Yue-Wei Zhang

Gating policy distillation with directional agreement between early and late RL shifts unlocks more effective knowledge transfer, boosting LLM performance on reasoning tasks.

Broader tokenizer alignment in on-policy distillation can actively harm performance by introducing noisy supervision signals.

Past post-training successes often turn toxic when reapplied to drifted checkpoints, making pre-training experience authorization essential to prevent autonomous self-improvement loops from wasting compute and destabilizing model weights.

IRIS can detect model substitutions and routing dilutions in LLM gateways with unprecedented accuracy using only the output text, challenging the reliability of commercial AI services.

Forget static hyperparameters: DVAO dynamically adjusts reward weights based on variance, leading to more stable and effective multi-objective RLHF.

Stop relying on stochasticity: this work shows how distilling experience into hierarchical knowledge unlocks more effective and stable LLM-based agentic search.