acceptodds
Under review as a conference paper at ICLR 2027

Close Enough is Nowhere Near: Fidelity-Consumability Decoupling in KV Caches

Abstract

Multi-agent LLM systems communicate in text, which is lossy and wasteful: the sender throws away the reading it already did. A growing literature instead transmits the KV cache, the natural carrier of that reading. We study the most appealing form of the idea: both models stay frozen and the sender’s cache is substituted for the reader’s with nothing trained. It does not work, and the obstruction belongs to the reader. A frozen transformer reads only caches close to the one it would have written itself, a region we call its tolerable set. We show that geometric similarity to that cache does not predict whether the reader can use it, and the failure is not gradual, so a more faithful substitute is not the remedy it appears to be. We explain this phenomenon with competitive attention. Experimentally, we test eight diverse substitution methods against two measured anchors: an uninformative-cache baseline and the reader’s own accuracy from having read the passage. All sit at the baseline. That turns a long tail of separate negative results into one outcome with a stated condition for overturning it. A different design does work: if the reader keeps its own cache and attends over the sender’s as extra context, adapting only its queries, it beats its own baseline, and also clears the baseline with eight senders pooled where undirected attention over the same union falls below it. We offer this as a characterization on one probe, not a leaderboard claim. Our central contribution is a redirection: pursuing faithful substitute caches is the wrong target; adding a cache to what the reader already has is the promising one.

Primary area:
foundation or frontier models, including LLMs
Keywords:
KV cache, key-value cache reuse, frozen language models, attention softmax, representational alignment, latent communication, multi-agent systems, negative results, model stitching

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_oELSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the feasibility of direct KV-cache communication between frozen LLMs and identifies a fundamental limitation of cache substitution: geometric similarity does not guarantee that a receiving model can effectively use the transferred information. Its central contribution is to distinguish cache fidelity from consumability and explain their divergence through the competitive nature of attention, where small perturbations can disrupt critical attention patterns. Through controlled experiments on a SQuAD-based factual recall task, the authors show that even highly similar caches can cause severe performance degradation, while eight diverse substitution methods fail to meaningfully exceed the baseline. In contrast, retaining the receiver’s own cache and incorporating the sender’s cache as additional context substantially improves accuracy, with adaptation further enabling effective transfer from multiple senders. These findings motivate a shift from cache reconstruction toward attention-based augmentation, although the evidence remains largely limited to a single model architecture and task setting.

Critical strengths
  1. Insightful problem formulation. The paper reveals a fundamental disconnect between KV-cache fidelity and consumability, supported by a compelling mechanistic explanation based on competitive attention.
  2. Systematic validation and actionable insights. Consistent failures across eight substitution methods strengthen the central claim, while the success of cache augmentation offers a promising alternative for latent communication.
Critical weaknesses
  1. Insufficient mechanistic evidence. The proposed explanation based on competitive attention is not directly validated through attention-margin or ranking analyses. Moreover, the key fidelity comparison is confounded by answer availability, and the on-manifold experiments do not preserve coherent key–value pairs. These limitations weaken the causal interpretation of the central findings.
  2. Limited evidence for the broader design conclusions. The negative results exclude previously successful fitted-bridge approaches, while the positive augmentation results are restricted to the same model checkpoint and a single factual recall task. Given that prior work has already identified attention-sensitive reconstruction errors and explored cache augmentation, the paper does not yet sufficiently establish the generality or distinctiveness of its proposed design principle.
Questions for authors
  1. Can the authors directly establish that changes in attention margins or key rankings explain the performance collapse, rather than merely ruling out error amplification? In particular, would the fidelity–consumability gap persist under matched, answer-bearing cache contents and controlled perturbations?
  2. Since the nearest-neighbor experiments replace individual keys while retaining the original values, how would the conclusions change when transferring coherent key–value pairs? What evidence establishes that failure arises from the receiver's tolerable set rather than inconsistencies introduced by the substitution protocol?
  3. How would a fitted bridge trained on independent paired-cache calibration data perform on the same split-context task? Since such methods have demonstrated successful cache substitution in prior work, this comparison seems essential for determining whether the observed limitation is intrinsic to substitution or specific to the tested constructions.
  4. Section 5.2 attributes the 0.695 accuracy to query-side LoRA, whereas Appendix F reports 0.600 for query-only adaptation and 0.695 when adapting queries, keys, and values. Which configuration produces the reported result, and does the claimed effectiveness of query-only adaptation still hold under the corresponding negative controls?
8 comments

---- Comment 1/6: Overview ----

This was a thoughtful review, and its questions identified the weakest links in our argument. We are glad the reviewer found the problem formulation insightful. In response we ran new experiments, corrected an error the reviewer identified, and sharpened the central claim.

Your questions about the mechanism and about why some methods work while others fail shaped the main change. The paper now argues one thesis from start to finish: success is decided by the objective a method optimizes, either closeness to the receiver's cache (fidelity) or the receiver's behavior (consumability), and the mechanism in Section 4 explains why only the second reliably works. We also rewrote the text for clarity and tone. New results and passages are marked in blue in the revised PDF. A summary of the full revision is in the general response posted at the top of the forum, and we answer each of your points below.

Summary of the main changes

  1. One thesis, stated in Section 2. The paper now sorts methods by their training target. Fidelity methods aim for a cache close to the receiver's own; consumability methods target what the receiver does with the cache. Under C1, substitution aimed at fidelity stays at the floor, even when it is fitted against the receiver. C2 states that methods trained on the receiver's behavior face no such limit, and C3 that the objective by itself determines which case holds. This is also why earlier positive results are consistent with ours (W2, Q3).
  2. A fitted bridge under two objectives (new Section 5.4). The same map and data give 0.078 when trained for reconstruction and 0.613 when trained through the receiver, averaged over three independent sets of training passages (floor 0.180, reference 0.665).
  3. Direct mechanism evidence (Section 4.1) and a confound-free fidelity comparison (Section 3.1, Appendix B).
  4. A correction to Section 5.2 (Q4).

---- Comment 2/6: Weakness 1 (a, b) ----

W1(a). The competitive-attention explanation is not validated by margin or ranking analyses.

  • Response: We agree that ruling out amplification is not the same as showing the mechanism, and the revision adds direct tests. To state the scope plainly: we measured rankings (whether the most-attended position changes) but did not measure attention margins.
  • New results: The strongest evidence is an intervention on the receiver's own full-passage cache (Section 4.1, "Direction, not distance"; Figure 2). Moving only its keys 3% of the way toward those of DeepSeek-R1-Distill-Qwen-1.5B, with the receiver's answer-bearing values untouched, lowers accuracy from 0.665 to 0.160, although the blended keys differ from the receiver's by a relative error of only 0.0002. The same blend applied to the values alone is harmless, and replacing all of the receiver's keys with those of Qwen2.5-1.5B-Instruct leaves accuracy at 0.685. Both changes are tiny by any distance measure; what separates them is the direction in which the keys move.

Second, at matched per-layer cosine on the answer-bearing cache, we aim noise at the directions the receiver's queries are most sensitive to, or at random directions. Because we choose the direction and observe the effect, this is also an intervention rather than only a correlation. Geometric distance alone would predict equal damage:

Per-layer cosineRandom directionsQuery-sensitive directions
0.9990.1050.040
0.9950.0600.010
0.9900.0250.005

Third, for rankings: at the same per-layer cosine, noise changes the most-attended position more often than the genuine cache does (0.30 versus 0.21 of contests; 0.15 versus 0.08 for sharply focused heads). Across 58 perturbations, this change rate predicts accuracy better than cosine (rank correlation 0.91 versus 0.80 to 0.87), and within one condition it modestly identifies failing questions (AUC 0.66 versus 0.48). Section 4.1 also states its limits: the relation works mostly through the first layer; in later layers genuine keys move attention less than noise yet are no more usable; and the change rate does not predict which fitted bridge works. These measurements are correlational and taken at the first decoding step.

The better behavioral test is the receiver's output divergence. Across 49 conditions, the KL divergence at the first answer token has a rank correlation of 0.955 with accuracy and explains 93% of held-out variance, against 0.81 and 45% for cosine.

  • Where in the paper: Section 4.1 ("Direct evidence"; "Direction, not distance"), Figure 2, Section 5.6, and Figure 5.

W1(b). The key fidelity comparison is confounded by answer availability.

  • Response: The reviewer is right.
  • New results: We add noise directly to the receiver's cache of the full passage, which contains the answer. Accuracy still falls from 0.665 to 0.105 at per-layer cosine 0.999, so the fragility does not depend on whether the answer is present. Noise on the answer-bearing cache, calibrated globally, gives 0.682, 0.467, and 0.003 at cosines 0.9999, 0.999, and 0.995; calibrated layer by layer, a per-layer cosine of 0.999 already gives 0.100. We also found that the original genuine-cache blend was position-misaligned after the masked span. Once filler tokens stand in for the answer one for one and the positions line up, the genuine full-passage cache (mean per-layer cosine 0.982) recovers 0.665. Noise at that cosine in every layer, applied to keys and values separately, gives 0.122 ± 0.013, below the 0.155 floor of that setup. This is the matched, answer-bearing comparison the reviewer asked for in Q1.
  • Where in the paper: Section 3.1 and Appendix B.

---- Comment 3/6: Weakness 1 (c) and Weakness 2 (a) ----

W1(c). The on-manifold experiments do not preserve coherent key–value pairs.

  • Response: This question led us to a real error. On auditing the code, we found that the key and value banks were subsampled independently, so each selected key was paired with the value of an unrelated token (Appendix E, formerly Appendix D, had incorrectly said that only keys were replaced). We reran the experiment with matched pairs.
  • New results:
Substitution (nearest genuine entries, Qwen2.5-1.5B)AccuracyKey / value cosine
Original run (keys with unrelated values)0.0050.82 / 0.10
Matched key–value pairs0.0050.83 / 0.39
Matched pairs, positional rotation made consistent0.0150.89 / 0.51
Keys only (receiver's own values kept)0.1700.83 / 1.00
Values only0.0051.00 / 0.39

Floor 0.180. The key and value cosines in the last column are reported with the corrected results in Appendix E. With matched pairs the conclusion is unchanged: genuine entries from other passages cannot stand in for the right ones, and the same holds on four further receivers. The rerun also shows that most of this effect comes from the values, which carry the content; genuine keys alone leave the receiver at the floor. We have revised Section 3.3 to say so, and we now treat this experiment as weaker evidence than we originally did.

  • Where in the paper: Section 3.3 and Appendix E (corrected).

W2(a). The negative results exclude previously successful fitted-bridge approaches.

  • Response: We agree, and the revision now tests a fitted bridge directly and explains why earlier ones succeeded (Appendix K). Closed-form ridge maps (Heo et al.) work within one model family, whose caches start much closer together than those of unrelated models. Our account predicts that a fidelity-trained map works only when the sender's cache already lies near the receiver's tolerable set. Among our senders, Qwen2.5-1.5B-Instruct is that case: its raw cache already scores 0.630, and every bridge leaves that accuracy in place. DeepSeek-R1-Distill-Qwen-1.5B and Qwen2.5-0.5B are not: with them the fidelity bridges stay below the floor (ridge 0.078 with the DeepSeek-distill sender), while the same bridge trained through the receiver reaches 0.613 and 0.318, respectively. So fitting alone does not decide the outcome; the objective does, given where the sender starts (Q3).
  • So that this condition is not circular, whether a cache lies in the tolerable set can be checked without accuracy labels, through the receiver's output divergence (Section 5.6), which separates working from failing caches where cosine cannot.
  • Where in the paper: Section 2 (C1–C3), Section 5.4, Table 3, Section 5.6, and Appendix K (Table 11).
Reviewer_8JdROpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates whether that KVC generated by one LLM can be consumed by another LLM without retraining. It distinguishes cache fidelity with consumability. The paper then tested controlled perturbations and different substitution methods, found that geometrically similar KVC can produce very different downstream accuracy.

Critical strengths
  1. This paper potentially provides important insights into KVC transferability.

  2. The paper proposes multiple substitution methods and tested all of them, which also contributes information to the communitiy.

Critical weaknesses
  1. the experiments use one small model, qwen2.5-1.5B, and only 200 evaluation examples. This coverage is too small to say the result is meaningful. Small models and large models can have very different response to KV geometry.

  2. Incomplete comparison with prior work, such as "Latent Space Communication via K-V Cache Alignment" and "Cache-to-Cache: Direct Semantic Communication Between Large Language Models".

Questions for authors

N/A

6 comments

---- Comment 1/5: Overview ----

Thank you for your time and for identifying the two issues that most limited the paper: coverage, and the comparison with prior work. We are glad the reviewer sees potential insight into KV-cache transferability. We have addressed both with new experiments.

The revision is substantial. Beyond the new experiments, we restructured the paper around a clearer central claim, namely that what decides whether a shared cache works is the objective used to produce it (closeness to the receiver's cache, or the receiver's behavior). The measurements showing that closeness and usability come apart (calibrated noise against a genuine cache, and nearest genuine keys) now cover five receivers rather than one, and we add four senders with different weights. The direct test of the objective (the fitted bridges) and the augmentation results still use the Qwen2.5-1.5B receiver only, as we state under W1. Text that is new to this version appears in blue in the revised PDF. Our top-level general response gives an overview of all changes; the comments below respond to your two weaknesses in detail.

Summary of the main changes

  1. More models (W1). Larger receivers (Qwen2.5-3B, Qwen2.5-7B), a receiver from another family (Mistral-7B-v0.3), and an instruction-tuned receiver (Qwen2.5-1.5B-Instruct). Four senders whose weights differ from the receiver's (Qwen2.5-1.5B-Instruct, DeepSeek-R1-Distill-Qwen-1.5B, Qwen2.5-Coder-1.5B, Qwen2.5-Math-1.5B).
  2. A controlled comparison with the approach of prior work (W2). We train one fitted bridge under two objectives on our task. This helps explain why Cache-to-Cache and latent KV alignment succeed where reconstruction-based substitution does not.
  3. One organizing question (Section 2). The paper now asks what each method is trained to optimize: a cache close to the receiver's own (fidelity), or the receiver's behavior when it reads that cache (consumability).

---- Comment 2/5: Weakness 1 (receivers) ----

W1. One small model (Qwen2.5-1.5B) and 200 evaluation examples are too limited; small and large models may respond differently to KV geometry.

Response. We agree that a single 1.5B receiver was too narrow, and the point about scale is well taken. We address model coverage and sample size separately.

Model coverage: new results.

  • Larger and different receivers. We repeat the core measurements (calibrated noise, the single-layer corruption sweep, and selection from a bank of genuine keys) on Qwen2.5-3B, Qwen2.5-7B, Mistral-7B-v0.3, and Qwen2.5-1.5B-Instruct, each with its own floor and reference measured in the same run:
ReceiverFloor / referenceNoise on own full cache, cos 0.995Genuine cache (cosine)Nearest genuine key–value pairs
Qwen2.5-1.5B-Instruct0.095 / 0.5900.0000.240 (0.995)0.000
Qwen2.5-3B0.305 / 0.7950.720 (0.005 at cos 0.95)0.515 (0.925)0.000
Qwen2.5-7B0.365 / 0.7700.1200.600 (0.989)0.010
Mistral-7B-v0.30.515 / 0.6450.0500.490 (0.846)0.000

Substituting the nearest genuine key–value pairs from other passages gives accuracy near zero on every receiver. The cosine inversion also holds on each receiver, though not equally cleanly. On Qwen2.5-7B and Mistral-7B, noise at cosine 0.995 falls well below a genuine cache at lower cosine. Qwen2.5-1.5B-Instruct shows the same ordering (0.000 against 0.240), but its genuine cache is itself weak (floor 0.095). As the reviewer anticipated, the threshold depends on the model: Qwen2.5-3B tolerates noise down to cosine 0.99 (0.720 at 0.995, above its genuine cache's 0.515) and fails only at 0.95 (0.005), so for that model the inversion appears only against stronger noise. The genuine-cache column also uses the position-by-position blend that Appendix B flags as misaligned; we have the aligned control only for Qwen2.5-1.5B. Mistral-7B leaves only a small gap between floor and reference, so its row is weaker evidence. One result from the original paper does not replicate. When noise is matched layer by layer, corrupting a single layer is not as damaging as we reported (for Qwen2.5-7B, 0.710 at one layer against 0.055 with noise matched only on average), and we have revised Section 3.2 accordingly. Bridges and augmentation were run only with the Qwen2.5-1.5B receiver, and we did not test any receiver above 7B.

The next comment continues W1 with senders and sample size.

---- Comment 3/5: Weakness 1 (senders and sample size) ----

Model coverage, continued.

  • One model reading another model's cache, which the original paper did not test. When Qwen2.5-1.5B reads the cache of DeepSeek-R1-Distill-Qwen-1.5B in place of its own, accuracy is 0.000 to 0.005 across runs (Tables 2 and 3), even though the keys of the two caches have cosine 0.983. Results for the four senders, which share the receiver's architecture but not its weights:
Sender (receiver Qwen2.5-1.5B)Key / value cosineRaw cache substitutedAugmentation, no trainingAugmentation, adapted
Qwen2.5-1.5B-Instruct0.9999 / 0.9880.6300.5900.810
Qwen2.5-Coder-1.5B0.9949 / 0.7200.0000.0200.155
Qwen2.5-Math-1.5B0.9818 / 0.5000.0000.0250.185
DeepSeek-R1-Distill-Qwen-1.5B0.9813 / 0.4670.0050.0100.135

This widens coverage on the sender side as well. The closest sender, the lightly instruction-tuned Qwen2.5-1.5B-Instruct, produces a cache within the receiver's tolerable set, and the receiver can read it as is. The other three continued training for longer and land outside that set. For them, raw substitution gives near-zero accuracy, augmentation without training falls below the floor, and adapting the receiver lifts it only to around floor level. A bridge trained on the receiver's behavior does succeed, at 0.613 with the DeepSeek-distill sender. Sections 2 and 5.3 therefore claim augmentation only for senders whose weights are the same as, or near, the receiver's.

Sample size: new results.

  • The comparisons that carry the argument are large relative to sampling error. Because questions share passages, we use passage-clustered bootstrap intervals, which are wider than a binomial standard error: the reference's interval is [0.59, 0.73] and the floor's [0.11, 0.26]. The key contrasts lie far outside them: 0.078 versus 0.613 for the same bridge under two objectives, stable over three independent sets of training passages, and 0.000 versus the reference of 0.665.
  • Because the 200 questions come from only ten passages, we now report confidence intervals from a bootstrap that resamples whole passages, which is more conservative than treating questions as independent. For the bridge comparison, the passage-clustered intervals for the gap between the behavior-trained and reconstruction-trained bridges are [0.41, 0.62] and [0.46, 0.67] (the two training sets with per-question logs). We reran the bridge comparison on the first 500 validation questions (30 passages, rather than 10). The result holds: floor 0.166, reference 0.626; the raw DeepSeek-R1-Distill cache gives 0.002, the ridge bridge 0.066 [0.04, 0.09], and the KL bridge 0.542 [0.44, 0.63], a gap of 0.48 [0.38, 0.56] with passage-clustered 95% intervals (Appendix G; summarized in Section 6). Only the bridge comparison was rerun at this size; the noise and cosine experiments, including those on the four new receivers, remain at n = 200 over ten passages.
  • A second task. On HotpotQA, where the answer requires two paragraphs (n = 200, floor 0.025, reference 0.435), noise that keeps the cache highly similar to the receiver's own drives accuracy to the floor or below, and the same linear bridge gives 0.005 when trained for reconstruction against 0.423 when trained through the receiver (Appendix M, Table 13). This uses one sender and one receiver.

Where in the paper: Sections 3.1 to 3.3, Sections 5.3 and 5.4, Section 6 (scope), Appendix G (n = 500), Appendix M (HotpotQA), and Appendix N (Table 14, with the layer-matched result for Qwen2.5-7B given in the text).

Reviewer_qqN4OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper asks whether a frozen transformer can use a KV cache produced by another model in place of its own. Using one reader (Qwen2.5-1.5B) and one probe (SQuAD short-answer recall with the answer masked in the reader’s copy of the passage, n=200), it reports that (i) isotropic noise on the reader’s own cache at cosine 0.995 drops accuracy below the no-information floor, while the same model’s cache on the unmasked passage transfers at 0.445 from a lower cosine (Table 2); (ii) corrupting a single layer of 28 with noise collapses accuracy, whereas replacing early-layer keys with nearest keys from a bank of other passages is tolerated for several layers (Table 3); (iii) substituting the nearest genuine key from a 6,000-key bank transfers at 0.000 (Table 4); (iv) eight constructed substitution methods all score at or below the floor of 0.180, while letting the reader keep its own cache and cross-attend over the sender’s cache scores 0.555 with no training and 0.695 with query-side LoRA, against a native reference of 0.665 (Table 1); and (v) the adapted cross-attention arm remains above the floor with eight pooled senders (Table 5). The proposed mechanism is that the softmax readout is competitive, so small displacements that flip attention winners matter more than geometric distance.

Critical strengths
  • The paper states its scope, anchors, and falsifiers explicitly (Sec. 6), re-measures the floor and reference inside every run, and reports negative controls and per-run spreads in the appendices. This is more careful than most negative-result papers.

  • The depth sweep (Table 3) is a clean design: holding 27 of 28 layers bit-exact and showing a step rather than a slope rules out a compounding-error account, and the key-versus-value sparsity experiment gives a prediction of the mechanism that the data support.

Critical weaknesses
  1. The sender and reader are the same model: every augmentation experiment uses a second copy of Qwen2.5-1.5B as the sender (App. A), and the only cross-model arm fails on a dimension mismatch (App. F). So the paper never tests one model reading another model’s cache, which is what the title and introduction are about. The positive result mostly shows that a model can read its own cache of the full passage

  2. The on-manifold vs. off-manifold comparison is not fair: the two arms in Table 3 are matched on average cosine across all layers. But bank keys taken from other passages are almost identical to the true keys in early layers and very different in later layers, while noise is spread evenly. Replacing early layers with bank keys is therefore close to doing nothing, which also explains why accuracy goes above the native reference at ℓ=2–3. Per-layer cosines are not reported.

  3. The substitution sweep is missing basi details. Table 1 does not say what the eight methods take as input, how they were trained, or on how much data. Two of them are regressions onto the reader’s cache, which is the “fitted bridge” case the paper says it does not run. One arm varies from 0.085 to 0.010 between runs (App. F), so the uniform failure could be unstable training rather than a property of the reader.

Questions for authors
  • For each substitution row in Table 1: what is the sender, what are the training inputs and targets, how many items, and what capacity? Is the sender the same checkpoint on the unmasked passage? If so, why does a regression toward the reader’s cache not recover the identity map’s 0.445?

  • Are the LoRA training items disjoint from the 200 evaluation items? Please report held-out accuracy.

  • Report per-layer achieved cosine for the on-manifold arm in Table 3. Does the step-vs-slope contrast survive if noise is scaled per layer to match the bank-key cosine at that layer?

  • Qwen2.5 keys have large-magnitude outlier channels and attention-sink behavior in early layers; App. H finds layers 0–1 fatal and layer 27 tolerable. How much of the “tolerable set is narrow” result is specific to those channels? Does the cosine-calibrated noise result change if noise is excluded from the top-k magnitude channels, or on a model without this property?

  • Can any of the cited fitted bridges (Heo et al., Dery et al.) be run on this probe? The paper’s position depends on their boundary and currently rests on citation only.

6 comments

---- Comment 1/5: Overview ----

We thank the reviewer for a careful and generous reading, and for questions concrete enough to answer with new experiments. We appreciate the recognition of the explicit scope, reference points, and falsifiers, and of the depth and key-versus-value experiments. Your observation that two rows of the sweep are regressions onto the receiver's cache led to the main change in the revision.

Building on that observation about the sweep, we reorganized the whole paper around a single question: what does a method optimize? A cache trained to be close to the receiver's own (fidelity) and a cache trained on the receiver's behavior (consumability) lead to sharply different outcomes, and every section now builds toward that contrast, written in more measured language than before. The blue text in the revised PDF marks what is new, and all results below appear there. For the revision as a whole, please see our general response at the top of the forum; here we take your points one by one.

Summary of the main changes

  1. A sharper central claim (Section 2). We no longer draw the boundary at whether a map is "fitted." Methods are now organized by what they optimize: closeness to the receiver's cache (fidelity), or the receiver's own behavior (consumability). C1: substitution that optimizes fidelity stays at the floor, fitted or not. C2: methods that optimize the receiver's behavior are not limited this way. C3: the objective alone decides which applies.
  2. A direct test of C3 (new Section 5.4). The same linear bridge, trained on the same data, scores 0.078 when fitted for reconstruction (key cosine 0.9988) and 0.613 when trained on the receiver's outputs (key cosine 0.9935), averaged over three independent sets of training passages. Floor 0.180, reference 0.665.
  3. Senders with different weights (new Section 5.3) and larger and different receivers (Section 3, Section 6, Appendix N).
  4. Corrections prompted by your questions and by an audit of our code (W2, W3, Q4 below).

---- Comment 2/5: Weaknesses ----

W1. The sender and receiver are the same model.

  • Response: We agree that this was the most important gap. We now test four senders that share the receiver's architecture and tokenizer but have different weights, so no dimension matching is needed: Qwen2.5-1.5B-Instruct, DeepSeek-R1-Distill-Qwen-1.5B, Qwen2.5-Coder-1.5B, and Qwen2.5-Math-1.5B.
  • New results: Substituting DeepSeek-R1-Distill-Qwen-1.5B's cache into Qwen2.5-1.5B gives 0.000 to 0.005 across runs, although its keys have cosine 0.983 to the receiver's own. Fitted reconstruction bridges from this sender reach key cosine 0.999 and stay below the floor, while the same bridge trained through the receiver reaches 0.613 (three training sets). Results for all four senders:
Sender (receiver Qwen2.5-1.5B)Key / value cosineRaw cache substitutedAugmentation, no trainingAugmentation, adapted
Qwen2.5-1.5B-Instruct0.9999 / 0.9880.6300.5900.810
Qwen2.5-Coder-1.5B0.9949 / 0.7200.0000.0200.155
Qwen2.5-Math-1.5B0.9818 / 0.5000.0000.0250.185
DeepSeek-R1-Distill-Qwen-1.5B0.9813 / 0.4670.0050.0100.135

The light instruction-tuned model's cache lies inside the receiver's tolerable set and works directly. The three models that continued training further fall outside it: their raw caches give accuracy near zero, augmentation with them falls below the floor without training, and adapting the receiver brings it back only to about the floor. For these senders, the bridge trained on the receiver's behavior is what succeeds (0.613 for the DeepSeek-distill sender). The sender rows are single runs. The bridge contrast also holds on a second task, HotpotQA, where the answer depends on two paragraphs (0.005 for the reconstruction bridge against 0.423 for the one trained through the receiver; Appendix M). We have scoped our augmentation claim accordingly: it holds when the sender's weights are the same as or close to the receiver's (Sections 2 and 5.3).

  • Where in the paper: Sections 5.3 and 5.4; Tables 1 to 3.

W2. The on- and off-manifold arms are matched only on average cosine.

  • Response: The reviewer is correct. Matching was on one mean key cosine (0.822). Our audit also found a second asymmetry: the genuine-key arm replaces keys and values, while the noise arm perturbs keys only. As the reviewer suspects, genuine keys are much closer to the originals in early layers, which explains why those layers do little harm. The values above the reference at ℓ = 2–3 (0.740 and 0.745) lie just outside the reference's passage-bootstrap interval [0.59, 0.73] and are most likely sampling variation over ten passages.
  • New results: On Qwen2.5-1.5B, genuine keys have cosine above 0.99 to the true keys in the first layers and about 0.75 to 0.85 in later layers. Once noise on the keys is matched to that cosine layer by layer, corrupting one to eight layers has little effect on this receiver or on four further receivers (Qwen2.5-3B, Qwen2.5-7B, Mistral-7B, Qwen2.5-1.5B-Instruct; for example, 0.710 at ℓ = 1 for Qwen2.5-7B, against 0.055 with average matching). The single-layer "step" in our original paper therefore came from average matching, and we have revised Section 3.2 and removed the claim from the abstract and introduction. On Qwen2.5-1.5B, with keys alone, genuine keys and noise matched layer by layer decline together (0.700 versus 0.682 at ℓ = 1, 0.170 versus 0.188 with all layers replaced), while average matching gives 0.040 at ℓ = 1. Noise matched layer by layer on keys and values together does more damage, but it declines gradually with depth (0.698, 0.343, and 0.027 at ℓ = 1, 6, and 12) rather than in a single-layer step. Corrupting one layer at a time locates the sensitivity: noise at cosine 0.999 in layer 0 alone gives 0.07, while every other layer except layer 15 tolerates cosine 0.822 (accuracy 0.60 to 0.73). While rerunning, we also found that the original genuine-key condition paired keys with unrelated values; we disclose this and report a corrected version (Appendix E). The conclusion you drew from the depth sweep, that errors do not compound across layers, is still supported by a direct measurement of how a perturbation propagates (Appendix D), and the key-versus-value sparsity experiment is retained (Appendix I).
  • Where in the paper: Section 3.2 now states the caveat and no longer claims that "the distinguishing property is neither geometric distance nor depth." Appendix C reports the new results.

W3. Missing details for the substitution sweep; two rows are fitted regressions; one row is unstable.

  • Response: This point led to the reframing in Comment 1. Regressions onto the receiver's cache are fitted, but they still optimize fidelity, and that is now what C1 is about. Answers on details and instability are under Q1.
  • Where in the paper: Section 2 (C1–C3), Section 5.1, Table 1, Appendix G.

---- Comment 3/5: Questions 1 to 3 ----

Q1. For each substitution row: sender, inputs, targets, data, capacity; and why does regression not recover the identity map's 0.445?

  • Why regression does not reach 0.445. The 0.445 "identity" figure comes from a position-by-position blend of the receiver's masked prompt with the full prompt. Because the masked span and the answer usually differ in token length, positions after the mask hold different tokens. When the prompts are aligned, the identity cache is the receiver's own full-passage cache and scores the reference, 0.665. We disclose this misalignment and add an aligned version. In the aligned version, each answer token is swapped for a filler token, so both prompts hold the same token at every position. There the genuine full-passage cache sits at mean per-layer cosine 0.982 and scores 0.665. Noise calibrated to that cosine layer by layer, for keys and values separately, scores 0.122 ± 0.013, under the floor of 0.155 for this setup.
  • Per-row details. For the original sweep, the sender is the same Qwen2.5-1.5B checkpoint reading the unmasked passage. The rows that remain in Table 1 are:
RowInputTarget or selection ruleTraining dataCapacity
Regressionsender's full-passage cachereceiver's cache; loss on key logits only (values untrained)200 passages, about 800 stepslearned map
By constructionsamesame, output constrained to the column space of W_Ksamesame
Codebook selectionreceiver's masked cachenearest entry in one of three codebooks (genuine-key bank, 2048-entry VQ, rank-16 PCA), by raw, logit-space, or soft top-8 matchingnone6,000 keys per layer
Oracle nearest pairsamenearest genuine key-value pair from the banknone6,000 pairs per layer
Fitted bridges (new)DeepSeek-R1-Distill or Qwen2.5-0.5B cacheridge or MSE onto the receiver's cache, or KL through the receiver3,000 to 3,300 SQuAD training passageslinear, per layer and KV head, about 3.7M parameters

The regression row is a weak instrument: its function class contains the identity, so its failure reflects optimization (values were never trained) rather than the limit C1 describes, and we no longer rest the argument on it. The closed-form ridge bridges have no training instability and still stay below the floor: 0.078 and 0.070 (ridge, MSE) with the DeepSeek-distill sender, 0.040 and 0.055 with Qwen2.5-0.5B.

  • Removed rows. Four rows of the original table (manifold loss, row-space target, worst-case hinge, high-capacity MLP) were produced on a server that is currently unavailable, so we could not re-verify their code and results. We have removed them rather than report numbers we cannot reproduce. The fitted bridges now carry this part of the argument, which also addresses the reviewer's concern about the unstable row. If we can re-verify the original runs, we will restore them with full details.
  • Where in the paper: Section 3.1 and Appendix B (misalignment); Section 5.1, Table 1, and Appendix G (per-method details and removed rows); Section 5.4 (bridges).

Q2. Are the LoRA training items disjoint from the 200 evaluation items?

  • Response: Yes. The adapters are trained on 200 questions from the SQuAD training split (different Wikipedia articles), and the checkpoint is selected on a held-out slice of that split. Evaluation uses 200 validation questions, so every augmentation number is held-out accuracy.
  • New results: We verified disjointness by hashing: no question, passage, or Wikipedia article is shared between training and test. A three-seed rerun of the adaptation ladder gives 0.613 ± 0.018 (query only), 0.665 ± 0.083 (keys and values), and 0.693 ± 0.093 (all three), with every control at or below the floor. The rerun also exposed a bug in our original training code, now fixed and disclosed in Appendix G: during training, the current token could not attend to its own key and value.
  • Where in the paper: Section 5.2 and Appendix G.

Q3. Per-layer achieved cosine for the on-manifold arm; does step versus slope survive per-layer matching?

  • Response and new results: Please see W2. Per-layer cosines of the genuine keys are 1.000 (layer 0), 0.997 (layer 1), 0.920 (layer 2), and 0.74 to 0.86 in most later layers. With noise matched to these values, the step disappears (Appendix C).
  • Where in the paper: Section 3.2 and Appendix C.
Reviewer_LQGkOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper investigates whether a frozen LLM can effectively consume KV caches produced by another model, focusing on direct cache substitution versus augmentation. Using Qwen2.5-1.5B and a SQuAD-based factual recall task, the authors show that high geometric similarity between KV caches does not necessarily translate into successful task performance. Several substitution methods fail to improve upon the baseline, which the authors attribute to the sensitivity of competitive softmax attention. They introduce the concept of a reader-specific "tolerable set" to characterize usable caches. As an alternative, they propose retaining the reader's original cache while attending to the sender's cache as additional context, demonstrating improved accuracy with optional LoRA adaptation.

Critical strengths

The paper provides empirical evidence that geometric similarity between KV caches does not reliably predict downstream task performance. The comparison of multiple substitution methods and the inclusion of baseline controls offer useful observations about the limitations of KV-cache sharing, although their generality remains limited.

Critical weaknesses
  1. The paper does not convincingly establish why KV-cache communication is preferable to sharing task-relevant text. The proposed augmentation approach resembles existing memory-attention mechanisms, and its practical advantages remain unsubstantiated.

  2. The interpolation experiments mix caches from different input contexts, potentially introducing token misalignment and inconsistent internal states. Moreover, the comparison between random perturbations and genuine sender caches confounds representation similarity with differences in information content, weakening the evidence for the proposed fidelity–consumability mechanism.

  3. The proposed softmax-competition explanation lacks direct experimental validation. The tolerable set is primarily defined through observed task performance rather than independently characterized, limiting its explanatory value.

  4. The main experiments rely on a single model and a narrow factual-recall task. The proposed augmentation increases the effective attention context, but computational cost, memory usage, and communication overhead are not evaluated against practical alternatives.

  5. The central concepts and experimental procedures are difficult to follow, and the distinction between empirical observations, mechanistic explanations, and broader conclusions is not sufficiently clear.

Questions for authors
  1. What concrete multi-agent scenario requires KV-cache communication rather than directly sharing answers, summaries, or relevant text? What practical advantage does the proposed approach offer over these alternatives?

  2. Since the sender and reader use the same checkpoint, what motivates the partial cache substitution and interpolation setup? How do the authors disentangle the effects of token alignment, inconsistent cache states, and differences in information content from the proposed fidelity–consumability mechanism?

  3. Can the authors provide direct evidence that competitive softmax attention is responsible for the observed failures? How is the proposed tolerable set characterized beyond its behavioral definition?

  4. The proposed augmentation increases the reader's effective attention context. Can the authors demonstrate an end-to-end advantage over text-based communication when accounting for computational cost, memory usage, and communication overhead?

7 comments

---- Comment 1/6: Overview ----

We are grateful for this detailed review. The concerns about confounds, the mechanism, and clarity were well founded, and addressing them has improved the paper. We summarize the revision first and then respond to every weakness and question in turn.

To address the concerns about structure and clarity, we rewrote the paper from the ground up. It now has one thesis, that the objective a method optimizes (fidelity to the receiver's cache, or the receiver's behavior) decides whether a shared cache is usable, and its sections separate observations (Section 3), mechanism (Section 4), and conclusions (Section 5), in a more measured tone throughout. All new material is highlighted in blue in the revised PDF. We point you also to the general response at the top of the forum, which summarizes the revision for all reviewers; below we respond to each weakness and question in turn.

Summary of the main changes

  1. Clearer structure, plainer language, and a more measured tone (W5). Observations (Section 3), the proposed mechanism (Section 4), and the conclusions drawn from them (Section 5) are now kept separate, and each section opens by stating what it sets out to establish and why. Section 2 now describes the concrete setting before defining any terms.
  2. A sharper central claim with a direct test (Sections 2 and 5.4). Each method is classed by its objective: fidelity (a cache close to the receiver's own) or consumability (the receiver's behavior). One bridge trained under both objectives scores 0.078 when trained for reconstruction and 0.613 when trained on the receiver's outputs, consistently across three independent sets of training passages (floor 0.180, reference 0.665). The successful bridge is the less similar of the two.
  3. Confound controls (W2, Q2) and direct mechanism measurements (W3, Q3).
  4. More models and different senders (W4).

---- Comment 2/6: Weaknesses 1 and 2 ----

W1. Why KV-cache communication rather than text? Augmentation resembles memory attention; practical advantages are unsubstantiated.

  • Response: This is a fair concern; the original paper motivated it too briefly. Our paper characterizes when cache sharing can work; it does not argue that caches should replace text everywhere. The motivation comes from work reporting concrete gains. Cache-to-Cache reports 3 to 5 points higher accuracy than text-based communication at 2.5 times lower latency, and concurrent work (SwaT) reports that a trained cache translator lets the receiver skip reading the context while keeping 92 to 108% of its own accuracy. We agree that augmentation resembles memory attention and do not claim it as a new method. In our paper it serves as a test of the account (see W3).
  • New results: A cost comparison on our task (see Q4 below; Appendix O, Table 15).
  • Where in the paper: Appendix K (Table 11) for the figures reported by Cache-to-Cache and SwaT; Section 1 ("What a method optimizes") for how these methods fit our two questions.

W2. Interpolation mixes contexts (token misalignment, inconsistent states); random-versus-genuine comparison confounds similarity with information content.

  • Response: The reviewer is right on both points.
  • New results:
    • Misalignment. Our audit confirms that the genuine-cache interpolation blends positions one by one, so after the masked span the two prompts hold different tokens. We disclose this and add a position-aligned version. With the answer replaced token by token by a filler, so that every position aligns, the genuine full-passage cache (mean per-layer cosine 0.982) reaches 0.665, while noise matched to the same cosine in every layer, separately for keys and values, reaches 0.122 ± 0.013, below this setup's floor of 0.155.
    • Information content. To separate similarity from content, we perturb a cache that already holds the answer, namely the receiver's own cache of the full passage. Even so, noise on its keys alone at per-layer cosine 0.999 lowers accuracy from 0.665 to 0.105, so missing information does not explain the loss. A single global calibration at cosines 0.9999, 0.999, and 0.995 gives 0.682, 0.467, and 0.003, and matching cosine layer by layer gives 0.100 already at 0.999.
    • Inconsistent states. A cache produced by one forward pass over one context is internally consistent by construction, and both ends of our comparison include such a cache. At full weight (α = 1), the aligned blend above is the receiver's own cache of the full passage, a single real forward pass, and it reaches the reference of 0.665. Adding noise to that same single-context cache, with no second context involved, lowers accuracy to 0.100 at per-layer cosine 0.999. The failure therefore does not require mixing contexts.
  • Where in the paper: Section 3.1 and Appendix B (aligned blend, answer-bearing noise, and the controls for the genuine-cache comparison).

---- Comment 3/6: Weakness 3 ----

W3. The softmax-competition explanation lacks direct validation; the tolerable set is defined only through task performance.

  • Response: We agree that ruling out alternatives is not the same as direct evidence. The revision adds measurements of the competition itself, and we answer the two parts of this weakness separately.
  • Direct validation (new results, Section 4.1, "Direct evidence" and "Direction, not distance").
    • Aimed noise. Holding per-layer cosine fixed on the answer-bearing cache, noise placed along the directions the receiver's queries respond to most is far more harmful than noise in random directions: 0.040 versus 0.105 at cosine 0.999, 0.010 versus 0.060 at 0.995, and 0.005 versus 0.025 at 0.99. Distance alone would treat the two alike.
    • A real neighboring model. Instead of synthetic noise, we move the receiver's keys a small fraction of the way toward the keys of DeepSeek-R1-Distill-Qwen-1.5B and leave the receiver's answer-bearing values in place. A 3% move lowers accuracy from 0.665 to 0.160, although the keys change by a relative error of only 0.0002. Moving the values by the same fraction does no harm, and replacing the keys entirely with those of Qwen2.5-1.5B-Instruct leaves accuracy at 0.685 (Figure 2). Both changes are tiny in distance; they differ in whether they shift which tokens win attention.
    • Most-attended position. At the same per-layer cosine, noise changes the position that receives the most attention from the question more often than the genuine cache does (0.30 versus 0.21 of contests). Over 58 perturbations, this change rate follows accuracy more closely than cosine (rank correlation 0.91 against 0.80 to 0.87).
    • Caveats stated in the paper. The change-rate relation comes mostly from the first layer, and the change rate does not predict which fitted bridge works. These measurements are correlational and taken at the first decoding step, so we present them as support for the account, not as proof.
  • Characterizing the tolerable set (W3b; see also Q3). The reviewer is right that the tolerable set is defined behaviorally. This is by design: it is the set of caches the receiver can use, and we do not construct it explicitly. What we can say about its shape comes from the aimed-noise result above. At matched per-layer cosine, the set is narrow along the eight most query-sensitive key directions per layer and wider along random ones, which is the uneven-tolerance account of Section 4.1. That account also predicts the bridge result of Section 5.4, since a reconstruction loss weights all directions equally and a loss through the receiver does not.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Revision Summary

---- Comment 1/3: Summary of the revision ----

We thank all four reviewers for their careful and constructive reviews. Several concerns were raised by more than one reviewer, so we summarize here how the revision addresses them. Detailed point-by-point responses appear under each review. New material is shown in blue in the revised PDF, and the text has been rewritten throughout for clarity and tone.

1. A sharper central claim (qqN4, oELS, 8JdR). We now organize methods by what they optimize: closeness to the receiver's own cache (fidelity), or the receiver's behavior (consumability). Substitution methods that optimize fidelity stay at the floor whether or not they are fitted against the receiver, while methods that optimize the receiver's behavior can succeed. This account also explains why published methods such as Cache-to-Cache and latent KV alignment succeed. Table 11 (new, Appendix K) places prior methods on these two axes. (Sections 1 and 2; Appendix K, Table 11.)

2. A direct test of the claim (oELS Q3, qqN4 Q5, 8JdR W2). We train one linear bridge, on the same paired caches, under two objectives:

Bridge (DeepSeek-R1-Distill-Qwen-1.5B → Qwen2.5-1.5B)AccuracyKey cosine
Reconstruct the receiver's cache (ridge)0.0780.9988
Reconstruct the receiver's cache (MSE, SGD)0.0700.999
Match the receiver's outputs (KL)0.6130.9935

Floor 0.180, reference 0.665; averages over three independent sets of training passages. The successful bridge is the less similar of the two. (Section 5.4; Table 3.)

3. More than one model (all reviewers). We add four senders whose weights differ from the receiver's (Qwen2.5-1.5B-Instruct, DeepSeek-R1-Distill-Qwen-1.5B, Qwen2.5-Coder-1.5B, Qwen2.5-Math-1.5B) and four receivers (Qwen2.5-3B, Qwen2.5-7B, Mistral-7B-v0.3, Qwen2.5-1.5B-Instruct). A light fine-tune of the receiver (Instruct) can share its cache directly, while models that trained further (Coder, Math, DeepSeek-distill) cannot: their raw caches give accuracy near zero, and augmentation with them falls below the floor, so we now scope the augmentation claim to senders close to the receiver. For those distant senders, the bridge trained on the receiver's behavior succeeds. On the four new receivers, the nearest genuine key–value pairs give accuracy near zero, and noise at high cosine sharply reduces accuracy, although the cosine at which a receiver breaks varies by model (Qwen2.5-3B still scores 0.585, against its reference of 0.795, with noise at cosine 0.99); the bridges and augmentation were run with the Qwen2.5-1.5B receiver only (Appendix N, Section 6). We also repeat the headline experiments on a second task, HotpotQA, where the answer depends on two paragraphs. On HotpotQA (n = 200, floor 0.025, reference 0.435), both headline results replicate: noise at high cosine drives accuracy to the floor or below, and the same linear bridge scores 0.005 when trained for reconstruction and 0.423 when trained through the receiver (three seeds: 0.430, 0.405, 0.435). A bridge trained on SQuAD also transfers to HotpotQA without retraining (0.385, against 0.005–0.015 for the SQuAD-trained ridge bridges). (Sections 3, 5.3, and 6; Appendices M and N.)

AuthorsOpenReview ↗

Summary: Mechanism and Confounds

---- Comment 2/3: Mechanism and confounds ----

4. Direct mechanism evidence (oELS, LQGk). At matched per-layer cosine, noise aimed at the directions the receiver's queries are most sensitive to does much more damage than random noise (0.040 versus 0.105 at cosine 0.999). At the same per-layer cosine, noise changes the most-attended position more often than the genuine cache (0.30 versus 0.21), and this rate predicts accuracy better than cosine across 58 perturbations (rank correlation 0.91 versus 0.80 to 0.87). Across 49 conditions (noise, genuine caches, selected keys, quantization, other models' raw caches, bridges, augmentation), the divergence of the receiver's next-token distribution from the one it produces with its own cache predicts accuracy far better than cosine: rank correlation 0.955 against 0.81, and 93% against 45% of held-out variance (new Figure 5, Section 5.6). (Sections 4.1 and 5.6.)

5. Confounds (LQGk, oELS, qqN4). Noise on the receiver's own answer-bearing cache still lowers accuracy from 0.665 to 0.100 at per-layer cosine 0.999, so the fragility does not depend on whether the answer is present. With every position aligned and every layer matched, the genuine cache reaches 0.665 while matched noise reaches 0.122. We also disclose and correct for position misalignment, per-layer cosine differences, mismatched key–value pairs in our original bank experiment, outlier channels, and the attention sink. (Sections 3.1–3.3; Appendices B, C, E, and L.)

AuthorsOpenReview ↗

Summary: Corrections, Clarity, and Further Addition

---- Comment 3/3: Corrections, clarity, and further additions ----

6. Corrections. An audit of our code, prompted by the reviews, led to the following changes:

  • The 0.695 augmentation result adapts the query, key, and value projections, not the query projection alone (oELS Q4). Over three seeds the values are 0.613 ± 0.018 (query only) and 0.693 ± 0.093 (all three). We also found and fixed an error in our original training code, disclosed in Appendix G.
  • The 200 test questions come from only ten passages; we now report passage-clustered confidence intervals.
  • Four rows of the original sweep table could not be re-verified and have been removed. The fitted bridges now carry that part of the argument (qqN4 Q1).
  • Global cosine is dominated by the first layer, which we now state, and we report per-layer measurements (qqN4 Q4).
  • In our original bank experiment, selected keys were paired with the values of unrelated tokens. With matched pairs the result is unchanged (0.005), but most of the effect comes from the values, so we now present that experiment as weaker evidence (Section 3.3, Appendix E). The bank comes from 28 distinct passages, and several appendix descriptions have been corrected.
  • Matching noise layer by layer (qqN4 W2) changed one of our original conclusions: a single corrupted layer is not as damaging as we reported. We have revised Section 3.2 and the abstract accordingly.

7. Clarity and tone (LQGk, oELS, qqN4). We rewrote the paper so that observations, the proposed mechanism, and conclusions are kept separate. We replaced coined terms with plain descriptions and stated categorical claims as findings with their conditions. The closing paragraph of Section 1 ("Our Contribution: Fidelity is not consumability") explains the distinction between fidelity and consumability.

8. Further additions. We add a quantization analysis (Appendix L), a cost comparison with text communication (LQGk Q4), and confidence intervals from a bootstrap over passages. We also reran the headline comparison at n = 500 over 30 passages (8JdR; Appendix G, summarized in Section 6): the reconstruction bridge scores 0.066 and the behavior-trained bridge 0.542 (floor 0.166, reference 0.626; gap 0.48, 95% interval [0.38, 0.56]). The cost comparison shows the trained bridge is much faster than a written summary, while passing the full text remains cheapest in our small setting; we make no efficiency claim (Appendix O).

We thank the reviewers again. Their questions led to the most informative results in the revision, and we would welcome any further discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.