Advancing LLM capabilities through specialized post-training techniques
Post-training of Large Language Models (LLMs) has emerged as a transformative approach for developing specialized AI systems. This research hub serves as a comprehensive repository showcasing our cutting-edge work in LLM adaptation techniques.
Our research encompasses innovative post-training methodologies, including novel data curation strategies, advanced model training recipes, and pioneering adaptation algorithms. We focus on bridging the gap between general-purpose LLMs and specialized excellence through systematic research and practical implementations.
Comprehensive experiments on domain-adaptive post-training of financial LLMs. We aim to answer the following questions: Given a strong general-purpose LLM (e.g., Llama3-8b-inst), how can we effectively adapt it to a target domain (e.g., finance) through post-training? What criteria are desirable for successful adaptation? What are the most effective training recipes with respect to data and model?
Published: EMNLP, 2025 🏅 Oral (Top 50% of accepted papers, ARR best paper nomination)
A comprehensive tutorial on adaptation of Large Language Models, covering both parametric and semi-parametric knowledge adaptation techniques. Presented at NAACL 2025.
Tutorial: Saturday May 3, 2:00-5:30pm @ NAACL25, New Mexico Convention Center
While Retrieval-Augmented Generation (RAG) has achieved remarkable success, the retriever is always designed for humans rather than for LLMs, even though the LLM is the actual user of the retrieved information. This paper investigates this mismatch and proposes a novel SFT+RL bridge mechanism to better connect the retriever and the LLM, from retriving human-"friendly" to assembling a LLM-"friendly" context.
Published: ACL, 2024
One of the earliest works on continual pre-training of language models. We propose a post-training algorithm with adaptive soft-masking mechanism that selectively updates LM parameters based on the post-training corpus to minimize catastrophic forgetting and enhance knowledge transfer.
Published: ICLR, 2023
