Overview
- A team at MIT CSAIL published a paper and GitHub materials that call the idea Weight‑Space Programming and show LLMs can be prompted to inspect raw weight tensors and produce edits intended to change model behavior.
- The paper reports small 'no‑tools' wins, such as correctly naming which of four 10×10 grid games a 17K‑weight policy was trained on and identifying heads in small transformer slices without running the networks.
- Critics note a sharp methodological split because the headline large‑model results, including a claimed Qwen3.5‑0.8B math improvement, use Python, code execution and iterative model runs rather than pure text‑only weight reasoning.
- Independent tests by outside researchers found the no‑tools prompts can work on other models but often need very long reasoning runs, heavy compute and show inconsistent success, so broad tool‑free self‑repair remains unproven.
- If the methods prove robust, direct weight edits could speed debugging and targeted fixes for models, but the community is calling for clear separation of experiment types, open reproductions of the code‑assisted procedures, and transparent benchmarks before judging practical impact.