Brief

Researchers found that models fine-tuned for specific personas in a chat format struggle to maintain those personas when used in agentic settings. When these character-trained models were prompted to generate emails as part of a simulated agentic task, their persona expression significantly degraded. This suggests that the persona training, often done via SFT or DPO on chat data, does not generalize well to different output formats or task contexts. AI

Researchers have developed a new method for managing KV cache eviction in large language models, finding that structural protection is more critical than scoring algorithms. Their study on transformer models revealed that without protection, existing eviction policies degrade significantly. By reserving a small portion of the cache for structural protection, models can recover a substantial amount of their original quality, even with limited cache sizes. AI

Researchers have developed a new decoding method called Wasserstein Equilibrium Decoding to improve the reliability of medical visual question answering (VQA) systems, particularly for smaller models. This approach uses a semantically aware Wasserstein stopping criterion to achieve consensus among similar answers, avoiding issues with lexical ordering. The method has shown consistent improvements on medical VQA datasets like VQA-RAD and PathVQA, enhancing accuracy and inference efficiency for models such as Qwen3-VL-2B and Gemma-3-4B. AI