Initialization Improves LLM-Driven Discovery: early parallel exploration predicts success (Vector Institute)
vector-institute · hf · 2026-10-13
Research from the Vector Institute and University of Toronto (including Colin Raffel) systematically studies LLM-driven iterative discovery of algorithms, theorems and drugs.
- Brittle success: A suite of 12 'Modular' harnesses tested across 5 diverse discovery tasks shows discovery success is brittle and highly sensitive to harness design.
- Mode collapse: A dramatic drop in iterate diversity is a common failure mode; popular SOTA harnesses and diversity-inducing interventions yield inconsistent gains.
- Key finding: Early discovery performance is predictive of eventual success.
- Intervention: A universally applicable initialization step — an initial stage of parallel exploration before iterative optimization — delivers consistent gains across harnesses and tasks.
Related event: Warm-Start Initialization Fixes Mode Collapse in LLM-Driven Discovery(2 posts)→
More from Research
- DOE puts $159M into Genesis projects, expanding OSPREY agent that runs particle accelerators — omnivaughn · 2026-10-14
- Real2Sim gets easy with coding models, but sim-to-real physics and contact solvers lag — jonstephens85 · 2026-10-14
- MIT CSAIL's InstructMesh makes genAI 3D models you can actually 3D print — MIT_CSAIL · 2026-10-14
- Swing Choreo launches with a self-built motion benchmark: 61/100 vs 38 for the next best — andrew_n_carr · 2026-10-13
- No public benchmark for motion generation, so Swing built one: Choreo scores 61 vs 38 — andrew_n_carr · 2026-10-13
- Swing shows compute-quality frontier for motion models: gains came from data selection, not scale — andrew_n_carr · 2026-10-13