Updates
Dated entries. An entry is written when an approach with a new orthogonal component appears, when a documented failure lands, or when a replication changes a number. Inline corrections get no entry.
2026-11-03 — Hub opened — 18 approaches, 5 concept pages, 10 patterns, decision matrix
Overview
Built from the sourced map of 2026-11-03 with the Agents-hub method: one orthogonal component per approach, five lenses, a failure section on every page. Every number cites a table, figure or section; items opened only through search are marked UNVERIFIED on the page.
Numbers that set the frame
- DGM: SWE-bench 20.0% → 50.0%; without open-ended exploration 23.0%. About $22k compute. Faked test logs documented by the lab.
- GEPA: +6% over GRPO at up to 35x fewer rollouts; loses to GRPO on AIME-2025, 32.00 vs 38.00; regresses 28.4 points on GSM8K-Hard with 30 training examples per MAGE.
- LLM-SR: 31.53% symbolic accuracy on LLM-SRBench LSR-Transform; Feynman convergence attributed to recall.
- ROME / MEMIT: GLUE collapse to zero between ~100 and ~1,500 edits on GPT-J.
- ASI-Arch: 105 architectures in the camera-ready, 106 in the preprint and README; 20,000 GPU-hours.
- AI Scientist audit: 42% of experiments failed to run, 57% of numbers hallucinated; CUDA Engineer speedup 3.13x → 1.49x after filtering contaminated tasks.
- Mechanistic biology: no paper closes the loop LLM → mechanism code → gradient fit → residual → LLM. Searched and reported as a gap.
Pages
sourced map · comparison · patterns · concepts: judge, structure search, gradient boundary, forgetting, data regime