Idea basins: why autonomous agents get stuck, and how frontier models explore differently
AlexGDimakis · x · 2026-11-13
Alex G. Dimakis describes "idea basins," independently discovered by multiple labs: autonomous agents iterating on a problem fixate on previously tried methods, staying in localized regions of solution space. Techniques like Fork-and-Flush and "Scaling Discovery requires forgetting" counter this by periodically deleting session context.
The team studies how frontier models explore basins on AutoResearchExam — e.g. for accelerating transformer decoder graphs on CPUs, models may try NumPy+asm kernels, native BLAS, JIT compilers, and more. The accompanying idea maps show Opus 5.5 (test AUARC 0.87) heading straight to the best idea and exploring it deeply, Qwen3.8 Max (0.63) trapped in one basin, and Gemini 3.8 Flash (0.78) switching often but scattered. Key finding: most runs stick to only one or two ideas, with models differing in how much they switch versus how fast they exploit the best direction.
More from coding & agent
- JoMoHa jointly RL-trains model and harness, hits 97.0% on ALFWorld — omarsar0 · 2026-11-13
- YC F26 team open-sources runtime to run 10,000 agents in one session — KlausCodes · 2026-11-13
- Open-Source Pipeline Turns LiDAR + OpenStreetMap into Isometric Pixel-Art Cities with FLUX.2 — zsakib_ · 2026-11-13
- X, MIT, UW & Stanford Launch WSDM Cup for AI Agents Writing Community Notes, $10k Prizes — NathanpmYoung · 2026-11-13
- Benchmarks won't move users off Claude Code — Antigravity needs real product uplift, dev argues — Justin_Halford_ · 2026-11-13
- Why AI labs must own meeting software: spoken context is the missing fuel for agents — signulll · 2026-11-13