acceptodds
Under review as a conference paper at ICLR 2027

Written as a Record, Read as an Address: What a Forward Pass Leaves in an Operation's KV Cache

Abstract

When a language model reads an operation such as "Swap the contents of Box F and Box B”, its forward pass writes keys and values for those tokens into the KV cache. Prior work on entity tracking establishes what models *use*: bindings are resolved at query time rather than stored as explicit latent state. We ask what they *write* at the operation span and how it is accessed. We split a forward pass into a frozen writer and a reader: the writer's cache is recomputed without gradients, while the reader sees only the instruction and operation tokens, with all state descriptions hidden, and is trained in isolation. Anything the reader recovers was therefore already present in the unmodified cache. On a synthetic boxes task, a base reader recovers of queried bindings against - after training, and recoverability tracks the operation's read/write footprint. We find two modes of access. Across Llama-3.1-8B and Mistral-7B, operation-span transplants causally redirect which visible state is read even when the two worlds hold identical values, revealing a *routing* record. Isolation training preserves routing and adds direct access to the *payload*, the value the operation read, from the single operand-name token in a narrow mid-depth band (layers 12-15 of 32 in Llama-3.1-8B, 14-17 in Mistral-7B) — the same site that holds the routing record. The same recipe extends to further operations, ToMi and GSM8K, but is bounded by training coverage and costs open-book accuracy. Operation tokens thus leave localized, causally recoverable records that support both routing and direct payload access, though the model that writes them reads mainly the address they carry and not the value.

Primary area:
interpretability and explainable AI
Keywords:
mechanistic interpretability, entity tracking, variable binding, KV cache, activation patching, interchange interventions, in-context state tracking, language model reasoning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_a1dNOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper asks what information a decoder-only transformer writes into the KV cache at an operation span, as distinct from what the unmodified model later retrieves. It freezes a base model as a "writer," recomputes its cache without gradients, masks state-description spans from a "reader," and trains only a LoRA reader at query/answer positions. On boxes and code state-update tasks, the trained reader can recover information from operation-span KVs that the base reader seldom outputs. Carrier transplants, visibility masks, and token/layer-specific interventions are then used to distinguish a native routing/address record from a payload that becomes directly accessible after isolation training. The paper reports replication across Llama-3.1-8B and Mistral-7B, additional results on ToMi and GSM8K, and several limitations and capacity controls.

Critical strengths
  1. A clean experimental separation between writing and reading. Freezing the writer, recomputing its cache with LoRA disabled, and varying only the reader's access is a sensible design. The transplant and masking controls make the routing claim more causal than a standard probe result.

  2. Several informative controls and transparent negative results. The paper tests irrelevant donors, a filler span, localized interventions, and multiple reader seeds. It also reports value-range failures and a limited Mistral replication. These checks support the conservative claim that operation-span K/Vs contain exploitable information, although not the stronger payload interpretation.

Critical weaknesses
  1. The method does not identify the central claim. My main concern is that the paper defines information as "stored" whenever it is recoverable by a trained reader, although that reader changes approximately 13.6M parameters. Freezing the writer shows that the signal originates in the cache; it does not distinguish an existing payload from a new decoder learned over distributed features. Neither ordinary fine-tuning nor the permutation control resolves this ambiguity. There is no constrained readout or transfer test, and performance collapses when answer values move from 10-25 to 30-45. I therefore do not find the headline payload claim established.

  2. The study appears exploratory and insufficiently curated. The pre-specified endpoint was null, and the emphasized contrast was selected afterwards. One reader seed per model family is excluded, while only Mistral passes the screen for the broader model experiment. The authors also disclose generative-AI assistance in experiment and figure code, and the resulting workflow appears insufficiently supervised. Section 2 reads as a checklist of short run-in paragraphs rather than a complete methodology. The Introduction contains five overlapping contribution bullets. Figures 1, 2, and 6 are difficult to parse, and all 14 result tables are placed after the references. Together with the post-hoc analysis and screening decisions, this gives the work the appearance of auto-research rather than a mature, hypothesis-driven study.

Questions for authors
  1. How can the current experiments distinguish a pre-existing payload representation from a new computation learned by the LoRA reader? Results with a more constrained readout or transfer to unseen values and operations would materially affect my assessment.

  2. The pre-specified endpoint was null, while the reported footprint contrast was selected afterwards. Could the authors clarify how this contrast was chosen and whether it has been confirmed on an independent run or held-out setting?

  3. Were the reader-exclusion and model-eligibility thresholds fixed before observing the outcomes? How sensitive are the main conclusions when the failed reader seeds are retained and the unsuccessful model families are considered?

  4. Do the token- and layer-localization results remain similar with bf16 writers or with counterfactual patches that stay closer to the model's normal activation distribution?

2 comments

We thank the reviewer for the detailed critique and concrete questions, which we answer with new experiments.

Q1: Pre-existing payload or new computation?
We trained constrained readers and tested transfer (G2).

  • Lower capacity. A rank-1 reader (0.85M parameters, vs. 13.6M) reaches .235 payload.
  • Transfer. Every constrained reader stays well above the untrained model (.040–.058) on unseen values and unseen operations.
  • The q/k-only reader is the most informative. Its W_V, W_O, MLPs and unembedding are frozen, so its answers must pass through the base model’s own value pathway, and it transfers with almost no loss (.322 trained, .328 unseen values, .333 / .252 unseen operations).
  • Linear probe. A probe on the untrained model’s operand-token K/V decodes the value at .995 on the trained operation and (.740 / .745) on the unseen ones.
  • Value-range collapse. This turned out to be largely a demonstration confound (G2).

We treat this as evidence that the reader exposes a representation already present in the carrier. It does not, however, separate a lookup from a computation by re-weighted attention and frozen MLPs. We will therefore soften “direct payload access” to “payload readout”.

Q2: How was the contrast chosen, and is it confirmed?
It was chosen after the prespecified endpoint came out null. That endpoint could not succeed under any outcome, because the hypothesis predicts chance on sw (G1); we reported it in Appendix D because it was prespecified. With the contrast locked in advance, five new seeds per model give +.593

+.521,+.674+.521, +.674

on Llama-3.1-8B and +.275

+.174,+.365+.174, +.365

on Mistral-7B. The Mistral replication is weaker, with one seed below base.

We agree with the broader point behind this question. The prespecified endpoint was mis-specified (G1), and the value-range experiment contained a demonstration confound (G2); we thank the reviewers, whose questions led us to find the second. All new experiments in this response use analyses fixed before training, and the revision holds the paper’s claims to the same standard.

Q3: Sensitivity to excluded seeds and families.
Reader exclusion (≥ .5 on both addressed roles under OPEN) was fixed before running any experiment. All seeds (except one in Mistral) improve: a = b + 1 binding becomes +.493 / +.570 (sx / sy) on Llama-3.1-8B and +.292 / +.347 on Mistral-7B, against ≈ .000 for the uninvolved variable (G1). From Table 5 in paper, the excluded seed binds shows +.305 / +.290 in Mistral and +.365 / +.385 in Llama. They exhibit relatively weakers result comparing to included seeds, but consistently follow the improving pattern over the base model.

The model gate (base OPEN > .5) was fixed before running any experiment. We have now trained readers on every family (G3). All gain payload, including DeepSeek-7B (.042 → .185), whose base model almost never executes the operation. Gains are smaller than for Llama, and Qwen2.5-7B in bf16 gains payload but not accuracy. We will drop the gate and report all families in the revised paper.

Q4: bf16 and on-distribution patches.
In a FRESH overwrite, we replace the keys and values at the operand token with those from the same prompt but the operation reads an unassigned value (Section 4.4 in paper), so they stay within the range of states the model normally produces. Table 9 in paper compares FRESH, MEAN and ZERO, and all three localize to the same band. We also reran localization with a bf16 writer and five bf16 readers:

Transplant siteNF4 writer (5 readers)bf16 writer (5 readers)
Layers 8–11+.113
+.021,+.206+.021, +.206

| +.098

+.039,+.158+.039, +.158

| | Layers 12–15 | +.223

+.119,+.328+.119, +.328

| +.162

+.104,+.220+.104, +.220

| | Other six bands (max) | .000 | +.007

−.005,+.018−.005, +.018

|

The site is unchanged: only layers 8–15 are sufficient. The effect at 12–15 is somewhat smaller in bf16.

Necessity agrees: in bf16, overwriting the operand token’s K/V at layers 8–11 or 12–15 lowers the readers’ own accuracy by .178 and .217, against at most .013 for any other band. G3 also shows similar payload gains under NF4 and bf16 across five families.

Presentation**.** We agree and will make the following changes: rewrite Section 2 as a continuous methodology section, simplify Figures 1, 2 and 6, and move the central result tables into the main text.

Generative-AI use. As disclosed in our AI-use statement, generative AI assisted with experiment and figure code and with language editing. The authors reviewed all code and are responsible for the design, analyses and claims. We take the reviewer’s concern to rest on the analysis decisions discussed above, which we address directly: all seeds retained, thresholds’ timing, and the independent replication.

Thank you for the detailed responses and additional experiments. The constrained readers, transfer tests, fresh-seed replication, and bf16 controls substantially address my concerns about recoverability and robustness. However, the evidence supports recoverable operand information rather than an explicit post-operation state, and the methodological presentation still requires substantial improvement. I raise my score to 2.

Reviewer_yDo8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Prior work has shown that operation related representations can support later addressing. This paper asks whether the operation KV cache also retains information about the value even when the base model does not use it. The authors train a Reader to recover this information. Their setup separates a frozen Writer that produces the KV cache from a trained Reader that answers using the cache. The results suggest that the base model mainly uses the operation cache for later addressing, while value information related to the operation can also be recovered by a trained Reader. This information is localized to specific tokens and middle layers.

Critical strengths
  1. The setup to split Writer and Reader is clean and helps separate information already present in the cache from abilities learned during Reader training. The Writer is always the frozen base model. Reader LoRA is applied only at query and answer positions, so Reader training cannot write new answer information back into the prefix cache. This provides strong evidence that the signal used by the Reader already existed in the original cache.

  2. The paper provides direct causal evidence. In the a=b+1 example, transplanting the operation carrier from a world with a=c+1 redirects the model toward the source for c. Masking that source removes the effect. This is stronger than showing that a hidden state can decode a value because changing the relevant KV actually changes where the model reads and what it outputs.

Critical weaknesses
  1. The evidence is not sufficient for the stronger mechanistic interpretation. The paper shows that the cache contains value related information that a trained Reader can use. It does not yet show that this information forms a stable value representation that can be read out directly, rather than complex traces that a trained Reader learns to exploit. Some claims in the paper are stronger than the evidence supports. The authors also acknowledge that the experiments do not establish whether the value information is complete or discrete, or whether the Reader performs a lookup or a learned computation. This distinction is central to the main conclusion. A weaker explanation is still possible. Normal Transformer computation may leave distributed value related traces in the operation KV, and supervised Reader training may learn how to reconstruct the answer from them. Under distribution shift or a new operation, these traces may not be usable in the same way, and Reader performance already drops sharply in such settings. Under this interpretation, the mechanistic and generalization significance of the result is much weaker.

This concern is also consistent with the generalization results. The Reader fails when answer values move outside the training range. Expanding the training range helps on the newly covered values, yet it still fails on substantially farther values. It also does not transfer directly to a genuinely unseen operation. These results do not yet establish a general and stable interface for reading values.

The authors could test a linear or much lower capacity Reader and evaluate transfer across operations and value ranges without retraining. This would help distinguish the two explanations.

  1. The main footprint evidence has a post hoc analysis issue. The paper presents the footprint result as a held out prediction, but Appendix D states that the prespecified endpoint did not produce a positive result and that the contrast emphasized in the main text was selected after seeing the result. Some Reader seeds that did not meet the performance threshold were also excluded. This makes the evidence look more like an analysis found after the experiment than a prediction fixed in advance and then confirmed on new data. The paper should clarify whether the seed filtering rule was also fixed before the results were observed and should report the analysis with all seeds. Otherwise it is difficult to rule out selection bias. Because footprint selectivity is important evidence for the paper, this issue materially weakens the conclusion.
Questions for authors
  1. Please report the footprint results under both the original prespecified endpoint and the final contrast, with all training seeds included. Please state when the final contrast and the 0.5 OPEN accuracy exclusion rule were fixed, and clarify what s1 through s5 refer to across experiments. In particular, why do other experiments report s1, s2, and s3, while the swap experiment reports s3, s4, and s5? Were there any unreported s1 or s2 runs for the swap experiment? If the final contrast or seed filtering was chosen after observing the results, please provide an independent replication using a fixed metric and inclusion rule.

  2. Can the authors provide more direct evidence that separates reading an existing value representation from learning to recover the answer from complex residual information? For example, can a much lower capacity or linear Reader recover unseen values, and can the same Reader transfer across value ranges or operations without retraining? If only a full LoRA Reader trained for a specific operation and value range succeeds, I would favor the weaker interpretation over the stronger direct payload access claim.

  3. Were isolation Readers trained on model families that were excluded because their base OPEN accuracy was below the threshold? If so, please report the results. If not, please explain why poor base execution is sufficient to conclude that there is no useful preexisting record. The paper asks whether failure to read information implies failure to write it, so restricting the experiment to models that already execute the operation well limits the scope of that claim. Successful recovery on weaker models would strengthen the generality of the conclusion. Failure would define its boundary more clearly.

1 comment

We thank the reviewer for the careful review. The distinction the reviewer draws, between an existing value representation and traces that a trained reader learns to exploit, is the right test for our payload claim, and we ran the suggested experiments (tables in G1–G3).

Q1: Existing representation or learned reconstruction?
Lower capacity. A rank-1 reader (0.85M parameters, 16× fewer) reaches .235 payload, and a q/k-only reader, which changes only attention patterns, reaches .322, against .058 for the untrained model. A linear probe on the untrained model’s operand-token K/V decodes the value at .97–1.00.

Unseen values. The collapse on 30–45 that you referred to was largely a confound: the demonstrations, visible under OP_ONLY, always used values 10–25 no matter what the range being evaluated was (we thank the reviewer for pressing on this failure, which led us to find the confound). With matched demonstrations, payload readers trained on 10–25 recovered 30–45 at the in-range rate (q/k-only .322 → .328, rank 1 .235 → .307). Swap accuracy rises from .090 to .559, though it stays below the in-range .803.

Unseen operations. Scored on the operand, the q/k-only reader gives .333 / .252 on (a = 1 − b / a = 2 * b), against .040 for the untrained model.

We consider these results as favoring an existing representation. However, re-weighting attention is itself a computation, therefore we will soften “direct payload access” to “payload readout” and state that our evidence does not separate a lookup from a learned computation over the carrier.

Q2: Prespecification and seed selection.
The reviewer is right that the footprint contrast was chosen after the prespecified endpoint came out null (in the revision we will say so in the main text). That endpoint required an exact match on sx, sy and sw, but the hypothesis predicts chance on sw, so it could not succeed either way. To test the contrast on data it wasn’t chosen on, we performed the replication experiment in G1. Additionally, both .5 thresholds were fixed before running any experiment, in order to ensure that every model we analyze can perform the basic task.

With all seeds included, mean a = b + 1 binding is +.493 (sx) / +.570 (sy) on Llama and +.292 / +.347 on Mistral, against ≈ .000 for the uninvolved sw.

With the contrast locked in advance, five new seeds give +.593

+.521,+.674+.521, +.674

on Llama (base +.060) and +.275

+.174,+.365+.174, +.365

on Mistral (base +.135, one seed below base). We will report the Mistral result as a weaker replication.

Regarding seed labels, “sk” is the k-th seed within each experiment. The swap s1/s2 used an earlier prompt format and were never evaluated on the test set. After retraining them, they score .667 / .658, and we will add them to the appendix.

Q3: Families that failed the gate.
We have now trained readers on every model family (G3). Every excluded family gains payload over its untrained model, including DeepSeek-7B whose base model almost never executes the operation (.02 OPEN, NF4), yet its payload rises from .042 to .185. Gains are smaller than for Llama-3.1-8B. Poor execution therefore does not imply that no record was written. We will drop the gate and report all families in the revised paper.

Reviewer_7dHhOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies the performance bottleneck of the attention module. Specifically, the authors investigate whether the origin of the failure in attention is in writing to KV or in reading from KV. The designed experiments reveal that reading from the KV limits performance, and this can be mitigated by additional training.

Critical strengths
  1. I really like the question and ideas investigated in the paper. This work investigates an important question: what is the origin of the attention failure?

  2. The paper identifies that the core bottleneck is reading information from the KV cache. The designed experiments are very interesting and novel in the literature.

  3. The authors show that LoRA on some parameters can greatly improve the found performance bottleneck.

Critical weaknesses
  1. The main concern in the experiment design is the set of trainable parameters. The authors train all the attention params WQW_Q, WKW_K, WVW_V, and WOW_O. I think it will be more beneficial to conduct a fine-grained analysis of these parameters. For example, whether QQ and KK are the main contributors or VV and OO.

  2. In addition, it is beneficial to discuss whether the benefits of tuning these parameters result from learning some knowledge via associative memory. The previous works, Birth of a Transformer: A Memory Viewpoint and Muon Outperforms Adam in Tail-End Associative Memory Learning, show that W_V and W_O are also associative memories. It is better to discuss whether the benefit comes from learning the factual knowledge in these parameters.

Questions for authors

See the weaknesses.

1 comment

We thank the reviewer for the encouraging review and the two concrete suggestions.

Q1: Which attention parameters matter?
We report this ablation for accuracy in Appendix J and we now also report payload readout (G2). Neither half dominates. With all description lines masked, a q/k-only adapter recovers .322 and a v/o-only adapter .275, against .432 for full LoRA and .058 for the untrained model, and a rank-1 adapter still reaches .235. We will move this decomposition into the main text.

Q2: Could the gain come from associative memory in W_V/W_O?
We agree that W_V and W_O could act as associative memories (as shown in Bietti et al., 2023; Wang et al., 2026). However, three results argue against this as the source of the payload:

  • q/k-only training. The q/k-only adapter leaves W_V, W_O and all MLPs at their base/original weights, so it cannot store new associations there. Yet it recovers about three quarters of the full payload.
  • Transplants. The reader’s weights are identical across donors, yet transplanting a donor’s operation-span K/V switches the output to that donor’s value. A fixed map in W_V/W_O cannot know which value a given donor holds (at most it decodes what the cache provides).
  • Transfer. Readers trained only on a = b + 1 with values 10–25 transfer to values 30–45 and to unseen operations (G2), for which no association was ever trained.

The above results rule out memorized answers in W_V/W_O. They do not show that the reader performs a pure lookup, since re-weighted attention followed by frozen MLPs can still compute. We will add this discussion to the Limitations section.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Response #1

We thank the reviewers for their feedback. All three valued the writer/reader design, and two noted that the transplants give causal rather than probing evidence. Three concerns were shared: the analysis protocol including post-hoc contrast and seed labels (G1), whether payload access reflects an existing representation or a computation learned by the reader (G2), and the model families below our screening threshold, together with precision (G3). While investigating G2 we found a confound in our value-range experiment, which we describe and correct below. In every table presented, intervals are 95% bootstrap CIs. While some results in paper are reported with three seeds, as requested by reviewers, we report every result in this response over five seeds.

G1. Analysis protocol

The reviewers are right that the Section 4.2 contrast was chosen after the prespecified endpoint came out null, and in revision we will mention it in the main text. However, the endpoint was mis-specified because under OP_ONLY the sw line is masked, and the hypothesis being tested says the operation carries nothing about sw. The endpoint could therefore not succeed whether or not the hypothesis held.

To test the contrast on data it wasn’t chosen on, we ran a fresh replication with five new reader seeds per model and everything else unchanged:

ModelPrespecified endpointContrast, trainedContrast, baseContrast, trained, per seed
Llama-3.1-8B−.010
−.030,+.010−.030, +.010

| +.593

+.521,+.674+.521, +.674

| +.060

+.018,+.105+.018, +.105

| .542 / .493 / .738 / .573 / .617 | | Mistral-7B | +.015

−.015,+.050−.015, +.050

| +.275

+.174,+.365+.174, +.365

| +.135

+.080,+.193+.080, +.193

| .282 / .415 / .255 / .097 / .325 |

Llama replicates clearly. Mistral replicates weakly: the mean exceeds the base model, but one seed (.097) falls below it. We will report it as a weaker replication.

Readers and model-eligibility. Both .5 thresholds (per reader and per family) were fixed before the first footprint run. Including the excluded seeds, mean binding is +.493 (sx) / +.570 (sy) on Llama and +.292 / +.347 on Mistral, against ≈ .000 for the uninvolved sw. We drop the family threshold entirely (G3).

Seed labels. “sk” denotes the k-th seed within each experiment, so labels are not shared across experiments (i.e. the same label in two experiments does not mean the same reader). In the revised paper, we won’t distinguish them by name and report the statistics for all seeds being trained and evaluated. The swap s1 and s2 were trained under an earlier prompt format and never evaluated on the test set. We retrained them in the final format, and they scored .667 / .658 (s3–s5: .627 / .688 / .610).

AuthorsOpenReview ↗

General Response #2

G2. Existing representation or learned computation?

Following what reviewers asked, we trained constrained readers on a = b + 1 (values 10–25) and evaluated them without retraining on unseen values (values 30-45) and unseen operations (subtraction and multiplication). On unseen operations we score the operand b, so the test measures access to the payload rather than the ability to compute a new operation (which is the case with scoring a).

Reader (trainable params)Trained settingValues 30–45 (unseen)a = 1 − ba = 2 * b
Untrained base.058.050.040.040
Full LoRA, rank 16 (13.6M).432.468.335.289
q/k only, rank 16 (6.8M).322.328.333.252
v/o only, rank 16 (6.8M).275.303.341.276
Rank 1, q/k/v/o (0.85M).235.307.323.236

Every reader stays well above the untrained model under both unseen values and operations. The q/k-only reader is the most informative: it changes only attention patterns, while W_V, W_O, the MLPs and the unembedding layer remain frozen, yet it recovers about 75% of the full payload and transfers with almost no loss. A decoder learned for one operation and value range would not be expected to do this. A linear probe on the untrained model’s operand-token K/V decodes the operand’s value at .97–1.00 (chance 1/16=0.0625) and at (.740 / .745) on the unseen operations, showing the value is linearly present before any reader is trained.

Value-range confound. In the in-context demonstrations, which OP_ONLY leaves visible, we always used values 10–25, including when the problem used 30–45. Matching the demonstrations to the problem largely restores the swap readers trained on 10–25:

Problem values (trained on)Demonstrations 10–25 (paper)Demonstrations 30–45
10–25.803 [.74, .86].312 [.17, .45]
30–45.090 [.00, .18].559 [.50, .61]

Matched accuracy on 30–45 stays below the in-range level, and a mismatch in either direction hurts, meaning that the reader is sensitive to the demonstration distribution. We will revise Section 5 and Appendix G, including the “answer side, not context side” conclusion, which this confound undermines, and highlight the reader’s sensitivity to the demonstrations’ distribution in the text.

G3. Families below the threshold, and precision

We trained readers on n = 100 samples with 5 different seeds on every family the gate excluded, with both NF4 and bf16 writers. Each cell reports the results for NF4 / bf16.

ModelBase OPEN accuracyOP_ONLY accuracy, base → trainedPayload, base → trained
DeepSeek-7B.020 / .180.090 → .254 / .110 → .260.042 → .185 / .058 → .172
CodeLlama-13B.200 / .320.110 → .342 / .100 → .302.042 → .190 / .033 → .202
Qwen2.5-7B.360 / .380.080 → .184 / .180 → .174.033 → .092 / .083 → .123
Yi-1.5-9B.530 / .540.070 → .380 / .090 → .312.025 → .202 / .008 → .160
Llama-3.1-8B (reference).730 / .570.070 → .598 / .050 → .556.058 → .432 / .125 → .387

Every family gains payload after training. This includes DeepSeek-7B, whose base model almost never performs the operation (.02 OPEN in NF4). Gains are smaller than for Llama-3.1-8B, and Qwen2.5-7B in bf16 gains payload but not accuracy. We will drop the gate and report all families in the revised paper.

Yi-1.5-9B was excluded in the paper at .500 (n = 40, gate >.500 OPEN). At n = 100 it sits at the threshold (.53 [.43, .63]), so its exclusion was a matter of sample size.

Finally, the table shows that payload gains follow similar pattern under NF4 and bf16 in every family.

We thank all reviewers again for their detailed feedback. We have uploaded a revised version of the paper that implements the changes promised in our responses.

p.s. On moving the central result tables into the main text, we moved the constrained-reader table (now Table 1), which carries the central evidence that a minimal reader recovers and transfers the payload, into the main text; the submitted version had no table there. The remaining result tables stay in the appendix because the revision already fills the 10-page limit. To compensate, the main-text figures now print the exact values instead of only bars, every figure caption points to the full table behind it (e.g., Figure 2 → Table 3, Figure 3 → Tables 5–7, Figure 6 → Table 13).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.