Tianhao Wu

Sample Complexity and Representation Ability of Test-time Scaling Paradigms

EmbedLLM: Learning Compact Representations of Large Language Models

RouteLLM: Learning to Route LLMs from Preference Data

From Crowdsourced Data to High-quality Benchmarks: Arena-Hard and Benchbuilder Pipeline

Bench-O-Matic: Automating Benchmark Curation from Crowdsourced Data

Thinking LLMs: General Instruction Following with Thought Generation

Meta-Rewarding Language Models: Self-Improving Alignment with LLM-as-a-Meta-Judge

R.I.P.: Better Models by Survival of the Fittest Prompts

DySpec: Faster Speculative Decoding with Dynamic Token Tree Structure