acceptodds
Under review as a conference paper at ICLR 2027

TRACE: Tracing Representation Ancestry for Hallucination Mitigation in Large Vision-Language Models

Abstract

Large Vision-Language Models (LVLMs) often generate fluent but visually unsupported content. Prior studies commonly explain such hallucinations through reduced visual attention or increased reliance on textual context, implicitly associating information with the modality of the token positions that carry it. However, Transformer layers repeatedly mix information across positions, allowing visual evidence to propagate into question and previously generated tokens. As a result, token modality does not always reveal where information originally comes from. We introduce TRACE, a provenance-based framework for analyzing and mitigating hallucination in LVLMs. TRACE traces information back to its source by propagating contributions from visual inputs, user queries, and generation history through intermediate representations. This analysis reveals a recurring pattern before hallucination, which we call provenance takeover: generation-history provenance becomes increasingly dominant over visual provenance as the model approaches a hallucinated token. Building on this observation, TRACE detects abnormal takeover during decoding and reinforces the corresponding visual-origin component of the current representation. This enables our method to intervene selectively when the model begins to drift away from visual evidence, without requiring additional training. Across many benchmarks, TRACE consistently reduces hallucination-related errors across LVLMs while largely preserving overall multimodal performance with modest inference-time overhead. These results establish information provenance as a useful perspective for understanding hallucination and a practical signal for improving visual grounding in LVLMs.

Primary area:
foundation or frontier models, including LLMs
Keywords:
VLM, Hallucination

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_pT86OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TRACE, a training-free method for mitigating hallucinations in LVLMs. TRACE propagates contribution-based provenance estimates for visual inputs, user queries, and generation history across layers. During decoding, it detects abnormal increases in history-relative-to-visual influence and selectively reinforces the estimated visual-origin contribution. The experimental results demonstrate the effectiveness of it in mitigating hallucinations.

Critical strengths
  1. Well-motivated perspective. Distinguishing information origin from its current token position is a useful perspective, since visual information can propagate through textual representations. This motivates looking beyond direct contributions from image-token positions.
  2. Coherent method design. TRACE connects provenance estimation, online detection, and adaptive intervention within a single framework, without additional training or an extra model forward pass.
Critical weaknesses
  1. Insufficient causal validation. The propagated weights provide attribution under the proposed rule, but do not establish that textual representations retain task-relevant visual semantics or that provenance takeover causes hallucinations.
  2. Unclear specificity of visual reinforcement. The paper does not establish that the gains are specific to reinforcing text positions assigned high visual-provenance weights, nor does it show that reinforcing randomly selected positions is ineffective.
  3. Reporting issues. Figure 1 lacks detailed model and experimental settings (e.g., sample and object counts), and the relative latency calculation for CRG in Table 4 is incorrect.
  4. Incomplete discussion of related work. Prior work on explicit [1, 2] and implicit [3] cross-modal information flow is not discussed. The paper should clarify how its provenance formulation advances beyond these studies.
Questions for authors
  1. Could the authors add activation patching experiments, using controlled corruption and restoration, to test whether text positions assigned high visual provenance causally contribute to visually grounded predictions?

  2. Could the authors compare against reinforcing randomly selected text positions, matching the number of contributing positions, intervention timing, and perturbation magnitude, to test whether this random control also reduces hallucinations?

References

[1] Cross-modal information flow in multimodal large language models. CVPR 2025.

[2] Same task, different circuits: Disentangling modality-specific mechanisms in vlms. NeurIPS 2025.

[3] Interpreting and Enhancing Emotional Circuits in Large Vision-Language Models via Cross-Modal Information Flow. ICML 2026.

3 comments
AuthorsOpenReview ↗

Reply to reviewer pT86

We thank the reviewer for the constructive feedback. We address the concerns and questions below.

W1 & Q1. Causal validation of provenance attribution

We appreciate this concern. TRACE estimates source-level contributions rather than claiming an exact causal decomposition of hidden representations. Our propagation mechanism tracks estimated information origins across Transformer layers, allowing visual-origin contributions to remain distinguishable even after being integrated into textual representations.

To strengthen the empirical validation, we have revised Figure 1 to present separate provenance trajectories for LLaVA-1.5-7B, Qwen-VL-Chat, and Qwen2.5-VL on 500 CHAIR images. Object mentions are extracted from each model's generated descriptions and classified as grounded or hallucinated using ground-truth annotations. We also provide real qualitative examples in Supplementary Section A.4, illustrating hallucinated object mentions and their correction after TRACE intervention.

More importantly, we have added an activation-patching experiment in Supplementary Section A.5 to directly examine the causal relevance of estimated visual provenance. Using 200 clean–corrupted image pairs from MS COCO, we restore clean activations at text positions selected by high visual provenance, low visual provenance, or random selection, while keeping the number of patched positions identical. High-visual-provenance positions consistently achieve substantially greater prediction recovery than both controls, particularly in middle-to-late Transformer layers. These results provide intervention-based evidence that TRACE identifies textual representations that causally contribute to visually grounded predictions, although they do not establish an exact decomposition of visual information.

W2 & Q2. Specificity of visual reinforcement

We appreciate this suggestion. We have added two ablations: random-token reinforcement and random gating, to examine whether TRACE's improvements arise from provenance-guided reinforcement and takeover-based intervention rather than arbitrary reinforcement.

Intervention strategyCHAIRs ↓Recall ↑
Regular decoding52.577.1
Random-token reinforcement49.673.5
Random gate + visual reinforcement46.576.8
Always-on visual reinforcement30.576.8
TRACE (takeover-guided)27.378.6

Random-token reinforcement yields only a modest reduction in CHAIRs while decreasing Recall from 77.1 to 73.5, suggesting that arbitrary reinforcement can suppress relevant visual content. Similarly, random gating achieves substantially weaker hallucination mitigation than takeover-guided intervention. These results support the importance of selecting representations according to their estimated visual provenance and using the takeover signal to guide intervention.

W3. Reporting issues

We thank the reviewer for identifying these issues. In the revised version, we have updated Figure 1 to specify the evaluated models, dataset, sample size, and object-mention analysis protocol. We have also corrected the relative-latency calculation for CRG in Table 4.

W4. Distinction from related cross-modal information-flow studies

We thank the reviewer for highlighting these works. While [1–3] also investigate cross-modal information flow, their attribution mechanisms and objectives differ from TRACE.

Specifically, [1] uses attention knockout to identify important information-transfer pathways between visual and textual token positions, measuring their importance through changes in model predictions. However, identifying information transfer between positions does not directly quantify its original source after cross-token mixing. TRACE instead propagates source-level attribution across layers, allowing visual-origin information to remain attributed to its original source even when carried by textual representations.

[2] compares computational circuits for equivalent visual and textual tasks, using back-patching to address differences in modality-specific processing. It does not estimate how source contributions evolve during autoregressive generation. [3] investigates emotion-related information flow using a steering-vector-based causal attribution framework, whereas TRACE estimates source provenance directly during decoding without constructing separate steering vectors.

More importantly, TRACE uses these source-provenance estimates to investigate provenance takeover and guide adaptive hallucination mitigation. Unlike [1–3], our method connects the evolving contributions of visual inputs, user queries, and generation history to an online detection signal and corresponding intervention strategy.

AuthorsOpenReview ↗

Follow-up on Rebuttal Responses

Dear Reviewer,

We have provided detailed responses to your concerns and would appreciate it if you could take a look when you have a chance.

As the discussion period is coming to an end, we would be grateful for any feedback on whether our responses have addressed your concerns. We would be happy to clarify any remaining questions.

Best regards,

The Authors

Thank you for the detailed rebuttal and additional experiments, which alleviate some of my concerns; I now consider the paper borderline, with no preference toward acceptance or rejection.

Reviewer_RNuLOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes TRACE, a training-free method for mitigating hallucinations in large vision-language models. It propagates token-level contributions across Transformer layers to estimate the provenance of hidden representations from visual inputs, user queries, and generation history. The authors observe an increase in generation-history provenance before hallucinated object mentions, termed provenance takeover, and use this signal to guide visual representation reinforcement during decoding. Experiments on CHAIR, AMBER-Generative, and MMHal-Bench report improvements on several LVLM backbones.

Critical strengths
  1. The paper offers an interesting perspective on the difference between the positions carrying information and its attributed input sources.
  2. The proposed approach is training-free, and the overall framework is clearly motivated.
  3. The results on CHAIR and AMBER-Generative show substantial improvements over regular decoding across three LVLMs. The ablations also suggest that provenance-based variants outperform the position-based alternatives considered in the paper.
Critical weaknesses

1. The proposed provenance-takeover phenomenon is not sufficiently validated.

Figure 1 is the main evidence for the claim that provenance takeover precedes hallucination. However, the figure and accompanying discussion do not clearly specify the models, datasets, or number of grounded and hallucinated mentions used in the analysis. The results are presented as averaged trajectories, without reporting their variability or how reliably the signal distinguishes individual hallucination events. In particular, it is unclear whether the same pattern holds across the different LVLMs and hallucination categories considered in the paper. The improvements reported in Tables 1 and 2 demonstrate the effectiveness of TRACE as an intervention, but do not by themselves validate the generality or predictive value of the proposed phenomenon. The paper also lacks real qualitative examples linking takeover signals to hallucinated tokens and the corresponding changes after intervention.

2. The interpretation of the estimated provenance and visual restoration is not fully justified.

TRACE propagates normalized token-contribution scores as estimates of information provenance (Eqs. 4–8). These scores offer a useful attribution heuristic, but their relationship to the actual influence of the input sources on model predictions is not evaluated. In particular, the formulation focuses on cross-token information transfer without explicitly accounting for how nonlinear transformations may change the relative contributions of already mixed information. Furthermore, previously generated tokens are treated as a separate history source, although they may already reflect visual information from earlier decoding steps. More importantly, Eq. (12) weights each full, potentially mixed contribution vector by a scalar visual-provenance score. This does not necessarily isolate the visual-origin component of that vector, as suggested by the notion of provenance restoration. Since this interpretation underlies both detection and intervention, its validity should be supported by more direct attribution or counterfactual analyses.

3. The contribution of takeover-guided intervention is not sufficiently separated from general visual reinforcement.

Table 3(b) shows that always-on restoration already reduces CHAIRs from 52.5 to 30.5, whereas the full adaptive takeover gate further reduces it to 27.3. The adaptive gate provides an additional improvement, including in Recall, but the results do not establish whether this gain comes from the specific takeover signal or simply from applying visual reinforcement more selectively. A comparison with alternative or randomized gates under comparable intervention strengths and frequencies would better clarify this point. The attribution component is also closely related to prior cross-layer attribution methods such as ALTI, while visual reinforcement has been explored by methods such as VISTA and AIR. The paper would benefit from a clearer demonstration of the advantage provided by the proposed combination over these established alternatives.

Minor Comments

  1. The role of query provenance is underexplored. Although TRACE explicitly tracks visual, query, and generation-history provenance (Eqs. 5–8), the takeover detector only considers the history-to-visual ratio (Eq. 9). The paper does not explain why query provenance is excluded from the detection criterion or examine whether query-driven hallucinations exhibit the same takeover pattern.

  2. The layer-wise intervention strategy lacks sufficient analysis. Figure 1 suggests that provenance takeover is most pronounced in middle-to-late Transformer layers, whereas TRACE evaluates the intervention gate independently at each layer. It remains unclear which layers contribute most to hallucination mitigation and whether interventions outside the layers exhibiting strong takeover are necessary.

Questions for authors
  1. Which models and datasets were used for Figure 1, and how consistently does the takeover pattern appear across models and hallucination types?
  2. What evidence supports interpreting the propagated attribution scores, particularly the restoration vector in Eq. (12), as visual-origin information?
  3. Does the takeover-based gate outperform simpler gating alternatives when the frequency and strength of intervention are controlled?
2 comments
AuthorsOpenReview ↗

Reply to reviewer RNuL (1)

We thank the reviewer for the detailed and constructive feedback. We address the weaknesses and corresponding questions together below.

W1 & Q1. Validation and consistency of provenance takeover

We appreciate the reviewer's concern. We have revised Figure 1 to present provenance trajectories separately for three LVLMs: LLaVA-1.5-7B, Qwen-VL-Chat, and Qwen2.5-VL, evaluated on 500 images from CHAIR. Specifically, we extract object mentions from each model's generated descriptions and classify them as grounded or hallucinated using ground-truth object annotations. We then analyze their provenance trajectories separately to examine the differences between grounded and hallucinated mentions across models.

The revised figure provides a clearer view of the consistency and variation of provenance takeover across different backbones, rather than relying on a single aggregated trajectory. Our analysis focuses on object hallucinations, and we do not claim that the same pattern necessarily holds for every hallucination category. Furthermore, we provide real qualitative examples in Supplementary Section A.4, illustrating hallucinated object mentions and how TRACE corrects them during generation.

We emphasize that provenance takeover is an observed pattern associated with hallucination, rather than a universal condition for every hallucination event.

W2 & Q2. Interpretation and validity of estimated provenance

We thank the reviewer for raising this important distinction. TRACE estimates relative source contributions rather than claiming an exact causal decomposition of hidden representations. Our approach propagates token-level attribution scores across Transformer layers to approximate how information originating from different sources contributes to subsequent representations.

Regarding nonlinear transformations, token-wise nonlinear operations do not directly transfer information between token positions, whereas cross-token information mixing primarily occurs through attention. Although nonlinear transformations may alter the relative influence of already mixed information, they do not independently introduce cross-token information transfer. Thus, our formulation approximates cross-token information provenance rather than modeling every transformation of source contributions.

Regarding generation history, we agree that previously generated tokens may contain information derived from visual inputs. However, TRACE separately tracks source contributions throughout decoding, allowing the estimated influence of original visual inputs to be distinguished from contributions associated with generated tokens. Therefore, the presence of visual information in previously generated representations does not imply that all their contributions must be attributed exclusively to generation history.

Finally, Eq. (12) performs visual-provenance-weighted reinforcement rather than exact extraction of isolated visual information. The provenance scores identify contributions estimated to carry greater visual-origin information and assign them larger weights during reinforcement. This interpretation is also supported empirically by our ablation studies, where provenance-based reinforcement outperforms position-based alternatives. While these results demonstrate the utility of the estimated provenance, we acknowledge that they do not establish exact causal attribution.

AuthorsOpenReview ↗

Reply to reviewer RNuL (2)

W3 & Q3. Contribution of takeover-guided intervention

We appreciate this concern. To better isolate the contribution of the proposed mechanism, we have added two additional ablations: random-token reinforcement and random gating. Random-token reinforcement examines whether reinforcing arbitrary token representations provides similar benefits, while random gating investigates whether randomly selecting intervention steps is sufficient compared with takeover-guided intervention.

Intervention strategyCHAIRs ↓Recall ↑
Regular decoding52.577.1
Random-token reinforcement49.673.5
Always-on visual reinforcement30.576.8
Random gate + visual reinforcement46.576.8
TRACE (takeover-guided)27.378.6

Our additional experiments show that random-token reinforcement can degrade model performance, indicating that arbitrary reinforcement does not provide the same benefits as visual-provenance-weighted reinforcement. The random-gate experiment further examines the contribution of the takeover signal independently of the general effect of selective reinforcement.

Regarding related work, VISTA employs visual steering vectors extracted from input-specific representations, whereas TRACE directly uses propagated source-provenance estimates without requiring separate steering-vector construction. Furthermore, TRACE adaptively determines when and how strongly to intervene based on the detected takeover signal.

For ALTI and AIR, we would appreciate it if the reviewer could provide the full paper titles or citations so that we can accurately identify the intended works and provide a detailed comparison.

Minor comments

M1. Role of query provenance. Our primary focus is visual grounding. Although TRACE tracks query provenance, we did not observe the same characteristic takeover pattern in query provenance as in visual provenance during our analysis. Therefore, we focus on the history-to-visual ratio as the detection criterion. We will clarify this motivation while acknowledging that query-driven hallucinations warrant further investigation.

M2. Layer-wise intervention strategy. In the revised Figure 1, provenance takeover can emerge across a broad range of layers, approximately layers 8–24, rather than consistently appearing at a fixed depth. Restricting intervention to predetermined layers may therefore overlook relevant provenance shifts. Since layer-wise detection follows the model's sequential processing and introduces limited additional overhead, we apply the gate across all layers to accommodate model- and sample-dependent variations.

Reviewer_rS7ZOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces TRACE, a training-free framework designed to mitigate hallucinations in Large Vision-Language Models (LVLMs) by tracking the provenance of their outputs. TRACE analyzes the contributions from visual inputs, user queries, and generation history across the layers of a Transformer, revealing a phenomenon termed provenance takeover. To address this, TRACE detects anomalous shifts at the source during the decoding process and dynamically reinforces information grounded in visual evidence. Experiments on three benchmarks demonstrate that TRACE consistently reduces hallucinations with only a modest increase in computational overhead.

Critical strengths
  1. The topic of this paper is important, as hallucinations can undermine user trust—particularly in high-risk applications such as medical analysis and various industrial settings.

  2. The paper is well-organized and highly readable.

  3. The proposed TRACE is lightweight and training-free.

Critical weaknesses
  1. The motivations and findings in this paper seem a little trivial, as previous studies have already discussed information flows in LVLMs [1,2,3].

  2. This paper only conducts experiments on hallucination-related benchmarks. It remains unclear whether the proposed TRACE mitigates hallucinations at the cost of general multimodal capacity.

  3. The paper fails to discuss or compare with several recent works that also incorporate visual information into the model's decoding process [4,5].

References:

[1] Understanding Information Storage and Transfer in Multi-Modal Large Language Models, NeurIPS 24.

[2] Towards Interpreting Visual Information Processing in Vision-Language Models, ICLR 25.

[3] Mitigating Hallucination in Large Vision-Language Models through Aligning Attention Distribution to Information Flow, AAAI 26.

[4] Look Twice Before You Answer: Memory-Space Visual Retracing for Hallucination Mitigation in Multimodal Large Language Models, ICML 25.

[5] Understanding and Mitigating Hallucinations in Multimodal Chain-of-Thought Models, CVPR 26.

Questions for authors
  1. Can TRACE be applied to more recent backbones, e.g., Qwen-3-VL-series?

  2. Can TRACE be used to improve multimodal reasoning capacity (as TRACE enables visually grounded decoding)?

3 comments
AuthorsOpenReview ↗

Reply to reviewer rS7Z (1)

We thank the reviewer for the constructive feedback and for recognizing the efficiency and training-free nature of TRACE. We address each concern below.

W1. Novelty and distinction from previous information-flow studies

We respectfully disagree that TRACE's motivation and findings are trivial or already established by previous studies. While [1–3] investigate information processing in LVLMs, they address different research questions and do not establish the provenance-takeover phenomenon or its use for adaptive hallucination mitigation.

Specifically, [1] employs causal tracing to identify components responsible for information storage and transfer, while [2] investigates how visual representations evolve and contribute to predictions across layers. Although [3] also analyzes information flow, it primarily examines how visual information propagates into semantic representations and addresses the mismatch between attention distributions and this flow. In contrast, TRACE tracks the estimated contributions originating from three distinct sources: visual inputs, user queries, and generation history, throughout autoregressive decoding. Importantly, TRACE propagates source attribution across layers, allowing visual-origin information to be tracked even after it has been integrated into textual representations, rather than relying solely on the current positions of visual tokens.

More importantly, our central contribution is not merely tracking information flow, but using this source-level perspective to reveal a previously uncharacterized hallucination-associated pattern, termed provenance takeover. This provides a different perspective on understanding hallucinations by examining how the relative contributions of different information sources evolve during generation, which is not explicitly investigated in [1–3].

Furthermore, TRACE translates this observation into a training-free, adaptive mitigation strategy. Unlike [3], which aims to align attention distributions with information flow, TRACE monitors provenance shifts during decoding and selectively reinforces visual-origin information when abnormal takeover is detected, with intervention strength determined by its severity. Thus, TRACE contributes a distinct hallucination-related observation and a corresponding inference-time intervention mechanism, beyond existing information-flow analyses.

W2. Preservation of general multimodal capability

We appreciate this concern. While our evaluation primarily focuses on hallucination benchmarks, several metrics provide evidence that TRACE preserves visual understanding and generation quality. CHAIR Recall and AMBER Cover are maintained or improved, suggesting that hallucination reductions do not simply result from omitting visual content. Average response length remains comparable to regular decoding, indicating that the improvements are not primarily due to shorter responses. Furthermore, MMHal-Bench shows improvements across counting, comparison, relationships, and holistic reasoning.

Although these metrics do not comprehensively evaluate general multimodal capability, they suggest that TRACE reduces hallucinations while preserving visual informativeness under the evaluated settings.

W3. Comparison with recent visual-grounding methods [4,5]

We thank the reviewer for highlighting these relevant works. Unlike [4] (MemVR), which re-injects visual tokens based on prediction uncertainty, TRACE reinforces visual-origin information already present in hidden representations based on detected provenance takeover. Regarding [5], this method targets hallucinations during multimodal chain-of-thought reasoning, whereas TRACE operates during standard autoregressive decoding without requiring explicit reasoning steps.

To further address this concern, we reproduced both methods using their publicly available source code and evaluated them on LLaVA-1.5-7B and Qwen2.5-VL. Since [5] is designed for multi-step reasoning, we adapt its intervention to a single reasoning step for comparison under our non-CoT setting. The results are presented below.

LLaVA-1.5-7B

Method(C_S\downarrow)(C_I\downarrow)F1 (\uparrow)
MemVR [4]45.212.878.4
Method [5]46.313.278.9
TRACE27.39.078.6

Qwen2.5-VL

Method(C_S\downarrow)(C_I\downarrow)F1 (\uparrow)
MemVR [4]27.37.370.9
Method [5]28.17.672.1
TRACE21.85.972.8
AuthorsOpenReview ↗

Reply to reviewer rS7Z (2)

Q1. Applicability to Qwen3-VL

We additionally evaluate TRACE on Qwen3-VL to examine its generalization to newer LVLM architectures.

Method(C_S\downarrow)(C_I\downarrow)F1 (\uparrow)
Regular56.110.574.3
VCD58.010.974.8
OPERA53.09.675.0
PAI55.410.575.0
VTI54.810.375.4
CRG53.310.075.8
Devils51.79.475.3
TRACE50.19.176.3

TRACE achieves the lowest hallucination rates and highest F1 among the evaluated methods, supporting its effectiveness on newer LVLM backbones.

Q2. Potential for improving multimodal reasoning

We believe TRACE may benefit multimodal reasoning by maintaining visual grounding throughout generation. Our MMHal-Bench results provide preliminary evidence, showing improvements in counting, comparison, relationships, and holistic reasoning. However, TRACE is primarily designed for hallucination mitigation, and we have not specifically evaluated its impact on dedicated multimodal reasoning benchmarks. We consider this a promising direction for future work.

References:

[1] Understanding Information Storage and Transfer in Multi-Modal Large Language Models, NeurIPS 24.

[2] Towards Interpreting Visual Information Processing in Vision-Language Models, ICLR 25.

[3] Mitigating Hallucination in Large Vision-Language Models through Aligning Attention Distribution to Information Flow, AAAI 26.

[4] Look Twice Before You Answer: Memory-Space Visual Retracing for Hallucination Mitigation in Multimodal Large Language Models, ICML 25.

[5] Understanding and Mitigating Hallucinations in Multimodal Chain-of-Thought Models, CVPR 26.

Thanks for the authors' detailed responses. My concerns regarding W1, W3, and Q1 are addressed. For W2 and Q2, I would like to see more results on multimodal perception and reasoning benchmarks to evaluate TRACE's generalizability. I'm now raising my scores to 3, as the authors have addressed most of my concerns. Since the ICLR policy allows revision of the paper, please consider adding these results directly into the revised PDF.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.