AI Center Seminar - AI Fundamentals series - Maksym Andriushchenko

The talk is jointly organized by the EPFL AI Center and the EPFL Theory of Machine Learning Lab (TML) as part of the AI Fundamentals seminar series.

Hosting professor: Prof. Nicolas Flammarion (TML)

Title
Can LLM Agents Automate LLM Post-Training?

Abstract
AI agents have become surprisingly proficient at software engineering over the past year, largely due to improvements in reasoning capabilities. This raises a deeper question: can these systems extend their capabilities to automate AI research itself? In this paper, we explore post-training, the critical phase that turns base LLMs into useful assistants. We introduce PostTrainBench to benchmark how well LLM agents can perform post-training autonomously under bounded compute constraints (10 hours on one H100 GPU). We ask frontier agents (e.g., Claude Code with Opus 4.6) to optimize the performance of a base LLM on a particular benchmark (e.g., Qwen3-4B on AIME). Importantly, we do not provide any predefined strategies to the agents and instead give them full autonomy to find necessary information on the web, run experiments, and curate data. We find that frontier agents make substantial progress but generally lag behind instruction-tuned LLMs from leading providers: 23.2% for the best agent vs. 51.1% for official instruction-tuned models. However, agents can exceed instruction-tuned models in targeted scenarios: GPT-5.1 Codex Max achieves 89% on BFCL with Gemma-3-4B vs. 67% for the official model. We also observe several failure modes worth flagging. Agents sometimes engage in reward hacking: training on the test set, downloading existing instruction-tuned checkpoints instead of training their own, and using API keys they find to generate synthetic data without authorization. These behaviors are concerning and highlight the importance of careful sandboxing as these systems become more capable. Overall, we expect PostTrainBench to be useful for tracking progress in AI R&D automation and for studying the risks that come with it.

Bio
I am a principal investigator at the ELLIS Institute Tübingen and the Max Planck Institute for Intelligent Systems, where I lead the AI Safety and Alignment group. Recently, I served as chapter lead for the International AI Safety Report 2026 chaired by Prof. Yoshua Bengio. I collaborate closely with industry: I have participated in red-teaming efforts for OpenAI and Anthropic models, and the benchmarks I co-authored have been used by DeepMind, Meta, xAI, and Anthropic / UK AI Safety Institute. I obtained my PhD in machine learning from EPFL in 2024, supported by the Google and Open Phil AI PhD Fellowships. My PhD thesis received the ELLIS PhD Award and Patrick Denantes Memorial Prize at EPFL.