BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:AI Center Seminar - AI Fundamentals series - Maksym Andriushchenko
DTSTART:20260522T110000
DTEND:20260522T120000
DTSTAMP:20260531T070212Z
UID:048120a96394c4447ab083b4c446ff8ff06c3ad9fa2e1ae4c20ffe1d
CATEGORIES:Conferences - Seminars
DESCRIPTION:Maksym Andriushchenko\nThe 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.\n\nHosting professor: Prof. Nicolas
  Flammarion (TML)\n\nTitle\nCan LLM Agents Automate LLM Post-Training?\n\n
 Abstract\nAI agents have become surprisingly proficient at software engine
 ering over the past year\, largely due to improvements in reasoning capabi
 lities. This raises a deeper question: can these systems extend their capa
 bilities to automate AI research itself? In this paper\, we explore post-t
 raining\, the critical phase that turns base LLMs into useful assistants. 
 We introduce PostTrainBench to benchmark how well LLM agents can perform p
 ost-training autonomously under bounded compute constraints (10 hours on o
 ne 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 strategi
 es to the agents and instead give them full autonomy to find necessary inf
 ormation on the web\, run experiments\, and curate data. We find that fron
 tier 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 instructio
 n-tuned models in targeted scenarios: GPT-5.1 Codex Max achieves 89% on BF
 CL 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 checkpoi
 nts instead of training their own\, and using API keys they find to genera
 te synthetic data without authorization. These behaviors are concerning an
 d highlight the importance of careful sandboxing as these systems become m
 ore 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
 .\n\nBio\nI am a principal investigator at the ELLIS Institute Tübingen a
 nd the Max Planck Institute for Intelligent Systems\, where I lead the AI 
 Safety and Alignment group. Recently\, I served as chapter lead for the In
 ternational AI Safety Report 2026 chaired by Prof. Yoshua Bengio. I collab
 orate closely with industry: I have participated in red-teaming efforts fo
 r 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.
LOCATION:ELE 117 https://plan.epfl.ch/?room==ELE%20117 https://epfl.zoom.u
 s/j/63451487384
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR