Scaling Discovery Requires Forgetting (AIDDA Seminar)

​Join us for our next seminar series talk!

​We have Shangyin Tan, a PhD student at UC Berkeley, advised by Koushik Sen and Matei Zaharia. He works on LLM agents and compound AI systems, particularly their automatic optimisation. He is a core contributor to DSPy, GEPA, and optimize_anything.

​He will be talking about his recent work, Scaling Discovery Requires Forgetting:

​https://arxiv.org/abs/2610.20563

​Abstract for talk:

​LLM agents can run research loops for hours, but running them longer, in parallel, or as teams all eventually plateau, because agents become anchored to their own history. We show that forgetting fixes this: periodically discarding every session and restarting from the best solution so far. At the same budget, a team of agents with forgetting beats parallel agents by up to 15% and a single agent by up to 39%.