LLM Post-training, Efficient Inference & Agent Systems

The Hong Kong Polytechnic University · 2025–Present
Ph.D. Researcher · Advisor: Prof. Jian Huang

On-Policy Distillation for LLM Self-Improvement

BOPD: A two-stage framework bootstraps a task-adapted teacher through reference-conditioned self-distillation, then trains an independent student via on-policy distillation. On Qwen3-4B, mean mathematical reasoning accuracy improves from 18.06% to 23.98%, surpassing the constructed teacher.

Adaptive Supervised Fine-Tuning

SPFT: Privileged-information-guided token weighting uses a frozen, reference-conditioned model without external teachers or on-policy rollouts, improving mean mathematical reasoning accuracy from 36.41% to 47.80% over standard SFT on Qwen3-8B-Base.

FamiFT: Familiarity-adaptive token weighting based on pretrained-model familiarity improves mean mathematical reasoning accuracy from 21.32% to 37.20% over standard SFT on Qwen2.5-Math-7B.

Efficient Decoding for Diffusion Language Models

CoDR: Training-free, sampler-agnostic remasking guided by confidence drift improves accuracy across two diffusion LLMs, three samplers, and four benchmarks, with 76% fewer forward evaluations than ReMDM in a matched setting. Paper

SWD: An information-theoretic connection between temporal token instability and masked-context dependency motivates stability-aware decoding. On LLaDA-8B, mean accuracy increases from 25.63% to 41.08% across four benchmarks while average forward evaluations decrease by 22.7%. Paper

LLM Agents for Data Science

DSAEval: A real-world data science agent benchmark with 641 tasks across 285 datasets, incorporating multimodal perception, multi-query interactions, and multidimensional evaluation of 13 advanced LLM agents. Paper

DARE: Distribution-aware retrieval integrates statistical data profiles into R package embeddings, achieving 93.47% NDCG@10 versus 79.32% for the strongest baseline and improving downstream agents on statistical analysis tasks. Paper