acceptodds
Under review as a conference paper at ICLR 2027

PIG: PARALLEL INCONSISTENCY GUIDANCE FOR ACCELERATING DIFFUSION LANGUAGE MODELS

Abstract

Diffusion language models accelerate generation by decoding multiple masked positions per forward pass. However, accuracy is known to drop when parallel decoding trajectories deviate from their sequential counterparts, which are slow and expensive "oracles" that adhere to the joint distribution of language. In this paper, we propose *Parallel Inconsistency Guidance* (PIG), a new paradigm that uses a trained auxiliary model to directly estimate the inconsistency between parallel and sequential generation from a single model forward pass. During inference, masked positions are traversed by decreasing confidence, and tokens are added to the parallel decode set one-at-a-time until the total estimated inconsistency between the parallel and sequential generation exceeds a threshold. Surprisingly, it suffices to train a lightweight M parameter auxiliary model on a small mixture – drawn from OpenCoder, GSM8K-train, MATH-train, and MBPP-train. For LLaDA-8B Instruct and NemotronDiffusion-3B Instruct, PIG extends the accuracy-TPF frontier on GSM8K, MATH500, MBPP, HumanEval, and MMLU, when compared to the industry standard Fast-dLLM and measures favorably on LLaDA-8B Instruct against DAPD, KLASS, EB-sampler, and other learned unmasking policies.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Diffusion, DLM, Diffusion Language Models, Parallel Decoding, Generative Models, Discrete Diffusion

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_cvf7OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This work proposes the method PIG. It decides which masked positions a diffusion LM commits in parallel by predicting sequential–parallel inconsistency: whether a parallel commit would differ from one-at-a-time greedy decoding. An approx 2M-parameter adapter on the frozen backbone predicts pairwise entropy finite differences from one forward pass.

Offline tables, fitted by sequential replay on GSM8K-train and OpenCoder, turn confidence plus aggregate coupling into a mismatch probability. PIG.INT commits the confidence-ordered prefix whose summed risk stays less than ϕ\phi. PIG.EXT widens the round until the predicted benefit of waiting exceeds θ\theta. On LLaDA-8B-Instruct and Nemotron-Diffusion-3B (GSM8K, MATH-500, MBPP, HumanEval; MMLU on Nemotron), PIG improves the accuracy vs tokens-per-forward (TPF) frontier over confidence thresholding, most at high parallelism.

Critical strengths
  1. Large gains over confidence thresholding at high parallelism, under a paired, fully tabulated protocol. For example, at 10 TPF on LLaDA PIG.INT gains +19.1 points on GSM8K and +11.4 on MATH-500, and on Nemotron PIG.EXT gains +27.8 on MBPP at 5 TPF (Table 2). These gaps are well outside the reported Wilson intervals.

  2. The mismatch estimate is validated during live decoding. Realized/predicted mismatch is 0.83–0.92, and realized per-round regret never exceeds 0.91φ (App. D.4).

Critical weaknesses
  1. In my understanding, the gains are not shown or cant be attributed to come from the learned predictor.

    • The coupling μ^i(S)\hat\mu_i(S) (Eq. 5) sums non-negative terms over every admitted token, so the table's coupling axis partly just tracks how many tokens are already in the round.
    • The only adapter ablation measures estimator quality (Fig. 6), not decoding. No confidence-only cumulative budget is tested, and EB-Sampler is not swept on the same setup.
    • So the improvement over Fast-dLLM may come from the cumulative-budget rule rather than the adapter. The other baselines don't settle this:
      • EB-Sampler and KLASS are single cells copied from a setup whose Fast-dLLM is 4-14 points weaker (App. D.7).
      • DAPD is a single-schedule re-implementation about 12 points below its published HumanEval accuracy (App. D.5).
      • DEMASK is not compared.
  2. "Acceleration" is measured only in TPF, and the headline claims exceed the paper's own tables.

    • There are no wall-clock results, although the adapter adds 12.8 ms per round (App. B). LLaDA TPF also counts EOS padding.
    • "PIG helps globally on the whole frontier" (Sec. 5) is contradicted by Nemotron GSM8K. There, Fast-dLLM's 0.896 at 2.56 TPF beats PIG.INT's 0.854 at 2.52 (Table 8).
    • The 1.7×1.7\times headline drops to 1.27×1.27\times if the MBPP reference moves by one problem (0.396 vs 0.394, Table 6). Some "matched-accuracy" ratios are taken at lower accuracy (HumanEval 1.29×1.29\times at −1.8-1.8 points).
Questions for authors

I have the following questions for the authors.

  1. What is the decoding frontier for PIG.INT with a confidence-only table (refitted), and for EB-Sampler/KLASS swept on your setup? If these match PIG.INT, then i think that the adapter cannnot be the source of the gains.
  2. I am curious what is the end-to-end wall-clock time at matched accuracy vs Fast-dLLM, including adapter, attention-map extraction and selection? What is LLaDA TPF excluding padding?
  3. If i understnad correctly, Prop. A.8 is no tighter than the trivial bound Rtotal≤N−NFER_{\text{total}} \le N - \text{NFE} for θ≥1\theta \ge 1, which covers most of the LLaDA sweep. Its non-negativity premise also fails for the deployed p^H\hat p_H (Fig. 10: 0.138 vs 0.287). Can you give a non-vacuous version, or scope the guarantee claims?
  4. Algorithm 2 returns S=∅S = \emptyset when the top candidate's risk exceeds φ\varphi (e.g. 0.021>0.0150.021 > 0.015, Fig. 8). Is the first token force-committed, and does Prop. A.4 still hold?

Minor:

  • Nemotron BL=64 results (Table 2, Fig. 5) are not tabulated.
  • App. B.1 says "nine" scalar inputs but lists eight, and "232×24≈3,300232 \times 24 \approx 3{,}300" is arithmetically wrong.
  • is Sec. 3's ordering condition is reversed?
8 comments

We sincerely appreciate the reviewer's careful evaluation and constructive feedback.

In response, we conducted the following additional experiments:

  • (i) To confirm PIG’s gains on the Accuracy-TPF frontier, we ran EB-Sampler and KLASS under our own harness directly (which is also Fast-dLLM’s harness) rather than using past paper’s results. [W1.2, W1.3, Q1]
  • (ii) To study the marginal effects of each of PIG’s components, we ran a training-free version of PIG and PIG with confidence as the sole feature for sequential-parallel inconsistency. [W1.2, Q1] This was intended to confirm the marginal gain of each of PIG’s components.
  • (iii) We ran a wall-clock experiment to evaluate the overhead of running PIG compared to native baselines. [W2.1, Q2]
  • (iv) We ran a preliminary ablation study on the adapter model sizes. [W2.1] We additionally addressed misunderstandings with regards to the prefix’s aggregate coupling representation [W1.1], the choice of metrics used for reporting performance [W2.3, Q1], and the interpretability of the theoretical analysis [Q3, Q4].

We would be happy to clarify any remaining questions and make every effort to conduct additional experiments if needed. The code is also provided to ensure reproducibility of our results and overall soundness of the paper.

To ground discussion, we first note references to Fast-dLLM [1], EB-Sampler [2], KLASS [3], and DAPD [4].

[W1] Effect of learned predictor

We justify that the gains can be attributed to the learned predictor in the following points.

[W1.1] Does the coupling just count tokens?

We wish to clarify that the table's coupling axis does not track how many tokens are already in the round. μ^i(S)\hat{\mu}_i(S) aggregates the estimated first-order effect of each member of SS on token ii. At most, the total number of tokens in a round can only track this in expectation. In Section 4.1 of the paper, we explain that inconsistency is governed by the slack in the top-1 vs top-2 predicted probabilities and by the size of the perturbation SS applies to it. Our feature μ^i(S)\hat{\mu}_i(S) approximates this total perturbation up to first order and is therefore relevant to the regret estimation.

[W1.2] No confidence-only cumulative budget, estimator-only ablations.

There are many ablations that are highly relevant to decoding, including Fig 6. Specifically:

  • Fig. 6: per-edge, aggregate and local-curvature Pearson correlation per held-out benchmark
  • Fig. 7: estimate against exact values, and the edge mass recovered.
  • Figs. 8–10: the calibrated tables p^ν\hat{p}_{\nu} and p^H\hat{p}_H, and the corpus transfer of p^ν\hat{p}_\nu.
  • Fig. 12: PIG's decoding width against local curvature along the decode.
  • Fig. 13: domain transfer to MMLU without refitting.
  • Figs. 14–15: realised against predicted regret on the decoders' own trajectories.

These give fine-grained analyses of decoding performance beyond surface-level accuracy or TPF metrics.

For this rebuttal, we have also added comparisons against EB-Sampler and KLASS sweeps, and a training-free PIG that runs the PIG.INT walk with inconsistency estimated as one minus confidence, and we have also run experiments with inconsistency measured from confidence alone.

We have updated the manuscript to show these in Fig 23 App. F. Notably, the training free PIG matches or slightly edges out against EB-Sampler (runs under our harness, code cloned from DAPD’s repo).

[W1.3] Baselines.

Our results in [W1.1] and [W1.2] show that the gains come from both the adapter and the cumulative budget rule: both components contribute.

Moreover, our implementation is sound: our harness matches Fast-dLLM's and our numbers are consistent with Jazbec et al. (ICML 2026 oral), and our EB-Sampler, KLASS, and DAPD evaluations used DAPD's released codebase. We suspect the gap in DAPD's reported HumanEval numbers is attributable to sensitivity to attention patterns, possibly due to different prompt setups. Finally, DEMASK does not have a public implementation.


References

[1] Wu et al., “Fast-dllm: Training-free acceleration of diffusion llm by enabling kv cache and parallel decoding” (ICLR 2026).

[2] Ben-Hamu et al., “Accelerated Sampling from Masked Diffusion Models via Entropy Bounded Unmasking”, NeurIPS 2026. (Notes: EB-Sampler)

[3] Kim et al., “KLASS: KL-Guided Fast Inference in Masked Diffusion Models”, NeurIPS 2025.

[4] Kim et al., “DAPD: Dependency-Aware Parallel Decoding via Attention for Diffusion LLMs”, ICML 2026.

[5] Jazbec et al. “Learning unmasking policies for diffusion language models”, ICML 2026.

[W2] Measure of acceleration

We next justify our metrics for acceleration and break W2 down into the following points:

[W2.1, Q2] Wall-clock.

Prompted by your review, we timed every operating point of every lane on LLaDA-8B-Instruct, with 64-token blocks, a 256-token canvas and batch size 1, on one H100 GPU with no other process (10 problems per point). Per round, PIG adds 4.0–4.3 ms for the adapter and 0.2–0.7 ms for the selection. That is a 9–10% wall-clock overhead against native selection. However, since the adapter's arithmetic is under 0.1% of a backbone forward (≈0.6 GMAC per round); the measured ~4 ms is dominated by kernel-launch latency in our unoptimised fp32, eager-mode implementation, so a fused implementation (e.g., CUDA graphs, bf16, selection on the GPU) should reduce the overhead further.

Therefore, further optimising PIG's feature extraction and selection code would reduce the overhead. In fact, preliminary experiment suggests that halving the adapter size would still maintain performance:

adapterparametersper-edge correlationaggregate coupling correlation
width 960.95M0.76–0.800.93–0.97
width 192 (paper)2.17M0.76–0.810.92–0.97

We will include further discussion of this in the revision. We also provide an accuracy-TPF frontier plot without padding calculated as (256 - Number of Pads) / Number of Forwards (see Fig 24 App. F). Our findings suggests that excluding pads does not alter our claims, as the inclusion of pads would have affected every baseline anyways.

[W2.2] Contradictory claims.

Our statement is not contradictory because we do not claim best performance at every configuration. However, thank you for pointing out an opportunity to improve clarity, and we will update the wording accordingly: "PIG improves the frontier at medium and high parallelism, with the largest gains as decoding becomes more parallel."

[W2.3] The 1.7× headline and matched accuracy.

The reporting is due to the granularity of our hyperparameter sweeps, which either matches or is finer than existing baselines. In more detail, Table 1 reports the fastest measured PIG cell whose accuracy reaches the baseline's most accurate cell (App. D.7). This is the same matched-accuracy speed-up convention as Fast-dLLM, EB-Sampler, and KLASS. This is why we additionally report full accuracy–TPF frontiers (Fig. 5, App. E) rather than single operating points.

[Q1] confidence-only table, EB-Sampler and KLASS on our setup.

Prompted by your review, we ran on LLaDA-8B-Instruct with 64-token blocks (Fig. 16's setting), on all four benchmarks:

  • PIG.INT with its table collapsed onto confidence;
  • the training-free unconfidence budget;
  • EB-Sampler and KLASS, both swept;

Results are in Fig 23 App. F for convenience.

We also provide an accuracy-TPF frontier plot without padding calculated as (256 - Number of Pads) / Number of Forwards (see Fig 24 App. F).

[Q3] Theoretical analysis.

We will scope the claim to clarify that the theorem is intended to be non-vacuous for θ≤1\theta \leq 1. More broadly, Prop. A.8. is intended to provide coarse regret scaling laws over the decode trajectory, and so may not have interpretations at all thresholds. We'd also like to point that our depth of theoretical analysis is rare compared to related works, and regret-based dynamics are notoriously difficult to analyze even with significant assumptions.

[Q4] The first candidate.

Yes, the most confident candidate is always committed, to ensure progress. Prop. A.4 still holds because that token's inconsistency is exactly zero: the sequential reference's first step reads the same canvas, so xs1⋆=x^s1x^\star_{s_1} = \hat x_{s_1} and R(S)R(S) is unchanged. We will correct Algorithm 2 to initialise S←{s1}S \leftarrow \{s_1\}. Thank you for pointing out this opportunity to improve clarity.

Minor Fixes

Thank you for noticing these, all fixed in the revision:

  • Nemotron BL=64 frontier tables are added.
  • App. B.1: nine scalars is correct, but the structural class is four-way (decoded, committable, outside the active block, canvas boundary), and the list now says so.
  • "≈232×24" becomes "≈136×24 ≈ 3,300", since there are 136 eligible receivers per state on average.
  • The ordering condition in Sec. 3 now reads i≤ji \le j.

References

[1] Wu et al., “Fast-dllm: Training-free acceleration of diffusion llm by enabling kv cache and parallel decoding” (ICLR 2026).

[2] Ben-Hamu et al., “Accelerated Sampling from Masked Diffusion Models via Entropy Bounded Unmasking”, NeurIPS 2026. (Notes: EB-Sampler)

[3] Kim et al., “KLASS: KL-Guided Fast Inference in Masked Diffusion Models”, NeurIPS 2025.

[4] Kim et al., “DAPD: Dependency-Aware Parallel Decoding via Attention for Diffusion LLMs”, ICML 2026.

[5] Jazbec et al. “Learning unmasking policies for diffusion language models”, ICML 2026.

Thank you for the detailed rebuttal and for the considerable effort behind the new experiments. I appreciate that Fig. 23 now includes the ablation I asked for, along with EB-Sampler and KLASS run in your own harness. It answers my main question: the adapter does contribute on GSM8K and MATH-500 at high parallelism. Fig. 24 also settles my concern about padding, and the clarification on the first candidate in Algorithm 2 is correct.

Also at this stage (closer to deadline), I am not asking for any new experiments. I would only appreciate it if you could clarify the remaining concerns below, which are about the existing results, the claims in the text, and the theory.

  1. First, Fig. 23 also shows that a large share of the improvement comes from the training-free budget itself. On GSM8K at 10 TPF, the confidence-only table already recovers roughly 70% of PIG.INT's gain over Fast-dLLM, and on HumanEval it matches PIG.INT. I think the paper's framing should reflect this. I was also unsure about the "knapsack refinement" label on the PIG.INT cells in Fig. 23/24, since Sec. 4.1, Algorithm 2, App. D.4 and pig/rules.py all describe a prefix walk that stops at the first candidate exceeding φ\varphi. Could you confirm whether this is a labelling error?

  2. On acceleration, if i undertand correctly, the reported 9-10% per-round overhead means the matched-accuracy ratios in Table 1 translate to only about 1.031.03-1.05×1.05\times in wall-clock time on GSM8K and MATH-500. The 1.7×1.7\times headline rests on a one-problem difference in the reference (198 vs 197 of 500). The revised wording is still contradicted by Nemotron GSM8K (Tables 2 and 8), and Table 1 and the abstract still rely on the copied EB-Sampler and KLASS numbers. I would really appreciate your view on how these claims will be stated in the revision.

  3. For theory, restricting Prop. A.8 to θ≤1\theta \le 1 does not quite resolve the issue. The bound is still no better than N−NFEN - \text{NFE} at θ=1\theta = 1, and its non-negativity premise does not hold for the deployed p^H\hat p_H (Fig. 10). The revised D.4 also says both quantities exceed θ\theta at every point, while reporting 0.99-2.00 at θ=1.5\theta = 1.5. Could you kindlyy clarify how the guarantee should be read for the deployed decoder? Finally, the Nemotron BL=64 tables mentioned in the rebuttal do not yet appear in the revision.

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

This work learns to estimate whether tokens generated in parallel would disagree with sequential decoding. It uses this signal to decide which tokens to commit together, aiming to reduce decoding steps while preserving accuracy.

Critical strengths
  1. I like the idea of predicting whether tokens can be generated together. The idea is interesting and the motivation is reasonable.
  2. The results on two models are encouraging.
Critical weaknesses
  1. PIG.INT and PIG.EXT work better in different settings. It is not clear how to choose between them on a new task.
  2. Predictor size is not systematically studied, so the trade-off between prediction quality and overhead remains unclear.
Questions for authors
  1. How much wall-clock speedup remains after including the predictor and selection overhead.
  2. Could a predictor trained on a larger backbone help a smaller model, perhaps with some adaptation?
2 comments

We sincerely appreciate the reviewer's careful evaluation and constructive feedback.

In response, we conducted additional experiments to study the optimal model size for the adapter [W2] and measure the wall-clock overhead [Q1]. We also discuss the optimal rule between our two proposed algorithms [W1] and how PIG can transfer from larger model backbones into smaller model backbones. We would be happy to clarify any remaining questions and make every effort to conduct additional experiments if needed.

[W1] Choosing between PIG.INT and PIG.EXT.

  • Conveniently, both rules read the same adapter and table, so choosing one per target TPF costs nothing. However, thank you for pointing out the opportunity for good selection rules. Specifically, across tasks, PIG.INT is the more accurate lane up to about 14 TPF on LLaDA-8B-Instruct (64-token blocks), and PIG.EXT from about 14 TPF on. For Nemotron, PIG.EXT is preferable from about 3 TPF on the code tasks. We believe that PIG.EXT is more performant at high TPF because it tracks the curvature of the decode trajectory (see App. D.2). Nevertheless, a principled single selection rule, for example, a Lagrangian combination of the two objectives, is a promising direction, and we will discuss it in the revision.

[W2] Predictor size.

  • We chose the current architecture because it predicts the statistics the decoders consume: a per-edge Pearson correlation of 0.76–0.81 and an aggregate coupling correlation of 0.92–0.97 on held-out test states (Fig. 6, App. B). An earlier model, which reads attention and confidence features only, underperforms it (per-edge correlation of 0.61–0.70, aggregate coupling correlation of 0.86–0.91) while costing more per round (App. B). Following the reviewer's suggestion, we performed a preliminary sweep of the width of the current architecture, with the same data and recipe:

    adapterparametersper-edge correlationaggregate coupling correlation
    width 960.95M0.76–0.800.93–0.97
    width 192 (paper)2.17M0.76–0.810.92–0.97

Surprisingly, halving the width had little effect on the feature extractor’s performance! This suggests that our predictor could be even smaller! We will add the sweep, further architecture ablations and a discussion of our choices to the revision.

[Q1] Wall-clock.

  • In response, we timed every operating point of every lane on LLaDA-8B-Instruct, with 64-token blocks, a 256-token canvas and batch size 1, on one H100 with no other process (10 problems per point). Per round, PIG adds 4.0–4.3 ms for the adapter and 0.2–0.7 ms for the selection. That is a 9–10% overhead against native selection. Moreover, the adapter's arithmetic is under 0.1% of a backbone forward (~0.6 GMAC per round); the measured ~4 ms is dominated by kernel-launch latency in our unoptimised fp32, eager-mode implementation, so a fused implementation (e.g., CUDA graphs, bf16, selection on the GPU) should reduce the overhead further. We will add a more thorough wall-clock overhead discussion in the revision.

[Q2] Transfer from a larger backbone.

  • This is an interesting idea. The adapter reads the backbone's own hidden states and predicts that backbone's own sequential trajectory. A predictor taken from a larger model would therefore need a map between hidden spaces of different widths. It would also have to cope with sequential trajectories the smaller model may not generate natively. In practice, fitting directly on the target backbone is cheap, requiring only about 1 H100 GPU hour of training once the data is curated (App. B); and this adapter is observed to generalize across different tasks.

Thank you for the efforts. The additional experiments have addressed most of my concerns, and the explanation regarding Q2 has also addressed my question. I'm satisfied with the responses and will keep my score.

Reviewer_9B7JOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Existing masked‑DLM parallel decoders select positions using token confidence or attention‑based dependencies, yet parallel commits cause sequential‑parallel inconsistency, where parallel predictions deviate from true sequential autoregressive outputs.

This paper presents PIG (Parallel Inconsistency Guidance). A small ~2M‑parameter offline‑trained adapter estimates pairwise entropy interactions from one DLM forward pass. Two decoding variants are built upon it:

• PIG.INT: limits within‑round inconsistency by a regret budget, adding confidence‑ranked candidates until predicted mismatch risk is too high.

• PIG.EXT: look‑ahead variant that stops expanding commits if deferring tokens would lower future errors.

Critical strengths
  1. Novel optimisation target: Instead of confidence or pairwise dependency, optimises for inconsistency vs sequential decoding, directly measuring the core failure mode of parallel DLM generation.

  2. Low compute overhead: Tiny ~2 M‑parameter adapter, works from a single backbone forward pass, no extra DLM evaluation at inference; only lightweight table look‑ups.

  3. Two complementary practical decoders: PIG.INT works well at lower‑speed high‑accuracy regimes; PIG.EXT shines for high‑throughput aggressive parallelism.

Critical weaknesses
  1. Offline data‑dependent components: Requires offline harvest of decoding states and sequential‑replay labels to train the adapter and calibration tables; per‑backbone fitting is mandatory, cannot zero‑shot apply to arbitrary new DLM.

  2. Approximate first‑order estimation: The finite‑difference estimator truncates higher‑order interaction terms; cross‑round prediction can over‑estimate future‑round benefits under large θ budgets.

  3. Relies on confidence ordering: candidate admission strictly follows confidence ranking; cannot re‑order masked positions, limiting solution space.

Questions for authors
  1. Can higher‑order interaction terms be incorporated into the inconsistency estimator while keeping inference cost low?

  2. Can we eliminate per‑model offline fitting, enabling zero‑shot PIG for arbitrary pre‑trained diffusion‑LM backbones?

  3. Would decoupling from confidence ordering (re‑rank positions by predicted inconsistency) bring further TPF improvements? What would the overhead be?

2 comments

We sincerely appreciate the reviewer's careful evaluation and constructive feedback.

In response, we conducted additional experiments that support our choice of confidence ordering [W3, Q3] and address the gains of training the adapter against a training-free version of our algorithm [W1, Q2]. We also discuss the incorporation of higher-order interactions into the framework [W2,Q1], and the training efficiency of the method compared to other baselines [W1]. We would be happy to clarify any remaining questions and make every effort to conduct additional experiments if needed.

[W1, Q2] Offline data-dependent components and training-free variant of PIG.

  • While our method requires fitting an auxiliary model, it is significantly more efficient than other learned unmasking policies, such as Jazbec et al [1] (ICML 2026 oral). In particular, they train a policy with reinforcement learning per benchmark (e.g. different trained policies for math, code, etc.) and per speed–up level. In contrast, we train the PIG adapter once per backbone and report favorable task transfer on unseen tasks like MMLU while Jazbec et al. observe a decrease in performance for unseen tasks. At matched block geometry, PIG matches their policies in the 2–8TPF band where those are reported (App. D.6). Our auxiliary model is therefore considerably more training efficient for comparable speed-ups.

  • To expand upon Q2, on a training-free variant of PIG: We find that a training-free variant of PIG compares favorably against other training-free methods, but underperforms the adapter variant, which underscores the benefits of offline training.

  • To back these claims, we implemented a training-free version of PIG based solely on confidence. We find that one minus confidence (the unconfidence) is naturally correlated with inconsistency (Pearson correlation 0.6) and use it directly as a training-free proxy. We provide a sweep of our algorithm’s thresholded regret lanes on LLaDA-8B-Instruct at Block size 64. We find that on Math500, GSM8K, HumanEval and MBPP, the training-free PIG variant compares favorably against Fast-dLLM and EB-Sampler, but underperforms against trained-PIG (see Fig 23 App. F).

  • We will include these sweeps alongside an accompanying discussion to the revision.

[W2,Q1] Approximate first-order estimation.

  • Fortunately, while the first-order approximation is a valid concern, we do not observe any significant issues in practice.

  • In particular, the realised true per-round regret used in our algorithm is observed never to exceed the specified regret threshold (see App D.4 Figure 14b), meaning that the regret prediction is strong enough such that we do not over-commit tokens when it is not safe to do so. Moreover, the higher-order interactions are implicitly used. Since we use first-order finite differences as features to a predictor for sequential-parallel inconsistency, the higher-order interactions are absorbed into it in expectation.

  • In general, there are major challenges for going beyond first-order estimation, in particular the need to track exponentially more terms at worse signal-to-noise ratios. Nevertheless, incorporating higher-order terms efficiently remains an interesting direction for future work.

[W3, Q3] Relies on confidence ordering.

  • We use confidence-based ordering for two reasons: (1) We specifically attempt to track a confidence-based sequential decoding order, and (2) it is an efficient search strategy for an otherwise NP-hard problem of selecting candidate subsets.

  • In particular for (1), we find that the frozen confidence order we use, and the re-ranked order from the sequential decode largely select the same tokens anyway: 88–95% of the positions a round commits are among the next |S| positions committed by the re-ranked sequential decode (LLaDA-8B-Instruct, BL = 64, four benchmarks, 50 problems each, φ∈[0.1,1]\varphi \in [0.1, 1]).

  • For (2), selecting candidate subsets is NP-Hard (App. A), and so confidence ordering gives an efficient and effective subset search heuristic. To check that the confidence ordering is not limiting, we implemented a more sophisticated knapsack-style implementation which enables different token orderings. We find that at significant wall-clock overhead, this only achieved a ~2% gain in TPF.

  • Overall, the reviewers' concern is valid and we will provide a detailed discussion that defends the choice of the ordering in the revision.

[1] Jazbec et al. “Learning unmasking policies for diffusion language models”, ICML 2026.

Dear Reviewer,

Thank you again for your feedback, which we believe has strengthened our submission.

As the rebuttal phase draws to a close, please let us know if you have any remaining concerns that we can address.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary for the AC

For the Area Chair's convenience, we summarize the primary review concerns, the evidence provided in our individual responses, and the corresponding manuscript updates.

Updates during rebuttal.

Compared with the original submission, we introduced new ablations (i) evaluating a training-free PIG variant and (ii) a confidence-only predictor to isolate the source of our performance gains. We also heavily optimized our codebase to improve wall-clock efficiency, successfully reducing per-round overhead to only approximately 1% on A40 GPUs. We re-ran baseline evaluations for EB-Sampler and KLASS natively within our own harness to guarantee fair comparisons (see Reviewer cvf7 [Q2.2] and Fig. 23). Furthermore, we provided interpolated matched-accuracy reporting and padding-free Accuracy-TPF (tokens-per-forward) frontiers to ensure robustness and strengthen our acceleration claims.

Summary of main concerns

  1. Source of gains and learned predictor contribution (cvf7, 9B7J). Reviewers asked whether the performance improvements stemmed from the learned adapter or merely the cumulative budget rule. We provided new ablations showing a training-free version of PIG (using unconfidence as a proxy) and a confidence-only table. While the training-free variant matches or slightly edges out other training-free baselines like EB-Sampler, the full adapter is necessary to achieve the substantial gains seen at high parallelism, particularly for the PIG.EXT algorithm. The adapter correctly captures first-order perturbations which are crucial for next-round regret estimation.
  2. Wall-clock overhead and acceleration metrics (cvf7, HgND). Reviewers asked how TPF improvements translated to actual wall-clock speedups given the adapter's overhead. Initially, our unoptimized implementation had a 9-10% per-round overhead. During the rebuttal, we optimized the code by reusing forward logits, keeping inputs on the GPU, and utilizing CUDA graphs. This reduced the overhead to approximately 1% of a backbone forward on an A40. We also demonstrated that excluding EOS padding from TPF calculations does not alter our claims relative to the baselines.
  3. Predictor size and offline training efficiency (HgND, 9B7J). Reviewers raised concerns regarding the necessity of offline fitting and the trade-off between predictor size and computational overhead. We demonstrated that training the adapter is highly efficient, requiring only about a single H100 GPU hour per backbone, and transfers favorably to unseen tasks like MMLU. A new ablation on adapter width revealed that halving the parameters (from ~2.17M to ~0.95M) maintains strong per-edge and aggregate coupling correlations, indicating our predictor could be made even smaller without sacrificing performance.
  4. Baseline comparability and matched-accuracy reporting (cvf7). To address concerns about copying single-cell baseline numbers and the robustness of our 1.7x speedup headline, we executed EB-Sampler and KLASS sweeps within our own harness. To resolve granularity issues in matched-accuracy reporting, we introduced an interpolated geometric mean TPF gain metric. This new reporting broadly confirmed PIG's reported gains across the frontier against Fast-dLLM, DAPD, EB-Sampler, and KLASS.
  5. Theoretical bounds and algorithmic design choices (cvf7, HgND, 9B7J). Reviewers requested clarification on theoretical guarantees and our reliance on confidence ordering. We clarified that Proposition A.8 is intended to be non-vacuous for threshold θ\theta < 1 and provided a tighter reading for the deployed decoder that does not rely on non-negativity premises. Regarding algorithmic choices, we defended confidence-based ordering as an efficient heuristic for an otherwise NP-hard subset selection problem, noting that it closely tracks the re-ranked sequential decode. We also showed that a more sophisticated re-ordering implementation through a knapsack-style implementation only yielded a ~2 TPF improvement at significant overhead. Finally, we clarified the selection rule between our two variants: PIG.INT is stronger for lower-speed, high-accuracy regimes, while PIG.EXT tracks decode trajectory curvature better at high TPF.

If the AC or Reviewers have any other questions or concerns, we would be happy to address them.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.