acceptodds
Under review as a conference paper at ICLR 2027

DELMOE: DUAL-SPARSE ROUTE DELETION WITH RECTIFICATION FOR EFFICIENT BATCHED DECODING

Abstract

Mixture-of-experts (MoE) models expand model capacity through sparse expert activation without proportionally increasing per-token computation. In batched decoding, different requests can collectively activate many experts, increasing latency without necessarily improving accuracy. Yet skipping experts can reduce accuracy, and surviving gate weights may no longer suit the reduced expert mixture. Existing batch-aware route replacement reduces active experts but largely preserves route counts and redistributes weights without calibrating them for accuracy under the altered routing state. We propose DelMoE with two modules for retraining-free batched MoE decoding. (1) Reuse-aware dual-sparse route deletion: We build a backbone from high-priority native routes, reuse its experts across requests, and supplement the retained set. Deleting routes outside this set reduces both active experts and token–expert routes. (2) Routing-aware gate-weight rectification: We calibrate one rectification coefficient per model offline and keep it fixed during inference to adjust surviving gate weights. To avoid executing deleted zero-weight slots, our custom vLLM CUDA backend processes retained routes as variable-length workloads. Across three MoE models and eight benchmarks, DelMoE achieves higher average accuracy and generally faster decoding than batch-aware route-replacement baselines, with particularly large accuracy gains at high sparsity. Compared with Vanilla, DelMoE achieves up to 2.49× decoding speedup at comparable or higher average accuracy while activating fewer experts.

Primary area:
infrastructure, software libraries, hardware, systems, etc.
Keywords:
Mixture-of-Experts, Efficient Inference, Batched Decoding, Dual-Sparse Route Deletion, Gate-Weight Rectification

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_ibXZOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper aims to enable MoE to generate responses in batches with fewer experts and routes, while still maintaining the quality of the responses. The author first retains the most important routes for each token, then utilizes the reuse of experts among the same batch requests, and deletes the remaining routes. Subsequently, the weights of the retained routes are adjusted, and a dedicated implementation is used to skip the deleted parts. The experiment compared the quality of the responses and the generation speed.

Critical strengths

S1. Reducing the number of active experts and routes from tokens to experts is a clear idea. It takes into account both the cost of accessing expert weights and the actual execution cost.

S2. The comparison of deleting routes and replacing routes is convincing. The author fixed the expert set and the original weights, showing that deleting routes alone can bring benefits, not just the expert selection or weight adjustment.

S3. The comparison of execution optimization maintains consistency in routes, weights, and outputs. This makes it easier to confirm that the additional acceleration comes from skipping invalid calculations rather than further sacrificing the quality of the responses.

Critical weaknesses

W1. Accuracy is measured under fixed batches, and speed is measured under dynamic service batches. Other requests in the same batch can affect the expert selection. The appendix also shows that batch size affects quality, so it is necessary to confirm whether the same answer quality can still be maintained when actual acceleration is achieved.

W2. The trigger condition for weight correction is determined before supplementing experts. Some tokens may have no routes deleted in the end and may still have their weights reduced. Therefore, it is not clear whether the benefit comes from fixing the changes after deleting routes or from ordinary output scaling.

Questions for authors

Q1. Can the same set of evaluation requests be used to measure the quality and speed of answers simultaneously under dynamic service batching?

Q2. With the option for supplementary experts enabled, using the same calibration method, how does the current correction compare to "only correcting tokens that are ultimately missing a route" and "uniformly scaling all tokens"?

3 comments
AuthorsOpenReview ↗

Response to Reviewer ibXZ (Part 1 / 2)

We sincerely thank the reviewer for the careful reading and the encouraging feedback. We are glad that you found reducing both active experts and token–expert routes a clear idea (S1), the comparison between route deletion and route replacement convincing (S2), and the execution comparison consistent in routes, weights and outputs (S3).

Your suggestions were highly constructive and helped us strengthen the paper: they led us to measure accuracy and speed on the same requests under dynamic batching, and to compare the paper rule against the two alternatives you proposed. We hope these results further strengthen your support, and we would be grateful if you would consider raising your score. Below, we provide detailed responses to all the comments:


W1: "Accuracy is measured under fixed batches, and speed is measured under dynamic service batches … it is necessary to confirm whether the same answer quality can still be maintained when actual acceleration is achieved."

Q1: "Can the same set of evaluation requests be used to measure the quality and speed of answers simultaneously under dynamic service batching?"

Answer: Thanks for the constructive suggestion. Yes. Qwen3-30B-A3B serves a stratified panel of 4,384 questions from the eight paper tasks, replayed as Poisson-arrival requests with the paper's prompts and a fixed per-request seed, so accuracy and TPOT come from the same responses. We test QPS 1.68 and QPS 4.0, which is about 80% of Vanilla's capacity. Vanilla runs first and last, and the two runs differ by only −0.20 points (95% CI [−0.73, +0.33]). Rescored on this panel, the paper's fixed-batch predictions reproduce the Table 1 differences (k=1k=1: −19.96 vs. −20.23; (1,10): −0.94 vs. −0.70).

MethodQPS 1.68: TPOT speedupΔ acc. [95% CI]QPS 4.0: TPOT speedupΔ acc. [95% CI]
DelMoE (2,0)2.08×−4.04 [−5.05, −3.01]1.54×−1.00 [−1.84, −0.13]
DelMoE Near-sat. (2,2)1.85×−1.47 [−2.39, −0.55]1.40×−0.56 [−1.37, +0.28]
SERE (k=2)1.56×−9.92 [−11.18, −8.61]1.17×−0.70 [−1.50, +0.12]
XShare (k=2)1.16×−5.96 [−7.05, −4.95]0.93×−0.48 [−1.24, +0.32]

Under realistic load, the near-saturation point keeps quality while accelerating: at QPS 4.0 it is within noise of Vanilla (−0.56) at 1.40× TPOT, with 9.1 ms lower TTFT. Both DelMoE points are faster than every baseline at both loads, and at QPS 1.68 Near-sat. is also far more accurate than SERE (−9.92) and XShare (−5.96).

At QPS 4.0, SERE, XShare and Near-sat. are all within noise of Vanilla in accuracy. Slower methods hold more concurrent requests and therefore decode larger batches, which retain a larger shared expert union and favor accuracy. XShare's 0.08-point edge over Near-sat. comes with a 50% larger decode batch (79.8 vs. 53.2) and a TPOT above Vanilla's. DelMoE reaches comparable accuracy while decoding faster with smaller batches.

Revision: We add this experiment to Section 4.3 (Table 3), with the full protocol and table in Appendix C.3 (Table C11).

AuthorsOpenReview ↗

Response to Reviewer ibXZ (Part 2 / 2)

W2: "The trigger condition for weight correction is determined before supplementing experts. Some tokens may have no routes deleted in the end and may still have their weights reduced. Therefore, it is not clear whether the benefit comes from fixing the changes after deleting routes or from ordinary output scaling."

Q2: "With the option for supplementary experts enabled, using the same calibration method, how does the current correction compare to "only correcting tokens that are ultimately missing a route" and "uniformly scaling all tokens"?"

Answer: Thanks for the detailed question. The gain requires rectifying every token that the deletion rule marks, and it is not ordinary output scaling. Calibration follows the paper's protocol (offline, b=0b=0, six held-out scores). Rectifying only tokens that finally lost a route coincides with the paper's rule at b=0b=0, so it receives the same coefficient. Uniform scaling, calibrated separately, selects 0.80 on Qwen1.5 and 0.95 on Qwen3. Results are eight-task averages, paired per question:

RuleQwen1.5 k1b5Qwen3 k1b10
No rectification (s=1)47.0381.55
Paper rule (dtd_t set before supplementation)49.16 (s=0.85)82.01 (s=0.95)
Only tokens finally missing a route47.34 (−1.82, p=0.001)81.25 (−0.76, p=0.032)
Uniform scaling, own calibration48.52 (s=0.80; −0.64)81.27 (s=0.95; −0.74, p=0.025)
Uniform scaling at the paper's s48.93 (−0.23, p=0.69)— (same s)

Rectifying only tokens that finally lost a route loses most of the gain on Qwen1.5 and all of it on Qwen3, where it falls below native weights. The tokens W2 asks about, whose routes were all restored by supplementation, matter: rectifying them alone gives +0.49 and rectifying only the other affected tokens gives +0.31, but rectifying both gives +2.13. Uniform scaling is never better than the paper rule. With its own calibration it is 0.64 points lower on Qwen1.5 and 0.74 points lower on Qwen3 (p=0.025), and even at the paper's s it does not improve on Qwen1.5 (−0.23, p=0.69). The paper rule has the highest accuracy on both models.

Rectification is also not an output-temperature adjustment. Dividing the logits by 1.25 or 1.5 at s=1s=1 does not raise accuracy (−0.31 and −0.90) and stays 2.4–3.0 points below rectification (p≤0.0001). Rectification acts on the affected tokens' routed outputs inside the network.

Revision: We add this comparison and the output-temperature control to Section 4.4.3 (Table 7), with the full table in Appendix C.6 (Table C25).

Thank you again for your valuable feedback.

Thank you for the clarifications. Most of my questions have been addressed, and I will keep my current score unchanged.

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

The paper tries to evaluate the latency and accuracy trade-off of dynamic expert skipping during batched decoding in MoE LLMs and is s well written in my opinion. Weight access cost and execution cost can increase if different experts are activated by different requests during batched serving, and previous methods mainly reduce #active experts by redirecting tokens from skipped experts to keep ones and transfer gate weights. The authors are basically saying that this route replacement still executes the same number of token-expert routes and can hurt accuracy. Thus, they propose DelMoE to keep each token's top-k routes and reuse these experts across the batch with a small number of supplementary experts.

Critical strengths

The idea of removing out-of-set routes is simple and pretty intuitive. The benchmarks also provided convincing evidence to show the benefits.

Critical weaknesses

The gate-weight rectification, which is calibrated offline, brings relatively small gain.

Questions for authors
  1. Why is d_t defined w.r.t. the initial backbone P rather than the final expert set H? How much accuracy loss would that lead to?
  2. Could you please clarify why the mean TTFT increases when using your method on Qwen1.5 and Qwen3?
5 comments
AuthorsOpenReview ↗

Response to Reviewer UCQz (Part 1 / 2)

We sincerely thank the reviewer for the positive and encouraging feedback, and for the time spent on our paper. We are glad that you found the paper well written, the idea of removing out-of-set routes simple and intuitive, and the benchmark evidence convincing.

Your questions were precise and insightful, and they helped us strengthen the paper: they led to a new token-set comparison that clarifies the role of dtd_t, and to a prefill fix that removes the TTFT increase. We hope these results further strengthen your support, and we would be grateful if you would consider raising your score. Below, we provide detailed responses to all the comments:


W1: "The gate-weight rectification, which is calibrated offline, brings relatively small gain."

Answer: Thanks for raising this point. Among the gate-weight rules in Table 6, rectification is the only one that improves on native weights, and it does so at no inference cost: a single scalar per model, calibrated once offline. Across the nine operating points it raises mean accuracy from 66.38 to 66.96 and improves seven of them, whereas renormalizing the surviving weights loses 12.8 points (53.61). It improves all three Qwen1.5 points, with the largest gain of +2.13 at the Sat. point (49.16 vs. 47.03, paired p=0.0003).


Q1: "Why is d_t defined w.r.t. the initial backbone P rather than the final expert set H? How much accuracy loss would that lead to?"

Answer: Thanks for the precise question. Defining dtd_t on PP lets one coefficient transfer to every (k,b)(k,b), and defining it on HH loses most or all of the gain. PP depends only on kk, so the trigger is independent of the selector and of the supplement budget bb. One coefficient calibrated at b=0b=0, where H=PH=P, can therefore be reused at every (k,b)(k,b). Defining dtd_t on HH instead, i.e., rectifying only tokens that finally lost a route with the same coefficient, loses 1.82 points on Qwen1.5 (47.34 vs. 49.16, p=0.001) and 0.76 points on Qwen3 (81.25 vs. 82.01, p=0.032). This removes most of the gain on Qwen1.5, and on Qwen3 it is even below native weights (81.55). The reason is that tokens whose deleted routes were all restored by supplementation still benefit from rectification (new Table 7 in Section 4.4.3).

Revision: We add this comparison to Section 4.4.3 (Table 7).

AuthorsOpenReview ↗

Response to Reviewer UCQz (Part 2 / 2)

Q2: "Could you please clarify why the mean TTFT increases when using your method on Qwen1.5 and Qwen3?"

Answer: Thanks for pointing this out, which led us to an implementation fix. The increase came from CPU-side overhead in our prefill implementation, not from the method. Splitting TTFT with paired runs showed that queueing time was unchanged (within ±0.5 ms) and the whole increase was a longer prefill step (+2.7 to +6 ms). The router selection kernel adds under 0.4 ms and expert computation is not slower. Profiling traced the rest to Python work repeated in every layer, such as configuration lookups and argument checks, which CUDA Graph replay hides during decoding but not during eager prefill.

We now cache these invariant computations and run prefill through a streamlined forward. Greedy tokens and log-probabilities on fixed inputs are bit-identical to before for both models and both values of kk. Re-measured under the Appendix C.3 protocol (A800, 128/32 tokens, 5,000 requests, Vanilla run before and after each group), DelMoE's mean TTFT now matches or beats Vanilla's in all 18 settings. It is lower in 16, by up to 8.02 ms, and the other two (+0.25 and +0.15 ms) are within Vanilla's own run-to-run drift of 0.45 ms:

ModelQPSVanilla TTFT (ms)Δ k=1Δ k=2Δ Sat.
Qwen1.5838.89−1.47+0.25+0.15
Qwen1.51646.65−3.74−2.01−2.57
Qwen1.52454.18−6.48−2.84−3.68
Qwen3887.83−4.23−4.84−1.48
Qwen316106.50−7.75−8.02−3.39
Qwen324123.13−6.49−6.45−7.60

QPS 24 rows average three repetitions. With 1,024–2,048-token prompts, Qwen3 DelMoE TTFT is 7–9 ms lower than Vanilla. DelMoE targets decoding latency, and these results show that it also does not slow the first token.

Revision: We add the prefill fix and re-measured TTFT to Appendix A.2 and Appendix C.3 (Table C8).

Thank you again for your valuable feedback.

Thank you for your detailed response and for adding new experiments. I keep my score because the benefit of gate weight rectification still looks small, and it is my main concern.

Reviewer_vGB2OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

In batched decoding, the tokens in a batch together activate a large number of experts. This increases the latency of MoE models without necessarily improving accuracy. To address this, the paper proposes DelMoE, a training-free method. DelMoE builds a batch-shared expert set from each token's high-priority experts, adds a small number of supplementary experts, and removes all routes to experts outside this set. For the affected tokens, the remaining gating weights are multiplied by a scaling factor selected offline. The authors implement a matching execution backend in vLLM and evaluate the method on 3 MoE models and 8 benchmarks. At a comparable number of activated experts, DelMoE achieves higher average accuracy than existing batch-aware methods, and it speeds up decoding by up to 2.49× over the original model.

Critical strengths
  1. The method is implemented in vLLM, and the paper reports end-to-end decoding latency under different request rates rather than only theoretical compute savings. Because it requires no retraining, it can be applied directly to existing models, which gives it good practical value.
  2. The paper finds that, after experts are skipped, appropriately scaling down the remaining weights can work better than either keeping the original weights or renormalizing them. This finding may be useful for future work on expert skipping.
Critical weaknesses

The models and benchmarks used in the evaluation are relatively weak, which makes it hard to show that the conclusions generalize. All three models have fewer than 30B parameters and were released before 2026, and Qwen1.5-MoE in particular is clearly outdated. The evaluation does not include more recent or larger MoE models. Most of the eight benchmarks are also close to saturation. For example, the original models score around 90% on GSM8K, HumanEval, and BoolQ, so these benchmarks can do little to distinguish accuracy losses. In addition, the paper's own results show that the loss caused by route deletion is concentrated in multi-step generation tasks. On harder tasks, the method's accuracy may therefore be substantially lower than what is currently reported.

Questions for authors
  1. Would the conclusions still hold when the method is evaluated on stronger models and harder benchmarks (e.g., GPQA-Diamond, AIME, SWE-Bench)?
  2. Could the authors give a more thorough theoretical analysis of why scaling down the remaining weights helps?
2 comments
AuthorsOpenReview ↗

Response to Reviewer vGB2 (Part 1 / 2)

We really appreciate the reviewer for the detailed and constructive feedback. We are pleased to see that you found DelMoE practically valuable, as it is implemented end to end in vLLM and requires no retraining, and that you found scaling down the remaining weights to be a useful finding for future work on expert skipping.

In response, we added harder benchmarks, a newer and stronger model, and a theoretical analysis. We believe our response has addressed your concerns, and we respectfully ask if you're comfortable to re-consider the rating in light of these new results and revisions. Below, we provide detailed responses to all the comments:


W1: "The models and benchmarks used in the evaluation are relatively weak … Most of the eight benchmarks are also close to saturation … On harder tasks, the method's accuracy may therefore be substantially lower than what is currently reported."

Q1: "Would the conclusions still hold when the method is evaluated on stronger models and harder benchmarks (e.g., GPQA-Diamond, AIME, SWE-Bench)?"

Answer: Thanks for raising this important point. Yes. At the paper's operating point, DelMoE shows no significant accuracy loss on harder benchmarks or on a newer, stronger model.

On saturation, the ~90% scores cited are Qwen3's. The original Qwen1.5 scores 51.3 on GSM8K, 48.2 on HumanEval and 80.1 on BoolQ. MATH is far from saturated on both models (13.9 and 72.5), and so is BBH on Qwen1.5 (35.1). The tasks also separate methods clearly: at k=1k=1 on Qwen3, the Table 1 average ranges from 8.54 (Direct) to 62.48 (DelMoE).

Harder benchmarks on Qwen3-30B-A3B: We use the paper's operating point, DelMoE (k,b,s)=(1,10,0.95)(k,b,s)=(1,10,0.95), without re-tuning:

BenchmarkVanillaDelMoE (1,10)
GPQA-Diamond (198 questions)54.5551.01
AIME 2024+2025 (60 problems, 32k output)41/6040/60
MATH Level 5 (1,324 problems)49.4050.30

On GPQA-Diamond, the 3.5-point gap is 7 of 198 questions and is not significant (Appendix C.2), whereas SERE at a comparable budget loses 11.6 points. AIME needs a 32k output limit because at the paper's 2,048 about 77% of outputs are truncated for every method. Even with a 32k limit, SWE-bench Verified is dominated by patch-format failures for every method, including Vanilla (58%), so we test multi-step reasoning with GPQA-Diamond and AIME.

A newer, stronger model, Qwen3.5-35B-A3B (February 2026): We ported the DelMoE selector to vLLM 0.19, validated it against our CUDA implementation (Appendix C.2), and reused (k,b)=(1,10)(k,b)=(1,10) and s=0.95s=0.95 from Qwen3 without recalibration:

Qwen3.5-35B-A3BVanillaDelMoE (1,10)
Active experts UdecU_{\text{dec}} / routes NdecN_{\text{dec}}69.23 / 12822.01 / 66.95
Eight-task average87.9887.89
GPQA-Diamond (16k output)83.8483.84
AIME 2024+2025 (32k output)37/6039/60

With 68% fewer active experts and 48% fewer routes, DelMoE matches Vanilla on the newer model without recalibration.

Revision: We add these results to Section 4.2 (Table 2), with all baselines, operating points and protocols in Appendix B and Appendix C.2 (Tables C3–C4).

AuthorsOpenReview ↗

Response to Reviewer vGB2 (Part 2 / 2)

Q2: "Could the authors give a more thorough theoretical analysis of why scaling down the remaining weights helps?"

Answer: Thanks for the insightful question. We give an analysis with an explicit assumption and then show that experiments support it. Consider a token marked by dtd_t in a routed layer. Let KK be its routed output after deletion with s=1s=1, and let MM be Vanilla's routed output at the same position.

Proposition 1 (same input): On the token's own input, M=K+RM=K+R, where RR is the output of the deleted routes. The least-squares scale ⟨K,M⟩/∥K∥2=1+⟨K,R⟩/∥K∥2\langle K,M\rangle/\Vert K\Vert^2=1+\langle K,R\rangle/\Vert K\Vert^2 is at least one when KK and RR are aligned, so attenuation does not reconstruct the deleted output locally. We measure 1.005.

Proposition 2 (across layers): Earlier deletions perturb the token's input. Assuming K=M+eK=M+e with ee uncorrelated with MM, the least-squares scale is ∥M∥2/(∥M∥2+∥e∥2)<1\Vert M\Vert^2/(\Vert M\Vert^2+\Vert e\Vert^2)<1, the standard shrinkage of a noisy estimate. Against Vanilla's trajectory we measure 0.961 for tokens whose deleted routes were all restored (B) and 0.934 for tokens that still lack a route (C), with 95% intervals [0.946, 0.978] and [0.922, 0.950].

Implication for the rule: A B token has all its routes and would need no rectification from a local view, but its input is still perturbed, so Proposition 2 applies. This is why dtd_t is set before supplementation and covers both B and C.

Empirical support: Experiments on Qwen1.5 k1b5 (s=0.85s=0.85, eight-task average, paired per question) support this analysis. Rectifying only B gives +0.49, only C gives +0.31, and both give +2.13 (p=0.0003) over native weights. The gain survives greedy decoding (+1.53, p=0.007) and requires rectifying in both prefill and decode and in all layers rather than either half.

The gain is not an output-temperature effect: dividing the logits by 1.25 or 1.5 at s=1s=1 gives −0.31 and −0.90, 2.4–3.0 points below rectification (p≤0.0001). It is also not a length effect: the degenerate-output rate is essentially unchanged, and length change does not track correctness (r=−0.04).

Revision: We add the token-set comparison and the output-temperature control to Section 4.4.3 (Table 7), and the derivation and measurements to Appendix C.6.

Thank you again for your valuable feedback.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers for their thoughtful and constructive feedback. We are encouraged by their recognition of the key contributions and strengths of our work: the paper is well written (UCQz), reducing both active experts and token–expert routes is a clear and intuitive idea (UCQz, ibXZ), the comparison between route deletion and route replacement is convincing (ibXZ), and DelMoE is practically valuable (vGB2).

We have carefully addressed each comment and believe our responses resolve the reviewers' concerns. In our revised manuscript, we have incorporated the suggested experiments, additional analyses, and relevant clarifications to further strengthen our work. Below, we summarize the core contributions of our study, the updates to our experiments, and the in-depth discussions included in our revision.


Core Contributions of Our Work

  1. Dual-Sparse Route Deletion: DelMoE builds a backbone from high-priority native routes, reuses its experts across requests in a batch, and supplements the retained set. Deleting routes outside this set reduces both active experts and token–expert routes, without retraining.
  2. Routing-Aware Gate-Weight Rectification: A single coefficient per model, calibrated once offline, rectifies the surviving gate weights at no inference cost.
  3. Dual-Sparse Execution and Extensive Evaluation: Our custom vLLM CUDA backend skips deleted routes. Across three MoE models and eight benchmarks, DelMoE achieves higher average accuracy than batch-aware route replacement and up to 2.49× decoding speedup over Vanilla at comparable or higher accuracy.

Updates of Experimental Results during Rebuttal

  • Section 4.2 (Table 2) and Appendix C.2: Added GPQA-Diamond and AIME 2024–2025, and the newer, stronger Qwen3.5-35B-A3B, on which DelMoE matches Vanilla with 68% fewer active experts without recalibration (vGB2).
  • Section 4.3 (Table 3) and Appendix C.3 (Table C11): Measured accuracy and speed on the same requests under dynamic batching. At QPS 4.0, DelMoE (2,2)(2,2) stays within noise of Vanilla at 1.40× TPOT (ibXZ).
  • Section 4.4.3 (Table 7) and Appendix C.6 (Table C25): Compared the paper's rectification rule with rectifying only tokens that finally lost a route, uniform scaling, and output-temperature scaling. The paper rule has the highest accuracy on both models (UCQz, ibXZ, vGB2).
  • Appendix C.3 (Table C8): Re-measured TTFT with a streamlined prefill. DelMoE's mean TTFT now matches or beats Vanilla's in all 18 settings (UCQz).

Updates of In-Depth Discussions during Rebuttal

  • Section 3.2: Clarified why dtd_t is defined on the backbone set, so that one coefficient calibrated at b=0b=0 transfers to every (k,b)(k,b) (UCQz, ibXZ).
  • Appendix C.6: Added a theoretical analysis of why the rectification coefficient is below one (vGB2).
  • Appendix A.2 and C.3: Explained the source of the earlier TTFT increase and the prefill fix, whose outputs are bit-identical (UCQz).

We believe these additions and clarifications comprehensively address the reviewers' concerns and enhance the quality of our manuscript. All revisions are highlighted in blue for ease of reference. We look forward to the reviewers' favourable consideration and remain grateful for their valuable feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.