acceptodds
Withdrawn from ICLR 2027

Can Circuit Analysis Predict Capability Transfer Across Tasks and Modalities?

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Circuit analysis is mainly used post-hoc to explain model behavior, but could it also be used to guide training decisions? We show that a model’s internal computation, measured before fine-tuning, can predict how training on one task will affect another. Across four vision-language models from two families and 18 tasks with textual and visual variants, we establish three findings. First, task circuits present in the base model largely persist after fine-tuning, meaning that circuits discovered before training remain representative afterward. Second, relationships between these base-model circuits (measured by cross-task circuit accuracy) can predict capability transfer: our circuit-based measure correlates with cross-task fine-tuning gains up to r = 0.77, and is competitive with an established black-box transferability estimation method while capturing non-trivial transfer across task types. Third, we characterize when these predictions succeed and fail, and find they are most reliable when task circuits are well localized, but break down when similar-looking circuits conceal conflicting computations. Together, these findings show how circuits measured once can inform capability transfer, turning circuit analysis from an explanatory tool into a predictive one.

Primary area:
interpretability and explainable AI
Keywords:
circuit analysis, interpretability, vision-language

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_aTMQOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper provides a thorough circuit analysis before and after finetuning of a vision-language model (VLM). They observe and concur with previous works that task circuits present in the base model, largely persist after finetuning. Since these circuits persist, they further use these to predict cross-task finetuning gains. Lastly, they run a failure analysis on the same to figure out when these predictions succeed or fail. The experiments are run reliably and consistently across model families, scale and seeds. The insights are quite interesting and actionable as well. However, the task selection seems a bit unintuitive with no explanation given regarding the chosen tasks and why the datasets were self-constructed. Moreover, it would be interesting to see how cross-task analysis relates to recent taskonomy works in VLMs which verify cross-task transfer empirically rather than task-relatedness measures, and can provide insights into why certain tasks are beneficial / detrimental for other tasks. Lastly, the writing can be improved by a few changes: highlighting the findings, minor typos (Figure 1; L755), labels unexplained (Figure 2(b)), etc. Given the weaknesses pertaining to reliability on scale and synthetic tasks, are addressed, I'd recommend the paper for acceptance.

Critical strengths
  1. The insights are actionable and relate well to previous work. Failure analysis (Section 5) is also appreciated and paves the way for future work for choosing training tasks for finetuning.
  2. The experiments are quite extensive, covering modalities, 2 model families, scale as well as multiple seeds. This is quite useful to validate and confirm findings.
Critical weaknesses
  1. The task selection seems quite narrow and moreover, no intuition is given why these specific tasks are selected. While synthetic tasks are useful for controlled experiments, it is difficult to gauge how these insights transfer to real-world tasks where different visual capabilities are required (spatial reasoning, scene understanding, etc.)
  2. The results in Figure 3 are quite heterogeneous, varying across models, tasks and modalities. While the paper addresses this, more explanation regarding these observations would strengthen the paper. Similarly, no discussion is provided for Figure 4 for the observations (e.g., why does the arithmetic training task predict to improve pattern matching). It also remains unclear how reliably XCA can be used to select training tasks for real-world use cases.
  3. The writing can be improved: main findings can be highlighted more clearly, labels and abbreviations in Figure 2(b) are not sufficiently explained, Figure 1 and line 755 contain typos.
Questions for authors
  1. Can you provide an intuition of why those specific tasks were selected? Moreover, why were they self-constructed?
  2. L431 claims that "in larger models from both families, the correlations to circuit localization metrics weaken". Does this suggest that circuit localization becomes a less reliable indicator of XCA's predictiveness at larger model scales, or does XCA's predictive performance itself also weaken with scale?
  3. Was there any correlation of task difficulty with the observations? In particular, could differences in task difficulty account for the varying observations in Figure 3?
Reviewer_tZkqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper investigates whether circuits identified in a pretrained model can predict how fine-tuning on one task affects performance on another. It introduces cross-task circuit accuracy (XCA), which evaluates the target task’s circuit on a candidate training task and subtracts the accuracy of a size-matched random circuit. Experiments cover four vision-language models and 18 synthetic tasks, each with textual and visual variants. The authors report that task circuits largely persist under LoRA fine-tuning and that XCA correlates with cross-task fine-tuning gains, although predictive performance varies substantially across tasks and models. XCA shows mixed performance relative to a baseline based on final-layer task representations. Further analysis associates stronger predictions with better-localized target circuits and identifies failures where tasks share functionality but require conflicting computations. The work positions circuit analysis as a potential tool for informing training decisions.

Critical strengths

The paper investigates whether circuits can inform training decisions before fine-tuning, extending their use beyond explaining existing behavior. XCA provides an intuitive way to assess functional compatibility between tasks. The study also evaluates transfer across four models and multiple task families and modalities, with random-circuit controls and a representation-based baseline. The analyses of circuit localization and conflicting counting tasks help characterize when the proposed predictor succeeds and where it fails.

Critical weaknesses
  1. The major concern is how significant and useful this discovery is. In Figure 3, the vision-language model does not show very strong correlation in many tasks. In Figure 5 and Appendix G, this method does not demonstrate obvious advantage over the baseline. Especially for Qwen models, the performance even lags behind the baseline in many tasks.

  2. There is also concern about the generality of the conclusions. This paper only uses two model families. It would be more conclusive if more models can be examined to see if the conclusions hold, and how comparable with the baselines. The tasks used for experiments also constrain the generality of this paper. If I'm not mistaken, circuit analysis is not inherently restricted to these more toy-like task families, so testing whether the findings extend to more realistic tasks would strengthen the paper. I think even some preliminary results would be helpful.

Questions for authors

Please see weakness which covers my questions. The general question I would like to figure out is how significant and generalizable these conclusions can be.

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

This paper investigates whether circuit analysis measured before fine-tuning can predict transfer between source and target tasks. It extracts target-task circuits using AP-IG attribution and mean ablation, then maps these circuits onto source-task token positions. Corss-task circuit accuracy measures how accurate the target circuite performs the source task, relative to a size-matched random circuit, showing that a model's internal compuatation can decide target-task improvement.

Critical strengths
  1. The empirical evaluation is carefully controlled on 4 models, 18 tasks, covering textual and visual variants, across hyperparameter sweeps and different seed, with negative results reported transparently.
  2. The methodology is fundamentally motivated, with clear presentation and writing style.
Critical weaknesses
  1. The baseline comparison is limited to one task-relateness estimator and does not include broader transferability-prediction methods or simple controls, such as task-family, prompt-similarity, and source-accuracy predictors. The general advantage and novelty against existing alternatives are not throughly discussed.
  2. The relationship between XCA and transfer remains correlational, and may not extend to general out-of-distribution tasks or domains, which may limit the proposed methodology's applicability in real-world scenarios.
  3. Circuit extraction is not cost-free, but runtime, break-even cost, and mapping sensitivity are not reported.
Questions for authors
  1. Can the authors discuss whether XCA predicts transfer to genuinely out-of-distribution tasks, such as held-out task families or natural datasets, or more difficult agentic tasks. The current experiments remain within the same synthetic benchmark and therefore do not establish OOD predictive generalization, which is common in certain post-training scenarios.
  2. What is the computational cost compared with directly testing candidate source-task fine-tuning? If the model is fine-tuned and evaluated with the same amount of train and validate data, can we extrpolate the performance improvement to predict model behavior at full training-set fine-tuning?
  3. Can the authors discuss if the XCA's predicibility extends to performance after reinforcement learning, than fine-tuning, as the two post-training methodology are both popular and theoretically correlated between on-policy and off-policy gradients.
Reviewer_D51EOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper asks whether looking inside a model before fine-tuning can predict whether training on one task will help another. The authors find each task’s circuit, the small set of model parts that does the task, in four vision language models across 36 simple tasks: arithmetic, counting, grid lookup, and pattern matching, each with a text and an image version.

They first show that circuits change little during fine-tuning. They then propose cross-task circuit accuracy (XCA) and run the test task’s circuit on the training task and measure how well it does. XCA correlates with real fine-tuning gains, up to r = 0.77, and performs roughly on par with an existing method that compares the model’s final outputs. Predictions work best when circuits are well localized and fail when similar tasks need conflicting computations, such as counting shapes vs counting colors.

Critical strengths
  1. The paper uses circuit analysis to make predictions rather than to explain behavior after the fact. It supports this with a full cross task fine-tuning experiment across four models and 36 tasks, with five seeds per run.
  2. The authors show where XCA fails and why. Localization can be measured before training, so it could tell practitioners when to trust the prediction.
Critical weaknesses
  1. R up to 0.77 is the best of all correlations in Figure 3 and there is no overall average.
  2. The training tasks cluster by family and modality, so the correlations may mostly reflect similar tasks help each other, which a simple same family indicator would also predict. I don’t see this control.
  3. Gamma is the only baseline, and it often does better. On Gemma 3 4B it beats XCA on image based grid and pattern tasks.
  4. The paper never tests whether XCA adds information beyond gamma.
Questions for authors
  1. What are the average correlations per model?
  2. Do the correlations hold after controlling for task family and modality?
  3. Does XCA add predictive power beyond gamma?
  4. Do circuits stay stable when fine-tuning on a different task?

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

Related papers

Discussion (0)

Sign in to comment.