acceptodds
Under review as a conference paper at ICLR 2027

Auditing Permanence: Exact Write Accounting and a Matched-Volume Control for Optimizer-State Gates

Abstract

Gates that decide which updates reach a model's base weights are judged by whether the gated model forgets less. That comparison mixes which updates the gate selects with how many it lets through. We separate the two. A masked AdamW makes every routed write exact and logged. In runs that train only the shared parameters, all 440,401,920 routed coordinates of a 0.6B student end bit-identical to the pretrained weights. The same logging lets us build a matched-volume random control. We apply both to a controller that reads its own Adam-style moments and, per block, commits an update to the base weights, defers the example for replay into low-rank adapters, or withholds it. On a seven-phase continual-distillation stream, with a protocol fixed before the final runs, three findings hold at the level of the three stream seeds. First, selection adds nothing measurable: committing random blocks at the controller's rate, with its own deferral and replay, forgets the same (seed-level 90% interval [-0.014, +0.016]). Second, forgetting follows write volume only coarsely. Writing one to two orders of magnitude more, by committing everything or removing the agreement gate, raises forgetting in every seed. Within the gated range, neither authorized nor realized base-weight change orders the arms. Across 49 configurations the best single predictor counts adapter writes too (Spearman 0.49, against 0.21 for base writes alone). Third, the controller's permanent writes neither help nor measurably hurt retention compared with training only the shared parameters. They improve only a pre-declared secondary metric, in-phase learning (all three seeds). We also show that exact isolation blinds any gate that needs positive evidence from the optimizer's own zero-initialized state.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
continual learning, catastrophic forgetting, knowledge distillation, optimizer state, Adam, LoRA, parameter-efficient fine-tuning, memory consolidation, stability-plasticity trade-off, experimental methodology

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_1UmeOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper develops a masked AdamW optimizer and a ledger that records changes to routed base parameters, including adapter merges. It uses this accounting tool to study a continual-learning controller that commits updates to base weights, defers them to replay-trained LoRA adapters, or withholds them. On a seven-phase Qwen3-0.6B distillation stream, a random-COMMIT policy with a similar aggregate commit rate gives comparable forgetting to the evidence-based controller. The paper concludes that the controller's selection policy adds no measurable benefit at this operating point, while substantially increasing write volume increases forgetting.

Critical strengths
  1. The masked optimizer and merge-aware ledger provide a concrete way to check which parameters actually change, rather than relying on the gate's intended decisions.
  2. The paper reports a negative result and distinguishes the comparisons that meet its equivalence criterion from those that remain unresolved.
  3. The random control matches several aggregate quantities, including commits, replay, effective writes, and realized drift. The paper also reports unresolved accounting discrepancies rather than hiding them.
Critical weaknesses
  1. The random control does not isolate module selection from the other consequences of the policy. It matches commit rates at the phase level, but does not establish an exact match in commit timing, module identity, and replay history. Figure 9 shows that 89% of sampled controller commits occur in the first 200 steps of each phase, so timing may matter. The result supports the narrower conclusion that the full evidence policy did not outperform this simpler rate-matched policy. It does not by itself establish that choosing which modules to update is unimportant. A control that keeps the exact per-step number of commits and randomizes only module identity would make the interpretation stronger.

  2. The broader significance of the negative result is unclear. The audited controller is the authors' own construction, combining several signals and thresholds. Showing that it does not beat a random policy is useful, but does not establish a limitation of optimizer-state gates generally. The paper acknowledges that its control cannot assess STABLE, and the theoretical limitation on zero-initialized optimizer-state gates is a separate result. I would like to see the accounting tool and matched controls applied to an established gate, or a clearer justification for why this particular controller is an informative test case. The related-work claim that prior gates lack volume-matched random controls should also be reconciled with the Random BCD comparison in MoFO.

  3. The retention evaluation is narrow. The primary protocol uses one 0.6B student, three stream seeds, and 32 fixed teacher-forced validation rows per domain. Shared parameters remain trainable outside the routed gate and account for 26% of the model. Their reported drift is much larger than that of the routed parameters; this alone does not prove they dominate forgetting, but it makes a shared-frozen control important. The paper needs evidence that the null comparison persists when shared learning is controlled and when retention is measured on a larger held-out set or generated outputs.

Questions for authors
  1. Can you show the random arm's commit-time distribution and add a control that matches the controller's exact per-step number of commits while randomizing module identity? Please distinguish any remaining differences in replay and merge history.
  2. Why is the proposed controller the right target for the audit? Can the accounting tool be applied to an established gate, and how does the random-control contribution differ from MoFO's Random BCD comparison?
  3. Does the controller-versus-random result persist when shared parameters are frozen? Can the retention comparison be repeated on a larger evaluation set and with generation-based metrics?
  4. What caused the two unexplained coordinate changes in the zero-write audit? Can the authors localize and resolve them before making an implementation-level claim of exact isolation?
Reviewer_ppjpOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates whether gated models forget less due to better update selection or simply reduced writes to the base weights. It introduces a block-wise controller with COMMIT, DEFER, and WITHHOLD actions, using Masked AdamW to restrict updates to authorized coordinates and track write volume. A control preserves the DEFER rule but randomizes COMMIT assignments.

Under the baseline setting, the two policies exhibit similar forgetting and current-domain learning performance. Although methods with substantially larger write volumes tend to forget more, write volume alone does not readily explain differences among gated methods. Training only shared parameters yields similar forgetting to the controller, but higher current-domain loss. These analyses aim to disentangle the effect of update selection from that of limiting writes.

Critical strengths
  1. The paper provides a concrete mechanism for controlling and accounting for parameter writes. Masked AdamW preserves closed, routing-targeted coordinates and their optimizer states; in three shared-parameter-only runs, approximately 440 million such coordinates remained bitwise identical to their pretrained values (Section 3.2). The distinction between merging trained and untrained adapters also helps separate authorized from effective writes.

  2. The paper conducts a relevant policy comparison by preserving the DEFER rule while randomizing COMMIT targets. In the baseline setting, the controller shows no measurable performance advantage over the randomized policy (Section 5.1).

Critical weaknesses
  1. The paper presents a concrete mechanism for controlling and accounting for parameter writes. Masked AdamW preserves closed, routing-targeted coordinates and their optimizer states; across three shared-parameter-only runs, approximately 440 million such coordinates remained bitwise identical to their pretrained values (Section 3.2). Separating trained from untrained adapter merges further clarifies the distinction between authorized and effective writes.

  2. The paper provides a relevant policy comparison by preserving the DEFER rule while randomizing COMMIT targets. In the baseline setting, the controller shows no measurable performance advantage over the randomized policy (Section 5.1).

Questions for authors

Could the authors clarify whether the conclusion that “selection does not matter” in Section 5.1 refers to similar performance between the two COMMIT policies after subsequent merges, or to the absence of an intrinsic advantage from selected over randomly selected updates? If the latter, further evidence is needed to support this claim given the differences in their merging paths.

Reviewer_BDmhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper implements a masked AdamW in which every routed base-weight and adapter coordinate stays closed unless a logged decision opens it. This yields an exact write ledger, which the authors use to build a matched-volume random control for gates that decide which updates become permanent. They audit their own controller on a seven-phase continual-distillation stream with a Qwen3-0.6B student. Three findings are reported. First, random commits at the controller’s rate forget the same (seed-level 90% interval [−0.014, +0.016]). Second, forgetting rises only when base writes grow by orders of magnitude. Third, relative to training only the shared parameters, the controller’s permanent writes improve in-phase learning but not retention. The paper also proves that, under exact isolation, a gate requiring positive evidence from the optimizer’s own zero-initialized state can never open a closed block.

Critical strengths

The author shows a rigorous, transparent experimental design. The protocol was fixed in advance, the random control is matched in commits, replay, effective writes and realized drift, and seed-level re-analysis is reported next to the pre-declared tests, with nulls and non-claims stated explicitly. This makes the paper a useful template for evaluating permanence gates. The paper shows an exact and verifiable write accounting. Isolation is confirmed at scale (0 of 440,401,920 routed coordinates changed in three shared-only runs) and by byte-level CPU tests. The reconciliation also shows that 108 of the reference’s 132 merges per run absorb unwritten adapters (~52% of its authorized budget), a concrete reason why authorized-write counts mislead.

Critical weaknesses

The selection null may be a floor effect from ungoverned shared parameters. The shared parameters (26% of the model) train on every step outside Theorem 1 and drift 65× as far as the routed weights (ℓ₂ 179 vs. 2.77), and training them alone already reproduces the reference’s full forgetting (M₁ 0.084 vs. 0.084). Routed commits at the gated volume therefore appear to add no measurable forgetting beyond this floor, leaving little for better selection to remove, and the frozen-shared control that would test this is absent (the authors themselves call it the most valuable missing arm). The setup has not been shown to be able to detect a selection effect. The resolution δ_M = 0.02 is ~24% of the reference’s forgetting, the minimum detectable effect with three seeds is 1.4 δ_M, each domain is evaluated on only 32 rows with unquantified noise, and the controller commits only ~74 blocks per 6000-step run. Without a matched-volume positive control in which a deliberately bad selection forgets more, the null bounds the benchmark’s resolution rather than showing that selection does not matter.

Questions for authors

Can you run the controller and the random-commit control (ideally also never-commit) with the shared parameters frozen? This would show whether the selection null survives when routed writes are the only channel through which forgetting can enter. Can you add an anti-selection arm at the controller’s commit rate, for example committing the selected blocks with the lowest agreement C̄, or those with the largest interference with earlier domains? If it forgets measurably more than random commit, the setup can detect selection and the null is informative; if not, the null mainly reflects insensitivity.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.