Technical Report · A4A-TR-2026-005

How Much of a Weight-Space Edit's Gain Does an Inference Configuration Recover? A Preregistered Check on Qwen3.5-0.8B

Theodore O. Cochran

AI for Altruism

Download PDF (60 KB)

Abstract

MIT CSAIL researchers reported in November 2026 that GPT-6 Astra, editing 131,072 weights of the small reasoning model Qwen3.5-0.8B without training, raised its MATH-500 accuracy from 7.9 to 48.1 percent in thinking mode. The authors attribute the gain to repairing reasoning loops. Their evaluation did not report the model with thinking off, which is its default mode. We preregistered and ran that comparison. Both of the authors' thinking-on scores reproduced. Under the model card's non-thinking benchmark configuration, the unedited model reached 43.4 percent, recovering 82 percent of the accuracy gap between the unedited and edited thinking-on configurations. The edited thinking-on configuration kept a 7.85-point advantage over it, beyond the preregistered 5-point margin, while generating 5.9 times as many tokens per answer. With thinking off, edited and unedited models were statistically equivalent within that margin. On MATH-500, the edit's measured benefit is concentrated in the tested thinking-enabled configuration. Most of the reported improvement can be approached without modifying weights, by a change of inference configuration.

Record

History

• 2026-11-08: Version 1.

Cite as

Cochran, Theodore O. 2026. “How Much of a Weight-Space Edit's Gain Does an Inference Configuration Recover? A Preregistered Check on Qwen3.5-0.8B.” Technical report A4A-TR-2026-005, AI for Altruism. https://reports.ai4altruism.org/tr-2026-005/