Han-Jia Ye

Synthesizing data environments from explicit world specifications, not just files, unlocks scalable training for data agents and reveals a fundamental computation-insight gap in frontier models.

Later Transformer layers in tabular foundation models aren't always the smartest; revisiting earlier thoughts dramatically boosts performance.

Treating all latent tokens identically bottlenecks multimodal test-time compute—disentangling hidden-state updates into visual-grounding and high-entropy reasoning channels beats standard output-space search under matched sample budgets.

CausalArena is introduced, a unified and evolvable benchmark for causal discovery under a common protocol, and substantial ranking shifts across SCM families and protocols are revealed, showing that strong performance in one benchmark regime does not reliably transfer to others.

Strong logical planning in LLM agents doesn't ensure safe execution—PeakBench reveals critical flaws in existing benchmarks that could lead to resource overflows.

Game solvers can teach LLMs how to make better decisions in long-horizon tasks by providing actionable turn-level feedback, leading to superior performance in complex environments.

RouteJudge reveals that user preferences can significantly inform the effectiveness of LLM routing strategies, transforming how we evaluate model performance.

Sparse updates in on-policy distillation can match full performance with significantly reduced training overhead, challenging conventional wisdom about dense parameter updates.

LLMs still struggle to go beyond simple lookups when answering questions about tables, especially when prediction and reasoning about unobserved data is required.

Forget hand-tuning rollout budgets: $V_{0.5}$ dynamically allocates compute to sparse RL rollouts based on a real-time statistical test of a generalist value model's prior, slashing variance and boosting performance.