Research / Repairing frozen models
CertiPatch: Specification Repair for Frozen Language Models with Replayable Empirical Certificates
Introduces CertiPatch, a train–certify–verify protocol that repairs frozen language models against deterministic specifications with gated low-rank patches and replayable empirical certificates, reaching zero failures on three enumerable specs.
- Published
- 9 February 2026
- Author
- Ali Uyar
Abstract
We study specification repair for frozen language models: given a deterministic prompt family and a computable Yes/No labeler, learn an inference-time patch that enforces the specification within a certified scope while explicitly measuring collateral change. We present CertiPatch, a train–certify–verify protocol that (i) learns a gated low-rank hookpoint patch under a constraint-first objective (minimize collateral subject to zero in-scope violations), and (ii) emits a replayable empirical certificate with fail-closed verification semantics. In a reduced-compute, single-seed study on EleutherAI/pythia-410m-deduped, CertiPatch attains zero failures on three fully enumerable specs (COMPARE-2D: 0/10,000, PARITY-4D: 0/10,000, BALANCE-PAREN-14: 0/32,767) and achieves 86.22% pass rate (68,974/80,000) on a large-domain, coverage-bounded specification (COMPARE-6D-STRAT). Where baseline cells are available (PARITY-4D), CertiPatch closes the spec with substantially lower measured collateral (KL_S = 0.00168) than completed budget-matched PEFT baselines. We further provide a six-condition compositionality study showing that naive additive patch composition fails (e.g., 5,003 violations on spec B), while sequential and joint repair recover feasibility. All claims are explicitly scope-qualified: single seed, reduced compute, and incomplete baseline coverage where artifacts are missing.
Why it matters
Each repair comes with evidence that can be rerun: replaying the certificate recomputes hashes, regenerates datasets and re-evaluates metrics, and any mismatch fails verification.
Cite
@misc{uyar2026certipatch,
title = {CertiPatch: Specification Repair for Frozen Language Models with Replayable Empirical Certificates},
author = {Uyar, Ali},
year = {2026},
doi = {10.5281/zenodo.18541322},
publisher = {Zenodo},
url = {https://doi.org/10.5281/zenodo.18541322}
}