cs.LGNov 3, 2026

Where Did the Update Go? Certifying Declared Change Envelopes for Model Updates

Authors: Muhammad Usama, Muhammad Saad Qadir

Organizations: Braindeck Inc. Seoul, South Korea · Department of Computer Science Bahria School of Engineering and Applied Sciences Bahria University, Islamabad, Pakistan

Abstract

Deployed language models are updated continually, and regulators increasingly accept changes that were declared in advance. Showing that an update changed behavior only where it was declared to is a different statistical problem from bounding how much it changed: a useful update is meant to change some decisions, so global disagreement and regression bounds cannot certify it. We formalize the problem with declared change envelopes. Before training, the developer commits to a region of inputs, and every decision that changes outside it counts as leakage. Leakage compares two models, so it can be certified from unlabeled traffic with anytime-valid tests whose sample cost we bound from above and below, and per-update certificates compose into a bound on lifetime drift outside all declared scopes. We measure leakage for LoRA and full fine-tuning, preference optimization, reinforcement learning, and knowledge editing on models from 1.5B to 72B parameters. Narrow updates leak widely: domain and policy LoRA updates keep only about 7-36% of the decisions they change inside the declared scope. Across 256 trained variants, leakage tracks the update's KL divergence to the incumbent on out-of-scope inputs (Spearman 0.94). None of the customary remedies we tested makes a rule-like update certifiable without giving up much of its intended change. Distilling the incumbent's decisions on unlabeled prompts in the deployment format does: anchored LoRA, GRPO, and (at a small learning rate) full fine-tuning updates certify 5% leakage from a few hundred to a few thousand unlabeled samples, and the certificates replicate on held-out traffic and on written answers. The certificates cover decision-like outputs only; on open-ended chat, re-batching alone changes 11-21% of responses and no update certifies. Code: https://github.com/Usama1002/declared-change-envelopes

Explore similar work

Sep 26, 2026

When Can Old Evaluations Certify a New Model? Label-Efficient Release Decisions under Evaluator Drift

Releasing a model update requires certifying that its current-population risk stays below a threshold. Trusted labels are expensive, while a cheap evaluator, such as an LLM judge, scores every example. Reusing evaluator errors from earlier audits is tempting, but when may such evidence replace current labels? It depends on the status of history. If the errors can change invisibly, no label-free test detects the change, and every valid, useful certifier must keep buying labels at a rate we characterize; if a bound on the change is assumed, label-free certification is valid at an explicit error cost. For the middle ground, where history is informative but untrusted, we propose portfolio vigilance, a sequential certifier mixing a betting expert guided by history with one that learns only from current labels; history affects only how it bets, so validity holds for any history. The contribution is not prior-informed betting or expert mixtures, but separating history that may enter validity from history that may only guide label collection. In a canonical model, accurate history shortens decisions but never raises the evidence growth rate; stale history can destroy it. On held-out CIFAR-10N and DICES-990 data, portfolio vigilance needs 0.465 (95% CI [0.327,0.575][0.327,0.575]) and 0.740 ([0.618,0.877][0.618,0.877]) times the labels of a matched prediction-powered monitor, with no observed false certification, and fewer labels on all six external blocks. Under corrupted advice it stays within 8.0% of its better component, while trusting history alone costs up to 1.66 times as much. In post-confirmatory repeated-judge experiments on DICES-990 and ToxicChat, changing a fixed LLM judge's rubric moves its scores beyond run-to-run variation; the portfolio then needs 0.790 ([0.667,0.909][0.667,0.909]) and 0.631 ([0.520,0.770][0.520,0.770]) times the labels of the matched monitor, and fewer than trusting history alone.
Oct 26, 2026

Beyond Update Counts: A Control Map for Evaluating Test-Time Model Updates

Test-time updating lets a language model change persistent state after training---such as model weights, adapter parameters, or a memory module---before or while answering queries. Existing evaluations often compare systems using update count or frequency and performance measured immediately afterward. These summaries omit what is updated, where, when, and in what order. The same total can therefore hide an important difference: one system may answer after one update while another answers after two. We introduce the Control Map, an evaluation framework that records the committed updates preceding each scored answer and separately evaluates behavioral changes, immediate learning (acquisition), later recall (retention), preservation of unrelated behavior (locality), and performance in new settings. We prove when counts alone determine final state and show, under stated assumptions, how differences between pre-answer histories limit differences in average scores. Using ControlMap, our experiments reveal three insights. First, update scheduling matters: at a 760M-parameter checkpoint, 24 schedules with the same update count yield 19.7%--34.2% accuracy. Second, learning an edit does not guarantee that it lasts or remains contained: all six selected CounterFact LoRA settings learn the new information, while ROME retains edits at all 13 operating points but fails locality at each. Third, no editor performs best consistently across models, benchmarks, and tests. Evaluations should therefore report update histories and assess acquisition, retention, and locality separately.
Nov 9, 2026

FireGate: Bounding Behaviour Change Before Training in Continual RL Fine-Tuning

Fine-tuning a deployed language model on one task after another can change behaviour that should stay fixed, such as earlier skills, general chat or refusals. Today, such forgetting is found only after training, on the benchmarks someone chose to run. This work bounds, before training, how often its outputs change on chosen prompt distributions. FireGate places each task's update, such as a new LoRA, behind a gate, which is a classifier on the model's frozen lower layers that decides for each prompt whether the update is used. Where no gate fires, the model computes exactly what the old one computes. As a result, on any prompt distribution, every bounded behaviour score changes by at most the gates' firing rate there over a whole task stream. Calibrating each gate on unlabeled prompts of what must not change fixes that rate before training, with a PAC guarantee, whatever weights training produces. Forgetting is thus bounded by the gate's false-alarm rate, and its price is the gate's power: no update, gated or not, that changes as few protected outputs can change new-task outputs more often than the best test of that level tells new-task prompts from protected ones. On reinforcement-learning streams of up to five tasks, gated training keeps the earlier tasks and changes 3.3% of held-out chat responses, whereas ungated training changes 91.8%. Across seven models, the gain kept behind the gate follows its firing rate on new-task prompts, measured before training. The certificate covers the gated model as deployed, on the prompt distributions sampled for calibration, not against an adversary who searches for prompts that fire the gate.