acceptodds
Under review as a conference paper at ICLR 2027

Causal Readout Dependence in Multimodal Language Models

Abstract

A multimodal model can identify an object when choosing among named alternatives yet fail to mention the same object when asked to describe an image. We ask whether this difference arises only from competition among possible outputs, or whether image-conditioned internal states themselves have different causal effects across readouts. Using Qwen, Mistral, and Llama vision–language models, we identify hidden states at text-token positions that change substantially with the image while preserving the concept named in text. We then transplant these image-conditioned states into otherwise image-absent runs and measure their effects across readouts. Under forced choice, late-layer transplantation recovers 89.1% of the image effect in Qwen, 82.8% in Mistral, and 48.0% in Llama. At the next-token position of a generation prompt, the same intervention recovers only 14.8%, 0.3%, and 7.4% of the image-induced increase in probability for the depicted object. Transplanting states across layers 0–31 increases recovery but leaves a large gap, and the depicted object is detected in none of 464 neutral transplanted captions per model family. Controls show that this difference is not explained by answer-space size, candidate naming, scoring procedure, or transplant coverage. Conversely, removing continued access to the image sharply disrupts generation while preserving much of the relative structure of scored decisions. We call this causal readout dependence: an internal intervention can have a large effect under one readout and a much smaller effect under another. The gap is already present before the first token is generated, and competition during generation widens it. Causal effects established under one multimodal readout therefore should not be assumed to transfer to another.

Primary area:
interpretability and explainable AI
Keywords:
vision-language models, mechanistic interpretability, causal interventions, activation patching, multimodal generation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_cpYCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

Multimodal models can recognise objects when answering multiple-choice questions, but are unable to mention the same object when asked to describe an image. This paper argues that this is related to differences in the model’s internal computational mechanisms resulting from the differing task instructions. To investigate this issue, the paper designed an ‘Activation Transplantation’ experiment across three models. The results demonstrate differences in the ability to reconstruct images between multiple-choice tasks and free generation. The paper also designed another set of experiments to demonstrate the differences between the two task instructions by removing the attention to the image tokens after prefilling. The paper defines this phenomenon and its explanation as ‘Causal Readout Dependence’: an internal intervention may have a significant impact under one readout condition, while having a lesser impact under another. A broader lesson is that causal effects across different tasks are likely to be inconsistent; therefore, hallucination mitigation methods effective in a specific task context should not be directly transferred.

Critical strengths

The research raises a very good scientific question. The phenomenon observed is interesting.

Critical weaknesses

(1) The scientific question raised in the paper is: ‘Multimodal models are able to recognise objects when answering multiple-choice questions, but are unable to mention the same object when asked to describe an image.’ This is a good question. However, there is a lack of correspondence between the experiments and the motivation. Evidence:

  1. The conclusions drawn from the main experiment based on the THINGS dataset do not address this question, nor do the explanations provided in the paper fully account for this phenomenon.
  2. The analysis based on the COCO dataset focuses more on the issue of hallucinations, and the authors also acknowledge that the results do not yet fully explain the POPE-CHAIR disagreement.

(2) The main experiment did indeed observe low recovery in generation position. However, it cannot yet be conclusively demonstrated that this result is due to the cause stated in the paper rather than other factors in the experiment. Evidence:

  1. There may be confusion between the ‘prompt’ and ‘readout’ factors. ‘Forced choice’ and ‘generation’ differ not only in terms of readout, but also in whether the prompt and answer are explicitly presented, as well as in the donor hidden states that are extracted.
  2. The introduction of grey images may have an impact. The recipient does not actually lack visual input; rather, it is fed a grey image; its visual tokens, KV cache and so on are not replaced alongside the hidden states; consequently, the generation may still be influenced by that visual context.
  3. The technical robustness of ‘generation-position recovery’ is questionable. The authors have demonstrated that whether the transplant covers the scored position can increase the open vocabulary recovery rate from near zero to over 70–80 percent. Therefore, it is necessary to verify more directly that the low recovery in generation positions is not caused by information propagation or the conditions under which the intervention covers the position.
Questions for authors

I would like the author to provide clearer and more robust evidence, as well as to improve the writing and the way in which the evidence is presented.

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

The paper asks why an object in an image can change answer scores yet rarely appear in a caption. The authors copy text-position hidden states from an image run into a blank-image run and measure how much of the image's effect returns. Across Qwen, Mistral, and Llama, this restores much of the multiple-choice score change but little of the object's probability at the start of captioning. The authors call this gap causal readout dependence and find that generation also relies on continued visual access.

Critical strengths
  1. Activation transplants show a large recovery gap between multiple-choice scoring and the start of generation in all three models. The multiple-choice result does not rest on one architecture or a linear probe.
  2. The first open-vocabulary test looked weak because the transplant missed the position being scored. Recovery rose when that position was included, and a different scoring method under the same prompt gave similar results.
Critical weaknesses
  1. The two tasks use different prompts, so the states copied into each task are different. The gap could come from those states, rather than only from how each task uses them.
  2. The authors copy text states but leave image states, attention caches, and masks from the blank-image run unchanged. This may hinder later captioning even if it changes an immediate score; it cannot explain the whole gap because the gap appears before the first generated token.
  3. The multiple-choice test scores full answer strings, but the caption test scores only the first token of an object name after a fixed prefix. The recovery percentages may differ partly because the tests measure different things, even with the same-prompt scoring control.
Questions for authors
  1. Could you compare text-state transplants, transplants with matching cache updates, and transplants with continued visual access? This would show how much of the missing object mentions comes from mixed states.
  2. Could you start from the same donor state, then branch into candidate scoring and free-form captioning? Changing only the later question or output requirement would show whether the gap remains.
Reviewer_AQKWOpenReview ↗
Rating
1(−1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The study uses activation transplants to compare recovery of an image's effect under different output measurements. Image-conditioned text states supply the donor activations, and blank-image runs receive them, across three model families. Recovery is high for forced-choice margins and much lower for source-token probability at a generation position. Captioning and continued-visual-access experiments further show how intervention results depend on the evaluation setting.

Critical strengths

The transplant results connect the internal intervention to behavior. In Qwen, restored forced-choice margins closely track image-present margins across items, even after controlling for the blank-image margin. This provides evidence beyond representation similarity. The concept-span comparison also shows that the effect depends on the broader text context.

The authors revise their initial open-vocabulary interpretation after finding incomplete transplant coverage. They report controls across model families and negative findings about their proposed explanation for benchmark disagreement. I find these qualifications helpful, particularly the caution about transferring conclusions between evaluation settings.

Critical weaknesses

I am not sure the recovery percentages establish different amounts of restored source information. The forced-choice outcome in Section 3.4 is the difference between source and target log-probabilities. A transplant could restore that margin through target suppression, source promotion, or both. Section 3.5 instead measures only the source's first-token log-probability at a generation position. High margin recovery alongside low source recovery therefore does not by itself establish that comparable source evidence reaches the two readouts differently.

Section 6 acknowledges the endpoint mismatch, but the stronger interpretation of the recovery gap still needs a matched outcome analysis. The same-prompt scoring control does not separate source promotion from target suppression for the primary 464-pair cross-context comparison. The promotion/veto analyses in Appendix C.3 concern separate direct-edit experiments and leave this question unresolved.

Questions for authors

In the primary 464-pair experiment, how much of the recovered forced-choice margin comes from source promotion, and how much from target suppression? Please report the image-present and transplant changes for both terms. Could both prompt contexts then be compared with the same source-only measure and tokenization convention? This would help determine whether the gap reflects differences in restored source evidence or, in part, the definitions of recovery.

1 comment

Given the lack of response from the reviewer during the rebuttal period, I am revising my final score to 1 (Reject).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.