Multimodal Hallucination Mitigation

A decoding-time intervention that penalizes outputs from a flawed multimodal model by contrasting its logits against a weaker, text-only baseline. By amplifying signals unique to the visual input, this method suppresses statistical biases learned from pure text corpora.

• Mechanism: Subtracts the log-probability of a text-only expert from a vision-language expert.
• Use Case: Preventing a VLM from defaulting to common-sense text answers when the image shows an anomaly.
• Trade-off: Increases inference latency due to dual forward passes.

Techniques that force the model to distribute cross-attention weights uniformly across relevant image regions rather than collapsing to a single token. This prevents the model from ignoring visual evidence in favor of memorized text patterns.

• Implementation: Adds an auxiliary loss to maximize the entropy of the attention distribution.
• Result: Reduces object existence hallucinations where the model references entities not present in the image.
• Metric: Measured via CHAIR (Caption Hallucination Assessment with Image Relevance) scores.

Fine-tuning Vision-Language Models (VLMs) on curated datasets containing hallucinated responses paired with explicit rejection feedback. The model learns to output 'I cannot see that' or similar disclaimers when visual evidence is insufficient.

• Data Format: Triplets of (Image, Question, Refusal Answer).
• Benefit: Directly aligns the model's behavior with epistemic humility regarding visual content.
• Challenge: Requires careful balancing to prevent the model from becoming overly cautious and refusing valid queries.

A hybrid architecture that prepends the outputs of a frozen object detection model (e.g., bounding boxes and class labels) to the visual tokens before text generation. This provides an explicit, symbolic bottleneck of visual facts.

• Pipeline: Faster R-CNN → Semantic Labels → LLM Decoder.
• Strength: Guarantees that generated nouns correspond to high-confidence detected objects.
• Limitation: Cannot describe abstract relationships or actions not captured by the object detector's ontology.

A post-hoc method that identifies specific linear directions in the model's representation space corresponding to hallucinated concepts and subtracts them during inference. This surgically removes the model's propensity to generate specific false facts without retraining.

• Process: Uses linear probes to find the 'hallucination vector' and applies a negative projection.
• Advantage: Lightweight and does not require modifying the base model weights.
• Risk: Potential for collateral damage to semantically related but truthful concepts.

A constrained decoding strategy that restricts the next-token distribution to a subset of words that have high visual grounding confidence. It dynamically adjusts the nucleus sampling threshold based on the attention paid to image regions versus text context.

• Logic: If visual attention is low, the sampling pool is aggressively pruned to high-frequency visual concepts.
• Goal: Prevents the model from drifting into pure language modeling mode.
• Integration: Implemented as a logits processor in standard generation pipelines.

The systematic process of anchoring generated text to verifiable source data. In multimodal contexts, this involves explicitly linking generated claims to specific image regions or retrieved text chunks.

• Citation Attribution: Providing direct pointers to evidence.
• Provenance Tracking: Maintaining a chain of custody for information.
• NLI Scoring: Using Natural Language Inference models to detect contradictions between a statement and its visual evidence before output.

A neural network trained on 400M image-text pairs to learn a joint embedding space. It serves as a foundational scoring function for hallucination detection by measuring the cosine similarity between a generated caption and the actual image.

• Low similarity scores often flag potential hallucinations.
• Used in models like DALL-E 2 and Stable Diffusion for guidance.
• Enables zero-shot classification of visual accuracy.

The task of precisely localizing the image region that corresponds to a specific noun phrase or textual claim. It directly combats object hallucination (mentioning objects not present).

• Referring Expression Comprehension: Mapping "the red car on the left" to a bounding box.
• Phrase Grounding: Linking every noun phrase in a caption to image pixels.
• Failure to ground a generated entity is a primary signal of a hallucination.

A prompting strategy that forces the model to generate interleaved textual rationale and visual evidence before producing the final answer. This explicit reasoning path reduces logical leaps.

• Step 1: Identify relevant objects in the image.
• Step 2: Describe their spatial relationships.
• Step 3: Synthesize the final answer based only on the described evidence.
• Acts as a self-verification mechanism.

The core architectural operation where textual queries attend to visual keys/values. Analyzing cross-attention heatmaps is a key interpretability technique for hallucination mitigation.

• Low attention weight on a generated object's image region suggests the model is relying on language priors, not vision.
• Allows engineers to audit why a token was generated.
• Used to enforce visual fidelity during decoding.

An architecture that grounds generation in a trusted, external multimodal index. Instead of relying solely on parametric memory, the model retrieves verified image-text pairs.

• Index: A vector store of domain-specific, factual images and captions.
• Retrieval: Fetches top-k relevant items before generation.
• Grounded Synthesis: The model conditions its output on the retrieved evidence, drastically reducing factual drift.