Weight-Space Programming on Qwen3-ASR-0.6B: Nine Independent Agent Edits, Random-Direction Controls, and a Missing Attention Mask

Takehi et al. (2026) showed that an LLM agent can reduce a model's failure by reading its weights and editing them by hand, with no gradients and a small forward-pass budget. We apply their protocol to speech recognition (Qwen3-ASR-0.6B) with budgets of 10, 25 and 50 forward passes. Nine independent agent arms (three per budget) each produce a hand-derived edit, paired with five random-direction controls matched in tensors, norm and number of changed weights. All runs are scored once on a never-seen FLEURS test set (6 languages). The target failure is that audible speech surrounded by long digital silence is tagged "language None" (30 of 40 stress clips). All nine edits reduce it with no newly broken clips (McNemar p < 0.01 per arm), and the real edit beats each of its five controls in 26 of 27 comparisons. Six of nine arms independently localized the same two decoder attention heads. Four arms pay a small clean-WER cost (+0.24 to +0.40 pp). Key caveat: the inference backend never applies the audio encoder's 8 s windowed attention mask outside FlashAttention-2; restoring the mask with no weight change fixes about as much as the best edit. The pathology is therefore largely a backend artifact, and the edits compensate for it in weight space. Bilingual code-switching was not fixed by any arm. Experiments were carried out with Claude agents; code, edits, controls and logs are released. Code, edits, controls and logs: https://huggingface.co/tahaenesaslanturk/Qwen3-ASR-0.6B-weight-space-programming