we adapted the proxy-anchor loss from its cosine-similarity form to Euclidean distance, where the signs flip. No experiment is affected, so no result changes and nothing needs to be rerun. The corrected loss is

\mathcal{L}_{PA} = {} & \frac{1}{|\Theta^{+}|} \sum_{\theta_q\in\Theta^{+}} \log\\!\Bigg( 1+\sum_{Z'\in \mathcal{Z}'^{+}_{\theta_q}} \exp\\!\big(\alpha(\lVert h_\theta(Z)-\theta_q\rVert_2-\epsilon)\big) \Bigg) \\\\ &+ \frac{1}{|\Theta|} \sum_{\theta_q\in\Theta} \log\\!\Bigg( 1+\sum_{Z'\in \mathcal{Z}'^{-}_{\theta_q}} \exp\\!\big(-\alpha(\lVert h_\theta(Z)-\theta_q\rVert_2-\epsilon)\big) \Bigg). \end{aligned}$$ This is combined with $\mathcal{L}_{manifold}$ (Eq. 2) as a sum. We have fixed Eq. (1) and re-checked the other equations in Section 2. ### W2 / Q3: Necessity of manifold modeling; comparison with simpler supervised projections at matched dimension and budget We added a compressed proxy-only ablation (Table: Compressed proxy-only vs. GA-ICL). Proxy-only is the simpler supervised projection: the same projection network and prototypes, trained with the proxy-anchor loss alone, at the same $Z'$ values as Table 10, on the same validation split, and with an identical 200-epoch budget as full GA-ICL. Full GA-ICL outperforms it at all 13 matched dimensions, by **1.8–3.3 points on Qwen3-4B QA** and **2.6–5.5 points on Mistral-7B QA**, and the gap widens as $Z'$ shrinks. Proxy-only also improves with compression (Qwen +2.4, Mistral +5.7 across the range), so part of the dimensionality gain comes from supervised compression alone, which we now state explicitly. The remainder, and a growing share at lower dimensions, is attributable to the manifold loss. On Mistral-7B QA, proxy-only is close to the random-projection control (Table 10), so most of the learned gain over random there comes from the manifold objective. We have scoped the claim in Appendix C.2 that proxy-anchor learning alone yields a substantial improvement to Qwen3-4B QA. In the random-projection control, both the projection and the prototypes are randomly initialized and then fixed, with no training. We did not run other supervised projections (e.g., LDA or PCA with a classifier head) and note this as a limitation. **Table: Compressed proxy-only ablation versus full GA-ICL at matched prototype dimensionality $Z'$ (validation accuracy, %).** Both variants use the same validation split and an identical 200-epoch training budget; proxy-only is trained with $\mathcal{L}_{PA}$ only, GA-ICL with $\mathcal{L}_{PA}+\mathcal{L}_{manifold}$. GA-ICL values are those of Table 10. $\Delta$ = GA-ICL $-$ proxy-only. | Model / Task | $Z'$ ($\times 10^3$) | Proxy-only | GA-ICL | $\Delta$ | |---|---:|---:|---:|---:| | **Qwen3-4B QA** | 2.5 | 65.2 | 67.2 | +2.0 | | | 2.3 | 64.8 | 67.5 | +2.7 | | | 2.1 | 66.3 | 68.1 | +1.8 | | | 1.9 | 66.5 | 68.7 | +2.2 | | | 1.7 | 67.1 | 69.5 | +2.4 | | | 1.5 | 67.6 | 70.9 | +3.3 | | **Mistral-7B QA** | 3.8 | 50.7 | 53.4 | +2.7 | | | 3.6 | 52.3 | 56.2 | +3.9 | | | 3.4 | 53.8 | 56.4 | +2.6 | | | 3.2 | 54.6 | 58.7 | +4.1 | | | 3.0 | 55.3 | 60.1 | +4.8 | | | 2.8 | 55.9 | 61.3 | +5.4 | | | 2.6 | 56.4 | 61.9 | +5.5 |
AuthorsOpenReview ↗

Response to Reviewer WTsa(W3/Q3)

W3 / Q3: Statistical protocol; what changes across the five seeds

We thank the reviewer for pressing on this, and we have corrected the manuscript. Given a demonstration order, selection and decoding are deterministic: rerunning a fixed configuration reproduces accuracy exactly.

The five "seeds" in Table 3 shuffle the order of the two selected demonstrations (positive/negative versus negative/positive) for every method. For GA-ICL and AdaICL they also vary the training seed. The reported ±\pm therefore reflects sensitivity to demonstration order and, where applicable, training seed, and not run-to-run noise.

With two demonstrations there are only two orders, so we now report both explicitly. Order sensitivity is large and method-dependent (Table: Demonstration-order sensitivity): on Qwen3.8-27B, BM25 on QA moves by 6.2 points between orders, while GA-ICL moves by 0.3; on Dialogue the swings are 1.0 (BM25) and 1.0 (GA-ICL).

GA-ICL's worst order exceeds BM25's best order in both tasks (81.3 vs. 76.8 on Dialogue; 89.5 vs. 86.0 on QA).

Table: Demonstration-order sensitivity on Qwen3.8-27B (single run per order, accuracy %). Order 1 and Order 2 are the two possible orders of the two selected demonstrations.

TaskMethodOrder 1Order 2Swing
DialogueBM2575.876.81.0
GA-ICL81.382.31.0
QABM2579.886.06.2
GA-ICL89.889.50.3
Reviewer_snq7OpenReview ↗

A note for Reviewer WTsa and the AC.

This reply says the five seeds change the demonstration order for every method and also the training seed for GA-ICL and AdaICL. The reply to me says the seeds only change the order, with no mention of training seeds. The order table here also doesn't match Table X in the reply to Reviewer rMup: there, BM25 is shown at its worse order and GA-ICL at its better one in every shared cell, even though the text says all methods use the same order. Since the question was about whether the reported numbers match the described setup, I think these differences are worth a closer look.

Reviewer_WTsaOpenReview ↗

Discussion about the experiment results

Seconding Reviewer snq717, this is the same thing I asked in W3. On QA, Order 1 gives GA-ICL a 10.0-point lead (89.8 vs 79.8); Order 2 gives 3.5 (89.5 vs 86.0). So which is in Table 3? If the table in the other thread puts BM25 at its worse order and GA-ICL at its better one, the reported margin isn't a same-order comparison at all. That's a reporting question, not a statistical one.

Two replies of authors also disagree on whether the five seeds include training seeds. With only two orders and deterministic decoding, that's not a detail, it decides what the ±SD means. I would like to know whether authors could provide a detailed discussion.

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

This paper proposes GA-ICL, a demonstration-selection method for hallucination detection. It learns a projection and class prototypes from frozen LLM representations using a proxy loss and a local manifold loss, then selects a fixed set of demonstrations for reuse across queries. Experiments on HaluEval and a binary version of FEVER compare the method with several retrieval strategies across different model families and sizes.

Critical strengths
  1. Selecting a small, reusable set of demonstrations is a useful approach to reducing repeated retrieval during inference. The method keeps the LLM frozen and trains only the retrieval module.

  2. The experiments cover multiple model families and include random selection and additional large-model baselines. Table 6 also shows a 2.1-2.7 percentage-point improvement from adding the manifold loss at full dimensionality, providing some evidence for this component.

Critical weaknesses
  1. The proxy loss in Eq. (1) is inconsistent with its description. Minimizing the positive term exp[-α(d−ε)] encourages larger same-class distances, while minimizing the negative term exp[α(d−ε)] encourages smaller different-class distances. This is opposite to the stated attraction and repulsion and to the nearest-prototype selection rationale. It is unclear whether this is a formula error or reflects the implemented objective, which is important for understanding the reported results.

  2. The contribution of the manifold loss in the final compressed model remains unclear. Table 6 shows gains at full dimensionality, but the larger gains follow dimensionality reduction, and a compressed proxy-only model is not included. The random-projection comparison in Appendix C.6 does not clearly state whether the randomly initialized prototypes are subsequently trained. A comparison at the same dimensionality and training budget would help determine whether the final gains require the manifold loss or mainly come from supervised dimensionality reduction.

Questions for authors
  1. Which signs are used in the implemented proxy loss? Would correcting Eq. (1) require rerunning the experiments?
  2. Are the prototypes in Table 10 optimized after initialization? If so, which objective is used?
  3. HaloScope [1] uses a latent-subspace score to learn a truthfulness classifier from unlabeled generations. Although its supervision setting differs from this paper's, it motivates a design question: with the same labeled data and frozen LLM features, how does GA-ICL compare with a lightweight classifier that predicts factuality directly?

[1] Du X, Xiao C, Li Y. HaloScope: Harnessing Unlabeled LLM Generations for Hallucination Detection[C]//Advances in Neural Information Processing Systems. 2024, 37: 102948-102972.

3 comments
AuthorsOpenReview ↗

Response to Reviewer YzVu(except Q3)

We thank the reviewer for the careful reading. Below we address the two weaknesses and the three questions together.

W1 / Q1: Sign of the proxy-anchor loss in Eq. (1)

The reviewer is right that Eq. (1) as typeset is inconsistent with its description. This is a typesetting error in the manuscript only. The implemented loss matches the textual description: same-class embeddings are pulled toward their proxy and different-class embeddings are pushed away. The error arose when we adapted the proxy-anchor loss from its cosine-similarity form to Euclidean distance, where the signs of the exponents flip, and we transcribed them incorrectly. No experiment is affected, so no result changes and nothing needs to be rerun. The corrected equation is

\mathcal{L}_{PA} = {} & \frac{1}{|\Theta^{+}|} \sum_{\theta_q \in \Theta^{+}} \log\Bigg( 1+\sum_{Z'\in \mathcal{Z}'^{+}_{\theta_q}} \exp\Big(\alpha(\lVert h_\theta(Z)-\theta_q\rVert_2-\epsilon)\Big) \Bigg)\\ &+ \frac{1}{|\Theta|} \sum_{\theta_q\in\Theta} \log\Bigg( 1+\sum_{Z'\in \mathcal{Z}'^{-}_{\theta_q}} \exp\Big(-\alpha(\lVert h_\theta(Z)-\theta_q\rVert_2-\epsilon)\Big) \Bigg). \end{aligned}$$ Minimizing the first term decreases positive distances, and minimizing the second increases negative distances. ### W2: Contribution of the manifold loss in the compressed model To isolate the manifold loss in the final compressed model, we added a compressed proxy-only ablation (Table: Proxy-only vs. GA-ICL). The projection and prototypes are trained with the proxy-anchor loss alone, at the same $Z'$ values as the random-projection control (Table 10), on the same validation split and with an identical 200-epoch training budget as full GA-ICL. Full GA-ICL outperforms proxy-only at all 13 matched dimensions, by **1.8–3.3 points on Qwen3-4B QA** and **2.6–5.5 points on Mistral-7B QA**, and the gap widens as $Z'$ shrinks. Proxy-only also improves with compression (Qwen +2.4, Mistral +5.7 across the range), so part of the dimensionality gain comes from supervised compression alone, which we now state explicitly. The remainder, and a growing share at lower dimensions, is attributable to the manifold loss. On Mistral-7B QA, proxy-only is close to the random-projection control (Table 10), so most of the learned gain over random there comes from the manifold objective. We have accordingly scoped the claim in Appendix C.2 that proxy-anchor learning alone yields a substantial improvement to Qwen3-4B QA. **Table: Compressed proxy-only ablation versus full GA-ICL at matched prototype dimensionality $Z'$ (validation accuracy, %).** Both variants use the same validation split and an identical 200-epoch training budget; proxy-only is trained with $\mathcal{L}_{PA}$ only, GA-ICL with $\mathcal{L}_{PA}+\mathcal{L}_{manifold}$. GA-ICL values are those of Table 10. $\Delta$ = GA-ICL $-$ proxy-only. | Model / Task | $Z'$ ($\times 10^3$) | Proxy-only | GA-ICL | $\Delta$ | |---|---:|---:|---:|---:| | **Qwen3-4B QA** | 2.5 | 65.2 | 67.2 | +2.0 | | | 2.3 | 64.8 | 67.5 | +2.7 | | | 2.1 | 66.3 | 68.1 | +1.8 | | | 1.9 | 66.5 | 68.7 | +2.2 | | | 1.7 | 67.1 | 69.5 | +2.4 | | | 1.5 | 67.6 | 70.9 | +3.3 | | **Mistral-7B QA** | 3.8 | 50.7 | 53.4 | +2.7 | | | 3.6 | 52.3 | 56.2 | +3.9 | | | 3.4 | 53.8 | 56.4 | +2.6 | | | 3.2 | 54.6 | 58.7 | +4.1 | | | 3.0 | 55.3 | 60.1 | +4.8 | | | 2.8 | 55.9 | 61.3 | +5.4 | | | 2.6 | 56.4 | 61.9 | +5.5 | ### Q2: Are the prototypes in the random-projection control (Table 10) optimized after initialization? **No.** In the random-projection control, both the projection and the prototypes are randomly initialized and then kept fixed; neither is trained and no objective is applied after initialization. We have clarified this in Appendix C.6. The control tests whether a generic reduction to the same dimensionality reproduces the gains; it does not (Table 10). The matched proxy-only comparison above (Table: Proxy-only vs. GA-ICL) separates the contribution of supervised compression from that of the manifold loss.
AuthorsOpenReview ↗

Response to Reviewer YzVu(Q3)

Q3: Comparison with a lightweight classifier on frozen features (HaloScope-style)

We thank the reviewer for the pointer to HaloScope. We did not include a classifier baseline because it addresses a different problem. A classifier trains on the full labeled set and needs white-box hidden-state access at every inference call. GA-ICL targets prompt-based detection with a frozen LLM: hidden states are used only offline to choose two demonstrations, and inference is a single forward pass with a fixed prompt and no additional trained classifier.

Because the two differ in supervision, access, and deployment, comparing accuracy would not tell us whether GA-ICL selects demonstrations well, which is the question the paper studies. The relevant comparisons are therefore among ICL selection methods (Table 3).

Our one supervised reference is LoRA fine-tuning of the LLM itself on Vicuna-7B (Table 5), where GA-ICL matches accuracy at orders-of-magnitude lower memory. We have clarified in Section 1 that this comparison is to LoRA-based fine-tuning, and we cite HaloScope in the related work.

Thank you for clarifying the equation and adding the matched compression ablation. This addresses part of my concern about the manifold objective. The ablation reports validation accuracy, so I view it as useful but limited evidence for the contribution of the manifold loss to test performance.

The two-order results are helpful and show that GA-ICL remains ahead of BM25 under both orders. However, the Dialogue values in Table X, 75.8 for BM25 and 82.3 for GA-ICL, correspond to different orders in the order-specific table. The explanations of how demonstration order and training seeds contribute to the reported variation are also not consistent across responses. These reporting issues limit my confidence in the comparisons. Based on the available evidence, I remain at weak rejection.

Reviewer_snq7OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes GA-ICL, a way to choose in-context demonstrations for hallucination detection. Hidden states from a frozen LLM go through a small projection network, which is trained with a proxy anchor loss and a local manifold loss. The two training examples closest to a learned prototype are then used as fixed 2-shot demonstrations for all test inputs. Experiments on FEVER and HaluEval with several LLMs report higher accuracy than retrieval baselines such as BM25, KNN, perplexity and SA-ICL.

Critical strengths

None.

Critical weaknesses
  1. The method does not match its motivation, and the gains have a simpler explanation. The paper argues that similarity-based retrieval is a poor fit for factual consistency. But GA-ICL does not retrieve per query. It picks one fixed pair of demonstrations for the whole test set. All hyperparameters are chosen on validation accuracy, and each setting yields a different pair. Tuning them is therefore close to trying many prompts and keeping the best one. Table 9 shows that random 2-shot prompts vary by 4 to 8 points, so this kind of selection alone could explain the gains. The paper also already uses labels and hidden states, so a simple classifier on those hidden states is the natural detector. There is no comparison to either baseline.

  2. The experimental reporting is not reliable. Decoding is at temperature 0 and the test split is fixed, so BM25, KNN and perplexity selection should be deterministic. Yet Table 3 gives them standard deviations of 0.7 to 1.7 over 5 seeds. Many results are at chance level, around 50% on a balanced binary task, and BM25 on Vicuna summarization is 26.5%, far below chance. This suggests that unparsed outputs are being counted as wrong. I am not alleging fabrication. But these numbers cannot be what the described setup produces, and as reported I cannot trust the tables.

Questions for authors
  1. What do the 5 seeds change for deterministic methods like BM25 and KNN?

  2. How many configurations were tried on the validation set? How does GA-ICL compare with picking the best of the same number of random 2-shot pairs?

  3. How are outputs that are neither yes nor no counted, and how often does this happen?

  4. How does GA-ICL compare with a linear probe trained on the same hidden states and labels?

2 comments
AuthorsOpenReview ↗

Response to Reviewer snq7

We thank the reviewer for the careful reading, and in particular for raising the question about the reported standard deviations. We answer that first, since it concerns the integrity of the reported numbers, and then the remaining points.

Standard deviations of deterministic methods (Q1 and the ethics flag)

The reviewer is right that, for a fixed configuration, BM25, KNN, and perplexity selection and temperature-0 decoding are deterministic, and that the manuscript did not explain what the five "seeds" vary. The explanation is as follows. Given a demonstration order, selection and decoding are deterministic: rerunning a fixed configuration, including repeated runs of the same order, reproduces accuracy exactly. We verified this directly.

The five "seeds" in Table 3 shuffle the order of the two selected demonstrations (positive/negative versus negative/positive) for every method. The reported ±\pm therefore measures sensitivity to demonstration order, not run-to-run noise.

With two demonstrations there are only two possible orders, so five seeds resample two orders. We now report both orders explicitly. Order sensitivity is large and method-dependent (Table: Demonstration-order sensitivity): on Qwen3.8-27B, BM25 on QA changes by 6.2 points between orders, while GA-ICL changes by 0.3; on Dialogue the swings are 1.0 (BM25) and 1.0 (GA-ICL).

We also withdraw the claims, in Section 3 and Appendix C.5, that GA-ICL is "more stable" than baselines on the basis of standard deviations, because those quantities arise from different sources for different methods.

Table: Demonstration-order sensitivity on Qwen3.8-27B (single run per order, accuracy %). Order 1 and Order 2 are the two possible orders of the two selected demonstrations.

TaskMethodOrder 1Order 2Swing
DialogueBM2575.876.81.0
GA-ICL81.382.31.0
QABM2579.886.06.2
GA-ICL89.889.50.3

W1: Do the gains have a simpler explanation (fixed pair plus validation-based selection)?

We agree that GA-ICL selects a fixed pair for the whole test set, which differs from per-query retrieval, and that our motivation text overstated the contrast. We have revised the introduction to say that GA-ICL learns a retrieval geometry used to choose a small set of demonstrations once per dataset–model pair, and that the claim concerns the quality of that selection and not per-query adaptivity.

We also agree that tuning on validation accuracy can resemble searching over prompts, and that the random baseline in Table 9 (std 3.5–7.8) shows that individual 2-shot prompts vary substantially. The following evidence bears on this, and we are explicit about what it does not show:

  • The gains are measured on a held-out test split that was used once per configuration, after all choices were fixed, and not on the validation split used for tuning.
  • GA-ICL's worst order exceeds BM25's best order on Qwen3.8-27B in both tasks (Table: Demonstration-order sensitivity), so the gains are not an artifact of a favorable order.

W1 / Q4: Comparison with a linear probe on the same hidden states

We did not include a probe, because it addresses a different problem. A probe trains a separate classifier on the full labeled set and needs white-box hidden-state access at every inference call. GA-ICL targets prompt-based detection with a frozen LLM: hidden states are used only offline to choose two demonstrations, and inference is a single forward pass with a fixed prompt and no additional trained classifier.

Because the two differ in supervision, access, and deployment, an accuracy comparison would not tell us whether GA-ICL selects demonstrations well, which is the question the paper studies. The relevant comparisons are among ICL selection methods (Table 3).

Our one supervised reference is LoRA fine-tuning of the LLM itself on Vicuna-7B (Table 5). We have clarified in Section 1 that this comparison is to LoRA-based fine-tuning and is limited to one model.

Thanks for the reply. It does help to know that the five seeds change the order of the two examples, and I appreciate that you dropped the claim that GA-ICL is more stable. But most of my concerns are still there, and the new numbers raise a new one.

First, the seed explanation isn't the same in every reply. To me, you said the seeds only change the order, for every method. To Reviewer WTsa, you said they also change the training seed for GA-ICL and AdaICL. It's also still unclear why the standard deviations drop to 0.1 to 0.4 for the larger models. Your own new table shows that just swapping the order moves BM25 by 6.2 points on QA, so I'd expect much bigger standard deviations there.

Second, the new results don't match how you describe them. Under Table X, you say every method uses the same order. But if I compare Table X with your order table, BM25 on Dialogue is shown with order 1 (75.8) while GA-ICL is shown with order 2 (82.3). In all four cells, BM25 gets its worse order and GA-ICL gets its better one. On QA this matters a lot: with the same order for both, the gap is 3.5 points (89.5 vs. 86.0), not 10 points (89.8 vs. 79.8).

Third, some of my questions weren't really answered. I asked how many settings you tried on the validation set, and how GA-ICL compares with simply trying the same number of random example pairs and keeping the best one. Saying the test set was used only once doesn't answer this. My question about how answers that are neither yes nor no are counted, which matters for the 26.5% on Vicuna, wasn't answered at all. On the linear probe, GA-ICL already needs the hidden states of the same model, and getting them during the normal forward pass costs almost nothing. You also compare against LoRA fine-tuning, which changes far more than a probe does. So I don't think it's a different problem.

I'm not saying anything was done on purpose. But since my original concern was that the reported numbers don't match the described setup, and the new table has the same kind of mismatch, I'll keep my score at 1 and keep the ethics flag. I'd ask the AC to look at the Table X order issue as well.

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

This paper proposes GA-ICL, a geometry-aware demonstration-selection framework for hallucination detection with in-context learning. GA-ICL extracts latent representations from a frozen LLM and learns a lightweight projection network together with class-aware prototypes. Its training objective combines a modified proxy-anchor loss with a manifold point-to-point loss based on local PCA approximations. Demonstrations are then selected according to their proximity to a learned momentum prototype, without updating the underlying LLM. The method is evaluated on FEVER and the HaluEval dialogue, summarization, and question-answering tasks. The main experiments cover six models, with additional larger-model experiments on Phi-14B and Qwen3-32B. Comparisons include lexical, perplexity-based, clustering, and adaptive ICL selection baselines. The paper also reports temperature, perplexity, shot-count, dimensionality, loss-component, and efficiency analyses. GA-ICL achieves the best accuracy in many configurations, although lexical retrieval remains competitive or superior in some smaller-model QA settings.

Critical strengths
  1. GA-ICL keeps the underlying LLM frozen and introduces only a lightweight learned selection module. This training-light setting is potentially useful when full model fine-tuning is undesirable.
  2. In addition to the main accuracy comparisons, the paper examines temperature sensitivity, demonstration count, prototype dimensionality, loss components, and inference efficiency. These analyses help characterize the method’s behavior and its boundary conditions, particularly the competitiveness of BM25 on evidence-grounded QA.
Critical weaknesses
  1. The evaluation does not cover recent state-of-the-art models. The main suite consists primarily of small or mid-scale models, while the larger-model study is limited to Phi-14B and Qwen3-32B. Both latent representation geometry and in-context-learning behavior can change substantially across model generations. Thus, the results establish effectiveness for the selected configurations, but do not yet provide convincing evidence that the advantages persist on current state-of-the-art open-weight architectures (e.g., recent Qwen 3.8, GLM 5.3 series, etc.). A stronger evaluation should include several recent SOTA models across different families and scales.
  2. The end-to-end method is not presented clearly enough. GA-ICL contains several interacting stages, including hidden-state extraction, projection, local manifold construction, optimization of two losses, momentum-prototype updates, and nearest-prototype demonstration sampling. A workflow diagram would substantially improve readability. It should distinguish the one-time training procedure from test-time inference, indicate which representations and prototypes are learned or reused, and show how selected demonstrations enter the final prompt.
Questions for authors
  1. Can you evaluate GA-ICL on several recent state-of-the-art open-weight LLMs?
  2. Can you clarify which stages are performed once per dataset–model pair and which are performed for each test instance?
2 comments
AuthorsOpenReview ↗

Rebuttal response

We thank the reviewer for the careful reading and for both requests.

Recent models. To test whether the gains persist on newer open-weight models, we evaluated Qwen3.5-9B and Qwen3.8-27B on HaluEval Dialogue and QA (and Summarization for Qwen3.5-9B) against BM25, KNN, AdaICL, and Top-k+Con-E (Table X). GA-ICL is the best method in every cell, with margins of +3.5 to +4.2 points on Qwen3.5-9B and +4.0 (Dialogue) and +8.6 (QA) on Qwen3.8-27B. The QA setting where BM25 was competitive at small scale (Section 4) no longer favors it at these scales. These are single-run results, using the same demonstration order for every method. All new models are from the Qwen family. We could not extend to other families, because the recent releases we considered are very large or exceed our compute, and the GLM series starts above 100B parameters. GA-ICL needs access to hidden states, so API-only models are also not an option. We state this limitation explicitly in Section 4.

Which stages run once and which run per test instance. Everything that involves learning happens once per dataset-model pair, before any test input is seen:

Extract final-layer, mask-weighted mean hidden states from the frozen LLM for the training split. Train the projection network and class prototypes with the proxy-anchor and manifold losses (about 3–4 hours and 640 MB per pair), including the local manifold construction and the momentum-prototype updates. Select the two demonstrations closest to the learned momentum prototypes (one per class).

At test time nothing is learned or retrieved. The same two demonstrations are inserted into the fixed prompt template (Appendix A) for every test input, so inference cost is one LLM forward pass per query, with no per-query hidden-state extraction, projection, or nearest-neighbor search. This is why the inference times in Appendix E are lower than for per-query retrieval baselines. We will add this description to Section 2.

Workflow diagram. We agree a figure would help. We will include one in the camera-ready version, separating the one-time training procedure from test-time inference, and marking which components are learned and which are reused.

Table(X): Accuracy (%) of demonstration selection methods on recent open-weight models (single run). GA-ICL achieves the best accuracy across all evaluated settings.

MethodQwen3.5-9B DialogueQwen3.5-9B QAQwen3.5-9B Summ.Qwen3.8-27B DialogueQwen3.8-27B QA
BM2571.182.865.775.879.8
KNN69.381.466.376.477.6
AdaICL71.680.265.078.381.2
Top-k+Con-E72.082.664.577.880.7
GA-ICL (Ours)76.187.069.882.389.8

Note: Qwen3.8-27B Summarization was not evaluated.

Reviewer_snq7OpenReview ↗

A note for Reviewer rMup and the AC about Table X.

The text says every method uses the same demonstration order. But the order table the authors posted in reply to Reviewer WTsa and me shows that this isn't the case for Qwen3.8-27B. On Dialogue, BM25 is reported with order 1 (75.8) and GA-ICL with order 2 (82.3). In all four cells that appear in both tables, BM25 is shown at its worse order and GA-ICL at its better one. With the same order for both methods, the QA gap over BM25 is 3.5 points (89.5 vs. 86.0) rather than 10 points (89.8 vs. 79.8). I'm not suggesting this was intentional, but the new-model margins may be smaller than Table X shows.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Final comments

For our submission we wanted the AC to have an accurate account of the reviewer discussion before we do.

Reviewer snq7 flagged an integrity concern about the standard deviations in our tables, and later noted inconsistencies between our replies to different reviewers. We want to be direct about what happened.

Standard deviations. For deterministic methods, the five “seeds” in Table 3 changed the order of the two selected demonstrations. Repeated runs with the same order give identical accuracy. We should have said this in the paper. Inconsistent replies. Our replies described the seed setting inconsistently, and the Qwen3.8-27B table in our reply to Reviewer rMup mixed demonstration orders: BM25 appears at its lower-accuracy order and GA-ICL Dialogue at its higher-accuracy order. This was an assembly error, and the entries are real runs. At matched orders, GA-ICL’s margin over BM25 is +5.5 (Dialogue, both orders) and +10.0 / +3.5 (QA, orders 1 / 2).

Verification. The supplementary material are also available. We would welcome the AC or reviewers checking the reported numbers against them. We did not see this offer taken up in the discussion, and we raise it only so the record shows the materials exist.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.