Technical Summary: CAVE – A Structured Credit Assignment Approach for Fragmented Visual Evidence Reasoning

1. Problem Definition: Fragmented Visual Reasoning (FVR)

While Vision-Language Models (VLMs) have advanced significantly in general multimodal reasoning, they exhibit brittleness when tasks require integrating nonlocal visual information with low semantic priors. The authors define this challenge as Fragmented Visual Reasoning (FVR). In FVR, task-critical evidence is spatially fragmented and semantically weakly separable across different image regions.

Current VLMs often fail in this setting because:

• Coarse Global Perception: Relying on global representations fails to capture long-range dependencies and subtle local differences.
• Verbalized Bias: Models tend to bias reasoning toward a verbalized subset of evidence, producing coherent but visually ungrounded chains.
• Error Propagation: Early errors in evidence selection lead to cascading failures in subsequent reasoning steps.

Existing interleaved reasoning methods (e.g., "Think-with-Image") allow for localized re-observation but lack action-specific process supervision. Current process-reward methods typically evaluate intermediate steps based on final-answer improvement or overall trajectory quality, failing to provide targeted guidance for specific reasoning behaviors like evidence acquisition or region focusing.

2. Methodology: CAVE (Credit Assignment for Visual Evidence)

To address these limitations, the authors propose CAVE, a structured process-reward method based on Group Relative Policy Optimization (GRPO) designed for interleaved visual reasoning.

2.1 Interleaved Reasoning Trajectory

The model's solving process is modeled as a trajectory $\tau = \{(M_t, a_t, o_t)\}_{t=1}^{T_\tau}$, where:

• $M_t$: Multi-modal context state (original image, question, history, local observations).
• $a_t$: Assistant action (textual reasoning, zoom-tool call, or final answer).
• $o_t$: Environment observation (e.g., zoomed sub-image).

2.2 Structured Process Credits

Unlike single state-utility rewards, CAVE decomposes the value of intermediate steps into three complementary, action-level signals:

1. Belief Update Credit ($c_{bu}$):

   • Goal: Measure progress toward the correct answer.
   • Mechanism: Calculates the gain in the conditional log-likelihood of the ground-truth answer body ($A$) between state $M_t$ and $M_{t+1}$ using teacher forcing.
   • Function: Rewards steps that move the model's state closer to the correct answer, constraining answer-oriented state updates.

2. Evidence Acquisition Credit ($c_{ea}$):

   • Goal: Capture precursor progress where critical visual clues become available before they are fully integrated into the answer belief.
   • Mechanism: Uses a set of $K$ textualized visual evidence units ($P = \{p_1, ..., p_K\}$) annotated for the task (e.g., "triangular building"). It measures the recoverability score gain of these units between states.
   • Function: Rewards trajectories that progressively acquire necessary visual details, distinguishing them from blind guessing.

3. Adaptive Focus Control Credit ($c_{af}$):

   • Goal: Constrain zoom-based visual exploration to balance evidence coverage and visual noise.
   • Mechanism:
     • Estimates state uncertainty ($u_t$) via the entropy of answer tokens.
     • Defines a target zoom scale $\hat{\rho}_t$ based on uncertainty.
     • Gates the reward: A zoom action is only rewarded if it yields positive belief or evidence gains ($g_t$) AND its scale matches the current uncertainty ($1 - |\rho_t - \hat{\rho}_t|$).
   • Function: Encourages the model to select local observations that induce state progress while avoiding irrelevant noise.

2.3 Optimization

The total reward is a weighted sum of the aggregated trajectory-level credits:
$$R_{CAVE}(\tau) = \lambda_{bu}C_{bu}(\tau) + \lambda_{ea}C_{ea}(\tau) + \lambda_{af}C_{af}(\tau)$$
The policy is optimized using GRPO, computing group-relative advantages based on $R_{CAVE}$. Optimization is applied only to assistant-generated tokens; tool observations serve as context but do not participate in the policy gradient.

3. Benchmark: TRACER-Bench

To enable controlled evaluation and process-level analysis, the authors construct TRACER-Bench (Text-Resistant Cross-Region Evidence Reasoning).

• Scope: 980 VQA samples covering four scenarios:
  1. Rule-Switching Navigation (VJump): Requires maintaining state across rule switches.
  2. Nonsemantic Tracing (LT): Tracing paths without semantic cues.
  3. Embedded Matching (Match): Fine-grained template-grid correspondence.
  4. Remote-Sensing Subimage Matching (MME-RS): Matching patches in real-world noisy images.
• Key Feature: Each sample includes a perception field containing key intermediate visual evidence (textualized) used for offline analysis and reward computation, but not provided as model input during inference.
• Isolation: The training corpus is strictly isolated from the test set regarding random seeds, layouts, and instances to prevent leakage.

4. Experimental Results

The authors evaluate CAVE (based on Qwen2.5-VL) against various baselines, including general VLMs, reasoning-oriented models, and tool-augmented methods.

• Fragmented Vision Performance: CAVE achieves the best overall average performance among comparable open-source baselines.
  • On TRACER-Bench, CAVE-7B improves the average score by +15.5% over the base Qwen2.5-VL-7B.
  • It significantly outperforms specialized reasoning models (e.g., ThinkLite-VL, VLAA-Thinker) and tool-assisted methods (DeepEyes).
  • An ablation study ("Belief-Only") shows that while answer-belief gain is effective, it is insufficient without evidence acquisition and focus control, confirming the necessity of the three-part structured credit.
• General Multimodal Performance: CAVE retains competitive performance on general benchmarks (MMStar, RealWorldQA, HallusionBench, etc.), achieving top or near-top results on 4 of 6 benchmarks. This indicates that optimizing for visual evidence exploration does not degrade general capabilities.
• Robustness: CAVE demonstrates slower performance degradation as reasoning depth (hops) and cross-region distance increase, suggesting improved stability in long-range dependencies.
• Efficiency: CAVE achieves higher accuracy with comparable or lower token/tool budgets than baselines, evidenced by a higher "useful zoom rate" (proportion of zooms yielding positive gains).

5. Significance and Claims

The paper claims that CAVE addresses a specific gap in VLM reasoning: the inability to reliably integrate spatially fragmented and semantically weak evidence.

• Structured Credit Assignment: The primary contribution is the decomposition of intermediate reasoning value into belief update, evidence acquisition, and adaptive focus control. This provides targeted, action-level supervision that existing trajectory-level rewards miss.
• Process-Oriented Optimization: The results suggest that the performance gains stem not from larger model scales or specific tool formats, but from the structured modeling of the visual evidence progression process.
• Generalizability: The method improves performance on fragmented tasks while preserving general multimodal capabilities, indicating that the learned ability to explore and integrate visual evidence is transferable.
• Benchmarking: TRACER-Bench provides a necessary framework for evaluating cross-region dependency reasoning with verifiable intermediate evidence chains, addressing the lack of process-evidence analysis in existing benchmarks like Tunnel Vision or BabyVision.

The authors conclude that CAVE effectively guides models to discover task-critical evidence, update reasoning states, and focus on informative regions, leading to more robust visual reasoning under conditions of low semantic separability and long-range dependencies.