Authors walk back: the 6x math jump came from fixing reasoning loops
anshulkundaje · x · 2026-11-03
In a follow-up thread, co-author a1zhang shares backstory on the Weight-Space Programming paper: the team originally thought GPT-6 Astra's weight edits on Qwen3.5-0.8B had engineered in a pure math capability gain — an even bigger claim that turned out not to be the case. The primary issue was reasoning looping.
He had suspected as much after seeing similar odd benchmark jumps when training RLMs earlier, traced to reduced harness-level looping. Astra turned out to use a trick loosely based on activation steering to suppress the effect, which also motivated the later ablation where the model is told looping is the issue but lacks access to activations.
More from Research
- Steering Qwen-Image-2.1 with Qwen3-VL's internal states, no training needed — zacharynado · 2026-11-03
- Report finds LLMs surprisingly good at reading and writing neural network weights — omarsar0 · 2026-11-03
- Researcher urges critics to read the prompts in the 'weight-space programming' post first — anton_lozhkov · 2026-11-03
- New paper: standard CoT monitorability metric is only an upper bound, improvable via self-distillation and RL — armancohan · 2026-11-03
- New Scaling Recipe for Data-Constrained Training Cuts Perplexity 14% via Hyperparameter Transfer and Ensembling — andrewgwils · 2026-11-03
- Purdue professor says arXiv banned him a year over 'hallucinated references' that weren't — RexDouglass · 2026-11-03