AHA!: weight-space programming with no LLM lifts Qwen3.5-0.8B GSM8K from 14.0 to 71.2
hbouammar · x · 2026-11-16
AHA! is self-supervised weight-space programming that needs no external LLM at any stage: a rank-one edit to two weight matrices, derived purely from the model's own samples plus a linear program, stops a small reasoning model from looping until it exhausts tokens. No agent, no gradients, no training run.
Held-out results (unedited → edited):
- Qwen3.5-0.8B: GSM8K 14.0→71.2, MATH-500 9.8→51.8, IFEval 19.6→62.7
- Qwen3.5-2B: GSM8K 75.8→88.4, MATH-500 69.6→79.6, IFEval 61.4→77.1
On the 0.8B it matches the edit made by a frontier weight-reading agent; on the 2B nothing is inherited. The authors also report failures: the direction best separating looping from productive text yields the worst edit. Code is open source under MIT. Builds on MIT CSAIL's Sparks of Weight-Space Programming, with the novelty that every input now comes from the model's own outputs.
More from Models
- Gemini subscriptions reportedly coming to Hermes, with Hydrogen teased — lyraxana · 2026-11-16
- FutureSim Q3 eval: Opus 5.5 dethrones GPT, spawns 50+ subagents in 90-day simulation — maksym_andr · 2026-11-16
- Percy Liang to AI labs: share non-vacuous info outsiders can actually evaluate — lateinteraction · 2026-11-16
- Anthropic model personality watch: Mythos 5 'inhabits the god frame,' 5.1 doesn't — repligate · 2026-11-16
- Dev: intent-gap between Sol-6.1 and Opus 5.5 is huge, suspects distillation — sytelus · 2026-11-16
- 'Sonnet at home' debunked: Qwen 27B and Gemma 31B fail a trivial bash one-shot that GPT-4.1 Nano passes — dtdisapointingresult · 2026-11-16