- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper investigates modality interference between textual reasoning and visual perception objectives in multimodal large language models (MLLMs). Through preliminary experiments, the authors observe that naively combining textual reasoning distillation with multimodal supervision produces non-additive performance gains and negatively aligned gradients.
To address this issue, the paper proposes UPRSD (Unified Perceptual and Reasoning via Staged Distillation), a three-stage training framework:
- Structural Priming: Separates reasoning and perception behaviors through prefix-conditioned supervision and bootstraps multimodal reasoning trajectories.
- Logic Integration:Unifies reasoning supervision through a shared reasoning interface while using task-dependent routing to avoid unnecessary deliberation.
- Reinforced Evolution:Applies GRPO with verifiable rewards and potency-aware data selection to improve reasoning performance.
The framework is evaluated using Uni-VL models initialized from Qwen3-VL at 4B and 8B scales. The reported results demonstrate substantial improvements over several open-source baselines on multimodal and textual reasoning benchmarks while maintaining competitive general multimodal understanding.
The paper further provides gradient-alignment analysis, stage-wise ablations, and training dynamics to support its proposed approach.
Critical strengths
Strength 1: Strong and broad empirical improvements across reasoning modalities.
The paper demonstrates substantial performance improvements across a diverse collection of textual and multimodal reasoning benchmarks.
In particular, Uni-VL-8B achieves a multimodal reasoning average of 57.0, compared with 46.2 for the strongest open-source baseline in the same model-size group. Its textual reasoning average also reaches 60.9, compared with 31.0 for the strongest corresponding baseline.
The improvements at both 4B and 8B scales strengthen the empirical contribution. These results suggest that the proposed training framework can effectively improve multiple capabilities without an obvious general degradation in the reported multimodal benchmarks.
Strength 2: A practically relevant training framework supported by informative diagnostics.
The paper addresses an important challenge in multimodal model training: simultaneously improving deliberative reasoning and visual understanding.
The combination of prefix-conditioned training, multimodal trajectory bootstrapping, unified reasoning supervision, and reinforcement learning provides a coherent practical training recipe.
The gradient-alignment analysis is also valuable. The reported change in mean gradient cosine similarity from approximately -0.12 under naive mixing to +0.18 after the proposed staged training is consistent with the authors' interference hypothesis.
Together with the stage-wise experiments, this provides a useful empirical starting point for understanding reasoning–perception interactions.
Critical weaknesses
Weakness 1: Insufficient causal isolation of the proposed modality-interference mitigation mechanism.
The paper attributes its improvements primarily to mitigating interference between reasoning and perception objectives. However, the experimental evidence does not sufficiently isolate this mechanism from other factors.
First, Stage 2 simultaneously introduces a unified reasoning interface and additional bootstrapped multimodal reasoning trajectories. Consequently, the improvements after Stage 2 cannot be attributed specifically to the proposed integration strategy.
Second, the gradient-alignment analysis compares different training checkpoints after different optimization histories. The observed shift from negative to positive gradient similarity is informative but does not establish that the reduced interference causes the downstream improvements.
Third, the paper does not clearly establish a training-token- and compute-matched comparison between the complete proposed pipeline and a conventional mixed-supervision pipeline using the same data and reinforcement-learning budget.
These limitations are consequential because the paper's central claim is not simply that staged training performs well, but that mitigating modality interference is the key mechanism responsible for its effectiveness.
Weakness 2: Limited independent validation of the routing and potency-aware reinforcement-learning components.
Two important components of the proposed framework remain insufficiently evaluated.
The inference-time router is a text-only classifier trained using domain-based reasoning/perception labels. Although the reported classification accuracy is high, this evaluation may not adequately capture visually dependent or ambiguous queries for which the appropriate reasoning behavior depends on the image content. A controlled comparison against always-on reasoning and direct-response inference would help establish its actual end-to-end contribution.
Similarly, the effectiveness of potency-aware data selection is not isolated from standard GRPO using random or difficulty-matched training examples. The description of the held-out mining pool and its relationship to the eventual RL training prompts also requires clarification.
Without these analyses, it is difficult to determine whether these components contribute the claimed efficiency and robustness improvements or whether most gains arise from additional supervised data and reinforcement-learning optimization.
This weakens the evidence supporting the distinct contributions of the complete framework.
Questions for authors
Q1. [Clarification and additional evidence — training-budget fairness]
Could the authors clarify the number of training examples, total training tokens, and optimizer updates used by the naive-mixed baseline and each stage of UPRSD?
In particular, is there a baseline that performs conventional mixed-supervision SFT followed by GRPO using the same backbone, training data, total optimization budget, and reward configuration as the complete proposed pipeline?
If such a comparison is available, reporting it would materially strengthen the evidence for the benefits of staged training.
Q2. [Additional evidence — mechanism attribution]
Could the authors disentangle the effects of the unified reasoning interface and the additional bootstrapped multimodal trajectories introduced in Stage 2?
For example, how does a model trained with the additional trajectories but without the proposed interface integration compare with the complete Stage-2 model?
It would also be useful to report gradient alignment for controls with more closely matched training histories. These results would materially affect my assessment of the central interference-mitigation claim.
Q3. [Clarification and additional evidence — routing and RL data selection]
Could the authors clarify how potency scores computed on a held-out mining pool are transferred to the selection of RL training prompts, given that the mining pool is described as disjoint from those prompts?
Additionally, could the authors report the performance of standard GRPO without potency mining under the same training budget, and compare the learned router against always-on reasoning on visually ambiguous queries?