TL;DR
This paper introduces HalluTrace, a novel framework that attributes object hallucinations in vision-language models to three distinct causes: visual grounding failure (VGF), language prior dominance (LPD), and cross-modal conflict (CMC). By understanding these sources, they develop Hallucination-Aware Decoding (HAD), a targeted intervention strategy that significantly reduces hallucinations and outperforms existing methods without retraining.
What Problem Does This Paper Address?
Large Vision-Language Models (LVLMs) are prone to "object hallucination," where they generate plausible but visually absent objects in image captions or answers. While metrics exist to quantify this problem (like CHAIR, POPE, and MME), the underlying mechanisms causing these hallucinations remain poorly understood. Existing mitigation techniques often apply interventions uniformly, which can be inefficient or even detrimental for certain object categories. This paper aims to bridge this gap by dissecting the causes of hallucination and developing a principled, targeted approach to address them.
Methods
The core of the paper is the HalluTrace framework, which causally attributes hallucinations to three sources:
- Visual Grounding Failure (VGF): The visual encoder fails to extract sufficient information to identify an object that is actually present.
- Language Prior Dominance (LPD): The language model, relying on its training data, predicts an object that is not visually present, overriding a correct visual signal.
- Cross-Modal Conflict (CMC): The visual and linguistic signals are contradictory, and the model incorrectly resolves this conflict.
To measure the contribution of each source, the authors employ causal component ablations. They independently "ablate" (i.e., disrupt) parts of the LVLM pipeline: the visual encoder (fvis), the cross-modal projector (fproj), and the language model (fLM). By measuring the resulting change in hallucination rate (using the CHAIR metric), they derive an "attribution score" for each source.
Based on these attribution scores, they propose Hallucination-Aware Decoding (HAD). This is a training-free decoding strategy that dynamically applies targeted interventions at each token generation step:
- For VGF, it uses visual amplification to boost the signal from the visual encoder.
- For LPD, it employs language prior suppression (similar to VCD) to reduce reliance on the language model's statistical priors.
- For CMC, it uses contrastive re-weighting to help resolve conflicting signals.
The key innovation is that HAD routes these interventions based on the predicted dominant hallucination source for the current token and object category.
Datasets
- COCO 2014 validation set: Used for evaluating hallucination metrics (CHAIR, POPE, MME) and for computing attribution scores. The evaluation is stratified by five common COCO supercategories: person, vehicle, animal, food, and furniture.
- COCO object categories (80): Used to define the object vocabulary for routing interventions in the HAD system.
The paper does not introduce a new benchmark but rather leverages existing ones to validate its approach.
Key Insights and Ideas
- Hallucinations are not monolithic: The paper convincingly shows that different object categories are susceptible to different types of hallucinations. This is a crucial insight that moves beyond simply measuring how much hallucination occurs to understanding why.
- Person/vehicle hallucinations are driven by language priors (LPD): These common objects are so well-represented in LVLM training data that the language model often predicts them even if they aren't visually present.
- Food/furniture hallucinations are often due to visual grounding failure (VGF): Smaller objects, those with high intra-class variation, or those frequently occluded in typical scenes pose challenges for the visual encoder.
- Animals present a mixed bag (VGF and CMC): This category is more complex, with hallucinations arising from both visual confusion and conflicting signals.
- Targeted interventions are more effective: The paper provides strong evidence that applying the correct intervention for the identified source of hallucination yields significantly better results than uniform application. This is formalized by Proposition 1, which shows a lower bound on improvement based on attribution scores.
Key Results
- Attribution consistency: Hallucination attribution patterns are remarkably consistent across different LVLMs and specific to object categories.
- HAD significantly reduces hallucinations: On the LLaVA-1.5-7B model, HAD reduced CHAIR₁ (hallucination rate) from 14.2 to 8.1, outperforming baselines like VCD (10.8) and ICD (11.3). It also improved POPE-Adv F1 from 85.9% to 89.6%.
- HAD generalizes: Similar improvements were observed across multiple LVLMs (LLaVA-1.5-13B, InstructBLIP, Qwen-VL, LLaVA-1.6-7B).
- Source-specific interventions matter: The diagonal entries in the intervention effectiveness matrix (matching intervention to source) were 2.4-2.7x more effective than off-diagonal entries.
- Attribution predicts effectiveness: The improvement in CHAIR from HAD was linearly predictable from the VGF attribution share (Pearson r = 0.86), validating the causal framework.
Limitations and Future Work
- Object Vocabulary: The current HAD implementation is limited to the 80 COCO object categories. Future work could extend this to open-vocabulary hallucination using a more general classifier.
- Attribution Computation Cost: Computing attribution scores requires multiple forward passes, making it computationally expensive for real-time deployment. Amortization via cached attributions is a practical workaround.
- Exhaustiveness of Sources: The VGF, LPD, and CMC taxonomy covers dominant failure modes but might not be exhaustive. Other types of hallucinations (e.g., attribute, relational) might involve additional sources.
- Human Evaluation: The study relies on automatic metrics. A human evaluation, especially for CMC, could provide further validation.
- Future Directions: The authors suggest exploring more sophisticated routing mechanisms and investigating how to further improve the efficacy of interventions.
Summary updated