acceptodds
Withdrawn from ICLR 2027

Plasticity-Compatible Representations Enable Fast Adaptation through Writable Interfaces

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Long-lived adaptive systems may retain useful internal computation even as their observation or output mappings change during deployment. A central question is therefore not only how to adapt, but what learned representations make simple, restricted adaptation effective. We study a two-timescale recurrent framework in which outer-loop task learning establishes persistent dynamics and a low-dimensional query representation, while deployment-time feedback updates only a small, resettable interface. We show that the geometry of the learned queries directly governs the functional consequences of local plasticity: a feedback-driven write at one state transfers to another in proportion to their query similarity, determining both useful generalization and collateral interference. This yields a mechanistic criterion for plasticity-compatible representations, which we validate through analytic characterization and targeted interventions on the learned write/read geometry. On held-out action remappings, the resulting reward-modulated interface adapts rapidly from sparse binary feedback and remains robust when feedback is corrupted, without modifying the recurrent backbone. We further test the principle on neural population recordings, where writable interfaces support retrospective adaptation to a contingency reversal and cross-day recording drift by targeting output- and observation-side mismatches, respectively. Together, these results show that learned representations determine not only what can be read out from a fixed computation, but also how subsequent local learning propagates through it, providing a route toward selective adaptation without relearning reusable dynamics.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Plasticity, Compatible Representation, Fast Adaptation, Writable Interface

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_jPHeOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper explores how fast-adaptation of input/output mappings can adapt a meta-trained recurrent integrator to shifts in both input distribution and output targets. The authors demonstrate that the representations learned by the integrator are designed to facilitate fast-adaptation. Furthermore, the framework is shown to be robust to noisy feedback, and capable of adapting to contingency-reversal and cross-day representational drift in real neural data.

This paper should be rejected foremost because 1) the absence of many methodological and experimental details hampers any assessment of the claimed results and 2) there are abundant issues with the presentation.

Critical strengths

None

Critical weaknesses

The most pressing omissions involve the following:

● The description of problem setting is so general that it gives no insight into the actual data scheme (153). For example, do the test-time target actions (Eq. 7) relate to the training labels trained in the outer-loop (Eq. 6)?

● A main result of the paper is the relationship between the query geometry and fast-weight updates. However, the relevant equation (Eq. 11) describes a quantity (\Delta \ell^\text{fast}) which is never defined. As a result, it is unknown how the RHS was arrived at, but it is likely missing factors of \lambda^{j-i} or similar.

● The function of \tau_C for the adaptation to representation drift is never explained, nor its set value justified (297). Furthermore, the object “V” in the equation for the interface (Eq. 16) is never specified.

● No description of the training scheme or model hyperparameter settings is given for any experiment.

● The experiments and figures refer to a variety of interventions (365), alternative adaptation schemes (376), one-hot feedback (403), and supervised tuning schemes (423) that are either inadequately described in the main text or, more often, not described at all - sometimes even including the appendices.

● The neural data for the contingency-reversal experiment is never described, nor is the specific data used for the adaptation results (416) explained. As a result, the scope of the claimed adaptability to representation drift is not clear - can the model adapt to reversed-contingencies given neural activity from the reversed setting or the original setting?

● The main text fails to mention, let alone describe, the second data set used for the neural representation drift experiments (429). Furthermore, the training data for the model is described in the appendices as including imaging FOV and recording dates (772) without justification, as well as an unspecified “offline analysis variable” (787).

Issues with presentation includes the following:

● The comparison of the proposed model to alternatives (Fig. 3) and the main text discussion (403) reference an unknown value K. In addition, the scalar values labeled “feedback corruption” are never explained (Fig. 3).

● The main text claims a relationship between query representations and stimulus identity (355) that is not evident from the data as presented (Fig. 2). In particular, no explanation is given for what the observed block structure of the learned queries (Fig. 2b) corresponds to. As an aside, why do random queries result in high overlaps?

● The interpretation of the connection between the query geometry and the plasticity reads as nonsensical (260)

● The contingency reversal results (Fig. 4) mention “ALM activity” without explanation (395), and include multiple plots that are never discussed (panel B, and the middle and right plots of C and D)

Questions for authors

Abstract:

● description of framework is unclear - how does the “query” relate to the interface?

● last sentence summarizing results seems vacuous

Introduction:

● italicized question is ambiguous - are you asking if such a system can be constructed, what properties of an architecture allow for such modularity, or something else?

● it would be helpful to clarify that “query” = “task representation”

● the first two points in the summary of contributions are almost identical (094, 097)

● Repeated references to “learned representations” are vague

Related work:

● the connection between the current work and the referenced literature is often weak

● reference to motor memories (140) seems irrelevant

Method:

● Description of problem setting gives no insight into the actual data scheme (153)

● In the training data, do the test-time target actions (Eq. 7) relate to the training labels trained in the outer-loop (Eq. 6)?

● Eq 10: write z_n and d_n explicitly to highlight the “three factor” structure

● Eq. 10: why choose to set the write and forget rates independently?

● Eq. 11: \Delta \ell^\text{fast} is not defined. It is not clear how the RHS was arrived at, but it appears to be missing factors of \lambda^{j-i} or similar

● Unclear interpretation of the query geometry (260)

● The description of the input representation shift is unnecessarily vague (266)

● Fig. 1a,b: could be replaced by an illustration of the general concept

● Fig. 1c: uses notation that is inconsistent with the maintext

● Fig. 2b: What are “read” and “write” trials? In the demonstrated block structure, what delineates blocks? Why does “random” show high overlap?

● Fig. 2 caption: Introduces unseen notation of B_t, and “post-support”

● The spike statistics \mu_s and \sigma_s are not defined (296)

● \tau_C and its set value are never explained (297)

● Eq. 15: The description of V is insufficient: what is the “source-only activity basis” that makes up V? Is V a matrix?

● The description of the input-adaptive update is overly complex and could be greatly simplified (320-323)

● Fig 3: What is K here? Is this the same K as in Sec. 3.6?

● Fig 3: What do the scalar values for “feedback corruption” correspond to?

Experiments:

● An overview of the training scheme and hyperparameter settings should be provided

● The claimed stimulus-related query geometry is not obvious from the presentation of the data in Figure 2 (355-358)

● The last two interventions are insufficiently detailed, and the figure shows an additional 2 interventions not mentioned in the maintext (365-370)

● The alternative adaptation schemes need a maintext description

● One-hot feedback is mentioned but never described (403)

● The discussion of corrupted feedback needs to reference Fig. 3c (407-412)

● Need to include a description of the dataset and setting in which it was recorded (416). It is unclear what data is used for both training and evaluation of the adaptation - is the model adapting to reversed contingencies given neural activity from the reversed-contingency setting, or the training data? This is important for interpreting the results.

● Fig 4: A discussion of the reward-noise robustness results is completely absent from the main text

● Fig 4: What is “ALM”? The caption is also missing an explanation of the middle and right panels.

● No description of the dataset used for neural population drift analyses (429)

● Fig 5c: No description of the 7 supervised tuning schemes shown on x-axis

● No explanation of the referenced numbers (453)

Discussion:

● Indecipherable explanation of plasticity compatibility (469)

● Why would training the model to map neural recordings to action outputs imply anything about the biological plausibility of the model itself? (481)

Appendix:

● The data-generating scheme for the synthetic data has a level of complexity that needs to be motivated

● The description of how the dataset was constructed from the neural recordings is completely unclear - why is imaging FOV and date included (772)? The choice of “offline analysis variable” for the target is completely unclear (787).

Reviewer_ED3UOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper studies test-time adaptation of a frozen recurrent decoder through a small, resettable "writable interface" targeting the input and output interfaces of the network, in an attmept to preserve core computational capabilities over the course of test-time adaptation. A GRU serves as the core computational unit, initially trained on a source task, then frozen. For output-side shifts, an additive matrix B is updated once per trial. A single write at trial "i" changes the fast logits at trial "j" in proportion to their inner product in normalized latent space. The authors call representations on which this transfer is selective "plasticity-compatible." For input-side shifts, a separate two-stage adapter first normalizes channels using unlabeled support trials. It then updates a low-rank input residual using gradient-derived eligibility through the frozen network.

Critical strengths

My main view of the paper's upside is the showcase of practical viability: On all held-out days, a small resettable input-side adapter driven by trial-level feedback improves on unsupervised alignment with feedback, coming quite close to fine-tuning performance. The binary-feedback version also beats supervised readout-only and LoRA adaptation, though attributing that to location is confounded.

While I like some parts of the experimental protocol and overall design (eg., the corrupted feedback test), I would not count them as a "reason to support acceptance" since they are generally good practices in the literature.

Critical weaknesses
  1. Framing and novelty.

The paper describes its training as an "outer loop" and "inner loop" on "two timescales" and positions itself against meta-learning and fast-weight work, but it is not meta-learning. The outer loop is ordinary cross-entropy on a single source task. There is no task distribution, the fast interface B is absent during training, the update rule is hand-designed, and the training objective never evaluates performance after adaptation, as MAML, OML and RL^2 each do in different ways. Nor is it a fast-weight model in the usual sense, since the slow weights are never trained with B in the loop. What the paper actually studies is how supervised source representations support restricted test-time adaptation. This question is well established: linear probing and few-shot fitting on frozen supervised features (Tian et al., 2020; Kumar et al., 2022), how class geometry affects transfer (Kornblith et al., 2021; Galanti et al., 2022), and which layers to adapt for a given shift (Lee et al., 2023). Studying it can still be a contribution, but only if the paper shows something these works do not.

The abstract and contributions also overstate what is shown. The principle is said to be tested on neural recordings, but no query kernels or interventions are reported for either dataset, and the input-side adapter has no corresponding analysis. The criterion is said to go "beyond source-task performance," which is never tested. Adaptation without deployment-time backpropagation holds only for B, a property readout adaptation shares; the input-side adapter backpropagates through the GRU on every trial.

  1. Positioning and related work.

The cross-day experiment is domain adaptation under covariate shift; the remapping experiments are readout adaptation from bandit feedback. The framing and related work should follow. The cited test-time training layers and Titans are sequence-model memory architectures, not adaptation under distribution shift. Missing work includes:

  • Test-time and domain adaptation: AdaBN (Li et al., 2016), TENT (Wang et al., 2021), test-time adaptation with binary feedback (2025), surgical fine-tuning (Lee et al., 2023), and output-side label-shift correction (Saerens et al., 2002; Lipton et al., 2018).
  • Bandit feedback and the update rule: REINFORCE (Williams, 1992) and the Banditron (Kakade et al., 2008), associative reward–penalty (Barto & Anandan, 1985). Delta-rule memories (DeltaNet, Gated DeltaNet) address the interference of Eq. 11 at the write rule itself.
  • Representations for later learning: OML (Javed & White, 2019), MER (Riemer et al., 2019), ANML (Beaulieu et al., 2020), Sorscher, Ganguli & Sompolinsky (2022), Chang et al. (2024), Sadtler et al. (2014), Golub et al. (2018), neural collapse (Papyan et al., 2020), the NTK, and local elasticity (He & Su, 2020).
  • Neural decoding: RL-based decoders that adapt from evaluative feedback to both input reorganization and output changes (DiGiovanna et al., 2009; Pohlmeyer et al., 2014; Iturrate et al., 2015), POYO unit identification (Azabou et al., 2023), per-session read-in layers, and ADAN (Farshchian et al., 2019). If the meta-learning framing is kept, the paper needs genuine meta-learned comparisons, such as training through the B update, OML, or an RL^2-style model with feedback as input. The only one included, ANIL, is undescribed and performs below chance.
  1. The central claim is that query geometry predicts adaptation "beyond source-task performance." To support this, experiments show that for a given source performance, geometry can (and must, in this case), and adaptation must follow the geometry - i.e., some representation geometries should adapt better than others. The experiments never separate the two. Solving the source task (identity over four stimuli) requires separating the four stimuli, and that separation is exactly what a permutation target needs. Source competence and favorable geometry are therefore entangled in this setup. The results are equally explained by "better source models adapt better," which is not a new claim or finding.
Questions for authors
  1. Can you show two representations with matched source accuracy but different kernels, and that adaptation differs as a kernel statistic predicts? A many-to-one source mapping followed by a target that splits the merged stimuli would be a direct test.

  2. The one-hot update is online gradient descent on the effective readout, and the binary update shares its direction with the adaptive-readout gradient. What distinguishes the method from readout adaptation, and what explains the gap between them?

These questions would clarify the contribution, but I do not expect answers to them to resolve my main concerns. Those concerns cannot be addressed in a rebuttal or with minor corrections: reframing the work away from meta-learning, positioning it against the test-time adaptation literature, and adding dissociation experiments would amount to a different paper. I am open to a rebuttal that identifies misunderstandings on my part; otherwise, I recommend resubmission after substantial revision.

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

This paper explores how learned representations can help recurrent neural networks to adapt quickly to new tasks without changing the network computations and representations. The main question the authors ask is whether representations developed during the initial training ,can support later learning through local updates.

The authors develop a recurrent framework with two components that function on two timescales, a recurrent backbone and a temporary interface that can be updated during use. During the initial training phase, a gated recurrent unit learns how to process sequences of inputs and to generate low-dimensional representations, that the authors call "queries". These queries guide action selection. When training has been completed, the recurrent network and its readout remain unchanged, and instead of updating the entire network, learning happens only through a small interface that changes from trial to trial based on feedback.

For action-remapping tasks this interface adjusts the output logits using the query representations, and these adjustments follow a three-factor plasticity rule involving the query, an action-dependent eligibility signal, and binary feedback indicating whether the selected action was correct or not.

The authors attempt to explain why some learned representations are better fit for this kind of adaptation than others and show that when an update is constructed using one query, its effect on another query depends on the inner product between their representations. Depending on whether the queries require to be associated with same or different actions/outcomes this may be beneficial or not. (This is the usual property of associative memories) The authors characterise those representations that allow useful generalisation while minimising interference as "plasticity-compatible representations".

The authors then examine how these representations change during the initial training process. They find that training makes query representations more consistent for repeated presentations of the same stimulus while increasing the separation between representations associated with different actions. This change in representation geometry appears to make later plasticity updates more selective. In controlled experiments, updates made after training produce less interference between different stimulus representations. The authors also test what happens when they deliberately disrupt the learned structure. When query representations are collapsed together, or when the coordinates used to write updates no longer match those used to read them, the network’s ability to adapt becomes substantially worse. These experiments suggest that the organization of the representations plays an important role in successful adaptation.

They test the proposed framework first on synthetic action-remapping tasks involving four stimuli and four actions, where the stimulus-action contingencies change between training and deployment. The proposed interface adapts to these changes within a small number of trials using only binary feedback about whether each response is correct, and remains effective when some of its feedback is corrupted.

The authors then test whether the same approach can be applied to neural population recordings, and in particular for a tactile discrimination task, they study how the model responds when the relationship between stimuli and actions is reversed. They demonstrate that the writable output interface can help the model recover decoding performance without changing the recurrent structure.

Critical strengths
  • The authors propose an intresting question to study, which representational properties make a local plasticity rule effective.

  • I like the attempt to connect the geometry of representations with the functional consequences of plasticity. The authors study how individual updates propagate through the states, and use interventions to establish the importance of the learned write/read geometry

Critical weaknesses
  • the cosine similarity of Eq. 12 does not include the query bias

  • The recurrent network is trained using ordinary supervised cross-entropy, thus it is unclear to me whether the observed geometry is specifically adapted for plasticity, or whether it simply reflects conventional discriminative representation learning

  • The authors do not provide enough evidence for preserved computation

  • the comparison wiht the Hebbian update is unfair, because it was expected to be poor as an unsupervised update

Questions for authors
  1. Can the authors give some justification behind the normalisation of the plasticity rule? As I understand it is a policy gradient rule with added normalisation. Can you characterise how this affects the leanning dynamics or convergence? Can you compare also with regular policy gradient updates?

  2. Since the outer loop optimises only the source task cross-entropy, how can the authors justify that that the learned geometry is specifically compatible with plasticity rather than it being discriminative?

  3. How does the proposed framework perform, when the deployment shifts require corrections that are inconsistent with the source-task similarity structure?

Comments on the paper

1 comment

Dear AC and Reviewers,

Thank you for your careful evaluation and constructive feedback. Your comments have helped us identify important directions for clarifying the paper’s positioning and strengthening its theoretical and experimental support. Given the scope of the revisions and additional validation needed, we cannot address these points thoroughly within the two-week rebuttal period. We have therefore decided to withdraw the submission and undertake a comprehensive revision. We sincerely appreciate your time and thoughtful suggestions, which will guide the next stage of this work. We look forward to sharing a substantially revised and strengthened version in the future.

Best regards.

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

Related papers

Discussion (0)

Sign in to comment.