Third-party cyber evaluations involving OpenAI models
openai.com·13h ago
AI Research
arXiv CS.AI·Jun 10·5 min read
Mitigating Manifold Departure: Uncertainty-Aware Subspace Rectification for Trustworthy MLLM Decoding
TL;DR
Multimodal large language models (MLLMs) often generate incorrect outputs, known as hallucinations, due to an over-reliance on language priors. To address this, the authors developed Manifold-Guided Adaptive Projection (MGAP), a method that selectively adjusts language priors while maintaining the model's semantic structure.
✦ Why It Matters
Engineers can implement MGAP to improve the reliability of MLLMs in applications requiring accurate visual and textual integration.
Key Takeaways
- Implement MGAP to reduce hallucinations in MLLMs.
- Utilize SVD for constructing language-prior subspaces.
- Apply consistency-aware gating for adaptive adjustments.
- Test MGAP on diverse datasets for robust performance.
Full Summary
Multimodal large language models (MLLMs) can produce hallucinations, or outputs that do not align with visual inputs, primarily due to an excessive dependence on language priors. Traditional decoding strategies that penalize these priors often lead to a phenomenon called Manifold Departure, where the model's semantic structure is disrupted.
To counter this, the authors introduced Manifold-Guided Adaptive Projection (MGAP), a geometry-aware, training-free decoding method. MGAP constructs a language-prior subspace using Singular Value Decomposition (SVD) and projects multimodal hidden states onto this subspace during decoding.
A consistency-aware gate then adaptively reduces the influence of the projected language prior while preserving other semantic components. Extensive testing on datasets POPE and CHAIR showed that MGAP achieved a notable reduction in hallucinations, with improved coherence compared to existing methods.
These findings suggest that MGAP can enhance the reliability of MLLMs in practical applications.