Chronological feed of everything captured from Fidji Simo.
This paper proposes a multi-agent LLM framework for clinical diagnosis that embeds counterfactual reasoning — explicitly modifying clinical findings and measuring how those changes shift diagnostic confidence — to mimic the differential diagnosis training used in medical education. A key contribution is the "Counterfactual Probability Gap," a metric that quantifies how strongly any individual finding supports a given diagnosis. These counterfactual signals drive iterative specialist-agent debates that prune unsupported hypotheses and improve reasoning interpretability. Benchmarked across three diagnostic datasets and seven LLMs, the approach consistently outperforms prompting baselines and prior multi-agent systems, with the largest gains on ambiguous cases.
MedConceal is a benchmark designed to evaluate medical dialogue systems on "hidden-concern reasoning" — the clinically realistic challenge of surfacing patient fears, misconceptions, and barriers that are not explicitly disclosed. Unlike prior benchmarks that expose full patient state to the model, MedConceal uses an interactive patient simulator that withholds latent concerns, forcing agents to elicit and confirm them through multi-turn dialogue. Evaluation across 300 cases and 600 clinician-LLM interactions reveals that frontier LLMs are competitive on confirmation metrics but human clinicians (N=159) remain superior at intervention — actually guiding patients toward appropriate care. This exposes hidden-concern reasoning under partial observability as a concrete, unresolved capability gap in current medical AI.
Diffusion policies for robotic manipulation face two compounding problems: hard negative class imbalance from uniform sampling slows training convergence, and undifferentiated noise scheduling causes inference timeout failures. VADF addresses both with a two-component, model-agnostic framework: an MLP-based Adaptive Loss Network (ALN) that performs difficulty-aware weighted sampling during training, and a Hierarchical Vision Task Segmenter (HVTS) that decomposes tasks into sub-instructions and assigns variable noise schedules based on action complexity during inference. The design is architecture-agnostic, enabling drop-in integration with any existing diffusion policy backbone.