JudgeMix uses judge-guided mixed-rollout data synthesis to ease the known train-test mismatch in supervised fine-tuning on expert trajectories for LLM agents.
This work helps models transfer across web platforms whose relational database schemas differ, tackling the generalization problem that mismatch creates.
This study finds that in structure-guided protein design, high structural confidence does not predict function, so it should not be the selection filter.
Drawing on metacognition, this architecture steers what to study next from the scientific understanding it has already built up, speeding discovery along.
NEAT uses a neighborhood-guided autoregressive Transformer to generate 3D molecules more efficiently than diffusion and flow models, offering a faster generative route for drug design.
HyBridge learns hyperbolic representations of hypergraphs with bridge-guided self-supervision, capturing groupwise relationships among nodes and tree-like structures.
OpenAI has rolled out a Compute Campus Policy Blueprint alongside its National Community Compact, aiming to protect local resources, create jobs, and steer AI infrastructure toward responsible growth.
The updated Model Spec from OpenAI sets out five core values to steer more capable models toward sound judgment, respect for human agency, and faithfulness to what people intend.