acceptodds
Under review as a conference paper at ICLR 2027

Source References Fail to Constrain Inference: Diagnosing Source Misbinding in Multi-Image Contexts

Abstract

In multi-image question answering, even when the question explicitly specifies the target image, the model may still answer based on other images in the context. We define this failure as source misbinding: the model answers correctly from the target image when it is presented alone but produces an answer consistent with a distractor source when other images are introduced. This discrepancy leads us to hypothesize that the model may fail to translate the specified source into a constraint on evidence use during inference. To test this hypothesis, we construct D-SrcBench and T-SrcBench and conduct source readout and intervention experiments. The results support the explanation that source information remains decodable within the model, yet it does not reliably constrain the evidence used for answering. Based on this diagnosis, we propose Source-First Inference: with the backbone frozen, the model first predicts the relevant image source and then downweights or masks attention to distractor-image tokens at sensitive layers. On the fixed confirmatory test set, internal hard masking (Hard) improves LLaVA-1.5-7B accuracy from 30.4% to 59.0%, outperforming soft biasing (Soft) with the same source reader (51.2%) and the best-performing evaluated baseline, our reproduction of MIA-DPO (36.0%).

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multimodal Large Language Models, Multi-Image Reasoning, Source Misbinding, Inference-Time Intervention

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_8rNDOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies source misbinding in multi-image VQA: cases where a question explicitly identifies a target image, the model can answer correctly from that image alone, but adding other images causes the answer to become consistent with a distractor image. The paper argues that this is not primarily a failure to represent source identity; instead, source information remains decodable internally but fails to sufficiently constrain evidence use during generation.

Critical strengths

The problem is formulated clearly, which is backed by some mechanistic analysis. The experiments shows improvements from the method.

Critical weaknesses
  1. Lack of baseline methods and recent backbones: the experiment is based on Llava 1.5, which is outdated. Only DPO is compared, other recent multi-image papers are neglected (there are quite some other multi-image papers). The central mechanistic claims are derived almost entirely from LLaVA-1.5-7B.
  2. Method and evaluation design is unsound: every benchmark question explicitly names the target source, e.g. “Answer using Image 1 only.” The target image index is therefore already available symbolically in the prompt. Yet Source-First Inference introduces a learned 35K-parameter reader whose purpose is to infer which image the question specifies. This creates an artificial intermediate problem that does not need to exist in the benchmark as formulated. Benchmark scope is narrow relative to the broad framing.
Questions for authors

See weakness.

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

The paper studies source misbinding: a model answers from the wrong image despite an explicit image reference. It introduces two benchmarks and tests whether source information remains encoded during these errors. The proposed method first predicts the target image, then masks attention to other images at selected layers. On the 322-question test set, it improves LLaVA-1.5-7B accuracy from 30.4% to 59.0%. The paper includes quantitative results, controlled interventions, and success and failure cases.

Critical strengths
  1. The benchmark separates source errors from recognition errors. Each image has its own answer annotations. Target-only tests check whether the model can answer the question without distractors. This makes the diagnosis more informative than reporting overall accuracy alone.

  2. The intervention results are substantial and carefully analyzed. Hard improves test accuracy by 28.6 percentage points, with a 95% confidence interval of [23.3, 33.8]. The paper measures both repaired errors and newly introduced errors. It also includes wrong-source controls, random-token controls, and informative case studies.

Critical weaknesses
  1. Simpler methods perform better and run faster. The target image is explicitly named in every evaluated question. A parser can therefore select it directly. Table 5 reports accuracy as:

Proposed Hard method 59.0%

Keep only the image predicted by the same reader 64.0%

Parse the image reference and keep only that image 81.1%

The shared-reader isolation baseline significantly outperforms Hard (Table 13). Parser-based isolation is also much faster in the reported timing study: 143.6 ms versus 694.3 ms (Table 34). These results leave the learned reader and internal masking poorly motivated for the tested task. The diagnostic study remains useful, but the method’s practical contribution is unclear.

  1. The added insight over prior work needs stronger evidence. MIA-DPO already studies answers drawn from the wrong image under explicit references. FOCUS already studies cross-image leakage and uses image isolation to reduce it. This paper adds stricter source attribution and mechanistic experiments. However, the main evaluation uses only LLaVA-1.5-7B and simple questions about two or three images. The small historical Qwen experiments do not validate the final method on the full test set. The probes also decode image origin and reference index separately. This does not show that the model correctly associates the reference with the image. The evidence therefore supports source-specific repair more strongly than the claimed failure mechanism.
Questions for authors
  1. What benefit does Hard provide over parsing the explicit image reference and running target-only inference? Is this benefit demonstrated in any experiment?

  2. Does the final method improve results on the full test set with a model that natively supports multiple images?

  3. What experiment shows that the correct source–reference association is present but unused, rather than incorrectly formed?

  4. Has the MIA-DPO reproduction been validated on an established benchmark? Could its high damage rate be affected by truncated outputs or the answer scorer?

  5. The Llava-1.5 is an outdated model. Do the findings and conclusions still hold on latest models?

Reviewer_3C2sOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates source misbinding in multi-image question answering, where models generate answers based on distractor images despite explicit references to a target image. The authors analyze this phenomenon and show that source information remains recoverable within the model but does not effectively constrain evidence selection during inference.

To address this issue, the paper proposes Source-First Inference, which first predicts the relevant image source and then restricts attention to non-target images at selected layers. The authors also introduce D-SrcBench and T-SrcBench to systematically evaluate source misbinding in two- and three-image settings.

Experiments on LLaVA-1.5-7B include source information probing and layer-wise intervention analyses. The proposed method improves accuracy from 30.4% to 59.0% on an independent test set, outperforming the evaluated baselines.

Critical strengths

Mechanistic Analysis: The paper provides a detailed analysis of source misbinding through probing and layer-wise causal interventions. These experiments offer useful insights into how source information can remain recoverable within the model while failing to effectively guide evidence selection during inference.

Substantial Performance Improvements: The proposed method significantly improves accuracy on the independent test set from 30.4% to 59.0%, while maintaining a relatively low damage rate on previously correct predictions. This demonstrates its effectiveness in the evaluated setting.

Critical weaknesses
  1. Novelty over Cross-Image Information Leakage: How is source misbinding fundamentally different from the previously studied cross-image information leakage problem [1]? The two appear to describe essentially the same failure phenomenon.
  2. Limited Model Evaluation: All experiments are conducted on LLaVA-1.5-7B. Evaluation on more recent and diverse MLLMs is needed to demonstrate generalizability.
  3. Limited Applicability: The proposed method focuses on a restrictive setting where only one target image is needed to answer the question. However, many real-world multi-image tasks require integrating information across multiple images. How well does the proposed method generalize to these broader settings?
  4. Computational Overhead: The proposed method requires two forward passes and additional attention manipulation. A comparison of inference time, memory usage, and computational cost against existing methods is needed.
  5. Data Overlap and Fairness: Does the Source Reader's training data overlap with the diagnostic evaluation set? If so, is the comparison fair to training-free baselines that do not have access to this task-specific training data?

[1] Park, Yeji, et al. "Mitigating cross-image information leakage in lvlms for multi-image tasks." arXiv preprint arXiv:2508.13744 (2025).

Questions for authors

In Section 3.3, the expected distractor attention share is already 75% due to having three distractors. How is the reported attention normalized to account for the number of images and visual tokens?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.