Autoresearch rediscovers optimization's greatest hits, from momentum to annealing
burny_tech · x · 2026-10-14
Noting that autoresearch practitioners are revisiting optimization's greatest hits — idea momentum, idea cross-entropy search, idea simulated annealing — the author observes these methods perform roughly the same, with genetic-style variants holding a slight edge, echoing non-convex optimization two decades ago.
He asks whether the field will repeat the old recipe of jointly parameterizing contexts and LLM transforms to traverse idea space parametrically — then answers: isn't that exactly what post-training already does?
More from Research
- TasteVal benchmark: AI 'research taste' doubles every 3 months, Opus 5.5 hits 2.3x expert level — kimmonismus · 2026-10-14
- Qdrant tests Perplexity's pplx-embed-v2-late: 21x smaller vectors, 99.9% retrieval quality kept — qdrant_engine · 2026-10-14
- A 0-parameter 'pick the longest option' model tops HLE, exposing benchmark flaw — antoine_chaffin · 2026-10-14
- Gemma 4 E4B + genomic model BOTANIC-1 pinpoints causal melon yield mutation in one afternoon — remilouf · 2026-10-14
- ETH/EPFL release 134-page fully open Apertus v1.5 multimodal reasoning report — LucaAmb · 2026-10-14
- RUC and Kling AI propose EventLens: event-structure RL to fix video LLMs' temporal understanding — Guowen Zhang · 2026-10-14