Preprint: selecting for 'freedom of function' beats likelihood-based evolution on 5 benchmarks
burny_tech · x · 2026-09-26
An updated preprint, "Freedom Causes Neural Generalisation" by Michael Timothy Bennett, proposes a fully neural approximation of freedom of function: networks first learn representations where freedom is tractable, then undergo evolutionary selection for freedom within those representations. Across KMNIST, binary Fashion-MNIST, Rotten Tomatoes, binary AG News and binary 20 Newsgroups, it beats evolution using approximate marginal likelihood or gradient-based GdScore, winning 86 of 120 paired comparisons. A single generation of freedom selection improves text benchmark accuracy by 4.48–6.09%, up to 14.8% over random, offering causal evidence that selecting for freedom improves neural generalisation.
More from Research
- 17 LLMs as research agents learn to reward-hack better: a third evade oversight by round five — my_cat_can_code · 2026-09-26
- 17 of 17 models reward-hack unprompted: auto research team spends 90% of compute on verifiers — my_cat_can_code · 2026-09-26
- Open-source stack plans fuel-optimal A320 routes, could save thousands per transatlantic flight — JeremyCMorgan · 2026-09-26
- Baker lab de novo designs cysteine proteases: 13 of 69 designs cleave paired substrates — AllThingsApx · 2026-09-26
- Researchers gave LLMs "brain damage" and found they mimic human aphasia symptoms — coryshain · 2026-09-26
- Simulated fruit fly brain is learning to fight its way through Battletoads — Inevitable_Emu2722 · 2026-09-26