acceptodds
Withdrawn from ICLR 2027

DataValve: Selecting-while-Training for Efficient Visual Instruction Tuning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Visual instruction tuning has become increasingly expensive as multimodal large language models (MLLMs) and training corpora scale. Existing data selection methods mostly follow an offline select-then-train pipeline: they score a large candidate pool, often with an additional target MLLM traversal, and then fine-tune on a fixed subset. This pipeline is both costly and blind to the evolving training state, so it cannot avoid samples whose marginal utility has already decayed. We propose DATAVALVE, a selecting-while-training framework for fixed-pool visual instruction tuning. A lightweight router selects training samples from pre-extracted features and adapts through delayed golden-set loss feedback. To counter the tendency of loss-only feedback to repeatedly favor high-loss sources, an online LoRA-influence gate modulates credit using relative update strength. This coupling prioritizes currently useful samples and suppresses redundant low-impact updates without external teachers, downstream validation sets, or explicit diversity losses. Experiments on LLaVA-v1.5-7B show that DATAVALVE routes only 20% of training data through target MLLM training while achieving 98.4% of full-data fine-tuning performance, outperforming strong baselines under the same routing budget with the best efficiency–accuracy balance. Further experiments demonstrate transfer across candidate pools, model scales, and MLLM families. Our code is included in the supplementary material.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Data Selection, Selecting while training, Visual instruction tuning

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

This paper proposes DATAVALVE, a framework that selects visual instruction-tuning data during model training. It contains a lightweight router and an online LoRA-influence gate, which prioritizes currently useful samples and suppresses redundant ones. Experiments show that it can achieve 98.4% with 20% of the data.

Critical strengths
  1. The paper addresses the limitations of traditional select-then-train paradigms by introducing an adaptive data selection framework that accounts for evolving sample utility during multimodal instruction tuning.
  2. DataValve integrates lightweight routing into the training process, avoiding an additional full-pool target-model scoring stage and achieving a favorable trade-off between computational efficiency and downstream performance.
  3. The combination of golden-set loss feedback and LoRA-based gradient information provides an interesting approach to adaptive sample selection, mitigating the over-selection of high-loss data sources without explicit diversity regularization.
  4. The method achieves 98.6% of full-data performance using only a 20% training data budget, with additional experiments demonstrating its applicability across different datasets and model architectures.
Critical weaknesses
  1. The proposed framework mainly combines existing techniques, including lightweight routing, held-out loss feedback, gradient-based reward modulation, and policy-gradient optimization. The core technical contribution appears largely heuristic.
  2. Golden-set loss changes are measured after multiple model updates, making it difficult to attribute performance improvements to individual selected batches. This may introduce noisy or misleading rewards during router optimization.
  3. Gradient energy does not necessarily reflect actual sample utility or redundancy. The paper lacks sufficient theoretical and empirical evidence establishing a reliable relationship between the proposed gradient-based gate and marginal training contributions.
  4. The paper lacks comparisons between frozen and continuously updated routers, as well as systematic evaluations against simpler gradient-based reward alternatives. These experiments are important for validating the necessity and effectiveness of the proposed components.
  5. The claimed "selecting-while-training framework" has been widely investigated by dynamic data selection methods. Thus, the novelty degrades.
Questions for authors

See weakness.

Reviewer_8oVpOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DataValve, an online data selection framework for efficient visual instruction tuning. A lightweight router selects training examples using pre-extracted multimodal features and is updated through delayed loss feedback from a held-out subset. A gate based on relative LoRA gradient strength modulates this feedback to mitigate over-selection of particular data sources. On LLaVA-v1.5-7B, the method achieves 98.6% of the average relative performance of full-data tuning while using a 20% training-data budget. Additional experiments evaluate the approach across datasets and model backbones.

Critical strengths

Strengths

  1. The paper is generally clear and easy to follow. The framework and its main components are presented in an accessible manner.
  2. The problem is meaningful, given the computational cost of visual instruction tuning. The proposed online selection mechanism directly addresses the limitations of static data selection, and the motivation and overall design are logically connected.
  3. The method achieves better average relative performance than the compared selection baselines under the same training-data budget. Experiments across different datasets and model backbones provide additional evidence beyond the main LLaVA-7B setting.
Critical weaknesses

Weaknesses

  1. The evaluation is restricted to LoRA tuning. No full-parameter fine-tuning results are provided, although the selection mechanism explicitly relies on LoRA gradient statistics. It remains unclear whether data preferences learned under parameter-efficient tuning also hold under full-parameter tuning, where the trainable parameter space and optimization dynamics differ. Evaluating this setting would help establish the scope of the method.
  2. The training data and task coverage are limited. Both LLaVA-665K and Vision-Flan-186K are relatively old instruction-tuning datasets, and the evaluation primarily uses established benchmarks. The paper does not demonstrate whether its benefits extend to more recent training corpora or more demanding multimodal training tasks. This limits the evidence for its relevance to current training settings.
  3. The practical significance of the Qwen3-VL-4B experiment is unclear. Qwen3-VL-4B-Instruct has already undergone multimodal instruction tuning, yet the paper further tunes it on LLaVA-665K without reporting the performance of the original checkpoint. Consequently, it is unclear whether DataValve improves the model or merely causes less degradation than full-data tuning. The reported advantage over full-data tuning alone does not establish that this additional training is useful. A more informative test of generality would vary the vision encoder and language backbone within a controlled framework with comparable training stages.
Questions for authors

See weakness above.

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

This paper introduces DataValve, an online data-selection method for visual instruction tuning. A lightweight router selects training examples using pre-extracted multimodal features and learns from delayed loss feedback on a held-out golden set. A gate based on LoRA-gradient statistics modulates this feedback before the router update. On LLaVA-665K, DataValve achieves 98.6% of full-data performance, averaged across normalized benchmark scores, using a 20% training budget. Additional experiments cover a second data pool and multiple backbones, with ablations examining key components.

Critical strengths
  • The paper is overall clear and well-written.
  • The proposed method targets a practical limitation of offline selectors by integrating data selection into training and avoiding a separate target-model scoring pass over the full candidate pool.
  • Experiments cover three backbones and two data pools, providing evidence of applicability beyond a single training configuration.
Critical weaknesses
  • Comparisons on additional backbones are limited. On LLaVA-v1.5-13B and Qwen3-VL-4B-Instruct, DataValve is compared only with full-data training and random selection. It would be more convincing to compare them with 1-2 additional data selection methods.

  • In Table 2, the no-gate variant achieves the same average relative performance as random sampling in Table 1. This highlights the importance of the gate but leaves unclear whether golden-set feedback provides additional benefit. Further clarification is needed.

  • The experiments do not isolate the benefit of online adaptation or establish its robustness. Figure 3 shows changes in source proportions, but source imbalance alone does not demonstrate redundancy or explain the performance degradation. It remains unclear whether adapting the router during training offers benefits beyond obtaining a better data mixture.

Questions for authors

The OpenReview abstract reports 98.4% of full-data performance, whereas the manuscript reports 98.6%. Could you clarify this discrepancy and confirm which value is correct?

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

Related papers

Discussion (0)

Sign in to comment.