acceptodds
Under review as a conference paper at ICLR 2027

Dual Tuning: Joint Screening of Data Utility and Supervision Modes for Multimodal Fine-Tuning

Abstract

Multimodal fine-tuning requires decisions about both the training data and the form of supervision, such as chain-of-thought (CoT) or direct-answer (DA) supervision. These decisions are jointly shaped by the capability of the base model, the quality of CoT traces, and the characteristics of the target task, rather than by any fixed recipe. Consequently, isolated ablations insufficient for joint assessment. Here we introduce Dual Tuning, a screening protocol conditioned on the training configuration that addresses a complementary question: for a given base model and candidate dataset, which task groups benefit from training, and which supervision mode is better supported by the observed gains? It uses one joint supervised fine-tuning run per configuration to obtain gains over the base model and differences between CoT and DA modes for each task group. Experiments on spatial, mathematical, and multidisciplinary tasks reveal heterogeneous patterns across subgroups and models. Qwen2.5-VL-7B favors DA under the evaluated spatial configuration and CoT on MathVista. However, a CoT advantage over DA does not always coincide with a genuine improvement over the base model. The results on the Massive Multi-discipline Multimodal Understanding and Reasoning (MMMU) benchmark identify 9 subjects favoring CoT, 4 favoring direct answers, and 10 with negative gains in both modes. Finally, subset retraining experiments provide directional support for using these signals in data curation. Dual Tuning offers a common empirical basis for screening candidate data and assigning supervision modes before subsequent training.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multimodal Large Language Models, Post-Training, Fine-Tuning, Data Utility, Supervision Modes

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 3 reviews, above 27% of the others.
Reviewer_Ny9MOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Dual Tuning jointly assesses the utility of candidate training data and supervision preferences across task groups through a single multimodal fine-tuning run using both chain-of-thought (CoT) and direct-answer (DA) supervision. By combining performance gains over the base model with performance differences between the CoT and DA modes, the method provides a basis for data selection and supervision assignment. Experiments across tasks and models show that the screening outcomes depend on the specific training configuration, while subset retraining provides directional evidence supporting the use of these signals for subsequent data curation.

Critical strengths

The paper presents several interesting findings, suggesting that the advantages of CoT are strongly associated with task characteristics, data quality, and base model capabilities. On many spatial perception tasks, such as object counting and size estimation, as well as some multidisciplinary tasks, DA outperforms CoT while requiring less computation. These findings suggest that selecting appropriate supervision formats for different tasks may offer opportunities to improve model performance during post-training.

Critical weaknesses

The role of CoT data quality remains unclear. Could the performance degradation observed after training with CoT data be attributable to a distribution mismatch between the CoT data and the base model? It is also unclear how sensitive the experimental results are to differences in CoT data quality.

The transferability of screening outcomes across base models has not been sufficiently validated. The model comparisons do not cover different scales or versions within the same model family, leaving it unclear whether task-group supervision preferences and data selection decisions can be reused across such models.

Questions for authors

I would appreciate a detailed clarification of the following two issues:

  1. Changing the source of CoT data affects the screening outcomes, but it remains unclear whether these changes primarily arise from differences in reasoning trace quality or distribution mismatches between the teacher and the base model. Please clarify whether comparisons controlling for other variables have been conducted, and provide information on reasoning correctness and trace length to help explain the negative gains observed in the CoT mode on some tasks.

  2. Subset retraining provides directional evidence supporting the screening signals. However, the performance of a final model trained on the combined selected data, using the recommended supervision format for each task group, has not yet been evaluated.

I also consider the following question a worthwhile direction for further investigation:

  1. Different base models may exhibit different supervision preferences, but it remains unclear whether screening outcomes can be reused across models of different scales or versions within the same family. Please discuss this transferability and clarify whether experiments have examined whether using one model's screening outcomes to guide data selection and supervision assignment for another model preserves the corresponding performance gains.
Reviewer_NLz7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Dual Tuning, a configuration-dependent screening protocol for jointly assessing training-data utility and the choice between chain-of-thought (CoT) and direct-answer (DA) supervision. It jointly fine-tunes one model on paired CoT and DA examples, then evaluates the base and fine-tuned models under both prompting modes. Group-level improvements over the stronger base-model mode, together with differences between the trained modes, produce screening recommendations. Experiments cover spatial, mathematical, and multidisciplinary benchmarks with two model backbones. Additional subset-retraining experiments examine whether the initial screening signals agree with subsequent performance changes.

Critical strengths

(1) The paper studies a practically important problem which jointly assesses training-data utility and supervision choices for multimodal fine-tuning.

(2) The analysis across subjects and task categories reveals differences in training gains and supervision preferences that aggregate benchmark scores can obscure.

Critical weaknesses

(1) Evaluation independence is unclear. The screening criteria use gains measured on benchmark evaluation groups, but the paper does not specify whether screening and subsequent evaluation use disjoint examples. This raises concerns about evaluation-driven selection bias and leaves generalization to untouched data unverified.

(2) Comparisons with existing data-selection methods are missing. Related Work briefly mentions related work such as DoReMi, but offers limited discussion of recent multimodal data-selection approaches. No experimental comparisons establish the proposed protocol’s advantages over existing selection methods.

(3) The screening procedure needs clearer presentation. The notation BL,BS,DTL,DTSB_L,B_S,DT_L,DT_S is difficult to follow, and Algorithm 1 stops at assigning screening categories. How these decisions translate into concrete training subsets and supervision assignments should be explained more explicitly.

Questions for authors

(1) Have you compared the resulting models with recent state-of-the-art multimodal models on the same benchmarks, and could you include representative mainstream models to contextualize the reported performance?

(2) See weaknesses.

1 comment

Thanks for the clarification, but the paper still lacks a direct experimental comparison with existing data selection methods, so I will keep my original score.

Reviewer_W5zoOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Multimodal supervised fine-tuning requires deciding which training data to keep and which supervision mode—chain-of-thought (CoT) or direct answer (DA)—to use, jointly. We introduce Dual Tuning, a screening protocol that runs one joint SFT on paired CoT/DA targets and evaluates the base and fine-tuned models under both prompts for each task group, yielding per-group gains over the base model and CoT–DA mode gaps. Experiments on spatial, mathematical, and multidisciplinary benchmarks (VSI-Bench, CV-Bench, MathVista, MMMU) with Qwen2.5-VL-7B and Ming-lite-omni v1.5 show that supervision preferences are highly heterogeneous across task groups: a CoT advantage over DA does not always imply improvement over the base model, and screening decisions vary with the base model, CoT data, and training stage. Subset retraining provides directional support for using these signals in data curation, offering a practical basis for jointly screening data and assigning supervision modes before training.

The paper frames a genuinely useful and well-posed problem, proposes a clean and cheap protocol, and contributes an honest empirical study of configuration-dependent supervision heterogeneity—these are real merits, and my critical strengths reflect them. However, the central practical promise (guidance for data curation) rests only on directional, exception-laden evidence without end-to-end validation; the experimental backbone is thin (two models, no variance or significance analysis) and includes a confounded comparison; and no code/data are released to let the community adopt the protocol.

Critical strengths
  1. A clean, cost-effective screening protocol that asks the right joint question. One joint SFT run per configuration yields both "gain over base" and "mode difference" signals at task-group granularity. This is simple to adopt, and the framing (data utility and supervision mode must be assessed jointly *given the training configuration) is a genuine and well-articulated corrective to isolated ablations.

  2. Honest, informative empirical study of configuration-dependent heterogeneity. The paper delivers concrete, sometimes counter-intuitive evidence (e.g., MMMU's 9/4/10 split; MathVista "Logical Reasoning" with positive mode gap but negative gains in both modes; decisions flipping under a different base model, different CoT data, and after RL) that aggregates systematically hide group-level supervision preferences.

Critical weaknesses
  1. The central claim of practical value is not end-to-end validated. The paper's motivation is using the screen for data curation, but the only supporting evidence is sign-agreement in subset retraining with admitted exceptions, and no final mixture is ever trained and compared against baselines. As it stands, we do not know whether following Dual Tuning's recommendations actually improves downstream model quality over simpler heuristics (e.g., just training on everything, or an aggregate mode choice). This is consequential because it is the difference between a tool with proven utility and a plausibly-arranged protocol.

  2. The evaluation depth is thin relative to the strength of the empirical claims. Two base models (only one MoE variant), no variance/repeated-seed analysis, no statistical testing, and ad-hoc thresholds with dichotomous handling of near-tie cases. Combined with the confounded spatial comparison (joint image+video vs. separate single-modality baselines), the evidence supports "plausible patterns exist" but not yet "these patterns are stable and reliable enough to act on." This limits confidence in the method as a deployable screening tool.

Questions for authors

Clarifications:

  1. Does the subset-retraining protocol ever train the final reassembled mixture (as the paper says a future config would)? If not, what is the actual expected benefit over simply training on all data or applying an aggregate mode choice?
  2. What is the variance/seeding policy? Are the reported gains stable across random seeds, given that several decisions rest on small differences (e.g., Object Size GAP after RL = 0.07, Route Plan 0.00)?
  3. How reliable are the externally-API-generated group labels, and how sensitive are the screening outcomes to label noise or to the choice of task→subject mapping?

Comments on the paper

1 comment

Dear Reviewers,

Please carefully review the authors’ rebuttal and engage in further discussion with the authors to clarify any remaining concerns and questions. Your active participation during the discussion period is greatly appreciated.

Best,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.