acceptodds
Under review as a conference paper at ICLR 2027

From Routing Signals to Selective Review: Visual regrounding in MoE VLMs

Abstract

Vision-language models (VLMs) may accept false visual premises, answering questions about a target object's color, count, location, or state even when it is absent. We call this reliability-critical behavior a target-absence grounding failure. Existing visual-grounding detectors primarily rely on generated responses, hidden states, or uncertainty measures. We present the first framework to leverage internal routing decisions in Mixture-of-Experts (MoE) VLMs to detect target absence before generation and guide selective correction. We extract target-token routing probabilities from Qwen3-VL-30B-A3B-Instruct and Gemma-4-26B-A4B-it, train a separate L2-regularized linear detector for each model, and use its predictions to selectively invoke a target-aware review prompt. Using routing alone, the Qwen and Gemma detectors achieve ROC-AUCs of 0.9988 and 0.9956 on GQA-Inpaint and retain 0.8095 and 0.7781 on the external OBER dataset, respectively. The resulting routing-gated policy improves end-to-end accuracy on GQA-Inpaint and OBER by +22.25% and +12.17% for Qwen, and by +13.42% and +1.39% for Gemma, without modifying model weights. Further analysis shows that the signal is localized to the target-object token, emerges in early MoE layers, and is distributed across partially substitutable experts. Although cross-dataset threshold shifts require recalibration, false-positive review causes limited harm overall, suggesting that intervention risk can be controlled through joint selection of the detector threshold and review prompt. Overall, we show that routing probabilities alone preserve actionable information about visual perception, allowing computation already produced by an MoE VLM to support low-cost detection and selective visual regrounding.

Primary area:
interpretability and explainable AI
Keywords:
Vision-Language Models (VLMs), Mixture-of-Experts (MoE), Expert Routing, Visual Grounding, Hallucination Detection

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_LDH4OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces a detect-then-review framework designed to mitigate target-absence grounding failures in Mixture-of-Experts (MoE) Vision-Language Models (VLMs). Rather than relying on generated responses or external verifiers to detect hallucinations, the authors extract target-token routing probabilities during the standard prompt prefill to predict if a queried object is missing from the image. An L2-regularized linear detector scores these routing features, and if target absence is predicted, the system selectively triggers a target-aware review prompt to correct the premise. Evaluated on Qwen3-VL-30B-A3B-Instruct and Gemma-4-26B-A4B-it, the routing-gated policy significantly improves end-to-end accuracy on the GQA-Inpaint dataset (+22.25% for Qwen, +13.42% for Gemma) without altering model weights or router parameters.

Critical strengths
  1. A useful connection between detection and intervention. The framework goes beyond reporting detector AUROC by connecting a pre-generation signal to a concrete corrective action. It requires neither backbone retraining nor an auxiliary visual encoder.
  2. Strong In-Domain Performance. The routing probes achieve AUROCs of 0.9988 and 0.9956 on GQA-Inpaint. Scene-disjoint splits and independent processing of paired images strengthen the detector evaluation, while OBER provides an informative test of distribution shift.
  3. Informative controls for the selective-review policy, particularly for Qwen. Always-review and generic-review comparisons, together with results stratified by true target presence (Tables 2, 11, and 12), help distinguish the contributions of gating and prompting. The paper also measures correct-to-incorrect changes following false-positive triggers (Sec. 4.2.4, Fig. 5, and Table 13) and reports pair-clustered bootstrap confidence intervals for the selective-versus-always accuracy differences, although corresponding intervals are not provided for the headline gains over standard prompting or the detector AUROCs. Together, these analyses help readers weigh correction benefits against intervention-induced harm.
Critical weaknesses
  1. Routing is not shown to beat hidden states. Appendix Table 8 reports single-layer hidden-state probes reaching 0.9987 AUROC, compared with 0.9988 for the full routing representation; one also achieves higher thresholded accuracy, 98.69% versus 98.54%. These results establish that routing contains predictive information, but provide limited evidence that it is a preferable monitoring signal. A validation-selected hidden-state probe is not compared against routing on OBER or within the complete selective-review policy. Such comparisons are particularly important because routing is the paper’s central methodological distinction.
  2. The gate adds little beyond always-review. Selective vs. always target-aware review is 89.00% vs 86.25% on GQA (+2.75, [1.17, 4.58]) and 95.57% vs 95.30% on OBER (+0.26, [0, 0.70]), so ~11.9 of the +12.17 OBER points come from the prompt alone. On OBER the frozen threshold triggers review on 83.91% of inputs at 31.30% specificity (Table 7): the policy is nearly always-review. Gemma's always-review accuracy is not reported.
  3. Cost Analysis. Avoiding review calls does not by itself establish lower total inference cost. The selective policy requires a standard-prompt prefill and an additional prefill when triggered, whereas directly applying the target-aware prompt to every input could avoid the initial standard prefill. End-to-end latency or compute measurements are needed to substantiate the efficiency claim.
  4. Some interpretive claims exceed the supporting analyses. The abstract describes the signal as localized to the target-object token and distributed across partially substitutable experts. However, selecting target-token features does not establish localization without comparisons to other token positions. Likewise, coefficient distributions and top-k feature subsets demonstrate predictive redundancy more directly than expert substitutability. These claims should be narrowed or supported with additional controls.
Questions for authors

Please address the concerns raised in the Weaknesses section. In particular, please clarify the comparative advantages of routing over hidden-state probes, the incremental accuracy and computational benefits of selective review over always-review, and the empirical support for target-token localization and expert substitutability. I am especially interested in the practical latency implications. Although the monitor uses prefill signals, could the authors demonstrate an online implementation in which the target-absence flag is available and acted upon before any answer token is emitted? It would be helpful to report the time required to obtain the flag and compare total response latency, including review when triggered, with standard generation and direct target-aware prompting.

2 comments

Thank you for the careful review and for recognizing the value of pre-generation detection. We have added the requested comparisons and timing measurements to clarify the choice of routing, the gate's benefit, and how the signal is used online.

W1

We added the requested OBER comparison with hidden-state probes selected on GQA validation. Each policy uses its detector's fixed threshold, and comparisons are made within the same model and generation batch.

Model / batchRouting AUCSelected hidden AUCRouting policy accuracyHidden policy accuracy
Qwen original OBER0.80950.6382, Layer 395.57%85.74%
Gemma supplementary OBER0.78020.9055, Layer 1693.83%94.35%

Routing performs better for Qwen, and hidden probing performs better for Gemma. To test input selection at equal coverage, we also fixed both Qwen review rates at 83.91%. Routing then reaches 95.57% accuracy and hidden gating 90.61%.

Routing also supports layer-expert analysis. The Top-500 experiment in W4 tests whether a selected feature subset retains detection performance. This makes the detector's reliance on routing features experimentally testable.

W2

Thank you for pointing out the missing Gemma always-review comparison. We evaluated all Gemma policies on the same supplementary batch of OBER answers.

Gemma supplementary OBERCorrect / TotalAccuracy
Standard1068/115092.87%
Always target-aware1082/115094.09%
Routing selective1079/115093.83%
Hidden selective1085/115094.35%

These policies are close in accuracy, with always-review slightly ahead of routing-selective review. For Qwen, the following table separates the prompt's contribution from the gate's additional gain by using the same target-aware prompt.

Qwen original cohortStandardAlwaysSelectivePrompt gain, ppGate gain, pp [95% CI]Selective review rate
GQA, 120 pairs66.75%86.25%89.00%+19.50+2.75 [1.17, 4.58]50.42%
OBER, 115 pairs83.39%95.30%95.57%+11.91+0.26 [0, 0.70]83.91%

The gate's additional benefit is clearest on GQA. On present inputs, selective review reaches 95.50% accuracy compared with 89.33% for always-review. By applying review to 50.42% of GQA inputs, the gate preserves more correct answers and achieves higher overall accuracy.

W3 / Q1

We measured both the time to obtain the flag and total response time on 200 images per model under deterministic generation. Total latency includes the original-question prefill, routing extraction, probe scoring, and answer generation.

ModelPre-answer branchFlag mean / P90, sStandard mean, sAlways mean, sSelective mean, s
Qwen200/2001.260 / 1.3511.7972.2333.169
Gemma200/2000.187 / 0.1910.5700.5070.681

The current implementation runs detection and generation separately and repeats prefill without KV-cache reuse. Selective review is therefore slower in this implementation. We will correct the cost claim and report the measured latency alongside the accuracy benefit described in W2.

For the online-decision part of your question, the implementation computes the routing score and selects the standard or target-aware prompt before answer generation. Runtime checks confirm that branch selection occurs before the first answer-generation forward pass completes on all 200 requests for each model. This establishes that the routing signal can actively determine the response path before an answer is emitted.

W4

Thank you for asking us to distinguish these interpretations more carefully. We compared token positions on the same 3,000 GQA test inputs, with equal feature dimensions and a separately trained probe at each position.

Token positionQwen AUCGemma AUC
Target span0.99770.9920
Non-target content0.99960.9987
End of question0.99950.9985

Detection is strong at the target token and at other question positions. We will replace “localized to the target-object token” with “detectable from routing at the target-object token.” The layer-wise probes also show that the signal is available in early layers.

For expert analysis, retraining on the top 500 routing features retains 0.9984 GQA AUC compared with 0.9988 for all features. We will replace “partially substitutable experts” with the finding that a subset of routing features retains nearly the full probe's predictive performance. These analyses establish that target absence is available in early routing signals and can be detected using a compact feature subset.

Thank you for the detailed response and additional experiments. Most of my concerns have been addressed. However, the advantage of routing over hidden-state probing remains model-dependent, and selective review offers limited gains over always-review at a higher inference cost. I will therefore maintain my original rating.

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

This paper investigates target-absence grounding failures in Mixture-of-Experts (MoE) Vision-Language Models (VLMs), where the object mentioned in the query is absent from the image, yet the model follows the false premise and hallucinates its attributes. The proposed method consists of three steps: 1) During the prefill stage, it extracts the full router softmax probabilities for the target noun token across all layers and experts, yielding a 6,144-dimensional feature vector (48×12848 \times 128) on Qwen3-VL-30B-A3B and a 3,840-dimensional vector on Gemma-4-26B-A4B; 2) It trains an L2-regularized linear probe on these routing probabilities to predict whether the target object is absent; and 3) When the probe detects target absence, the model switches to a target-aware prompt that instructs it to first verify the object's existence before regenerating the answer. Regarding the main results, the detection AUC reaches 0.9988 for Qwen and 0.9956 for Gemma on GQA-Inpaint, while dropping to 0.81 and 0.78, respectively, on the out-of-distribution OBER dataset. In terms of end-to-end accuracy, Qwen achieves improvements of 22.25 and 12.17 percentage points on GQA and OBER, respectively, whereas Gemma improves by 13.42 and 1.39 percentage points. Further analysis indicates that this grounding signal becomes decodable as early as Layer 2 and is distributed across a large number of experts.

Critical strengths

Practical problem formulation and clear design: The paper cleanly decouples when to intervene (detection) from how to intervene (review prompting). To the best of my knowledge, reading out router probabilities on the target token in MoE VLMs specifically to detect target absence is a novel combination that has not been explored in precisely this form before.

Rigorous experimental protocol and measured claims: Standardization statistics, regularization hyperparameters, and decision thresholds are selected strictly using the training and validation splits, with scene-level partitioning to prevent data leakage. The authors include a shuffled-label control and transparently report the anomalous 0.4323 value observed in the initial control run. Furthermore, the paper proactively clarifies the boundaries of its conclusions across multiple sections, for instance, explicitly noting that probe weights do not imply causality, five sampling runs do not constitute independent replications, and the confidence interval for differences on OBER includes zero.

Transparent reporting of limitations and trade-offs: The authors carefully distinguish between the transferability of AUC rankings and the non-transferability of fixed thresholds in Table 9. They also candidly report the performance costs incurred by false triggers in Figure 5 and Table 13, rather than selectively reporting only the gains.

Critical weaknesses

Confounding in the detection task: the probe may be detecting image editing artifacts rather than target absence. GQA-Inpaint originates from Inst-Inpaint (2023), where all target-absent images are synthesized via inpainting while all target-present images are unedited originals; consequently, the ground-truth labels are perfectly collinear with editing artifacts. Two pieces of evidence point to this risk: (1) In Table 8, a probe on Layer 0 hidden states already achieves 0.9948 AUC, reaching near-perfect separability after just a single attention layer; and (2) when transferred to OBER, the AUC drops sharply to 0.78-0.81. This hypothesis can be tested directly: OBER presumably comes from the OBject-Effect Removal dataset introduced in ObjectClear (CVPR 2026), which contains both real-captured and simulated object-removal image pairs. Because the real-captured target-absent images do not contain inpainting artifacts, the authors could simply report the AUC separately on the two subsets, yet the paper fails to do so. Additionally, three control experiments are needed: (i) applying equivalent inpainting to irrelevant regions of target-present images; (ii) evaluating on natural target-absent benchmarks in the style of POPE or NOPE; and (iii) querying an edited image about an object that remains present in both the original and edited versions.

Routing probabilities offer no distinct advantage over hidden states, limiting the novelty of the claimed "first framework." Router probabilities are merely the softmax of a linear projection of hidden states and therefore cannot contain more information than the hidden states themselves. Tables 1 and 8 corroborate this: mean-pooled hidden states achieve 0.9979 AUC, and single-layer hidden states at Layers 3-5 reach 0.9987, which is indistinguishable from routing's 0.9988. At Layer 2 specifically, routing reaches only 0.9836, underperforming the hidden state's 0.9986. Moreover, the core components of the approach have direct precedents: MoEE (ICLR 2025) already showed that routing weights can serve as semantic features; and HALP (EACL 2026) predicted VLM hallucinations during the prefill stage. While the authors pivot to emphasizing the interpretability of layer-expert coordinates, the supporting evidence is weak: the largest coefficient accounts for only 0.164% of the total weight magnitude; inspection of the supplementary CSV reveals that approximately 16.6% of the features have inconsistent signs across the three random seeds; and no causal intervention experiments are conducted to verify the functional role of these experts.

End-to-end gains stem primarily from the review prompt rather than the gate, and the gating framework is nearly identical to RSP. On GQA, simply applying the review prompt to all inputs ("always-review") already brings Qwen to 86.25% (+19.5 percentage points), with the routing gate contributing only an additional 2.75 points. On OBER, selective gating outperforms always-review by a mere 0.26 percentage points (95% CI: [0,0.70][0, 0.70]), because under the fixed threshold, specificity drops to 31.3% and 84% of all inputs trigger the review prompt. Furthermore, using a prefill signal to gate an existence-verification prompt is virtually identical in framework to RSP (arXiv:2605.28123, which the paper cites), differing only in the choice of gating signal, yet no empirical comparison is provided. Missing baselines include: training-free RSP-style signals (e.g., attention entropy, 1−ptop11 - p_{\text{top1}}, yes/no logits), gating via a hidden-state probe, a two-step prompting baseline that first asks "Is there a [X]?", and established hallucination mitigation methods such as VCD and OPERA.

Figure 1 is blurred and has uneven background colors, which is obviously AI-generated, while the AI Use Statement hides this.

Questions for authors

Q1. What are the separate detection AUCs on the real-captured versus simulated image pairs in OBER? Furthermore, what is the detection AUC when equivalent inpainting is applied to irrelevant regions of target-present images, as well as on natural target-absent images (e.g., POPE/NOPE-style subsets)?

Q2. Under the exact same data splits and review prompt, how does the proposed router probe compare against gating with RSP's training-free signals, a hidden-state probe, or a simple two-step prompting baseline that explicitly asks whether the target exists first?

Q3. What is the overall policy accuracy when evaluated under more realistic target-absence priors of 5%, 10%, and 20%?

Q4. If you perform expert ablation or forced routing on the top-ranked experts identified in Table 3, does it causally alter the model's responses to false-premise queries?

2 comments

Thank you for the detailed review and weight checks. Your controls help clarify how routing supports absence detection and review.

W1/Q1

Routing can support absence detection without inpainting. We tested the original frozen gates on 57 pairs photographed before and after physical removal. Qwen and Gemma reach 0.9049 and 0.8549 AUC, respectively. On 600 official POPE queries, they reach 0.6747 and 0.7154.

To assess sensitivity to editing, we also tested unrelated object removal. In 52 groups where the target stays present, removing another object raises false positives from 0 to 45 for Qwen and from 3 to 39 for Gemma. These existing edits are not matched for editor and settings.

We therefore trained new probes entirely on unedited COCO images to remove that training confound. They reach 0.9780 and 0.9665 OBER AUC. This training experiment shows that routing-based detection can be learned without inpainted images.

W2

Routing provides layer-expert coordinates that let us test which inputs support the probe. Retraining with the top 500 features retains 0.9984 GQA AUC, close to 0.9988 with all features. For detection on OBER, Qwen routing outperforms the selected hidden probe, whereas Gemma's hidden probe is stronger.

To assess ranking consistency, we checked three training seeds. As you observed, 16.63% of coefficients change sign. All 500 features selected by absolute mean weight have consistent signs, and independently selected Top-500 sets share 359–398 members. The high-weight features therefore show substantial agreement across seeds.

Building on MoEE's routing representations and HALP's prefill detection, our contribution connects absence detection to tested layer-expert analyses and selective review.

Q4

We ablated Table 3's eight experts at target-token prefill positions, without replacing selected experts or renormalizing their weights. Baseline, noop, ranked ablation, and three matched-random controls produced 2,100 manually checked answers on 350 images.

Ranked ablation changes some outputs and leaves correctness unchanged on all 175 absent images. Random controls cause at most one absent answer to become wrong. We agree that the ranking does not establish causal necessity and will remove that implication. Thank you for prompting this check. It separates the causal question from our supported Top-500 finding that a ranked feature subset retains nearly all probe performance.

W3/Q2

Like RSP, we use a signal to decide whether to review. We compared routing with other gates using the same prompt and splits. Prompt and utility thresholds were selected on GQA validation, then fixed for 60 test pairs and 115 OBER pairs. Answers were manually checked.

PolicyGQA accuracyOBER accuracy
Routing86.50%95.91%
Hidden probe87.33%85.13%
Attention entropy83.33%95.91%
Two-step existence query81.50%96.35%
Always-review83.17%95.91%

For yes/no logits and inverse confidence, validation selects always-review, giving the last row's results. Entropy and confidence are RSP-style adaptations. The two-step method uses an extra existence query. Routing and hidden probes read the original-question prefill.

We also evaluated VCD separately with one answer per image. It improves OBER accuracy from greedy's 83.04% to 86.96%, with GQA unchanged at 66.67%.

The policy comparison shows where selection helps. In this new batch, routing adds 3.33 points over always-review on GQA [95% CI 1.33, 5.50]. Present accuracy rises from 87.33% to 94.00%, while absent accuracy stays at 79.00%. Routing and always-review perform similarly on OBER. The gate's clearest added value is therefore preserving correct present-object answers while retaining correction on absent inputs.

Q3

We tested lower absence rates by reweighting the original results at the fixed threshold.

Absence rateGQA standard/selectiveOBER standard/selective
5%93.37% / 94.85%98.34% / 95.33%
10%90.41% / 94.20%96.68% / 95.35%
20%84.50% / 92.90%93.36% / 95.41%

OBER loses accuracy at the two lowest absence rates, so its threshold needs adjustment for this setting. The GQA gains persist across all three rates, extending the benefit beyond a balanced dataset.

W4

We used AI to generate the cats, edited the text and layout in slides, and then used AI to refine the figure's overall appearance. Thank you for flagging the omission. We will disclose both uses and fix the blurred parts and uneven backgrounds, clarifying how AI was used in the method illustration.

Thank you for the additional experiments. The rebuttal strengthens the empirical evidence, but my concerns about target specificity and the routing-specific contribution remain, so I retain my rating.

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

The paper considers a particular kind of hallucination in vision language models. Given an image and a question about the image with a false premise (e.g. “what color is the dog?” given an image without a dog), models will sometimes answer the question rather than pointing out the false premise. The authors call this failure mode “target-absence grounding failure”.

The authors show that the routing probabilities of an MoE can be used to predict when a model is going to suffer from this failure mode. Furthermore, detecting the potential failure and re-prompting the model can improve results on several visual reasoning benchmarks.

Critical strengths

It’s interesting that this kind of failure mode can be reliably detected from MoE routing probabilities alone.

Critical weaknesses

My main concern is significance. I don’t believe that others have shown this particular result before (detecting target-absence grounding failure from MoE probabilities); however prior work has already shown that some kinds of model hallucinations can be detected from internal model activations. The routing probabilities in an MoE are just another kind of hidden state, so it’s not that surprising that this can be done in light of prior work.

The fact that the “Always-Review” baseline (Table 2) performs so well also undercuts the significance of the method. This baseline can be seen as a form of prompt engineering, specifically asking the model to avoid this kind of failure mode; this already recovers most of the performance on GQA, and matches the proposed method on OBER. This makes me wonder whether another round of prompt engineering could entirely outperform the proposed method, while being much simpler.

Questions for authors

How significant are these results given the strong results from the Always-Review baseline?

1 comment

Thank you for the thoughtful review. We appreciate your recognition that MoE routing probabilities can detect target absence. Your questions help us explain why the layer-expert analysis matters and what the gate adds to a strong review prompt.

W1

Routing connects detection to an analysis of specific layers and experts. We first use the routing probabilities to detect target absence, then examine which features support that prediction. Ranking features by probe weight and retraining on the top 500 retains 0.9984 GQA test AUC, close to 0.9988 with all features. Separate layer-wise probes test where the signal becomes available.

These experiments give the layer-expert coordinates a practical role. We can identify a feature subset, test whether it preserves detection performance, and use the resulting detector to select inputs for review. Our contribution brings together a predictive signal, an experimentally tested analysis of its features, and a downstream use that improves answers.

W2 / Q1

Prompt design determines what the model is asked to do, and the gate determines which inputs receive that instruction. We isolate the gate's contribution by keeping the prompt fixed.

Qwen original cohortStandardAlwaysSelectivePrompt gain, ppGate gain, pp [95% CI]Selective review rate
GQA, 120 pairs66.75%86.25%89.00%+19.50+2.75 [1.17, 4.58]50.42%
OBER, 115 pairs83.39%95.30%95.57%+11.91+0.26 [0, 0.70]83.91%

The original Qwen results show a clear additional gain on GQA. On target-present inputs, selective review reaches 95.50% accuracy versus 89.33% for always-review, explaining how selection protects correct answers. The OBER policies perform similarly under the frozen threshold.

To test your suggestion about further prompt design, we evaluated target-aware, evidence-first, and balanced-check prompts. We selected the prompt on the original 60 GQA validation pairs, then evaluated 60 held-out test pairs and 115 OBER pairs. Target-aware had the highest validation accuracy. The following table uses a new batch of manually checked answers and the original routing threshold.

Qwen supplementary promptGQA always / selectiveOBER always / selective
Target-aware83.17% / 86.50%95.91% / 96.00%
Evidence-first70.33% / 73.33%81.83% / 82.52%
Balanced-check81.00% / 83.67%88.09% / 89.04%

For the selected prompt, the additional GQA gain is 3.33 percentage points [pair-clustered 95% CI 1.33, 5.50]. Present accuracy rises from 87.33% to 94.00%, and absent accuracy remains 79.00%. Across the three tested prompts, selective review has higher observed accuracy than always-review, with the clearest evidence on GQA.

These comparisons show that prompt design and input selection can work together. A good prompt provides the correction, and the gate helps preserve correct answers by choosing when to apply it.

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

This paper studies false-premise answering in vision-language models: a model may describe an object’s attributes even when that object is absent from the image. The proposed method extracts target-token routing probabilities during prefill and trains a model-specific linear probe to predict target absence. This prediction gates a target-aware review prompt, while the VLM and router parameters remain unchanged. The probes achieve near-perfect detection on GQA-Inpaint for both Qwen3-VL and Gemma, although ROC-AUC drops to 0.8095 and 0.7781 on the external OBER dataset. Routing-gated review improves final-answer accuracy for both models. Additional experiments show that absence information is already linearly accessible in early Qwen layers. The most interesting finding is the gap between information readable from internal representations and information reflected in the model’s answers.

Critical strengths

The problem is well scoped, and the intervention follows naturally from it. The paper focuses on target-absence failures rather than treating all hallucinations as one phenomenon. Target-token features, absence prediction, and target-aware prompting form a clear, testable pipeline. The method also requires neither backbone updates nor an external vision model.

The detection study includes useful controls and reasonably substantial in-domain evaluation. Source scenes remain within a single split, and the probe does not receive paired-image differences. Hidden-state comparisons provide relevant context. Separately trained single-layer probes also support the early-layer observation more convincingly than coefficient visualizations alone would. Results from two MoE model families strengthen the basic empirical finding.

The reporting is unusually candid about limitations. The authors distinguish ranking quality from threshold transfer, quantify harmful interventions, and acknowledge that repeated generations from one image are correlated observations. The pair-clustered bootstrap analysis is helpful. In particular, the manuscript appropriately recognizes that the selective-versus-always-review advantage on OBER is small and uncertain.

Critical weaknesses

1, The distinctive value of routing over existing representations remains unclear. The all-layer routing probe achieves 0.9988 ROC-AUC, but several single-layer hidden-state probes in Appendix Table 8 already reach 0.9987. Layer 3 hidden states achieve 98.69% accuracy, slightly exceeding the routing probe’s 98.54%. The evidence therefore establishes that internal representations predict absence more clearly than it establishes a specific advantage of routing.

2, Most answer-quality gains come from the prompt; the incremental benefit of routing is more modest. For Qwen on GQA, always using the target-aware prompt raises accuracy from 66.75% to 86.25%; selective review raises it further to 89.00%. The gate contributes, but the full 22.25-point improvement should not be interpreted as the contribution of routing detection. On OBER, always-review accuracy is 95.30%, versus 95.57% for selective review, with a clustered bootstrap interval that includes zero. Qwen also reviews 83.91% of inputs, limiting selectivity under distribution shift. Matched-trigger-rate comparisons against random and hidden-state gates are needed. A direct probe-based absence response would also help establish whether prompted review is the best use of the detector.

3, The end-to-end policy evaluation and deployment coverage remain limited. Review is evaluated on 120 GQA pairs and 115 OBER pairs. Half of the GQA cohort comes from the validation split used for model and threshold selection. This is explicitly disclosed and should not be characterized as undisclosed leakage, but an entirely held-out policy evaluation is still needed. Answer scoring is performed by one researcher, without reported blinding or inter-annotator agreement. Target extraction is validated only on controlled templates, leaving complex references, multiple targets, and abstention coverage unresolved. Finally, balanced present/absent cohorts do not establish the intervention’s risk when absent-target queries are uncommon.

Questions for authors

1, Could you compare routing, the validation-selected best hidden-state probe, and relevant pre-generation detectors on OBER, including AUC, calibration, and downstream gating performance?

2, Could you add edited-but-target-present controls and naturally absent, unedited images? How much better are target-token features than non-target or end-of-question features?

3, What computation does selective review save relative to using the target-aware prompt in a single initial pass? Please provide end-to-end measurements that include prefill, routing extraction, and cache reuse.

4, At matched trigger rates, how does routing gating compare with random and hidden-state gating? What happens if predicted absence directly produces an explicit absence response?

5, Could the final policy be evaluated on a larger, entirely held-out cohort, with blinded scoring or independent annotation to assess scoring reliability?

6, Thresholds are currently selected for detection balanced accuracy. How would the policy change if thresholds instead optimized final-answer utility, intervention harm, and computational cost? Does the benefit persist across different absence prevalences?

2 comments

Thank you for the thoughtful review. Your questions help us clarify how routing supports detection and when using it improves answers.

W1/Q1

Routing links detection to specific layers and experts. Retraining on its top 500 features gives 0.9984 GQA test AUC, close to 0.9988 with all features. This tests whether the features identified by the probe retain its predictive performance.

We compared detectors on OBER within each model and batch, selecting hidden layers on GQA validation.

Model/batchGateAUCPolicy accuracy
Qwen originalRouting0.809595.57%
Qwen originalHidden0.638285.74%
Qwen originalYes/no0.994396.26%
Gemma supplementaryRouting0.780293.83%
Gemma supplementaryHidden0.905594.35%

Routing outperforms hidden probing for Qwen, while hidden probing is stronger for Gemma. Yes/no is also strong and uses an extra existence query.

Alongside AUC, which measures ranking, we evaluated calibration to assess the reliability of predicted probabilities. Qwen routing's OBER ECE is 0.3133 versus 0.2488 for hidden probing, where lower is better. Routing's specific value is that it supports detection while making the probe's reliance on layer-expert features directly testable.

W2/Q4

The gate helps preserve correct answers while selecting inputs for review. Holding the prompt fixed isolates this benefit from prompt improvements.

DatasetAlwaysSelectiveGain, pp [95% CI]
GQA86.25%89.00%2.75 [1.17, 4.58]
OBER95.30%95.57%0.26 [0, 0.70]

These original Qwen results show a clearer gain on GQA, where selective review maintains higher present-input accuracy.

To test the selection itself, we compared gates at the same 83.91% OBER review rate. Routing reaches 95.57% accuracy, versus 90.61% for hidden gating and 93.37% averaged over random gates. We also tested returning an absence statement whenever the gate triggers. This reaches 96.92% on GQA and 65.65% on OBER. These controls support using the gate to select inputs for visual review.

W3/Q5

To separate policy evaluation from model and threshold selection, we evaluated the original 60 GQA test pairs alone. Accuracy rises from 68.00% to 89.00%. One researcher checked answers against images, so these are single-annotator results. The improvement therefore also holds on inputs excluded from selection.

For extraction coverage, all 240 controlled-template questions yield the correct target phrase and token span. The extractor abstains on unsupported or ambiguous templates and empty phrases. This establishes coverage for the evaluated templates.

Q2

We first tested the original frozen gates without inpainting. On 57 pairs photographed before and after physical removal, Qwen and Gemma reach 0.9049 and 0.8549 AUC. We also tested sensitivity to unrelated edits. In 52 target-present controls, removing another object raises false positives from 0 to 45 for Qwen and from 3 to 39 for Gemma.

To separate the potential of routing from this training confound, we trained new probes entirely on unedited COCO images. They reach 0.9780 and 0.9665 OBER AUC. These results show that absence detection can be learned without edited training images.

Separately trained probes also detect absence well at question end. We agree that 'localized' was too strong and will revise it. Our finding is that routing can reveal absence before generation.

Q3

We timed 200 images per model, including prefill, routing extraction, scoring, and generation. Selective review takes 3.169 seconds versus 2.233 for direct target-aware prompting on Qwen, and 0.681 versus 0.507 on Gemma. The implementation repeats prefill without KV-cache reuse, so fewer review calls do not produce a speedup here.

We will correct the cost claim accordingly. The demonstrated benefit remains that reviewing only 50.42% of original Qwen GQA inputs gives higher accuracy than always-review.

Q6

We selected a threshold for final-answer accuracy on GQA validation. It yields the same test decisions and 89.00% accuracy as the original threshold. We then reweighted the original results to test 5%, 10%, and 20% absence. At 5% absence on OBER, selective accuracy is 95.33% versus 98.34% for standard answering, showing the need for a threshold suited to that distribution. On GQA, the benefit persists at all three absence rates.

The author's reply partially solved my problem. However, according to the results provided by the author, the simpler yes/no detector is actually more powerful. The author states that yes/no requires additional queries, which is a reasonable distinction; however, their own routing strategy also incurs additional computations and is slower than directly using target-aware prompts. To explain the value of the method based on computational cost, a complete time consumption comparison with yes/no needs to be conducted.

Additionally, the low missing rate experiment exposed the risks of actual deployment. When the OBER missing rate is set at 5%, the selective review decreased by 3.01 percentage points. Although the author acknowledges the need to adapt to the distribution of the threshold, they have not demonstrated whether it is possible to select an effective threshold without accessing the test labels and restore the benefits. Re-weighting is only a prevalence-based analysis under fixed conditions and is not equivalent to real deployment verification.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.