Gated Repair of Language-Model Binding Errors in Synthetic JSON Tasks
Year: 2026
Type: preprint
Abstract: Language models that fill structured answers sometimes copy a neighbouring value into the wrong field. On synthetic JSON tasks, we ask whether a frozen model can repair such errors in its own output without an oracle, a stronger model, a trained probe or gold labels at inference time, with certified harm to already-correct outputs. A gate reads the model's digit probabilities under masked counterfactual views of the prompt, writes one digit, lets the model continue and keeps the result only if the continued output agrees with itself; a second round acts only on rows that still disagree. In a preregistered confirmation, two-round repair beat one round on natural whole-value copy errors of Mistral-7B-Instruct-v0.3 (10:0 discordant pairs), with end-to-end harm on 1/360 clean sources, certified at a target rate of 1/40. The model's own gate also repaired binding errors of Yi-1.5-6B-Chat (59:0 against no repair). Program-assisted, synthetic-error and untestable cases are separated. Preprint, version 1 (2026-10-31); not peer reviewed. A revised version will follow on this record (known layout issues to be fixed: table numbering, two blank section pages). SHA-256 of the PDF: ffe005185584cd78b09bcae9d9b2265983c1c6a11b1047e8748764ef933ff2ed
Author: Zhao, Luoqing
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7224993145