Responsible Finetuning of Large Language Models

Ling Liu

Conference on Machine Learning and Systems 2025 · Day 4 · Invited Talk

This article delves into the critical and evolving challenges surrounding the responsible finetuning of Large Language Models (LLMs), with a particular emphasis on ensuring their safety and robustness in real-world applications. Presented by Professor Ling Liu of Georgia Tech at MLSys 2025, the talk underscores that while LLMs are becoming increasingly powerful and ubiquitous, their inherent limitations and risks, especially concerning safety, remain significant. The core message is that safety alignment, though crucial, is a brittle and complex process, prone to issues like catastrophic forgetting, contextual dependencies, and vulnerabilities introduced during downstream user finetuning.

AI review

Professor Liu covers genuinely important ground — embedding drift, downstream finetuning vulnerabilities, vaccination via min-max optimization, and post-pruning with Antidote — but the talk as described is an academic survey delivered from 30,000 feet. The core ideas are real, but there's no code, no reproducible setup, no model names tied to specific benchmark numbers, and no way for a working engineer to act on any of it. The article itself flags that 'specific, detailed experimental setups are not exhaustively covered,' which is a polite way of saying: you'll need to track down the papers…