acceptodds
Under review as a conference paper at ICLR 2027

CoPA: Compressing Visual In-Context Learning for Parameterized Test-Time Adaptation

Abstract

We study visual few-shot adaptation from labeled support examples. CoPA processes a support set with a frozen vision-language model, compresses the resulting activations with a Perceiver resampler, and uses an episodically meta-trained hypernetwork to generate low-rank updates for query inference. The backbone remains frozen and no gradients are taken at inference time. Across 14 datasets and four backbones, we evaluate predictive performance for up to 16 support examples. In this range, CoPA outperforms the frozen prompting and one-step test-time optimization baselines evaluated here. These comparisons do not isolate the effect of the architecture because CoPA also receives supervised episodic training. A separate profile up to 64 supports shows that query-phase memory and latency are approximately independent of support-set size, while support processing still grows with that size. We characterize CoPA as an amortized, support-conditioned few-shot adapter, rather than a training-free in-context learning method or an unlabeled test-time adaptation method.

Primary area:
foundation or frontier models, including LLMs
Keywords:
In-context learning, test-time adaptation

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_65ikOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the visual few-shot adaptation from labeled support examples in language-model in-context learning. The authors propose CoPA to compress activations via a Perceiver Resampler and use an episodically meta-trained hypernetwork to generate low-rank updates for query inference. Experiments are conducted across 14 datasets and 4 backbones. The results demonstrate that CoPA outperforms frozen ICL and one-step test-time training.

Critical strengths
  1. The authors honestly state that CoPA is not training-free and the end-to-end single-query latency is not better than ICL.

  2. The experiments cover 14 datasets, 3 task families, and 4 backbones.

  3. Section 4.5 and Appendix G provide useful information about transfer and negative transfer.

Critical weaknesses
  1. There is no comparison against methods that use the same labeled training data and the same adaptation. This is necessary to check whether the Perceiver resampler plus hypernetwork is needed.

  2. One-step TTT matches frozen ICL to two decimal places across nearly all Table 1 entries, suggesting negligible update effects. Since these baselines are neither tuned nor multi-step, they are weak proxies for test-time optimization, undermining the claim of outperforming TTT.

  3. Efficiency claims are overstated. Although query-phase cost is constant in K, support processing is not, and single-query latency exceeds ICL at both K = 16 and 64. The key scenario of multiple queries sharing one support set is unbenchmarked, and the absence of accuracy results at K = 64 leaves the long-context claim unverified.

  4. Experimental reporting limits confidence in the results. The main table lacks variance estimates and multiple seeds, and regression MAEs across different scales are averaged together. Cars’ zero-shot accuracy of 0.00 suggests a possible output-format failure that exaggerates the apparent benefit of supervision. Moreover, the authors acknowledge that the FFN down-projection is suboptimal (Appendix H), yet use it for the main results.

If the authors can solve my concerns during the rebuttal, I am happy to raise my scores.

Questions for authors
  1. What are the results of a LoRA/adapter baseline trained on the same labeled episodes and evaluated at the same shot counts?

  2. How do the Perceiver and hypernetwork components compare to simpler aggregation?

  3. Do the conclusions hold across seeds and across the other three backbones on more than one dataset?

  4. What is the amortized latency when many queries share one support set?

5 comments

Thank you for the constructive review and for recognizing our training assumptions, evaluation breadth, and transfer analysis. We address your four weaknesses and questions below.


W1:

We agree that the comparisons do not isolate architecture from supervision. CoPA receives labeled episodic meta-training; frozen ICL does not. Our goal is support-to-weight adaptation without gradient-based training at test time, not a training-free method. We will qualify the results and acknowledge that superiority over supervision-matched static LoRA remains unestablished.

The Perceiver provides fixed-length support representations and the hypernetwork generates low-rank updates, but their necessity is not isolated. A preliminary fast-weight implementation inspired by “Linear Transformers Are Secretly Fast Weight Programmers” performed poorly, motivating hypernetwork exploration inspired by Text-to-LoRA. This is design context, not a controlled ablation; attention visualizations likewise do not establish component superiority.

W2:

These baselines use one AdamW update with learning rate 10−410^{-4}, matching CoPA’s adapter rank and site but not its supervision or optimization budget.

We will consistently say “the evaluated one-step TTT baselines,” document their settings, and remove broader superiority claims. Their near-equality with frozen ICL does not establish that tuned or multi-step TTT is ineffective.

W3:

Query-phase independence from support count does not imply constant end-to-end cost. At 16 supports, CoPA totals 1596 ms versus 972 ms for ICL; at 64 supports, 6882 versus 4542 ms, also with higher overall peak phase memory.

The demonstrated property is approximately support-count-independent query-phase cost after adaptation, not lower single-query latency or overall peak memory. Multi-query amortization is unbenchmarked; Q4 provides a conditional estimate. Predictive evaluation extends to 16 supports, while profiling extends to 64; we will not claim accuracy scaling beyond the evaluated range.

W4:

Evaluation uses seed 42 and up to 1,000 queries per dataset, reusing the query subset across shot counts. For each query and shot count, one random support set is sampled without replacement from the training split, without repeated draws. We will report valid-prediction counts, acknowledge missing across-seed uncertainty, and remove the cross-scale regression “Avg.” column.

We double-checked the Cars code and found no implementation error. Our interpretation is that producing highly specific category names without examples is difficult under exact matching; demonstrations clarify the label vocabulary. We will not interpret 0.00 as absent recognition capability or attribute the entire gain to improved recognition.

Down-projection is a fixed default, not the best tested site. Up-projection improves Age MAE from 5.21 to 5.04 and EmoSet accuracy from 76.4% to 78.1%. We will acknowledge this without assuming generalization to other tasks.


Q1:

We do not have this baseline. It would help separate supervised task specialization from support-conditioned generation; one-step TTT is not a substitute. We will acknowledge this unresolved comparison.

Q2:

We lack controlled comparisons against pooling or another aggregator with a comparable weight generator. The preliminary alternative under W1 does not replace these controls. Figure 5 will remain descriptive, not evidence that the Perceiver causes the gains.

Q3:

Single-seed evaluation does not establish across-seed robustness. We will report evaluation and training-seed counts and avoid treating small differences as statistically established.

The other three backbones are evaluated on VQAv2 only; the 14-dataset study primarily uses Gemma 3 4B-it. We will clarify that this is not a complete 14-dataset-by-four-backbone evaluation.

Q4:

For QQ serial queries sharing reusable supports, modeled average CoPA cost is

T‾(Q,K)=SK/Q+CK,\overline{T}(Q,K)=S_K/Q+C_K,

using reported support- and query-phase times. At 16 supports this is 1480/Q+1161480/Q+116 ms; at 64, 6765/Q+1176765/Q+117 ms. Against repeated uncached ICL at 972 and 4542 ms per query, respectively, both yield a modeled crossover at two queries.

These are estimates, not throughput measurements. They assume adapter reuse, exclude batching and scheduling, and do not establish an advantage over prefix-cached ICL or apply directly when retrieval changes supports.

Thank you again for the helpful feedback. We hope these clarifications and revisions address your concerns and would appreciate reconsideration of the score if you find them sufficient.

Thank you for the response. I appreciate your agreement that the current comparisons do not disentangle architectural impact from supervision, as well as your plan to tone down and qualify the claims regarding TTT and efficiency.

However, most of my primary concerns have been acknowledged rather than resolved, as the rebuttal does not present empirical results for Q1, Q2, or Q3. Consequently, the actual necessity of the Perceiver and hypernetwork remains unverified. Furthermore, the clarification that TTT baselines rely on a single AdamW update at a learning rate of 1e-4 suggests their performance would naturally mirror frozen ICL by design, severely limiting their informative value as baselines. Because the core methodological questions remain unaddressed empirically, I am maintaining my score at 2.

Two additional concerns:

  1. Regarding Cars, Appendix F notes that candidate categories are explicitly provided in the prompt. This detail makes an exact zero-shot accuracy of 0.00% difficult to justify, especially given that Food-101 achieves 35.1% under the identical prompt template. Could you provide sample raw predictions and evaluate performance using constrained label matching?

  2. The amortization analysis in Q4 is benchmarked solely against uncached ICL and omits the kNN setting reported in Table 1, where support sets vary dynamically across individual queries.

Thank you for requesting the Cars diagnostic. Preliminary results show that our earlier assurance about the evaluation implementation was too strong.

On the same 100 Cars queries using Gemma 3 4B-it, with all 196 candidate categories provided:

ProtocolAccuracy
Original free generation, 15-token cap and scorer0/100
Same outputs, separator-normalized candidate parsing0/100
Candidate-constrained greedy decoding53/100

Of the unrestricted outputs, 98 reached the generation limit without a recognized ending token. Examples include “Based on the image, the vehicle is a Ford Edge SUV (” for Ford_Edge_SUV_2012, and “Based on the image, the car is a Ferrari 458 for Ferrari_458_Italia_Coupe_2012.

Thus, the reported zero is highly sensitive to the output protocol and cannot establish absent recognition capability. Separator normalization alone does not resolve it. Constrained decoding also allows complete category names, so this comparison does not isolate decoding constraints from generation length. We retract our earlier vocabulary-based explanation and will reassess the zero-shot comparison. The 53/100 result is a subset diagnostic, not a replacement for Table 1.

Additional evaluation-seed experiments are running; we will report their results separately. These do not resolve the missing supervision-matched or component baselines.

We also agree that shared-support amortization does not apply when kNN supports change per query. We withdraw the crossover estimate as evidence for Table 1’s setting and claim no advantage over prefix-cached ICL.

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

The paper proposes CoPA, which converts a labeled visual support set into temporary low-rank weights for a frozen vision-language model. A no-gradient support pass caches activations at one adapted layer, a Perceiver resampler compresses them into a fixed set of latent tokens, and a hypernetwork meta-trained on labeled episodes emits a LoRA update for the query pass, after which the supports are discarded. On 14 regression, classification, and VQA datasets with Gemma 3 4B-it, CoPA meta-trained on each target dataset outperforms frozen in-context learning and two one-step test-time training baselines for up to 16 shots under both random and retrieved supports, with VQAv2 gains also shown on three other backbones. A study of meta-training mixtures shows positive transfer within a task family and negative transfer across modalities, and a cost profile shows that query-phase memory and latency stay flat up to 64 shots while support processing grows. The paper states explicitly that its comparisons do not separate the architecture from supervised meta-training.

Critical strengths

The method is described clearly, and when meta-trained on the target dataset it improves over frozen prompting on all 14 datasets, under both random and retrieved supports, and on VQAv2 across four backbones.

Critical weaknesses

The central claims are left untested. The main CoPA results come from hypernetworks meta-trained on each target dataset, while frozen ICL receives no training and the one-step TTT baselines nearly match it, often to the second decimal. The paper lists supervision-matched baselines as an open question, but runs neither a static LoRA trained at the same site on the same episodes nor any support ablation (for example, shuffled labels or supports from another dataset). Its own results leave the role of the supports unclear: CoPA is essentially unchanged between random and retrieved supports and flat from one to sixteen shots where plotted. The efficiency case is similarly limited, since CoPA is slower and uses more peak memory than frozen ICL for a single query at every profiled support size. Therefore, the evidence currently shows that supervised meta-training on the target dataset beats untrained prompting, not that the generated weights adapt to the support set.

Questions for authors
  1. Could you report a static LoRA at the same site and rank, trained on the same episodes and budget without support input, and CoPA evaluated with shuffled-label supports and with supports from another dataset, on a few datasets from each task family?
  2. Which backbone, prompt, selection strategy, and evaluation set produced Figure 1, and why does it differ from Table 1 and Figure 3? Are the Table 1 CoPA entries averages over shot counts or 16-shot values, which configuration does each main-paper row of Table 4 correspond to, and how many test queries and seeds underlie each entry?
  3. The title presents CoPA as visual in-context learning for test-time adaptation, but the abstract and Section 3.1 state that it is neither training-free in-context learning nor unlabeled test-time adaptation, since it needs labeled supports and supervised episodic meta-training. Why is the work framed this way rather than as a meta-trained few-shot adapter, which is how the paper itself characterizes CoPA and which would call for supervised few-shot baselines?
5 comments

Thank you for the detailed review. We address the weaknesses and questions below, grouping overlapping concerns.


W1:

We agree that the main comparisons do not isolate architecture from additional supervised training. CoPA receives labeled episodic meta-training; frozen ICL does not, and the TTT baselines use one inference-time update. Matching adapter rank and insertion site does not match supervision or optimization budgets.

A static LoRA trained on the same episodes, site, rank, and budget is an important missing control. We do not have this result and will not claim superiority over supervision-matched LoRA. We will qualify the results as comparisons between complete systems under different training protocols. Our contribution is the support-to-weight procedure and its evaluated predictive and computational behavior, not an isolated architectural explanation for the gains.

W2:

These results support only a limited conclusion about test-time optimization. They characterize our one-step implementations, not tuned or multi-step TTT. We will consistently use “the evaluated one-step TTT baselines,” remove broader superiority claims, and avoid attributing their small effects to text-token loss masking without direct evidence.

W3:

We agree that computing weights from support activations does not establish how much performance depends on the particular examples or their correct labels. The small random-versus-retrieved differences and flat shot-count curves admit multiple explanations, including early saturation or limited episode-specific support dependence.

We will report these as selection insensitivity and performance stability over {1,2,4,8,16}\{1,2,4,8,16\} shots, not proof of effective information aggregation or many-shot accuracy scaling, and revise the appendix interpretations. The results do not rule out a substantial contribution from task specialization learned during meta-training.

W4:

Your reading of Table 2 is correct: CoPA does not improve single-query total latency or overall peak memory over frozen ICL. At 16 shots, its phase times total 1596 ms versus 972 ms, with peak phase memory of 10.72 versus 9.83 GB. At 64 shots, the corresponding figures are 6882 versus 4542 ms and 17.65 versus 14.95 GB.

The demonstrated property is narrower: after support processing, query-phase memory and prefill time are approximately independent of support count under the fixed configuration. Support-processing cost still grows with support count. We will not present query-phase savings as end-to-end savings.


Q1:

We do not have these control results. A supervision-matched static LoRA would help separate task specialization from the benefit of support-conditioned generation. Shuffled-label and mismatched-support evaluations would help establish functional dependence on support examples and labels.

The transfer study changes meta-training data while retaining target-dataset supports at inference. It addresses a different question and does not substitute for inference-time support ablations. We will acknowledge these unresolved distinctions rather than present existing results as answering them.

Q2:

Figure 1 uses Gemma 3 4B-it, random supports, dataset-specific prompts, and seed 42, without repeated support-sampling runs. We will identify the exact prompts and evaluation splits in its configuration description.

Table 1 averages over {1,2,4,8,16}\{1,2,4,8,16\} shots; its entries are not 16-shot values. Figures 1 and 3 show shot-specific results. This explains the aggregation difference, but not discrepancies at identical shot counts. We will reconcile the configurations and plotting records and correct inconsistent labels or values.

The main configuration is layer-8 FFN down-projection with rank 8. The different values in Table 4’s “Mid (Main Paper)” and down-projection rows require clarification. We will replace the ambiguous label with the full configuration and explain or correct the discrepancy using the run records.

Our protocol uses one query per support episode and a single evaluation seed, 42. One query per episode differs from the total number of evaluated queries. We will report that total for each dataset, distinguish it from support and seed counts, and state that the evaluation does not establish across-seed robustness.

Q3:

CoPA is most precisely described as a meta-trained, support-conditioned few-shot adapter. In the title, “compressing visual in-context learning” refers to replacing labeled demonstrations in the query context with generated weights; “parameterized test-time adaptation” refers to producing temporary weights at deployment, not optimizing them through inference-time gradients.

We would prefer to retain the title while making this meaning explicit and consistently distinguishing CoPA from training-free ICL and unlabeled TTA. This clarification does not remove the relevance of supervised few-shot controls, whose absence we acknowledge above.

Thank you for identifying the distinction between supervised task specialization and support-conditioned weight generation. We have now added two static LoRA controls on Age:

  1. Query-trained LoRA: trained without support input on individual labeled image–answer pairs.
  2. Episodic LoRA: trained on labeled support–query episodes with query-completion supervision.

Both use rank 8 at the same layer-8 FFN down-projection. At evaluation, each fixed adapter receives support examples through the ICL prompt.

MethodK = 1K = 2K = 4K = 8K = 16
Frozen ICL————11.82 ± 1.01
Query-trained static LoRA + ICL — new6.89 ± 0.317.04 ± 0.819.25 ± 0.6413.31 ± 1.0914.33 ± 2.20
Episodic static LoRA + ICL — new6.49 ± 0.826.75 ± 0.536.58 ± 0.396.67 ± 0.626.14 ± 0.31
CoPA————5.03 ± 0.23

Entries are Age MAE ↓ under random support selection, using three evaluation seeds and 100 valid predictions per seed. Values are mean ± sample standard deviation across evaluation seeds, not independent training runs. Dashes indicate settings outside the additional three-seed evaluation.

The stronger episodic LoRA baseline substantially reduces the gap to CoPA. Its results also show why query-only-trained LoRA evaluated with long demonstration prompts should not be the sole supervised control. At K = 16, CoPA retains lower mean error than episodic static LoRA: 5.03 versus 6.14 MAE.

The two LoRA variants use 200 optimizer updates and 6,400 supervised query presentations. Although they share the task training split, they do not reproduce CoPA’s historical episodes or exactly match its training budget. We therefore present this as an additional supervised-control experiment, not a fully matched architectural isolation.

This experiment addresses part of Q1, but not the requested shuffled-label or cross-dataset-support evaluations. In particular, outperforming a fixed adapter does not by itself establish that CoPA functionally depends on the correctness of each support label. Those support-dependence questions, broader task-family coverage, and the configuration-reconciliation issues in Q2 remain separate limitations.

Thank you for the candid response and the new static LoRA experiment. I also appreciate the narrower claims on efficiency and the one-step TTT baselines.

  • My main concern remains open. The shuffled-label and cross-dataset support evaluations are still missing, although they need only inference with the existing checkpoints.

  • The new control is also limited. It covers one dataset and one training run. CoPA appears to have been trained for up to ten times as many steps (App. F), so part of its advantage may reflect training budget. The query-trained LoRA was also never evaluated without demonstrations.

  • The new CoPA result on Age at 16 shots is lower than the values the paper reports for the same setting, and those values already differ across figures and tables.

  • On the title, I understand the intended meaning, but readers will likely take "test-time adaptation" in its usual unlabeled sense.

Reviewer_Nrz7OpenReview ↗
Rating
2(−1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents CoPA, a support-conditioned adaptation method that compresses visual demonstrations with a Perceiver and uses a meta-trained hypernetwork to generate episode-specific LoRA weights. Support examples can be discarded before query inference, making query-side prefill cost independent of the number of shots.

Critical strengths

This paper turns context learning into lora mounting on the foundation model, eliminating the need for additional tokens for context learning. The experiments are also very comprehensive.

Critical weaknesses
  1. Training protocol mismatch between CoPA and the main baselines.
    As stated in Section 4.1 CoPA receives supervised episodic meta-training, while frozen ICL and one-step TTT do not. Although this limitation is acknowledged, it remains difficult to attribute the gains to the architecture rather than the additional supervision. A supervision-matched meta-trained baseline is needed for a more conclusive comparison.

  2. Missing comparison against standard episodic LoRA.
    CoPA generates episode-conditioned low-rank updates through the factorization in Eq. (2), but Table 1 does not include a standard LoRA model trained on the same episodic data. Such a baseline would clarify whether the Perceiver–hypernetwork design offers benefits beyond the low-rank parameterization itself.

Questions for authors
  1. Ablation of the Perceiver compression module is missing.
    The Perceiver resampler is introduced as the core fixed-length bottleneck, while Figure 5 provides only qualitative analysis. Comparisons with simpler aggregation schemes under the same latent budget would clarify the necessity and information cost of this compression.

  2. The boundary of the efficiency claim needs clarification.
    Table 2 shows that CoPA reduces query-side prefill cost but has higher end-to-end latency than standard ICL for a single query. Reporting the break-even number of queries would better establish the practical scope of this advantage.

3 comments

Thank you for the constructive review and for recognizing the support-to-weight formulation and the breadth of the evaluation. We appreciate your distinction between demonstrating the performance of the complete system and isolating the contribution of its architecture. We address each concern below.


W1:

We agree that the training protocols differ and that this limits attribution of the performance gains. CoPA learns a support-to-weight mapping through supervised episodic training, whereas frozen ICL receives no additional training and the evaluated TTT baselines perform only a one-step inference-time update. Table 1 compares complete adaptation systems under their stated protocols; it does not isolate architecture while holding supervision and optimization budget fixed.

We will make this distinction explicit in the contribution statements and main-results discussion, rather than relying only on the existing limitations. Our supported claim is that the complete meta-trained CoPA system outperforms the evaluated frozen-ICL and one-step TTT implementations. We do not claim that these comparisons demonstrate an architectural advantage under matched supervised training. The transfer experiments examine sensitivity to the source of meta-training data, but do not replace a supervision-matched baseline.

W2:

The reason we left LoRA out is because that the compute budget is substantially higher compared to existing TTT comparisons. Introducing LoRA finetuning will definitely perform better than CoPA which we initially thought adds little to no value to our comparisons as we are looking into TTT methods. However, we are happy to include those comparisons if we have the compute budget. We will also release out code (same as the one in the submission but organized) so that others with sufficient compute could attempt this as well.


Q1:

The Perceiver provides a fixed-length interface between variable-length support activations and the hypernetwork: the support sequence is mapped to M=128M=128 latent tokens before weight generation. This explains its architectural role, but does not establish that it is necessary or more accurate than pooling or another aggregation mechanism.

We agree that Figure 5 is not an ablation. Its attention maps are descriptive and cannot determine whether the Perceiver preserves the most useful information or causes the observed performance gains. We will remove remaining language implying that the visualization establishes selective information retention or the superiority of the resampler. The current evidence supports the behavior of the complete Perceiver–hypernetwork system.

Q2:

We agree. The efficiency claim concerns query-phase dependence on support size, not lower total latency for one query. For example, at K=16K=16, the reported CoPA phase times sum to 1480+116=1596 ms, compared with 972 ms for standard ICL. CoPA also does not have lower overall peak memory in this setting: its support phase reaches 10.72 GB, compared with 9.83 GB for ICL. The existing measurements permit a conditional amortization calculation. Let SKS_K denote CoPA’s support-processing time, CKC_K its query-phase time, and IKI_K the reported ICL time. For QQ serial queries sharing one reusable support set, the modeled costs are TCoPA=SK+QCK,TICL=QIK.T_{\mathrm{CoPA}}=S_K+QC_K,\qquad T_{\mathrm{ICL}}=QI_K. Under these assumptions, CoPA becomes faster when Q>SK/(IK−CK)Q>S_K/(I_K-C_K). The first integer crossover is 5, 3, 2, 2, and 2 queries for K=1,4,8,16,64K=1,4,8,16,64, respectively, using Table 2’s measurements. We will label this as an analytical estimate, not a measured multi-query benchmark. It assumes adapter reuse and repeated uncached ICL prefill; it does not establish an advantage over prefix-cached ICL or account for batching and scheduling. It also does not apply directly when query-specific retrieval changes the support set.

Thank you again for highlighting the missing supervised LoRA comparison. We have now added two static LoRA controls on Age: one trained on individual labeled image–answer pairs, and one trained on labeled support–query episodes with query-completion supervision. Both use rank 8 at the same layer-8 FFN down-projection. Their adapters remain fixed at inference, with demonstrations supplied through the ICL prompt.

MethodK = 1K = 2K = 4K = 8K = 16
Frozen ICL————11.82 ± 1.01
Query-trained static LoRA + ICL — new6.89 ± 0.317.04 ± 0.819.25 ± 0.6413.31 ± 1.0914.33 ± 2.20
Episodic static LoRA + ICL — new6.49 ± 0.826.75 ± 0.536.58 ± 0.396.67 ± 0.626.14 ± 0.31
CoPA————5.03 ± 0.23

Entries are Age MAE ↓ under random support selection, reported as mean ± sample standard deviation across three evaluation seeds, with 100 valid predictions per seed. Dashes denote settings not covered by the additional three-seed evaluation, not missing results in the original paper.

Episodic training substantially strengthens static LoRA and narrows the gap to CoPA. CoPA retains lower mean error at K = 16, but this is a single-task comparison, not evidence of universal superiority. We also retract our earlier unsupported statement that LoRA fine-tuning would “definitely perform better than CoPA”; this required measurement rather than speculation.

Both new LoRA controls use 200 optimizer updates and 6,400 supervised query presentations. They use the same task training split, but do not reproduce CoPA’s historical episodes or exactly match its training budget. This adds the requested type of supervised episodic control, while the simpler-aggregator ablation remains unresolved.

Regarding efficiency, our earlier crossover calculation remains conditional on reusable supports and repeated uncached ICL. It is not a measured advantage, does not apply when retrieval changes supports per query, and establishes no advantage over prefix-cached ICL.

Reviewer_Nrz7OpenReview ↗

Response to the authors

Thanks for the response. However, for most of my concerns, the authors did not provide meaningful results, which largely lower my confidence of this work. Taking all factors into consideration, I decide to lower the rating score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.