acceptodds
Under review as a conference paper at ICLR 2027

Target Correction for Diffusion Sampling and Speculative Decoding

Abstract

During LLM inference, control algorithms decide which tokens to resample for diffusion models and which request tokens to verify for speculative decoding. For the sake of efficiency and simplicity, these algorithms do not attempt long-horizon planning to hit a specified target—fidelity for diffusion sampling, or goodput for speculative decoding. This paper presents *target correction*, a technique for noninvasively imbuing simple control algorithms with target awareness. The approach (1) interprets the algorithm as myopically playing a betting game in which wealth multiplies according to random outcomes, (2) devises a target-aware utility function of wealth, (3) derives a target-aware value function from a continuous-time approximation of the game, and (4) uses that value function to modify the original algorithm. Our correction of the Entropy-Bound diffusion sampler in DiffusionGemma achieves 4% lower NFE at comparable or better fidelity. To illustrate how surgical target correction can be, we improve DSpark's goodput robustness under skewed workloads by up to 10% with just a single-line change.

Primary area:
foundation or frontier models, including LLMs
Keywords:
diffusion sampling, speculative decoding, inference, planning, stochastic control, betting

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_Ro8POpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on a specific and important problem: how inference-time controllers decide which tokens to process in LLM speculative decoding and diffusion language model sampling. These decisions involve trade-offs between computational efficiency and objectives such as throughput and generation fidelity, making the problem worth studying. The proposed target-correction framework is intuitive and supported by experimental results. However, I have some concerns about the experimental model selection: why not also evaluate on models such as LLaDA to further verify the generalizability of the method?

Critical strengths
  1. The problem addressed is specific and important, and the paper is properly structured.

  2. The proposed idea is straightforward and clear, requiring only relatively small modifications to existing algorithms.

Critical weaknesses
  1. Many dLLM papers use LLaDA for evaluation, whereas this paper only adopts DiffusionGemma for its diffusion sampling experiments. From a generalization perspective, evaluating additional models, potentially including different sizes of LLaDA, would help demonstrate the transferability of the method. Although the sampling mechanisms differ, the authors could also clarify why LLaDA was not selected and whether these differences limit the applicability of their approach.

  2. Although the paper includes DSpark and Entropy-Bound as two direct baselines, it does not sufficiently compare against compatible competing approaches or demonstrate that its core construction is more effective than simpler alternatives. In other words, the concern is not that the paper lacks baselines, but that it mainly demonstrates improvements over the original controllers without fully establishing the advantages of the proposed unified framework. It would be helpful to compare against simpler alternatives, such as different request-throughput weighting strategies for speculative decoding or a history-based penalty without the wealth term for diffusion sampling. Ablations isolating the roles of wealth, churn divergence, and terminal utility would also help clarify which components are responsible for the observed gains.

  3. The connection between the theoretical framework and the concrete algorithmic choices is not always sufficiently clear. The paper provides useful high-level intuition, but the logical chain from the motivating problem and modeling assumptions to the derivation of the value function and the final algorithmic modifications could be better established. As a result, it is sometimes difficult to understand why each design choice is necessary rather than simply a reasonable heuristic. I mean this paper makes me confusing.

Questions for authors
  1. Could the authors explain why LLaDA was not selected for the experiments? Are there particular differences in its sampling mechanism that would limit the applicability of the proposed method?

  2. Could the authors provide some qualitative examples of actual generations, particularly for the diffusion sampling experiments, to illustrate the differences between the original and corrected methods?

  3. The 4% NFE improvement limited, please show the calculating progress.

4 comments

Thank you, we agree with all this feedback about rigorous comparison with mechanistic insight. We updated our paper to improve in all these areas.

Q3: NFE Improvement. Great news: we substantially improved SEB. We now see 17% reduction in NFE (not just 4%). We accomplished this through a non-homogeneous terminal utility and a new mode churn statistic.

Q2: Generations. Please see page 23 for an interesting pair of real trajectories produced by EB and SEB on MMLU Pro Philosophy. You can see how EB gets stuck on an ultimately-irrelevant token. On this example, SEB improved NFE by 60%.

W2: Baselines. We now compare against four baselines (not just one): EB, CreditDecoding, SWD, and KLASS (on pages 9, 21 and 22). SEB improves upon all of them by more than 12% in NFE. Importantly, these new experiments provide interesting mechanistic insight. The methods based on KL penalties (SWD uses forward, KLASS uses reverse) increase NFE rather than decrease it. Using KL to measure mode churn has the unfortunate property of penalizing sharpening on a correct mode (which is what a denoising algorithm is supposed to do!) Our new SEB avoids this with a different DiD_i (equation 6 on page 7).

W1, Q1: LLaDA. There are two reasons we didn't use LLaDA. SEB is designed for uniform renoising diffusions as in DiffusionGemma. LLaDA 1 is a research model (4K context length); we focus on production codebases, where keeping code changes small is important. LLaDA 2.1's production (SGLang) sampler is much different (based on token-to-token editing).

If you want, we can try target-correcting the LLaDA 2.1 sampler. (After some analysis, we may have identified a linear myopia in it.) If so, please let us know soon, to give us time to implement and run this application. However, please also take into account the following point.

W3: Applying Target Correction. We added extensive new applications to KV cache eviction and chunked prefill. Target correction is not purely "plug and play" (there is indeed some technical work required to design the game and utility for each application), but we hope these additional examples offer more guidance to researchers and inference engineers.

Here is an overall summary for SEB. It averages across the 12 datasets and 7 hyperparameter values (16K+ runs per method).

SamplerStability / History SignalAccuracy (%)Mean NFE (↓\downarrow)Tokens / NFE (↑\uparrow)
Entropy-Bound (EB)Stateless (HiH_i only)59.7542.2713.93
SWDForward KL weight exp⁡(−λ KL(pi−∥pi))\exp(-\lambda\,\mathrm{KL}(p_i^-\|p_i))60.0642.7513.71
KLASSReverse KL gate KL(pi∥pi−)<ϵKL\mathrm{KL}(p_i\|p_i^-) < \epsilon_{\mathrm{KL}}59.6643.0813.45
CreditDecodingMulti-step argmax logit credit (α,β,γcd)(\alpha,\beta,\gamma_{\mathrm{cd}})59.5540.1514.56
Steady Entropy-Bound (SEB)Target-corrected wealth V′(Wi)V'(W_i) + mode churn DiD_i59.5035.0815.99

But even in individual hyerparameter-paired comparisons, SEB demonstrates much better mean NFE. Here is the more fine-grained comparison of SEB to KLASS. In this sweep, we matched KLASS' hyperparameters to SEB's γ\gamma in a strong manner consistent with KLASS' choices.

Budget γ\gammaKLASS Acc (%)SEB Acc (%)KLASS Mean NFE (↓\downarrow)SEB Mean NFE (↓\downarrow)KLASS Tok/NFE (↑\uparrow)SEB Tok/NFE (↑\uparrow)
0.00360.1659.7449.0737.7212.4215.27
0.0159.9560.3748.2136.8812.6115.51
0.0359.6159.4947.5136.5712.7615.69
0.1 (default)60.7959.9144.4536.0313.3215.89
0.360.2459.8242.6235.0813.6916.14
1.059.6658.8139.9633.8514.1516.50
20.057.2158.3929.7229.4416.5717.23
All59.6659.5043.0835.0813.4515.99

Thank you for the response and additional experiments. For W2, could the authors compare SEB against a simpler baseline that directly incorporates the proposed mode churn statistic into EB, without wealth-based correction? This would help clarify whether the additional wealth mechanism is necessary for the observed improvements. I will also consider the other reviewers' opinions before making my final decision.

Certainly! Recall, in SEB, the target-corrected effective entropy is: H~iSEB=ν⋅(Hi+αiDi)\tilde{H}_i^{\text{SEB}} = \nu \cdot (H_i + \alpha_i D_i) Wealth WiW_i appears in the terms ν\nu and αi\alpha_i, not the mode-churn statistic DiD_i. The most natural ablation of wealth is to just add DiD_i directly to the entropy, yielding "EB + Churn": H~iEB+Churn=Hi+Di\widetilde{H}_i^{\text{EB+Churn}} = H_i + D_i

However, as seen in the table below, EB + Churn does not reduce NFE. Without wealth, Di≥0D_i \ge 0 acts only as a one-sided penalty (Hi+Di≥HiH_i + D_i \ge H_i): even when a frontier token is repeatedly corroborated with Di=0D_i = 0, its entropy HiH_i is never discounted. Wealth-based target correction enables that.

BudgetEB Acc (%)EB + Churn Acc (%)SEB Acc (%)EB Mean NFEEB + Churn Mean NFESEB Mean NFE
0.00360.2060.2059.7450.8653.0437.72
0.0160.2061.4260.3747.1548.5336.88
0.0360.2960.5859.4944.5245.7936.57
0.159.9960.2459.9142.9543.2536.03
0.360.3760.0359.8241.1741.6935.08
1.059.2860.4158.8139.2439.1833.85
20.057.8957.6858.3929.9830.4129.44
All59.7560.0859.5042.2743.1335.08

Please let us know if you'd like to see any other ablations.

Reviewer_7nm5OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a target-correction framework with applications to speculative decoding and diffusion sampling. For diffusion sampling, SEB augments entropy-bound sampling with consecutive-prediction KL divergence and an accumulated wealth variable.

Critical strengths

They extensively measure performance of entropy-bound sampler.

Critical weaknesses

My main concern is the distinction between SEB and existing stability-based sampling methods. KLASS (https://arxiv.org/abs/2511.05664) already uses changes in token distributions to guide sampling, but the paper does not discuss or compare against it. The sampling settings differ, but the relationship is close enough to warrant an explicit comparison.

It is also unclear whether the accumulated wealth mechanism is necessary. A comparison between entropy-bound sampling with a KL-based stability criterion and the full SEB method would help isolate its contribution. Without this ablation, it is difficult to tell whether the additional complexity is useful.

The reported NFE reduction is approximately 4%. Given this marginal improvement, the comparison with simpler stability-based alternatives is particularly important. These concerns primarily relate to the diffusion sampling component.

Questions for authors
  • How does SEB differ from KLASS beyond the difference in sampling settings?

  • How much improvement remains when the accumulated wealth variable is removed and only the entropy and KL criteria are retained?

4 comments

Thank you for this insightful and valuable feedback - we agree with those weaknesses of the initial application to diffusion. Accordingly, we have massively improved the paper with better algorithms, expanded comparisons and new application domains.

W1. Comparison to existing methods. We substantially improved SEB, incorporating a different terminal utility and moving beyond KL divergence. We now see 17% reduction in NFE (not just 4%) over EB (page 9). The effectiveness of SEB vs. EB is visualized using real generations on page 23. We additionally compare to CreditDecoding, SWD, and KLASS (pages 21-22) and achieve 12-18% improvements over those.

Besides reduced NFE, SEB differs from KLASS in two key ways. Primarily, it now doesn't use KL to measure instability. Our new experiments elucidated an important point: methods based on KL (SWD and KLASS) increased, rather than decreased, NFE relative to EB. KL has the unfortunate property of penalizing sharpening on a correct mode (which is what a denoising algorithm is supposed to do!). Our new mode churn DiD_i (equation 6 on page 7) avoids this. Furthermore, KLASS is designed primarily for masked diffusion rather than uniform renoising (as in DiffusionGemma).

W2. Ablation. Removing state (and therefore the nonlinear value function on that state) turns SEB back into EB; that is the intentional design feature of target correction. Nonetheless, we agree that providing more evidence for the core mechanism of target correction is important. Accordingly, we added new applications to KV cache eviction (pages 20-26) and chunked prefill (pages 26-30).

Here is an overall summary for SEB. It averages across the 12 datasets and 7 hyperparameter values (16K+ runs per method).

SamplerStability / History SignalAccuracy (%)Mean NFE (↓\downarrow)Tokens / NFE (↑\uparrow)
Entropy-Bound (EB)Stateless (HiH_i only)59.7542.2713.93
SWDForward KL weight exp⁡(−λ KL(pi−∥pi))\exp(-\lambda\,\mathrm{KL}(p_i^-\|p_i))60.0642.7513.71
KLASSReverse KL gate KL(pi∥pi−)<ϵKL\mathrm{KL}(p_i\|p_i^-) < \epsilon_{\mathrm{KL}}59.6643.0813.45
CreditDecodingMulti-step argmax logit credit (α,β,γcd)(\alpha,\beta,\gamma_{\mathrm{cd}})59.5540.1514.56
Steady Entropy-Bound (SEB)Target-corrected wealth V′(Wi)V'(W_i) + mode churn DiD_i59.5035.0815.99

But even in individual hyerparameter-paired comparisons, SEB demonstrates much better mean NFE. Here is the more fine-grained comparison of SEB to KLASS. In this sweep, we matched KLASS' hyperparameters to SEB's γ\gamma in a strong manner consistent with KLASS' choices.

Budget γ\gammaKLASS Acc (%)SEB Acc (%)KLASS Mean NFE (↓\downarrow)SEB Mean NFE (↓\downarrow)KLASS Tok/NFE (↑\uparrow)SEB Tok/NFE (↑\uparrow)
0.00360.1659.7449.0737.7212.4215.27
0.0159.9560.3748.2136.8812.6115.51
0.0359.6159.4947.5136.5712.7615.69
0.1 (default)60.7959.9144.4536.0313.3215.89
0.360.2459.8242.6235.0813.6916.14
1.059.6658.8139.9633.8514.1516.50
20.057.2158.3929.7229.4416.5717.23
All59.6659.5043.0835.0813.4515.99

Thank you for the KLASS comparison.

It seems the revised SEB changes quite many things, e.g., stability measure and the terminal utility, and it remains unclear how much the wealth correction adds beyond mode churn alone. Given the extent of these changes, I maintain my recommendation and believe the revised method needs a fresh round of review in a future submission, since it seems like a new paper.

Great point on the role of wealth; sorry we didn't provide this ablation earlier. Recall, in SEB, the target-corrected effective entropy is: H~iSEB=ν⋅(Hi+αiDi)\tilde{H}_i^{\text{SEB}} = \nu \cdot (H_i + \alpha_i D_i) Wealth WiW_i appears in the terms ν\nu and αi\alpha_i, not the mode-churn statistic DiD_i. The most natural ablation of wealth is to just add DiD_i directly to the entropy, yielding "EB + Churn": H~iEB+Churn=Hi+Di\widetilde{H}_i^{\text{EB+Churn}} = H_i + D_i

However, as seen in the table below, EB + Churn does not reduce NFE. Without wealth, Di≥0D_i \ge 0 acts only as a one-sided penalty (Hi+Di≥HiH_i + D_i \ge H_i): even when a frontier token is repeatedly corroborated with Di=0D_i = 0, its entropy HiH_i is never discounted. Wealth-based target correction enables that.

BudgetEB Acc (%)EB + Churn Acc (%)SEB Acc (%)EB Mean NFEEB + Churn Mean NFESEB Mean NFE
0.00360.2060.2059.7450.8653.0437.72
0.0160.2061.4260.3747.1548.5336.88
0.0360.2960.5859.4944.5245.7936.57
0.159.9960.2459.9142.9543.2536.03
0.360.3760.0359.8241.1741.6935.08
1.059.2860.4158.8139.2439.1833.85
20.057.8957.6858.3929.9830.4129.44
All59.7560.0859.5042.2743.1335.08

We significantly enhanced our applications during the rebuttal; we agree it’s a lot to review! Still, we maintain that the scholarly contribution hasn’t changed that substantially. The primary content (target correction) has not changed at all. The paper always anticipated many applications to control in inference engines. Would you mind telling us what aspects of SEB / mode churn you still find unsettled as a reviewer?

Reviewer_K8VEOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes to equip greedy LLM inference control algorithms (DSpark for speculative decoding, the EB diffusion sampler for DiffusionGemma) with target correction, a technique that allows these algorithms to have a sense of their progress toward the final objective while making local decisions, and experiments show improvement on final metrics.

Critical strengths

To the best of my knowledge, framing LLM inference controllers as myopic bettors in a multiplicative wealth game, and correcting them with a target-aware utility, is novel and interesting. The paper shows gains in two quite different settings (speculative decoding for autoregressive LLMs and sampling for diffusion LLMs), which suggests the idea is general. I also find it nice that the existing baselines (DSpark and EB) come out exactly as the linear-utility case (V(W) = W). This makes target correction a principled, minimal modification of existing controllers rather than a new algorithm.

Critical weaknesses
  1. Lack of ablations:
  • For comparisons with DSpark, evaluation is limited to one model (Gemma 4 12B), one batch size and α=1\alpha=1. Only relative goodput is reported.
  • For comparisons with EB, evaluation is limited to topk=32.
  1. Figure 4 shows that the SEB gains are modest on average (~4% NFE, +0.5 accuracy) and can vary a lot across benchmarks and γ\gamma values. Are these results averaged over multiple seeds? It's hard to assess how significant these results are.
Questions for authors
  1. Can the authors elaborate further on how the DStreak experiments were setup? How realistic are the workloads being experimented with?
2 comments

Thank you for the valuable feedback! We updated the paper with improved algorithms and a ~dozen new experiments.

W1a: Comprehensiveness. We expanded the DStreak experiments (page 6) to include 4 models (not just 1). Goodput is our main metric but we also added latency-attainment curves. We strengthened DStreak with a second-order term: now its performance is better than DSpark's not just in a specific regime, but in all settings. At tail latencies, it even beats DSpark in plain vllm bench serve, which doesn't involve a target SLO. This performs continuous batching at different requests/second, so it involves different batch sizes.

Q1: Realism. We couldn't evaluate our algorithm in a live production cluster (as DeepSeek did) but our experimental setup is realistic. SPEED-Bench is currently the best dataset for this evaluation. For each of the four models, w did three experiments (corresponding to rows). Each involved running with a sweep of hyperparameters (over the x-axis). We varied target SLO (first row) and fraction of high-entropy requests (second row). The third is standard vllm bench serve; we measure wall-clock time, which makes our gains seem less impressive (because the measurement doesn't isolate adaptive verification) but is as realistic as possible. These experiments are representative for a wide range of hardware; DStreak inherits DSpark’s adaptability in this respect since, at startup, they measure forward pass times at all possible sizes BB.

W1b, W2: SEB. We strengthened SEB so it no longer involves a top-kk approximation and takes self-conditioning into account. The gain is dramatic: we now see 17% reduction in NFE, not 4% (page 9). We added comparisons to three additional baseline algorithms (pages 21 and 22). Note that, since most of the algorithms share the γ\gamma hyperparameter, we do 7 properly paired runs/comparisons across the entire dataset. (16K+ paired comparisons, per metric, are used to assess one algorithm vs another). This all boosts confidence in the results well beyond any statistical uncertainty.

Ablations. Target correction is itself a small ablation of an original algorithm. To provide further support for the intrinsic mechanism of target correction, we added two entirely new applications to KV caching and chunked prefill. These are arguably even more compelling than our original applications.

Here is an alternative presentation of the DStreak results. We ran DStreak and DSpark in vLLM on the SPEED-Bench dataset across four models. First, we vary the request batch size (16, 32, 48, and 64 concurrent requests). Goodput is the percentage of requests that meet a target generation speed, and P90 TPOT is the 90th-percentile time per output token.

ModelBatch Size RRDSpark Goodput (%)DStreak Goodput (%)DSpark P90 TPOT (ms)DStreak P90 TPOT (ms)
Qwen 3 4B1675.5990.2316.5813.44
3264.3673.2426.0622.27
4861.5968.1635.8030.16
6458.4664.4545.2438.36
Qwen 3 8B1680.4792.5823.0018.51
3266.4175.3936.3630.06
4863.8070.3149.8441.65
6460.9466.2863.1753.09
Gemma 4 12B1669.9280.0833.6827.67
3258.6964.1653.5046.79
4856.1261.7273.0364.19
6453.9759.1192.4381.84
Gemma 4 31B1660.3569.9255.5349.48
3252.1558.8989.3681.68
4850.2656.05123.28116.35
6448.8354.23157.11150.27
Mean Across All 16 Runs--61.3769.6855.8748.49

Second, we fix the batch size at 16 and vary the fraction of easy (high-acceptance) prompts in the batch from 0% to 50%, averaged across all four models:

Easy (Low-Entropy) Request FractionDSpark Goodput (%)DStreak Goodput (%)
0%73.8375.34
10%74.5677.49
25%73.9379.79
50%69.2982.47
ModelDSpark Output Tok/sDStreak Output Tok/sDSpark Mean Decode Latency (s)DStreak Mean Decode Latency (s)DSpark P90 Decode Latency (s)DStreak P90 Decode Latency (s)DSpark P90 TPOT (ms)DStreak P90 TPOT (ms)
Qwen 3 4B166.68172.1763.6161.1083.4878.96327.27309.46
Qwen 3 8B131.96137.6479.5976.29102.9999.35403.69389.25
Gemma 4 12B103.99111.1691.0584.68127.49118.67499.81465.45
Gemma 4 31B46.1047.92207.08200.60300.56282.791178.551108.68
Mean Across All 4 Models112.18117.22110.33105.67153.63144.94602.33568.21


And here is an overall summary for SEB. It averages across the 12 datasets and 7 hyperparameter values (16K+ runs per method). But even in individual hyerparameter-paired comparisons, SEB demonstrates much better mean NFE.

SamplerStability / History SignalAccuracy (%)Mean NFE (↓\downarrow)Tokens / NFE (↑\uparrow)
Entropy-Bound (EB)Stateless (HiH_i only)59.7542.2713.93
SWDForward KL weight exp⁡(−λ KL(pi−∥pi))\exp(-\lambda\,\mathrm{KL}(p_i^-\|p_i))60.0642.7513.71
KLASSReverse KL gate KL(pi∥pi−)<ϵKL\mathrm{KL}(p_i\|p_i^-) < \epsilon_{\mathrm{KL}}59.6643.0813.45
CreditDecodingMulti-step argmax logit credit (α,β,γcd)(\alpha,\beta,\gamma_{\mathrm{cd}})59.5540.1514.56
Steady Entropy-Bound (SEB)Target-corrected wealth V′(Wi)V'(W_i) + mode churn DiD_i59.5035.0815.99

Comments on the paper

3 comments
AuthorsOpenReview ↗

New applications of target correction

We updated the paper with two new applications of target correction to KV cache eviction and chunked-prefill scheduling. We think these are exciting because:

  • the improvements observed are substantial (up to 6x KV-cache hit rate over LRU, and +18% request goodput over chunked prefill under decode contention), and
  • they improve not just small subsystems (diffusion and spec decoding), but core processes that affect essentially every inference request in every server configuration.

We hope you enjoy these applications. We think they substantially extend the audience for, and interest in, the paper.

Here are our primary results for target-corrected LRU KV cache eviction ("Walrus"). They are a series of vllm bench serve experiments, averaged across different runs having different query-per-second settings (1, 2, 4) on Mooncake agentic workload traces.

ModelLRU Hit Rate (%)Walrus Hit Rate (%)LRU Mean TTFT (s)Walrus Mean TTFT (s)LRU Output Tok/sWalrus Output Tok/s
Qwen 3.8 27B4.743.7113.080.213.016.9
Qwen 3.6 35B-A3B4.646.545.131.124.229.5
Gemma 4 26B-A4B10.239.718.814.335.939.5
LFM 2.5 24B-A2B4.845.827.721.731.134.5

And these are the results for the application to chunked prefill (CP -> RCP). They are also a series of vllm bench serve experiments on Mooncake agentic workload traces.

ModelCP Goodput (%)RCP Goodput (%)CP Median TTFT (s)RCP Median TTFT (s)CP Mean E2EL (s)RCP Mean E2EL (s)
Qwen 3.8 27B1.717.8429.3315.4491.7421.1
Gemma 4 31B5.117.6851.5613.8968.1833.1
Qwen 3.6 35B-A3B11.130.1204.8143.3238.0225.5
LFM 2.5 24B-A2B15.838.8130.4101.2154.1150.8
Mean Across All 4 Models8.526.4404.0293.4463.0407.6
AuthorsOpenReview ↗

Discussion conclusion

We thank the reviewers for their valuable feedback. Reviewers initially found our target correction framework interesting, but felt the experimental gains in our two applications were marginal and did not involve sufficient comparisons to previous algorithms/models (e.g. KLASS, LLaDA). In response, we substantially enhanced the applications by making better use of target correction, and ran multiple new experiments with many more baseline algorithms/models. We also added two new applications of target correction.

The main unresolved question is whether these improvements and additions constitute a "new paper". We indeed made a concerted effort to address reviewer concerns. But the main material on target correction did not change, so in our view the paper remains fundamentally the same. We continue to anticipate further new applications of target correction.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.