acceptodds
Withdrawn from ICLR 2027

Topology-Guided Adapter Lifecycles for Continual Test-Time Training

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Continual adaptation of a frozen language model commonly updates one low-rank adapter on an arriving stream. This design becomes fragile when domains recur: current updates overwrite earlier specialization, while periodic resets discard useful state. We treat adaptation as an adapter lifecycle problem and use the shape of the response cloud to decide when to route, spawn, or reuse an adapter. Two ingredients make this signal usable. First, a canonical response coordinate is invariant to the GL(r) gauge freedom of a low-rank factorization and preserves pairwise response distances; across our gauge sweep the resulting barcode drift stays below 2 × 10⁻¹², whereas the naive internal coordinate changes by more than 10². Second, a block comparison around an excluded transition band cancels any transient contained in that band while retaining the full separation between sustained regimes. On the three public continual learning protocols for language models, and without ever being told where one task ends, the resulting adapter bank reaches 80.0 on the Standard CL benchmark, matching the multi-task bound and exceeding every sequential published method, 78.6 on the fifteen-task Long Sequence benchmark against 73.6 for the strongest published replay-free method and 78.1 for per-task models, and 43.2 on TRACE against 40.0 and 32.9 for the strongest replay-free and boundary-free published results. Our pipeline reproduces every quoted baseline to within 0.5 points. Traversing each protocol a second time on disjoint shards separates the methods further: our score falls by 0.7 on TRACE where the strongest baseline falls by 5.5, and the bank holds nine adapters over sixteen stages of eight recurring tasks. A detection study on the same streams locates the cause, showing that the second-order signal used by the strongest boundary-free baseline collapses from 0.788 to 0.602 ROC-AUC on transitions whose moments agree, where the topological signal holds at 0.871. We also report an unlabeled test-time setting and a homogeneous stream on which topology offers no advantage.

Primary area:
interpretability and explainable AI
Keywords:
Continual Test-Time Training, Continual Learning, Parameter-Efficient Fine-Tuning, Low-Rank Adaptation, Adapter Routing, Persistent Homology, Topological Data Analysis, Task-Free Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ezm5OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies the continual adaptation capability of language models. The problem is potentially important, and the later sections appear to contain substantial experiments. Nevertheless, the introduction is hard to comprehend. The target problem is not well-specified, the limitations of and relationships with existing literature remain unclear, and numerous critical terms and symbols appear without definition. These issues prevent readers from proceeding to subsequent sections and assessing the paper’s technical contributions.

Critical strengths

The studied problem has practical significance.

Critical weaknesses
  1. The background information for the continual adaptation of language models is missing, which makes it hard for readers to understand the importance of this study. Also, the core problem studied in this paper is not sufficiently motivated. The narrative logic in introduction is confusing with many terms and symbols used without definition, e.g., coordinate, GAGA.
  2. The limitations of existing methods are insufficiently discussed, particularly those of the closely related FiUni. It is unclear under what circumstances or in which applications the second-order criterion would fail, apart from theoretically constructed transitions with matched moments. In addition, the related methods are organized without a clear internal logic, which makes it very difficult to follow.
  3. It is unclear how the proposed study addresses the problem stated in the introduction. As a result, the introduction does not provide valid evidence in support of the claimed contributions.
Questions for authors
  1. What is the core issue this study focuses on? Why is it important?
  2. Why do existing methods fail to address this issue in principle? What are the challenges?
  3. How does this work address these challenges in general? What innovations can this work bring to the research community to address the investigated problem?
Reviewer_x62sOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes TopoLoRA, a method for managing a bank of LoRA adapters during continual adaptation without task boundaries. It constructs low-dimensional coordinates from adapter responses, extracts persistent-homology signatures, and compares these signatures across temporal blocks. These signals, together with a frozen routing probe and recurrence tracking, are intended to guide adapter creation and reuse.

Experiments cover supervised continual learning, recurring task streams, unlabeled test-time adaptation, and synthetic changes designed to preserve first- and second-order statistics.

Critical strengths
  • The response-based representation addresses a concrete problem with LoRA coordinates. Equivalent LoRA factorizations can distort the intermediate representation while leaving the adapter’s function unchanged. Computing descriptors from the response, while preserving its pairwise distances, provides a sensible basis for comparing adapter behavior.

  • The evaluation considers recurrence and capacity use alongside final performance. The recurring streams, bank-size measurements, and reactivation ablations directly examine whether the method reuses existing adapters. This is useful evidence for the proposed lifecycle mechanism.

Critical weaknesses

The detection argument does not yet establish the claimed behavior of the online controller. Proposition 2 guarantees cancellation when a transient falls entirely inside the excluded band of a particular candidate window. During sliding detection, that same transient enters neighboring comparison blocks. With L=4L=4, an isolated signature deviation of magnitude MM can produce a score of M/4M/4 at four consecutive candidate positions. The proposition is valid under its stated condition, but the broader claim that the detector responds to duration regardless of amplitude does not follow. There is also an unresolved implementation issue: Eq. (1), with the default settings, requires windows through t+5t+5, while Section 3 describes allocation through consecutive geometric novelty scores. The paper needs to specify when the topological score becomes available and exactly how it affects allocation and reuse.

The baseline records contain inconsistencies that weaken confidence in the comparisons. Table 1 labels the results 75.9/74.3/36.5 as “O-LoRA + replay,” but the cited ELLA paper assigns these numbers to DATA + Replay. This leaves the identity of the corresponding baseline in Table 3 unclear. (ELLA, Table 1) There are also discrepancies between reported measurements: Table 2 gives TRACE scores of 23.4/39.5/33.2 for O-LoRA/ELLA/FiUni, whereas Table 9 gives 23.1/40.0/32.9 despite describing its entries as common-pipeline measurements. Finally, Section 5 claims that all ten reproduction checks differ by at most 0.5 points, although its reported FiUni result, 54.4 versus 55.2, differs by 0.8. These discrepancies need to be reconciled so readers can identify which runs support the reported gains.

Questions for authors

When is s(t)s(t) evaluated online, given its dependence on windows through t+5t+5, and how does it affect the allocation rule in Section 3?

Is the replay baseline in Tables 1 and 3 O-LoRA or DATA, and which implementation produced the recurring-stream results?

Reviewer_WJrQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper claims that continual test-time adaptation of a frozen language model becomes fragile when it leans on a single continually-updated low-rank adapter. This is because recurring domains make fresh updates overwrite earlier specialization while periodic resets throw away reusable state. The paper posits that the harder problem is deciding when to route, spawn, or reuse an adapter on a live unlabeled stream that offers neither task identities nor reliable boundaries. To overcome these limitations, this work presents TopoLoRA, which treats adaptation as an adapter lifecycle problem and uses the shape of the adapter's response cloud. The key idea involves two ingredients: (i) a canonical response coordinate that is invariant to the GL(r) gauge freedom of the low-rank factorization and preserves pairwise response distances, so the topological descriptor measures the model rather than the arbitrary internal bookkeeping; and (ii) a block comparison around an excluded transition band that cancels any transient inside that band exactly, while retaining the full separation between sustained regimes, so short bursts do not open unnecessary adapters. Experiments are conducted across three public continual-learning protocols (Standard CL, Long Sequence, TRACE) on T5-Large and LLaMA-3.1-8B, a set of author-constructed recurring variants, an unlabeled continual test-time setting on AdaptEval, and a synthetic moment-matched detection study, where TopoLoRA matches the multi-task bound on Standard CL (80.0), reaches 43.2 on TRACE against 40.0 for the strongest replay-free method, falls by only 0.7 under recurrence where the strongest baseline falls by 5.5, and attains 0.912/0.841/0.906 detection AUC against 0.727 for the best of twenty-four alternatives. Notably, the authors reproduce every quoted baseline to within 0.5 points in their own pipeline before comparing, and also report a homogeneous stream on which the method is beaten by plain replay.

Critical strengths
  • The central insight is genuinely novel. First- and second-moment detectors cannot see a change in shape, so promoting a topological descriptor to the control signal is well motivated. Specifically, the observation that such a descriptor must be computed on a gauge-invariant coordinate is both subtle and correct.

  • The work is backed by sound theoretical guarantees. Proposition 2 / Corollary 1 show that a transient contained in the excluded band contributes exactly zero to the score while a sustained change contributes the full regime separation, which is a clean property.

  • The experimental honesty is well appreciated. The authors reproduce every published baseline to within 0.5 points in their own pipeline before drawing any conclusion (Table 2, Figure 4), rebuild the TRACE stream from original sources when the official release was unreachable and recover the published O-LoRA number (23.4 vs 23.1).

Critical weaknesses
  • In all honesty, the practical payoff outside the recurrence regime is modest relative to the machinery introduced. On the single-pass public protocols TopoLoRA matches or only marginally exceeds the strongest replay-free methods (e.g. +0.1 over ELLA on Standard CL, and 80.0 is the multi-task ceiling rather than a margin), on the unlabeled streams the per-domain oracle is still better, and on a homogeneous stream it is beaten by both shared replay and task-aware SD-LoRA (0.771 vs 0.791 and 0.796).

  • The sensitivity sweeps (Table 33) are reassuring within this design, but the method is intrinsically tied to the low-rank gauge it exploits, so it is unclear how the approach would transfer to adaptation schemes that are not low-rank.

Questions for authors
  • The recurring variants are central to the contribution. Would it be possible to construct a recurrence setting from an existing published benchmark protocol, so that at least one comparison point is external rather than every method being re-run in your own pipeline on a stream you defined?

  • The canonical coordinate and the whole topological signal rely on the low-rank structure of the adapter. Can the approach be extended, or at least diagnosed, on a non-low-rank adaptation scheme (e.g. a small fully fine-tuned head or a prompt-based adapter), or is the gauge-invariance argument specific to LoRA-style residuals?

  • Practical value proposition. Given that the clear gains concentrate in the heterogeneous-and-recurring regime while simpler replay matches you elsewhere, in which deployment setting does the added lifecycle machinery and its five hyperparameters justify itself over a well-tuned replay baseline?

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

Related papers

Discussion (0)

Sign in to comment.