acceptodds
Under review as a conference paper at ICLR 2027

Memory Is Not Control: Recovering Explicit Forgetting in Reasoning-Distilled Models via Contrastive Routing

Abstract

Large language models (LLMs) have demonstrated strong reasoning abilities, which distillation transfers to smaller models. However, reasoning-distilled models recall earlier information better than their base models on a multi-turn dialogue benchmark while following earlier instructions less reliably, illustrating that memory does not ensure control. In this paper, we study this gap through explicit forgetting and find that reasoning-distilled models achieve higher immediate forgetting compliance than their paired base models, but this advantage shrinks or reverses on delayed probes. In contrast, base models often avoid forgotten content that the distilled models reuse. Yet injecting base-model activations or logits during generation fixes some answers while breaking others, since it acts before the original answer exists. Inspired by these findings, we propose Forgotten-State Contrastive Routing (FSCR), which turns the forgotten-content leakage of both completed answers into a contrastive routing logit. Concretely, FSCR switches to the base answer only when this logit is non-negative. The logit is the minimum of three margins computed from the lexical leakage of each answer and the gap between them. We evaluate FSCR on multiple explicit-forgetting benchmarks with several reasoning-distilled models, illustrating consistent improvements with both annotated and automatically extracted forgotten content, demonstrating that FSCR can achieve better forgetting compliance across models and dialogues. The gains are largest after intervening, showing the effectiveness and robustness of our analysis findings and the proposed FSCR. Our code and data are available at https://anonymous.4open.science/r/fscr-review-DEF8

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language models, multi-turn dialogue, dialogue systems

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

This paper studies explicit forgetting in reasoning-distilled language models, focusing on whether a forgetting instruction remains effective after intervening conversation turns. The authors introduce ICF-DSM, evaluate four base/distilled model pairs, and propose Forgotten-State Contrastive Routing (FSCR). FSCR generates an answer from each model and routes to the base-model answer when three conditions on lexical leakage scores are met.

The paper reports positive held-out gains across all four pairs, including experiments using dialogue-derived forgotten states. The problem is relevant, and the results suggest that complementary model outputs can be used to improve forgetting compliance. My main reservations concern the interpretation of the delayed-forgetting results and whether the proposed gate adds value over a simpler, calibrated selector.

Critical strengths

1、The consistent held-out gains suggest that the base and distilled models make complementary errors that can be usefully exploited. Table 2(c) reports improvements of 1.00–5.76 percentage points across four model pairs, with dialogue-clustered confidence intervals excluding zero. Using the same candidate answers across routing rules also makes the comparison cleaner. Together with the positive dialogue-derived-state results, this supports FSCR as a practical, training-free intervention that does not rely solely on manually annotated forgetting targets.

  1. The distinction between immediate and delayed forgetting is worth studying. ICF-DSM tests whether a forgetting instruction remains effective after intervening dialogue, while also evaluating retained-fact use and adversarial reconstruction. This is more informative than checking only whether the model suppresses a fact immediately after the request. The matched-query control and blind human audit are useful additions, even though their scope limits the conclusions that can be drawn from them.
Critical weaknesses
  1. I am not convinced that FSCR’s three-condition gate offers a better trade-off than a simpler, properly tuned selector. At the reported settings in Table 4, margin-only achieves larger net gains in every model pair, while FSCR routes fewer answers and introduces fewer regressions. Appendix D.4 makes the trade-off concrete: the extra checks avoid 8 regressions but also give up 88 fixes. This may be reasonable when regressions are particularly costly, but a higher margin-only threshold might achieve a similar trade-off. The current results do not establish whether FSCR’s additional conditions help beyond implementing a more conservative operating point. This matters for judging the method’s contribution.

  2. The evidence for a delay-related control deficit is narrower than the original comparison suggests. The immediate and delayed probes differ in both query wording and intervening dialogue. For Qwen-7B, base-model SMS rises from 5.43% to 39.64%, while distilled-model SMS falls from 34.00% to 22.43% (Table A4). The relative contrast therefore cannot be attributed solely to the distilled model losing the instruction over time. The matched-query control is a useful step, but it is limited to Qwen-7B; the zero-versus-two-exchange interaction is estimated at zero with a wide confidence interval (Table A6). This does not rule out a delay effect, but leaves open how much of the original result reflects intervening turns rather than differences between the probes. The other model pairs lack the same controlled comparison.

Questions for authors
  1. Could you tune the margin-only threshold on the development set to match FSCR’s coverage, then compare both methods on the held-out set? This would help determine whether the extra conditions improve the trade-off when both methods operate at a similar coverage level.
  2. How do you interpret the original two-exchange contrast in light of the zero estimated interaction in Table A6? Do matched-query results show a similar pattern for any of the other model pairs?
3 comments
AuthorsOpenReview ↗

Response to Reviewer YStk (1/2)

Thank you for your careful review and constructive suggestions. We have added routing comparisons calibrated on development data and extended the matched-query controls to address your questions.

W1 & Q1. Comparisons with calibrated margin-only routing

Following your suggestion, we calibrated margin-only routing to match FSCR's coverage on the development set and compared the frozen rules on held-out data. We also evaluated the more conservative operating point you proposed by matching FSCR's development regression count. Both thresholds were selected on ICF-DSM development dialogues and frozen before test evaluation: t = 0.32 for development coverage and t = 0.447 for the development regression count.

The table reports FSCR minus margin-only in net gain (pp), pooled over the four DeepSeek pairs, with dialogue-clustered 95% CIs and shared three-vote labels. The comparison with coverage matched on development data shows a small observed difference on ICF-DSM and a positive difference on MQuAKE. With the threshold matched to the development regression count, FSCR achieves larger net gains on both test sets.

Test set (evaluations)Matched development coverage (t = 0.32)Matched development regression count (t = 0.447)
ICF-DSM held-out (5,208)−0.21 [−0.48, +0.08]+0.73 [+0.36, +1.11]
MQuAKE transfer (3,998)+0.23 [+0.03, +0.45]+0.65 [+0.40, +0.93]

On ICF-DSM, FSCR routed 11.5% of answers and margin-only at t = 0.32 routed 11.3%. FSCR made 217 fixes and 17 regressions, compared with 226 fixes and 15 regressions for margin-only. At t = 0.447, margin-only made 166 fixes and 4 regressions: relative to FSCR, this more conservative rule gave up 51 fixes to avoid 13 regressions. The +0.73% difference therefore describes the held-out performance of rules matched on the development regression count, with their realized test regression counts reported explicitly.

On MQuAKE, the same frozen FSCR rule made 72 fixes and 1 regression with 144 switches. Margin-only at the threshold matched to development coverage switched 336 answers, producing 66 fixes and 4 regressions. We also swept every shared margin threshold on the test set as a descriptive check of the attainable trade-off. At the same observed regression cap of one, FSCR fixed 72 errors, whereas the best margin-only threshold fixed 47. Reaching 72 fixes with margin-only required at least six regressions. This comparison directly tests whether increasing a single margin threshold can reproduce FSCR's operating point on these transfer samples.

The small observed difference on ICF-DSM at matched coverage can be traced to the individual routing decisions. When the distilled answer leaks heavily and the base answer does not, their leakage gap is also large, so the two rules mostly select the same answers. Of FSCR's 600 switches, 501 are shared with margin-only and contribute the same 202 fixes and 12 regressions to both rules. FSCR's 99 additional switches have gaps between 0.15 and 0.32 and yield 15 fixes and 5 regressions. The 87 switches made only by margin-only have gaps of at least 0.32 but fail one absolute check; they yield 24 fixes and 3 regressions. These differing decisions account for the 11 net corrections out of 5,208 evaluations (0.21%), whose interval includes zero.

Together, these comparisons characterize both the shared decisions and the contribution of the absolute checks. The higher threshold matched to development regressions gives 0.73% fewer net corrections on ICF-DSM and 0.65% fewer on MQuAKE, while the MQuAKE threshold sweep shows that the additional conditions select an operating point that a single margin threshold does not reproduce under the same observed regression cap.

We have added these comparisons calibrated on development data, routing-decision analyses, and transfer results in Section 5.5, Table 4(c), and Appendices D.4 and F.3 of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer YStk (2/2)

W2 & Q2. Matched-query controls and the two-exchange contrast

To separate query wording from intervening dialogue, we expanded the matched-query analyses in Appendices B.3 and B.4. The zero interaction in Table A6 is a point estimate for one contrast in 100 dialogues; the expanded controls examine the trajectory across multiple delays while holding the final question fixed.

In the original matched-query control, Qwen-7B's distilled model scored 36%, 51%, 34% and 23% SMS after 0, 1, 2 and 4 intervening exchanges, respectively; its base model scored 11%, 8%, 9% and 7%. With the wording fixed, the base model no longer shows the rise from 5.43% to 39.64% seen in Table A4. From 0 to 2 exchanges, both models lose exactly 2%, yielding the zero interaction. This two-point contrast does not use the distilled model's rise to 51% at one exchange or its decline to 23% at four exchanges. Its 95% CI of [−13, +13]% also accommodates effects of a few percentage points. Using all four delays, the distilled model's SMS slope was 3.86% per exchange more negative than the base model's (p = 0.0094).

We then extended the control to 300 source dialogues with 0, 2, 4, 8 or 16 intervening exchanges, again keeping the final question fixed within each dialogue. Qwen-7B's distilled model lost 1.37% more SMS per exchange than its base model (95% CI [1.00, 1.73]). Under this fitted linear trend, a two-exchange contrast would be about 2.7%, within the original interval; this is an implication of the fitted slope rather than a separately measured two-exchange effect.

The expanded matched-query controls also identify faster declines in other model pairs. Across all four pairs, four of six prespecified slope comparisons survived Holm correction (adjusted p = 0.0003 each): Qwen-7B and Llama-8B with short exchanges, and Qwen-14B and Llama-8B with long exchanges. All six point estimates indicate a faster decline for the distilled model. The Qwen-1.5B estimate (−0.07% per exchange) and Qwen-14B estimate with short exchanges (−0.37%) were not significant.

We have added the interpretation of the two-exchange contrast and the expanded matched-query results in Appendices B.3 and B.4 of the revised manuscript. To make comparisons across model pairs easier, we revised Figure 2 to show the distilled–base gap on immediate and delayed probes (a) and the size and direction of each behavioral score difference (b).

Dear Reviewer,

We appreciate the time and effort you have devoted to reviewing our work. We would like to briefly follow up on our responses to your comments and ask whether you have any further feedback.

We hope our rebuttal has helped clarify the points you raised. If any aspects remain unclear or require additional discussion, we would welcome the opportunity to elaborate.

Thank you again for your thoughtful feedback and your time.

Best regards, The Authors

Reviewer_a1uYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates the fundamental distinction between inference memory and control dynamics within language models. The authors propose an approach called "Forgotten-State Contrastive Routing" (FSCR), integrated with "logit mixing" strategies, to decouple mere information storage from operational reasoning during complex generation.

Critical strengths

The intent to conceptually and architecturally separate 'memory' from 'control' offers a notable theoretical perspective that contrasts sharply with traditional, opaque end-to-end black-box models. Forgotten-State Contrastive Routing (FSCR) offers a compelling approach to managing reasoning trajectories without overburdening the model's decision-making dynamics.

Critical weaknesses

The primary weakness is the severe lack of clarity in the methodological sections. Numerous key passages, particularly the exact mechanics of 'logit mixing' and the practical implementation of the routing (Algorithm 1), are excessively convoluted and opaque. This forces the reader into an interpretive effort that severely penalises the manuscript's readability.

Furthermore, the ablation studies are incomplete. For an architecture proposing such specific, decoupled modules, it is essential to rigorously isolate the empirical impact of individual components. Without tests that selectively disable contrastive routing or logit mixing, it is impossible to verify how much each innovation contributes to the final performance, undermining the reproducibility and robustness of the results.

Questions for authors

Could you restructure and formalise the mathematical and logical steps for 'logit mixing' and FSCR with greater transparency to make the framework more accessible to readers?

Do you plan to include a comprehensive ablation study to isolate each module's effectiveness, ensuring that performance gains are attributable to the proposed routing rather than uncontrolled end-to-end dynamics?

2 comments
AuthorsOpenReview ↗

Response to Reviewer a1uY

Thank you for your thoughtful review and suggestions on the method description and component ablations. We have clarified the generation and routing procedures and added the requested ablations.

W1 & Q1. Clarifying logit mixing and the FSCR procedure

We have made the distinction between the two procedures explicit in the opening of Section 4. FSCR is a training-free selection rule over two completed answers; logit mixing is a separate decoding-time baseline. FSCR has no learned modules. Both models generate their answers independently, and FSCR returns one of them. The term "contrastive routing logit" in Section 4.3 denotes a scalar decision score computed from the completed answers.

Logit mixing (Section 3.2 and Appendix B.7). At answer step jj, this baseline decodes from (1−α)zD,j+αΠzB,j(1-\alpha)\mathbf{z}_{D,j}+\alpha\Pi\mathbf{z}_{B,j}, where Π\Pi maps base-model tokens to the distilled vocabulary by token string. With α=0.9\alpha=0.9, it produced 20 fixes and 56 regressions on 300 Qwen-7B forgetting probes. These regressions show that changing generation token by token can also damage answers that were already correct, motivating selection after both answers are complete.

FSCR (Section 4 and Algorithm 1). The distilled and base models answer independently. Given the forgotten facts FF, FSCR computes lexical leakage scores LDL_D and LBL_B from their visible final answers (Equations 3–4). With annotated states, it returns the base answer only when all three conditions hold: LD≥0.50L_D\geq0.50, LB≤0.40L_B\leq0.40, and LD−LB≥0.15L_D-L_B\geq0.15. Otherwise, it keeps the distilled answer. Model weights, activations, and token logits remain unchanged, and no judge or correctness label enters the routing decision at inference time.

We have added these clarifications to the opening of Section 4 and Section 4.3, with the complete routing procedure in Algorithm 1 (Appendix C). To make comparisons across model pairs easier, we revised Figure 2 to show the distilled–base gap on immediate and delayed probes (a) and the size and direction of each behavioral score difference (b).

W2 & Q2. Isolating the contribution of each routing condition

We added an ablation of each of FSCR’s routing conditions. Logit mixing is evaluated independently as described above; FSCR already operates without it. To isolate each condition, we removed it while keeping the generated answers, the other thresholds, and the three-vote labels fixed. We also disabled routing entirely by always retaining the distilled answer. The table reports all variants on the same 5,208 held-out ICF-DSM evaluations with annotated states.

Routing ruleFixesRegressions
Keep the distilled answer00
Full FSCR21717
Remove the distilled-answer leakage check24420
Remove the base-answer safety check26624
Remove the leakage-gap check21821

Each condition reduces regressions in this ablation. The leakage-gap check prevents 4 regressions at the cost of one fix. The distilled-answer leakage check and base-answer safety check prevent 3 and 7 regressions, respectively, while retaining fewer fixes. Because every row uses the same candidate answers and labels, these differences isolate the effect of the routing conditions.

We separately tested whether switching the same number of answers without leakage evidence could reproduce the gain. On the 434-dialogue held-out tail, same-coverage random routing changed SMS by −0.15, −1.08, −0.23, and −0.69% for the four pairs (Table 4(b)). On two prespecified 300-probe holdouts, FSCR recovered 12 and 5 base-only wins with no regressions. Random routing is expected to lose 1.92 and 2.42 answers on those holdouts, and it matches FSCR’s net gain with probability at most 3.19×10−53.19\times10^{-5}. These controls show that the observed gain depends on the routing evidence rather than switching alone.

We report the condition ablations in Appendix D.4 and the random-routing controls in Appendix D.3.

Dear Reviewer,

Thank you again for your careful review and helpful suggestions. We are writing to follow up on our rebuttal and would appreciate hearing your thoughts on the clarifications we provided.

When you have a chance, please let us know if there are any remaining concerns or points that would benefit from further explanation. We would be happy to address them and continue the discussion.

Thank you for your time and consideration.

Best regards, The Authors

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

Strong conversational memory does not ensure that reasoning-distilled models continue honoring explicit forgetting instructions after intervening dialogue. The paper proposes FSCR, which compares forgotten-content leakage in completed base-model and distilled-model answers and selectively returns the base answer. Experiments on explicit-forgetting and derived knowledge-update dialogues report improved compliance across model pairs, with the largest gains on delayed probes and fewer benefits under adversarial questioning.

Critical strengths
  1. The diagnostic separates immediate compliance from maintaining an updated instruction state across later turns.

  2. Held-out dialogue evaluation, clustered uncertainty estimates, and a blind human audit strengthen the empirical analysis.

Critical weaknesses
  1. Simpler routing rules obtain larger net gains, while FSCR trades lower coverage for higher precision. Without matching coverage or the allowed number of regressions, the benefit of its three-margin gate remains unclear.

  2. The immediate and delayed probes differ in wording as well as dialogue position, and the matched-query control covers only one model pair. Differences in training data and chat templates also prevent attributing the observed behavior specifically to reasoning distillation.

  3. Lexical scoring can miss paraphrased disclosures and penalize harmless overlap, with more regressions when the forgotten scope is extracted automatically. Human agreement on individual fixes and regressions is also limited, leaving uncertainty about the reported correction rates.

Questions for authors

When margin-only routing is tuned to the same coverage or regression budget as FSCR, does the three-margin gate still improve net corrections?

4 comments
AuthorsOpenReview ↗

Response to Reviewer gqGa(1/3)

Response to Reviewer gqGa

We thank the reviewer for the thoughtful feedback. We have added controlled routing comparisons, extended the fixed-question experiments, and evaluated scoring and judgment reliability. We address each point below.

W1 & Q1. Comparing FSCR with a calibrated margin-only rule

Thank you for suggesting these controlled comparisons. We calibrated margin-only on ICF-DSM development dialogues to match FSCR's coverage or regression count, then froze both thresholds before testing: t = 0.32 for development coverage and t = 0.447 for the development regression count. Against the rule calibrated to FSCR's development regression count, FSCR improves held-out net corrections by 0.73% on ICF-DSM and 0.65% on MQuAKE. The table reports FSCR minus margin-only in net gain, pooled over 4 DeepSeek pairs, with dialogue-clustered 95% CIs and shared three-vote labels.

Test set (evaluations)Development coverage matched (t = 0.32)Development regression count matched (t = 0.447)
ICF-DSM held-out (5,208)−0.21 [−0.48, +0.08]+0.73 [+0.36, +1.11]
MQuAKE transfer (3,998)+0.23 [+0.03, +0.45]+0.65 [+0.40, +0.93]

On ICF-DSM, FSCR routed 11.5% of answers and margin-only at t = 0.32 routed 11.3%. FSCR made 217 fixes and 17 regressions, gaining 3.84% over keeping the distilled answer (95% CI [3.23, 4.47]); margin-only made 226 fixes and 15 regressions. At t = 0.447, margin-only made 166 fixes and 4 regressions: relative to FSCR, it avoided 13 regressions while giving up 51 fixes. These are the held-out outcomes of matching the regression count on development data.

On MQuAKE, with the same frozen thresholds, FSCR made 72 fixes and 1 regression with 144 switches. Margin-only at the development coverage-matched threshold switched 336 answers, making 66 fixes and 4 regressions. We also conducted a descriptive sweep of every shared margin threshold on the test set. The best threshold with at most one regression fixed 47 errors, whereas reaching 72 fixes required at least six regressions. This comparison directly shows the benefit of the absolute checks at the same observed regression count on MQuAKE.

The small difference at matched coverage on ICF-DSM is explained by the overlap between the routed sets. When the distilled answer leaks heavily and the base answer does not, their leakage gap is also large. Consequently, 501 of FSCR's 600 switches are shared with margin-only and contribute the same 202 fixes and 12 regressions to both rules. FSCR's 99 additional switches have gaps between 0.15 and 0.32 and yield 15 fixes and 5 regressions. Margin-only's 87 additional switches have gaps of at least 0.32 but fail one absolute check, yielding 24 fixes and 3 regressions. The net difference is therefore 11 of 5,208 evaluations (0.21%), with an interval that includes zero. Together, the comparisons identify where the absolute checks contribute: they retain more fixes than the development regression-matched rule on both test sets and exclude regression-inducing switches on MQuAKE, where margin-only at t = 0.32 makes 4 regressions to FSCR's 1.

A regression occurs when routing replaces a correct distilled answer with an incorrect base answer. For FSCR, this occurred in 17 of 5,208 ICF-DSM evaluations (0.33%).

We have added these comparisons and the routing analysis to Section 5.5 and Table 4(c), Appendix D.4, and Appendix F.3.

AuthorsOpenReview ↗

Response to Reviewer gqGa(2/3)

W2. Fixed-question controls across model pairs

We have extended the control from one model pair to all, holding the final question fixed within each dialogue, and added a comparison with an Instruct model trained from the same base. The distilled-versus-base decline persists with fixed question wording.

The expanded experiments use 300 source dialogues, with 0, 2, 4, 8 or 16 intervening exchanges in the short-exchange setting. Across the short- and long-exchange settings, four of six prespecified slope comparisons survived Holm correction (adjusted p = 0.0003 each): Qwen-7B and Llama-8B with short exchanges, and Qwen-14B and Llama-8B with long exchanges. All six point estimates show a faster decline for the distilled model; for Qwen-7B, the gap is 1.37% per exchange (95% CI [1.00, 1.73]).

The same-base Instruct comparison provides a further control for the tested Llama models. From 0 to 16 short exchanges, state-management success (SMS) fell from 52.8% to 19.0% for R1-Distill-Llama-8B and rose from 82.7% to 90.3% for Llama-3.1-8B-Instruct. The slope difference was −2.34% per exchange (95% CI [−2.75, −1.92], Holm-adjusted p = 0.0001). Thus, the two checkpoints trained from the same Llama base show opposite trends: the decline observed in the reasoning-distilled checkpoint is not shared by its Instruct counterpart. Together, these controls establish the fixed-question delay pattern across the tested model pairs and distinguish the two post-training outcomes in the tested Llama comparison.

We have added the expanded fixed-question controls and the same-base Instruct comparison to Appendix B.4 and Appendix G. To make comparisons across model pairs easier, we revised Figure 2 to show the distilled–base gap on immediate and delayed probes (a) and the size and direction of each behavioral score difference (b).

AuthorsOpenReview ↗

Response to Reviewer gqGa(3/3)

W3. Leakage scoring, extracted states, and human judgments

We evaluated these three aspects directly by testing a stronger leakage scorer, automatically extracted forgotten states, and a second blind human audit. The routing rule gains with both lexical and model-based evidence, and retains a positive net gain with automatically extracted states.

For leakage scoring, the lexical failure check detected 2,692 of 3,657 judge-labeled misuses in distilled answers (73.6%). Replacing lexical overlap with Qwen3-14B raised development AUC on distilled answers from 0.85 to 0.90. With thresholds reselected on development dialogues, held-out net gain increased by 1.00% over lexical FSCR (95% CI [0.50, 1.54]). This experiment demonstrates that FSCR can use stronger leakage evidence through the same routing rule.

With automatically extracted forgotten states, FSCR made 203 fixes and 72 regressions, gaining 2.52% over the distilled model (95% CI [1.84, 3.21]). These results quantify performance when the forgotten scope is obtained from the dialogue itself, including the additional regressions in this setting.

In the second blind audit of 84 routed probes, the two annotators agreed on forgetting violations for 159 of 168 answers (κ = 0.893). Among the 69 audited transitions labeled as fixes or regressions by the judge, the annotators jointly identified 27 fixes and 3 regressions. Of the 17 judge-labeled regressions, they jointly confirmed 3, rated 8 as unchanged, and rated 1 as a fix. They also jointly rated 22 of the 52 judge-labeled fixes as unchanged. We report these transition-level results alongside the judge-based counts, providing a direct human assessment of the corrections and regressions within the audited sample.

We have added the extracted-state results, second blind audit, and stronger-scorer experiment to Appendix D.4, Appendix D.5, and Appendix G.

Dear Reviewer,

Thank you again for your time and thoughtful feedback. We would like to briefly follow up on our rebuttal, where we provided detailed responses to your comments.

When you have a chance, we would greatly appreciate your feedback on whether our responses have addressed your concerns. If any questions or concerns remain, we would be happy to provide further clarification and continue the discussion.

Thank you for your time and consideration.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.