Mao Zheng

NGM-RAG achieves superior performance in multi-hop reasoning by leveraging graph structures, outperforming traditional RAG methods and state-of-the-art graph-enhanced models.

EasyOPD unifies on-policy distillation methods, enabling seamless integration and superior performance across diverse tasks in large language models.

DCPM reveals that separating memory processes can boost LLM agents' ability to infer user intentions and beliefs across sessions, leading to a dramatic increase in personalization accuracy.

Current translation benchmarks miss critical real-world constraints: IFMTBench shows instruction following scales more sharply with model size than translation quality, and general instruction following rankings correlate weakly with translation behavior.

LLMs might seem fluent in Chinese-English translation, but HardMTBench reveals their surprising struggles with domain-specific knowledge, exposing weaknesses hidden by standard benchmarks.

A 440MB multilingual translation model now rivals commercial APIs, opening the door for performant on-device translation.

Achieve the best of both worlds in LLM policy optimization: SRPO combines the rapid gains of self-distillation with the long-term stability of group-relative methods, outperforming both by adaptively routing samples.

Factually dubious LLM outputs can be tamed by strategically penalizing high-confidence predictions at "risky" tokens during fine-tuning, guided by sentence-level factuality labels.

On-Policy Distillation could be the key to more robust and reliable LLM knowledge transfer, but the field is fragmented and lacks a unified theoretical understanding.

LLM-as-a-judge consensus is often an illusion: models agree on surface-level features, but diverge wildly when evaluating true quality, a problem fixable by injecting domain knowledge into rubrics.