Qwen3.5 "weight editing" hype debunked: it's just a steering vector plus trial and error
akbirthko · x · 2026-11-03
rikiyatakehi claimed GPT-6 Astra directly edited Qwen3.5-0.8B weights — changing only 0.016% of weights, fixing reasoning loops, improving math up to 6x and instruction following 3x with no gradients or training, dubbed "Weight-Space Programming." But @MiniUlisse notes it's unsurprising: the method is just deriving a steering vector from looping vs. non-looping reasoning examples plus trial and error, long known to work. The infographic breaks down the released logs: Astra writes example pairs, takes the activation difference for a "repeat yourself" direction, bakes it into layers 15/19 MLP weights as a rank-one edit, then tunes strength to α=6. MATH-500 accuracy went 7.9%→48.1%, but truncated answers fell 93%→33% — the model isn't better at math, it just stops looping; no gains on models that don't loop. 576 of 579 newly correct answers had been cut off.
More from Models
- Kimi K3.1 Pro open weights expected to drop this week — thursdai_pod · 2026-11-03
- Model swap cut costs 59% and latency 75% — and he still blocked the release — Fun_Employment6042 · 2026-11-03
- Perplexity distills 9B embedding teacher into 0.6B model sharing the same embedding space — antoine_chaffin · 2026-11-03
- Gemini keeps wiping users' chat histories, sparking reliability complaints — matrixmayhem · 2026-11-03
- Astra's Qwen weight edit was a published interpretability technique, not RSI — markjeffrey · 2026-11-03
- Yoav Goldberg: Astra's weight-editing feats are mostly known interpretability tricks — yoavgo · 2026-11-03