acceptodds
Withdrawn from ICLR 2027

Can Generative Image Models Reason Through Each Other’s Outputs?

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Can separately trained generative image models use each other’s outputs for further reasoning? We study this question with models trained from scratch to perform addition modulo 23 by taking images of two operands as input and generating an image of the answer. We split training, development, and test sets between the complete set of 276 possible unordered modulo-23 additions, then train five models with different seeds and select the three that converge. Those three image-to-image transformers trained from scratch generalize to correctly solve 89.1% of held-out addition problems on average. We then ask whether an answer generated by one model can be used by a separately trained model to solve another held-out addition problem. Among cases where both models can solve their respective arithmetic steps independently, passing the first model’s generated answer to the second attains the correct result in 210 of 215 cases. In the remaining five, the generated number is still read correctly but changes the second model’s answer. This suggests that chains of generative image models, by passing images between steps, can reason and solve multi-step arithmetic addition problems.

Primary area:
generative models
Keywords:
generative models, image generation, visual reasoning, model interoperability, multimodal reasoning

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_j3a3OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates whether independently trained image generation models can use each other's generated digits to perform arithmetic operations. The authors evaluate modulo-23 addition, binary addition, and multi-step multiplication.

Critical strengths

The research question is potentially interesting.

Critical weaknesses
  1. Very limited novelty and significance. The paper essentially demonstrates that models can recognize generated digits and perform simple arithmetic operations. It is unclear what fundamentally new capability or insight this provides.
  2. Extremely simplistic experimental setting. The experiments rely on fixed layouts, predefined digit positions, and a tiny arithmetic vocabulary. Even the largest cross-model multiplication experiment only evaluates 20 pairs of 8-bit operands. The results provide little evidence of meaningful generalization.
Questions for authors

See weakness.

Reviewer_i3ZHOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper investigates whether an image generation model can understand an image generated by another image generation model and then generate a new one. The authors conduct experiments using basic mathematical computations as image generation prompts.

Critical strengths

None

Critical weaknesses
  1. The writing is poor, making it difficult to understand the motivation.
  2. The paper presents insufficient experiments, with its main content spanning fewer than six pages.
Questions for authors

Please see weaknesses.

Reviewer_2vWmOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies whether independently trained image-output models can use one another's generated digits to continue arithmetic calculations. Experiments cover modulo-23 addition, binary addition, and program-controlled longer calculations. Clean-image controls and appearance changes assess transfer reliability and sensitivity to digit rendering.

Critical strengths
  1. The clean-image control helps distinguish upstream arithmetic failures from errors caused by replacing a standard digit with a generated drawing. Both overall and conditional success rates are reported.
Critical weaknesses
  1. The central question is insufficiently motivated. The experiments do not establish why reading another model’s generated digit constitutes a distinct reasoning problem rather than visual recognition under a changed input distribution. The first model’s arithmetic computation appears unnecessary; directly generated or rendered digit images could test much of the same behavior.
  2. The multistep results primarily demonstrate externally programmed composition. The program supplies the operation sequence, partial-product construction, and carry routing, while the models execute local arithmetic operations already covered during training. Long execution chains therefore provide limited evidence of additional reasoning capability.
  3. The findings offer little new insight. Successful transfer of recognizable digits and failures under appearance changes are consistent with ordinary visual recognition robustness. The experiments neither distinguish this explanation from a reasoning-specific effect nor establish a general principle governing cross-model transfer.
  4. The research argument is poorly developed. The abstract and introduction mainly summarize experiments and results, while Related Work does not clearly establish a meaningful research gap or distinguish the contribution from existing studies.
Questions for authors
  1. What specific hypothesis requires the first model to perform arithmetic before generating a digit? Would directly generated or rendered digit images test the same hypothesis, and what would be lost by removing the first computation?
  2. Which result cannot be adequately explained by visual recognition robustness and execution of previously trained local arithmetic operations? Could suitable controls distinguish the proposed interpretation from these simpler explanations?
  3. What motivates communicating scalar numerical values through images, and what scientific insight does this setting provide beyond existing visual arithmetic and recognition robustness research?

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

Related papers

Discussion (0)

Sign in to comment.