Parshin Shojaee Shares PhD Thesis on Open-Ended LLM Discovery

Parshin Shojaee announced that he has officially passed his PhD defense and shared his dissertation, *Steps Toward Open-Ended Reasoning and Discovery with Language Models*. The thread centers on a broader question: can language models actually participate in open-ended scientific discovery, rather than only restating what is already known? The topic matters because Shojaee shifts attention from static knowledge to the process of discovery itself.

Core question

Shojaee frames scientific discovery as a messy, multi-path process with dead ends, parallel hypotheses, and low-probability ideas that may still matter. In his telling, the issue is not simply whether LLMs have read enough papers or memorized enough science; it is whether they can keep searching beyond known answers and generate genuinely new directions.

Research directions he highlights

One direction is the balance between memory and exploration. Shojaee points to stronger-memory agentic setups as potentially helpful for sustained research, while also raising the risk that organizing too early around existing clues could narrow the search space. A second direction is adaptation: systems should be able to change course as intermediate results come in.

He also emphasizes diversity optimization, arguing that many current discovery systems stop exploring once they find a merely effective solution. To support more open-ended discovery, he presents search—and especially evolutionary-style search among multiple candidates, with generation, variation, comparison, and selection—as an important path forward.

2026-07-17 ~ 2026-07-17 · 5 related posts

1 near-duplicate retellings: ParshinShojaee