Technical Summary: MLLMs Get It Right, Then Get It Wrong: Tracing and Correcting Late-Layer Textual Bias

1. Problem Statement

Multimodal Large Language Models (MLLMs) exhibit a consistent bias toward textual information when it contradicts visual evidence. Even when an image provides clear proof of a fact (e.g., a red car), MLLMs often output the answer suggested by conflicting text (e.g., "blue car"). This "textual override" poses significant risks for safety-critical applications like medical diagnosis or autonomous driving, where trusting erroneous text over visual perception can lead to serious errors.

Prior work has documented this phenomenon but has not clarified the underlying mechanism. It remains unclear whether the failure stems from poor initial visual encoding, loss of information during cross-modal fusion, or the correct processing of visual data followed by its subsequent overwriting by textual priors in deeper layers.

2. Methodology and Analysis

2.1 Layer-wise Tracing and Modal Dominance Ratio (MDR)

To investigate the failure mechanism, the authors trace prediction trajectories across model layers. They project hidden states at each layer $l$ to output probabilities using the language model head. They introduce the Modal Dominance Ratio (MDR) to quantify the preference for visual ($w_{vis}$) versus textual ($w_{text}$) answers at each layer:
$$\text{MDR}^{(l)} = \log \frac{P^{(l)}_t(w_{vis})}{P^{(l)}_t(w_{text})}$$

• Positive MDR: Preference for the visual answer.
• Negative MDR: Preference for the textual answer.

2.2 The "Late-Layer Textual Override" Phenomenon

Analysis of Conflict-Incorrect samples (where the model fails) reveals a surprising pattern:

1. Initial Correctness: In intermediate layers, the model assigns a higher probability to the correct visual answer (Positive MDR).
2. The Override: In later layers, the MDR flips to negative, and the model shifts its preference to the incorrect textual answer.
3. Conclusion: The visual information is successfully encoded but fails to survive the full forward pass due to late-layer processing.

2.3 Directional Signature of Correctness

The authors observe that prediction shifts occur in both successful and failed cases. However, the direction of the shift distinguishes them:

• Failures (Conflict-Incorrect): 85% of transitions shift toward text ($\Delta \text{MDR} < 0$).
• Successes (Conflict-Correct & Non-Conflict): 89–91% of transitions shift toward vision or remain stable ($\Delta \text{MDR} > 0$).

This directional asymmetry suggests that late layers sometimes refine predictions (beneficial) but sometimes suppress visual evidence (harmful).

2.4 Conflict-Aware Layer Reference Decoding (CALRD)

Based on these findings, the authors propose CALRD, a training-free, inference-time intervention method. It does not require external knowledge or retraining.

Detection Mechanism:
CALRD identifies "harmful overrides" using two signals derived from the transition layer ($L_{trans}$), defined as the layer with the sharpest distributional shift (detected via Jensen-Shannon Divergence):

1. Anchor Confidence ($D_{conf}$): The probability of the top prediction at the transition layer. High confidence indicates a strong intermediate preference.
2. Prediction Retention ($\rho$): The ratio of the final layer's probability to the transition layer's probability for the same token. Low retention indicates the prediction was suppressed.

Intervention Logic:
A correction strength $\lambda$ is calculated as:
$$\lambda = D_{conf} \cdot (1 - \rho)$$

• If a confident prediction exists ($D_{conf}$ high) but is suppressed ($\rho$ low), $\lambda$ becomes large.
• If the prediction is stable or uncertain, $\lambda$ remains near zero.

Decoding Strategy:
The final output logits are adjusted by blending the transition-layer logits with the final-layer logits:
$$\phi'_t = (1 - \lambda) \cdot \phi^{(N)}_t + \lambda \cdot \phi^{(L_{trans})}_t$$
This restores the vision-grounded answer that was present in intermediate layers but overwritten in the final output.

3. Key Contributions

1. Mechanistic Discovery: The paper identifies "late-layer textual override" as the primary cause of visual-textual conflict failures, demonstrating that models often "get it right" initially before changing their minds.
2. Directional Predictor: The authors show that the direction of prediction shifts (toward text vs. toward vision) is a reliable predictor of correctness, distinguishing harmful overrides from beneficial refinements.
3. CALRD Method: A training-free decoding strategy that selectively recovers overridden predictions using internal model signals, preserving standard performance while correcting conflicts.
4. Conflict-VQA Benchmark: The construction of a new diagnostic benchmark (5,969 samples) with explicit visual-textual annotations to facilitate mechanistic analysis of conflict resolution.

4. Experimental Results

The authors evaluated CALRD on five MLLMs (InstructBLIP, LLaVA-1.5/1.6, Qwen2.5-VL, Qwen3-VL) across varying architectures.

• Conflict Benchmarks: CALRD achieved up to 9.4% absolute improvement on conflict tasks (e.g., PhD-icc and Conflict-VQA).
  • On PhD-icc, LLaVA-1.5 improved from 27.72% to 36.35% (+8.63%).
  • Improvements were generally larger for weaker models, as stronger models already resolve more conflicts correctly in their late layers.
• Standard Benchmarks (Non-Conflict): CALRD largely preserved or slightly improved performance on standard hallucination benchmarks (POPE and CHAIR), demonstrating that the method does not indiscriminately degrade performance on non-conflict tasks.
• Comparison: CALRD outperformed existing layer-based methods like DoLa and DeCo, which often apply uniform corrections and can degrade performance when late-layer processing is actually beneficial.
• Efficiency: CALRD adds minimal overhead (~1.5ms/token and 80MB memory) compared to vanilla decoding, making it practical for deployment.

5. Significance and Claims

The paper claims a shift in perspective for improving MLLM reliability: rather than focusing solely on teaching models to "see better" (better visual encoding), reliability may be improved by helping models retain what they have already seen.

The authors emphasize that their method is modest in scope:

• It addresses a specific failure mode (late-layer override) rather than all hallucinations.
• It relies on the model's own internal representations, requiring no external knowledge bases or additional training data.
• The intervention is adaptive; it only activates when the specific signature of harmful override (high confidence + low retention) is detected, ensuring that beneficial late-layer processing is not disrupted.

The work suggests that the "blind trust in text" observed in MLLMs is not necessarily a lack of visual understanding, but a failure of information preservation through the network depth. CALRD provides a mechanism to recover this lost information.