Technical Summary: EASE (Evidence-Anchored Spatial Attention) for Multimodal RLVR

1. Problem Statement

Reinforcement Learning with Verifiable Rewards (RLVR) has become a standard paradigm for improving Vision-Language Models (VLMs) by optimizing outcome rewards derived from final answers. However, the authors identify a critical limitation in outcome-only RL: it fails to supervise the visual evidence acquisition process.

When a VLM answers a question correctly, an outcome reward does not distinguish between a response grounded in relevant visual evidence and one derived from language priors, dataset biases, or lucky guesses. This creates a "perception credit assignment problem," where the model is not penalized for ignoring the image regions that justify the answer. Consequently, models may produce plausible text while remaining weakly grounded, leading to increased risks of hallucination, particularly in tasks requiring fine-grained perception, object counting, or multi-region reasoning.

2. Methodology: EASE

The authors propose EASE (Evidence-Anchored Spatial Attention), a framework that augments multimodal RLVR with visual-evidence process supervision. EASE operates under the principle that high-reward trajectories should not only yield correct answers but also demonstrate attention aligned with the visual evidence supporting those answers.

Core Mechanism

• Evidence Annotation: The method utilizes annotated bounding boxes (evidence regions) from training data. These annotations are treated as privileged training labels used solely to construct an auxiliary target distribution; they are not provided as input during inference.
• Smoothed Target Distribution: EASE converts annotated evidence boxes into a soft, smoothed target distribution over visual tokens. Each box is modeled as a Gaussian component centered at the box's midpoint, with a covariance that ensures most probability mass lies within the annotated region while tolerating boundary noise. For multi-evidence questions, an equal-weight mixture of these Gaussians is used, further smoothed with a uniform background distribution to avoid over-constraining context.
• Reward-Aware Gating: The attention supervision is applied only to trajectories that receive high verifier rewards (i.e., correct answers). The authors argue that low-reward responses often exhibit unreliable attention patterns, and forcing alignment on incorrect answers could introduce noise.
• Attention Supervision: During the actor update, EASE extracts the response-to-vision attention from a semantic middle layer of the transformer (specifically layer $\ell^* = \lfloor 2L/3 \rfloor$). It computes the Kullback-Leibler (KL) divergence between the model's attention distribution and the evidence-anchored target. This KL loss acts as an auxiliary term added to the standard RL objective (e.g., DAPO or GRPO).

Training and Inference

• Training: The model is trained on a balanced mixture of single-evidence and multi-evidence examples. The auxiliary loss guides the model to focus on relevant regions for successful trajectories.
• Inference: The inference process remains unchanged. The model receives only the original image and question, maintaining the same architecture and decoding procedure as the underlying VLM without requiring evidence metadata.

3. Key Contributions

1. Identification of a Missing Signal: The paper identifies visual evidence acquisition as a missing process signal in multimodal RLVR and introduces EASE to address the perception credit assignment problem without requiring evidence metadata at inference.
2. Process Supervision via Attention: The authors propose a method to map annotated evidence regions to a smoothed target distribution and apply this guidance specifically to high-reward trajectories, ensuring that correct answers are supported by correct visual focus.
3. Empirical Validation: EASE is validated across multiple VLM backbones (Qwen2.5-VL-7B, Qwen3-VL-4B, and Qwen3-VL-8B) and diverse benchmarks. Diagnostics confirm that EASE improves the alignment between visual attention and annotated evidence regions.

4. Experimental Results

The paper reports consistent improvements over outcome-only RL baselines (GRPO and DAPO) across three model scales:

• Performance Gains: On the Qwen2.5-VL-7B backbone, EASE improves the average score over DAPO by 2.9 points. On Qwen3-VL-4B and Qwen3-VL-8B, the improvements are 3.1 and 2.5 points, respectively.
• Benchmark Specifics: Gains are most pronounced in perception-heavy reasoning, hallucination robustness, visual math, and logic benchmarks. For instance, on Qwen2.5-VL-7B, EASE improves performance on MathVista by +6.5 points and WeMath by +4.9 points compared to DAPO.
• Competitiveness: EASE-7B achieves the best average score among compared public 7B-scale multimodal reasoning methods on shared benchmarks (e.g., MathVista, LogicVista).
• Diagnostics: Ablation studies and diagnostic metrics (Attention Mass, Pointing Accuracy, Multi-evidence Coverage) demonstrate that EASE successfully increases the amount of attention falling within annotated evidence regions and improves the coverage of multiple supporting regions.

5. Significance and Claims

The authors claim that EASE demonstrates that supervising visual evidence acquisition is a practical and effective complement to final-answer rewards for achieving more grounded multimodal reasoning. By encouraging models to "look" at the right parts of the image when they are correct, EASE reduces reliance on language shortcuts and mitigates hallucination.

The paper positions EASE as a method that improves the internal reasoning process (specifically spatial attention) without altering the inference interface. The authors note that while the method relies on evidence annotations during training, it does not require them at inference, making it applicable to standard VLM deployment scenarios.

Limitations Acknowledged

The paper modestly acknowledges several limitations:

• Data Construction Cost: Obtaining reliable evidence boxes requires additional annotation effort.
• Scope of Evidence: The method is best suited for tasks where evidence can be localized to specific image regions. It may be less effective for holistic, stylistic, or globally distributed cues.
• Causal Measure: The authors clarify that attention diagnostics serve as evidence of improved spatial grounding but are not a complete causal measure of visual reliance.

In summary, EASE represents a shift from purely outcome-based reinforcement learning to process-aware supervision, ensuring that the path to a correct answer involves the correct visual evidence.