Technical Summary: Instruction Lens Score for Object Hallucination Detection

Problem Statement

Multimodal Large Language Models (MLLMs) have demonstrated significant capabilities in visual understanding and reasoning. However, a critical barrier to their reliable deployment is object hallucination (OH), where models generate descriptions of objects that are not present in the input images. Existing detection methods face two primary limitations:

1. Auxiliary Model Dependence: Methods relying on external models (e.g., GPT-4) or dedicated hallucination detectors incur substantial computational overhead and require additional training.
2. Vulnerability to Misleading Visual Signals: Vision-based detection scores (e.g., attention weights, local similarity) often suffer from spurious high confidence when visual encoders introduce misleading information or when local visual patterns of hallucinated objects resemble real ones (e.g., similar textures for spoons and knives).

Methodology: Instruction Lens Score (InsLen)

The paper proposes Instruction Lens Score (InsLen), a training-free, plug-and-play detector that leverages the internal embeddings of instruction tokens to filter misleading visual information. The core insight is that instruction embeddings implicitly encode visual information grounded in the image while effectively suppressing erroneous cues introduced by visual embeddings.

The method utilizes the Logit Lens technique to project intermediate layer embeddings into the vocabulary space to analyze confidence distributions. InsLen comprises two complementary components:

1. Calibrated Local Score ($S_{cls}$)

This component addresses the issue of spurious overconfidence in vision-based scores.

• Mechanism: It introduces a Calibration Confidence ($S_{cafe}$) by measuring the maximum confidence assigned to an object token across all instruction embeddings.
• Logic: Since instruction embeddings filter out misleading visual cues, hallucinated objects receive lower confidence in this projection.
• Fusion: $S_{cafe}$ is multiplicatively fused with existing vision-based scores (e.g., Local Similarity Score, Internal Confidence) to calibrate local visual evidence:
  $$S_{cls}(o) = S_{cafe}(o) \cdot S_{local}(o)$$

2. Context Consistency Score ($S_{ccs}$)

This component addresses the limitation of patch-level visual embeddings, which may fail to distinguish objects with similar local textures.

• Mechanism: It aggregates instruction embeddings that assign high probability to the object token to form a global contextual representation ($\bar{z}$).
• Logic: Instruction embeddings capture global object context. The consistency between the generated object token embedding ($h_o$) and the aggregated instruction context ($\bar{z}$) is measured using a normalized $\ell_2$ distance.
• Weighting: The consistency score is weighted by the average confidence of the selected instruction embeddings to ensure reliability:
  $$S_{ccs}(o) = S_{con}(o) \cdot p(o|\{\hat{z}_1, ..., \hat{z}_m\})$$

Final Scoring

The final InsLen score is a weighted combination of the Calibrated Local Score and the Context Consistency Score:
$$S_{Ins}(o) = \omega \cdot S_{cls}(o) + (1 - \omega) \cdot S_{ccs}(o)$$
where $\omega$ is a hyperparameter.

Key Contributions

1. Analysis of Instruction Embeddings: The authors reveal that instruction embeddings in MLLMs implicitly encode reliable visual semantics and can effectively filter out misleading information originating from visual embeddings, a capability previously overlooked for hallucination detection.
2. Proposed Detector (InsLen): A novel, training-free detector that integrates calibrated visual evidence with instruction-based context consistency without requiring auxiliary models or additional training.
3. Comprehensive Evaluation: Extensive experiments across multiple benchmarks (MSCOCO, Objects365, POPE, CLEVR) and diverse MLLM architectures (LLaVA-1.5, InstructBLIP, mPLUG-Owl3, Qwen3-VL, LLaVA-OneVision) demonstrate the method's effectiveness.

Experimental Results

• Performance: InsLen consistently outperforms existing state-of-the-art methods (e.g., GLSIM, SVAR, Internal Confidence) across all tested benchmarks and model architectures.
  • On the MSCOCO benchmark with Qwen3-VL, InsLen improved AUROC by 7.70% over the strongest baseline (GLSIM).
  • On the POPE benchmark, it achieved up to 13.81% higher AUROC than baselines for LLaVA-1.5.
  • On the CLEVR benchmark (testing attribute and relational hallucinations), InsLen showed superior performance, particularly on stronger models like Qwen3-VL (77.72% AUROC vs. 73.01% for SVAR).
• Robustness: The method remains effective on post-trained models (e.g., LLaVA-RLHF, LLaVA-RLAIF-V) and demonstrates stability across different input instruction lengths and hyperparameter settings ($\omega$, $m$, $\tau$).
• Efficiency: While introducing slightly higher computational overhead than some baselines (e.g., ~104ms vs. 95ms for LLaVA-1.5), the cost is minimal compared to answer generation and significantly lower than methods requiring auxiliary models or zeroing operations (e.g., EAZY).

Significance and Claims

The paper claims that Instruction Lens Score provides a valuable and underexplored signal for improving the reliability of MLLM generation. By demonstrating that instruction embeddings can effectively suppress misleading visual information, the work offers a new perspective on hallucination detection that does not rely on external resources or retraining. The authors position InsLen as a robust, plug-and-play solution that enhances the trustworthiness of MLLMs in real-world applications where object hallucination poses significant risks.