GSAR: Typed Grounding for Hallucination Detection and Recovery in Multi-Agent LLMs
Multiple research papers released on arXiv address the challenge of hallucinations in large language and vision-language models. One paper introduces In-Context Visual Contrastive Optimization (IC-VCO) to mitigate multimodal hallucinations by using contrastive images within a shared context and a novel sample editing strategy. Another study investigates architectural factors influencing hallucination robustness, categorizing hallucinations and providing guidance on model design. Additionally, a new framework, BenHalluEval, is proposed for evaluating and detecting hallucinations in Bengali language models, highlighting the inadequacy of existing methods for low-resource languages. Other research explores reframing hallucination detection as out-of-distribution detection and examines how prompt toxicity affects factual reliability. AI
IMPACT These studies offer new techniques and benchmarks for improving the factual accuracy and reliability of LLMs, crucial for their safe deployment in sensitive applications.
- Answer-agreement Representation Shaping
- HalluScore
- Adaptive Detection Routing
- Qwen3-14B
- CuraView
- Adaptive Unlearning
- HalluScan
- LLM Ghostbusters
- Instruction Lens Score
- CAAFC
- LLaMA-70B-Instruct
- QAOD
- LLaVA-v1.5
- arXiv
- Qwen2.5-VL
- PubMed Central
- SIRA
- bioRxiv
- PCNET
- SSRN
- TokenHD
- BenHalluEval
- LLMs
- Llama
- Gemma
- MimicIV
- IC-VCO
- GPT-5.4