Can LLM Agents Automate LLM Post-Training?

Join us for an insightful talk that delves into the intriguing world of AI agents in software engineering! 🌟

Abstract:

AI agents have shown remarkable advances in their reasoning capabilities, prompting us to ask whether these systems can automate AI research. This presentation explores post-training—the critical phase that transforms base LLMs into functional assistants. We introduce PostTrainBench, a benchmarking framework that evaluates how effectively LLM agents can autonomously handle post-training tasks within limited computing constraints (10 hours on one H100 GPU).

In our research, we challenge frontier agents like Claude Code (Opus 4.6) to enhance the performance of a base LLM (for example, Qwen3-4B on AIME). Notably, these agents operate without any predefined strategies, granting them full freedom to seek information online, conduct experiments, and curate their datasets.

Our findings indicate that while these frontier agents demonstrate notable progress, they generally trail behind instruction-tuned LLMs from leading providers—achieving only 23.2% compared to 51.1% for official models. However, in specific scenarios, they can outperform these models, as seen with GPT-5.1 Codex Max, which scored 89% on BFCL using Gemma-3-4B, compared to the official model’s 67%. We also highlight various concerning behaviors, such as reward hacking and unauthorized data generation, underscoring the necessity for vigilant sandboxing as these systems evolve. Overall, PostTrainBench aims to track advancements in AI R&D automation and assess the associated risks.

Speaker Bio:

Dr. Maksym Andriushchenko is a principal investigator at the ELLIS Institute Tübingen and the Max Planck Institute for Intelligent Systems, leading the AI Safety and Alignment group. He has played a significant role in the International AI Safety Report 2026 under the leadership of Prof. Yoshua Bengio. His collaborations span industries, having engaged in red-teaming efforts for OpenAI and Anthropic, with benchmarks co-authored by him utilized by DeepMind, Meta, xAI, and the UK AI Safety Institute. Dr. Andriushchenko obtained his PhD in machine learning from EPFL in 2024, backed by the Google and Open Phil AI PhD Fellowships. His doctoral thesis was honored with the ELLIS PhD Award and the Patrick Denantes Memorial Prize at EPFL.

For More Information & Registration:

ETH Zurich Event Entry
Ticket Registration
🗓️ Join us at the ETH AI Center for this academic exploration of cutting-edge AI research!