GEASS: New Method Reduces Hallucinations in Vision-Language Models by Selectively Trusting Generated Captions
Researchers developed GEASS, a technique that improves how vision-language models use their own generated captions to reduce hallucinations of non-existent objects. The method works by determining on a per-query basis whether a caption helps or hurts accuracy, since the same caption can improve answers to broad questions while degrading answers to detailed ones. This matters because vision-language models are increasingly used in real applications, and reducing hallucinations is critical for their reliability.
A new study on arXiv presents GEASS (Gated Evidence-Adaptive Selective Caption Trust), a training-free approach to reduce hallucinations in vision-language models. The researchers discovered that simply appending a model's generated caption as auxiliary evidence can paradoxically lower accuracy—dropping performance on HallusionBench by nearly ten percentage points in some cases. Through diagnostic testing, they found that caption utility is query-dependent: the same caption helps when answering global questions about an image but harms performance on detail-focused questions. This occurs because captions compete with the image for the model's attention. GEASS addresses this by using a logit-level module that decides per query how much to trust the caption, gating decisions based on the model's confidence and the entropy reduction the caption provides. Tested across four vision-language models and two benchmarks, GEASS improved performance over both standard inference and existing contrastive decoding methods with minimal computational overhead.
What's missing
The study does not discuss potential limitations of the approach, such as computational overhead in real-world deployment scenarios beyond the stated 'two forward passes,' generalization to other types of hallucinations beyond object hallucination, or how performance scales with different model sizes and architectures beyond the four tested.
What different sources said
- arXiv cs.AICenter
GEASS: Gated Evidence-Adaptive Selective Caption Trust for Vision-Language Models
Related
Topology-Aware Thermodynamics Improves DNA Probe Specificity Design
Researchers developed a new framework for designing DNA probes that accounts for the spatial organization of matched sequences, not just overall thermodynamic stability. Traditional methods rely on scalar measures like melting temperature and free energy, which miss how mismatches are distributed along the probe. The approach could improve diagnostic accuracy in applications like HPV detection and gene expression profiling.
Study Identifies Optimal Thermal Dose for Combining Focused Ultrasound with Immunotherapy in Tumors
Researchers used multimodal PET imaging to identify an optimal thermal dose range for focused ultrasound ablation that destroys tumor tissue while preserving conditions for immunotherapy delivery. The study found that excessive heating collapses blood vessels needed for antibody access, while insufficient heating fails to adequately reduce tumor burden. The findings could guide clinical design of combination treatments pairing thermal ablation with immunotherapies.
Plant MSH1 Protein Functions as Mismatch-Directed Nuclease for Organelle Genome Maintenance
Researchers have identified the precise mechanism by which the AtMSH1 protein in Arabidopsis plants recognizes and cleaves DNA mismatches and lesions, preventing mutations in organellar genomes. The protein combines a DNA mismatch recognition module with a nuclease domain that makes staggered cuts at specific positions relative to DNA damage. This discovery explains how plants maintain unusually low mutation rates in their mitochondrial and chloroplast DNA compared to other eukaryotes.