acceptodds
Under review as a conference paper at ICLR 2027

The Half-Life of Repairs: Why Fixes Fail under Continued Training and How to Make Them Last

Abstract

Post-hoc repairs to a trained differentiable audio synthesizer—a frozen gain, an inserted filter, a reverb reparameterization—are routine and physically verifiable, yet continued training silently undoes them. We formulate repair persistence as an engineering objective and measure it with a pre-registered four-arm paired- continuation protocol and tri-state verdicts: unrepaired training, immediate re- pair, repair under continuation, and repair with behavioral anchoring. In a con- trolled cello-vocoder experiment, a +5.03 dB gain restores the median loudness drift from −5.03 to −0.0006 dB; after 15,000 continued-training steps the unan- chored arm degrades to −5.25 dB (R = −0.27, the repair-retention ratio; direc- tional, within the seed-variance band), the benefit halving within 500–1,000 steps (single-seed). Three instruments localize the cancellation to the harmonic and noise branches; output-level gradient pressure falls in all 12 instrumented runs. Anchoring is a fuse, not a retainer—no stable persistence advantage over plain repair (1/3 vs 1/3 seeds)—and the repair-plus-anchor package backfires on both URMP seeds. A 34-run, 13-arm, two-domain matrix shows erasure replicating at the gradient-instrument level; persistence tracks the calibration source (3/3 vs 1/3 seeds; not separable, Fisher p = 0.20). Claims cover the loudness-defect domain only. Contributions: a repair-visibility taxonomy; a persistence-time / compensation-path detection protocol; a constructive candidate route to durable repair.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Neural vocoder / DDSP, Speech & Audio, Music Generation, Neural Audio Generation, Continual Learning, Repair Persistence, Paired Continuation, Structural Reparameterization

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_7VRkOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies the persistence of post hoc repairs to differentiable audio models during continued training. The key observation is that a repair that is effective immediately after application may gradually be canceled by subsequent optimization, as other trainable components adapt around the intervention. The authors introduce a repair-retention metric and the notion of repair “half-life,” together with a four-arm paired-continuation protocol for evaluating whether a repair persists, is partially retained, or is “eaten.” Experiments on a DDSP-based cello synthesizer show that a fixed gain repair can lose most of its benefit after continued training, while behavioral anchoring does not consistently improve persistence.

Critical strengths
  • The paper highlights an important issue that can easily be overlooked in practice: a repair that works immediately does not guarantee that it will remain effective after continued training.
  • Extensive experiments are conducted to support the claim and the experimental design seems sound.
Critical weaknesses
  • The paper is quite difficult to follow. A large number of terms, arm identifiers, metrics, qualifications, and experimental details are introduced very quickly. For example, C3, R3, A1–A4, B0–B6, C1–C3, and U1–U3 are all used in different parts of the paper, which makes the main narrative unnecessarily difficult to reconstruct.
  • Also the abstract contains too many experimental details, numerical results, abbreviations, and qualifications before clearly communicating the basic problem, approach, and takeaway.
  • The empirical scope remains limited relative to the broader framing. The main evidence comes from one differentiable audio-synthesis system and primarily a loudness/level-defect setting.
  • Some conclusions are based on fairly limited statistical evidence. For example, the reported calibration-source difference of 3/3 versus 1/3 seeds is not statistically separable (p=0.20), and the authors acknowledge a confound between calibration source and initial mismatch. Similarly, the original single-seed EATEN results become considerably less stable across additional seeds.
  • While the paper also claims to study how to make repairs persist, the behavioral anchoring intervention does not consistently solve the problem.
Questions for authors
  1. How broadly do you expect the observed repair-erasure behavior to generalize beyond the current DDSP/loudness setting? Could the authors provide experiments on another architecture, repair type, or defect to support the broader applicability of the proposed protocol and taxonomy?
  2. The paper notes that the 0.5/0.25 thresholds are engineering choices rather than empirically calibrated values, and the alternative CI-based criterion changes several run-level verdicts. How robust are the PERSIST/PARTIAL/EATEN conclusions to the chosen thresholds?
  3. Since lightweight anchoring does not reliably preserve repairs and structural reparameterization has relatively limited evidence, what is the main actionable recommendation when continued training is unavoidable?
  4. Could the presentation be substantially simplified?
Reviewer_GCcCOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper examines whether repairs made to differentiable audio synthesizers last through further training. It introduces a controlled evaluation method and shows that training can undo a loudness correction, while behavioral anchoring does not always keep it in place.

Critical strengths
  1. The paired comparisons with unrepaired training help distinguish loss of the correction from ordinary model drift.
Critical weaknesses
  1. The paper is very difficult to read; the Introduction and Abstract directly start talking about results without motivating or introducing the research problem properly. Further, after going through the Introduction, it is hard to understand what has already been done in the field. The introduction mixes the practical problem, possible mechanisms, experimental results, and qualifications before clearly explaining why the problem arises.
  2. The abstract talks about Post-hoc repairs to the trained differentiable audio synthesizer; however, the problem statement in the introduction explains compensation around a fixed intervention.
  3. The distinction between what they observe, what they hypothesize, and what they establish is blurred.
  4. The key relationship between the training objective and the repaired defect remains insufficiently explained.
Questions for authors

Can the authors share audio samples from before repair, right after repair, and after further training?

Reviewer_u5cnOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper presents a protocol for differentiable audio synthesizer repair.

Critical strengths

The paper is so badly written that if there are any strengths, it is very difficult to distinguish them from all the noise.

Critical weaknesses

The paper is unreadable - a word salad that needs a lot of work on both improving the english and overall presentation.

Some examples:

"Citation delimitation." -> what does this mean?

"To our knowledge, the first fully machine-archived case set and multi-start matrix." -> the sentence makes no sense What is the reason to label the "Motivating Cases" as C3 and R3?

"Number-disclosure caliber." -> what?!

The problem statement: "A physical constraint that holds at inference need not remain a constraint on the equilibrium of the re-optimized model: the remaining trainable modules may form compensation paths that route around or cancel the repair." -> I did read this multiple times and still can't understand what it is about (and it is written in bold so it must be important).

Questions for authors

Please see above.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.