Who Checks the Weight Editors? Controls That Change the Reading of LLM-Authored Weight Edits and of Prompt-Rule Effects in Small Language Models
Year: 2026
Type: preprint
Abstract: AI agents have begun to write edits directly into the weights of language models. We argue that the bottleneck in this emerging area is the adjudication of claimed gains, and propose seven inexpensive controls that any such claim should pass: hash-verified replication, failure-mode decomposition, the cheapest alternative, dose arms with a random-direction control, a double placebo, manual reading of outcomes and pre-declared reading rules. The work began in late October 2026 with a harness built to test whether a prompt-level grounding rule acts through an identifiable internal direction (harness archive SHA-256 04e1e4d9…, dated 30–31 October 2026). When the MIT CSAIL report on LLM-authored weight edits appeared on 2 November 2026, the same harness was extended and used to cross-validate it independently. Case 1 re-examines the LLM-authored edit of Qwen3.5-0.8B (MATH-500 7.9% → 48.1%). We replicate the gain and (i) quantify that it is almost entirely the repair of a generation loop in reasoning mode; (ii) consistent with independent evaluations, show that switching reasoning off with no weight change recovers most of the gain with 6–9× fewer tokens; (iii) show that ablating the edited direction does not reproduce the gain, whereas reflecting it does. In an exploratory follow-up, a loop direction found in seven CPU minutes without any agent, written into the same 64 columns at the same norm, recovers about 63% of the agent's gain and beats an equal-norm random direction (60 problems). Concurrent work posted on 4 November 2026 reached convergent results; both timelines are reported. Case 2 evaluates a short grounding instruction in Qwen2.5-7B-Instruct against two length-matched placebos. Under partially blinded reading by a single LLM reader, the instruction reduces invented specifics relative to both placebos. Both placebos, however, increase fabrication relative to no system text, and an automatic extractor biased the comparison against the instruction. 70–90% of the rule-minus-placebo activation direction is shared with 'this is an instruction'. The text of the instruction, the placebo texts and the new items are withheld and committed by hash (PROVENIENCIA_SHA256.txt, anchored in Bitcoin block 974,965 via OpenTimestamps). Case 1 is fully reproducible from public artefacts.
Author: Campos, Vagner
Institution: Walsh College
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225798571
Cites:
Cited by:
Related to: