Taiqiang Wu

A conflict-aware approach to decoding can triple resistance to errors in LLMs while maintaining accuracy, fundamentally changing how we handle knowledge conflicts in AI.

Achieving up to 4.395x speedup in RL training for LLMs by smartly reusing shared prefixes could revolutionize how we approach large-scale model training.

Attention Sink, where Transformers fixate on seemingly irrelevant tokens, is more than just a quirk – it's a fundamental challenge impacting training, inference, and even causing hallucinations, demanding a systematic approach to understanding and mitigating its effects.

Autonomous driving's next leap hinges on reasoning, not just perception, but current LLM-based approaches are too slow for real-time control.

Training LLMs for efficient reasoning is best achieved by using easier prompts to ensure a dense positive reward signal, preventing undesirable length collapse.