Yuanda Xu

Smaller, cheaper models generate superior negative examples for preference distillation, significantly boosting student performance without requiring larger teacher models.

Identity drift in generative agents reveals that anti-self-deception is a dominant modification behavior, challenging assumptions about agent fidelity under pressure.

Simulating 8.3 billion diverse personas reveals nuanced user interactions that traditional evaluations miss, transforming how we assess AI systems.

TREK transforms the way models tackle challenging prompts by expanding their exploration support, leading to substantial performance gains even in the hardest task scenarios.

Overconfident tokens, often missed by entropy-based methods, carry surprisingly dense corrective signals in on-policy distillation, allowing for near-baseline performance with <10% of tokens.

Pinpointing performance bottlenecks in RAG pipelines just got easier: RAGPerf offers a modular benchmarking framework to dissect and optimize each component.

Stop wasting compute on easy and impossible examples: PACED distillation focuses your student model's training on the sweet spot where it actually learns.

Reasoning models aren't just verbose, they're actively *harmed* by their own verbosity, but a simple self-distillation trick can compress their outputs by up to 59% while boosting accuracy by up to 16 points.

Overconfident errors in RLVR monopolize probability mass and suppress exploration, but a confidence-aware penalty fixes this and boosts mathematical reasoning performance.