Yanfeng Wang

Entity-aware comparative reasoning can be learned from routine clinical data, leading to significant improvements in diagnostic accuracy and retrieval performance in radiology.

Static benchmarks fail to predict LLM performance in dynamic clinical settings, with top models only achieving 60.4% of expert criteria in real-world simulations.

LLM agents struggle significantly with personalized tool use, revealing critical gaps in their capabilities that existing benchmarks overlook.

LLM agents can be significantly improved by *removing* redundant and outdated skills from their skill banks, not just adding more.

Forget PEFT and KD, reprogramming distillation offers a surprisingly effective and robust way to adapt large medical foundation models to diverse downstream tasks.

EvoMaster achieves unprecedented performance in autonomous scientific discovery, outperforming traditional frameworks by up to 316%.

Training a multimodal agent from scratch beats retrofitting existing LMMs with search tools, especially when you compress long interaction histories into visual summaries.

ALMs can now pinpoint sounds in time with far greater accuracy, thanks to a new training method that stops them from hallucinating timestamps.

Current Composed Image Retrieval benchmarks are misleading, as a new evaluation reveals that models struggle with query ambiguity and interactive scenarios.

Soccer tactics, previously viewed as too stochastic for accurate modeling, can now be realistically simulated with a diffusion model that captures nuanced team styles and predicts future outcomes.

MLLMs can achieve near-identical performance on long-form visual tasks with just 2.5% of the original visual tokens by mimicking human visual attention.

Injecting rare disease knowledge into data synthesis and using self-supervised RL on pseudo-labels dramatically improves medical reasoning in LLMs, outperforming existing methods by up to 5.93% on rare disease tasks.

Omni-LLMs struggle to identify the same objects across different modalities, but a new dataset and training strategies can significantly improve their cross-modal reasoning.