acceptodds
Under review as a conference paper at ICLR 2027

Keep the Lineages Alive: Diversity-Preserving Resampling for Diffusion LLM Decoding

Abstract

Inference-time sequential Monte Carlo (SMC) decoding can improve the reasoning accuracy of masked diffusion language models (diffusion LLMs) without post-training by resampling a population of partial generations according to a reward. Its effectiveness therefore depends on the resampling step, which must concentrate computation on promising candidates while retaining enough diversity to explore alternative reasoning paths. The reward can be internal, the decoder’s own confidence on the committed prefix, or external, a process reward model scoring that prefix. Under either reward, multinomial resampling can disrupt this balance by duplicating high-weight particles and discarding lineages that may still reach a correct solution. The effective sample size (ESS) does not register this loss: multinomial resampling resets it to its maximum even as distinct lineages disappear. To retain such lineages, we introduce lineage-preserving SMC (LP-SMC), which combines Chopthin resampling with an input weight floor. Instead of equalizing weights, Chopthin bounds the ratio between the largest and smallest weights it returns and carries the unequal weights forward. The floor limits how far below the maximum a weight may fall before resampling, so that lineages with low early weight keep a chance to recover. Across mathematical-reasoning and code-generation benchmarks on two base models, LP-SMC raises oracle coverage, the fraction of problems with at least one correct particle, in every setting, by 1.8 to 25.6 points, and raises end-to-end accuracy under a common reward-model selector in eleven of twelve settings, by up to 7.9 points on LLaDA-1.5 and 18.3 on Dream-7B. The coverage gain persists under the external reward model, which suggests that diversity-preserving resampling and the guidance signal contribute separately to inference-time search.

Primary area:
foundation or frontier models, including LLMs
Keywords:
diffusion language models, inference-time scaling, sequential Monte Carlo, resampling, diversity preservation, process reward models, decoding

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_8UVyOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes LP-SMC, a diversity-preserving resampling method for diffusion LLM decoding. It combines Chopthin resampling with a weight floor to preserve promising reasoning lineages during SMC decoding.

Overall, the method is well motivated and experimentally well analyzed, but its practical quality–compute trade-off remains unclear due to the high particle-based inference cost and the lack of evaluation under smaller particle budgets.

Critical strengths

(1) The paper identifies the problem that existing SMC methods may lose potentially useful reasoning lineages, and uses Chopthin resampling and a weight floor to reduce lineage collapse caused by stochastic resampling and low weights, respectively.

(2) The method outperforms SMC on two models and multiple benchmarks, and validates the method through multiple metrics including surviving lineages, distinct outputs, oracle coverage, and final accuracy. The experimental analysis is detailed.

Critical weaknesses

(1) The overall inference cost is relatively high, and the practical quality–compute trade-off of the method is not sufficiently clear. Although LP-SMC itself does not introduce additional computational overhead compared with SMC, the whole method is built on SMC decoding with N=16N=16 particles, which requires substantially more computation than standard decoding. The paper mainly compares with SR-SMC under the same budget, but lacks comparisons with simpler methods such as best-of-NN.

(2) The main experiments are all based on N=16N=16, so it is unclear whether the advantages of LP-SMC can stably generalize to lower-compute settings such as N=4N=4.

Questions for authors

(1) The experimental setup mentions confidence-threshold =0.9=0.9, which I understand as corresponding to decoding with a non-fixed number of steps. However, the setup also mentions 8 decoding steps per block. Is 8 the maximum number of steps? How are these two hyperparameters made compatible?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 8UVy

We thank the reviewer for the review, which we found informative. The points on smaller budgets and best-of-NN led to the new experiments below.

Regarding the question (confidence threshold of 0.9 and 8 decoding steps):

Yes, the number of steps is not fixed, and our setup sentence was imprecise: it should say eight blocks, not eight decoding steps per block. The value 8 is not a maximum. We kept the settings identical to the baseline’s hyperparameters, as they report them, and we use its decoder unchanged, which follows Fast-dLLM’s confidence-based parallel decoding. In that decoder, the threshold rule takes over whenever a threshold is set: in every step, it commits all masked positions of the current block whose confidence is at least 0.9, always commits at least the most confident one, and repeats until the block has no masked positions left. A block of 32 tokens therefore takes between 1 and 32 steps.

The value 8 is the steps argument of the baseline’s configuration, which is the generation length divided by the block size (256/32256/32), or one scheduled step per block. It defines the fixed schedule used when no threshold is given. When the threshold is set, the decoder ignores the fixed token quota and continues until the block is complete. The two settings therefore do not conflict. We will clarify the setup description: the sequence is decoded in eight blocks of 32 tokens, and the number of steps within each block is set adaptively by the 0.9 threshold.

Regarding Weakness 2:

Under our rule κ=ln⁡N\kappa=\ln N and the gate ESS<N/2\mathrm{ESS} < N/2, resampling cannot trigger for N≤5N\le5. In that regime, particle states evolve independently, so with the same final selector the method reduces to best-of-NN independent decoding. We therefore ran the ablation at N=6N=6 and N=8N=8.

The following MATH-500 comparison uses LLaDA-1.5 with decoder-confidence guidance and everything else fixed:

NNSR-SMCLP-SMCDifference (95% confidence interval)
644.249.2+5.0+5.0 (+1.4,+8.8)(+1.4,+8.8)
845.249.2+4.0+4.0 (+0.6,+7.4)(+0.6,+7.4)
1649.252.8+3.6+3.6 (+0.4,+6.8)(+0.4,+6.8)

We also evaluated N=8N=8 on all eight model-benchmark settings under decoder-confidence guidance. Accuracy is measured after the final RM selector.

ModelBenchmarkSR-SMCLP-SMC
LLaDA-1.5GSM8K87.990.7
LLaDA-1.5MATH-50045.249.2
LLaDA-1.5HumanEval44.550.6
LLaDA-1.5MBPP46.548.3
Dream-7BGSM8K77.986.3
Dream-7BMATH-50039.847.2
Dream-7BHumanEval50.663.4
Dream-7BMBPP59.460.8

All rows in these response tables are scored with the same evaluation pipeline. Minor grading differences change the N=16N=16 MATH-500 results from 49.0 and 52.4 in Table 1 to 49.2 and 52.8 here; we will document the exact conventions in the appendix.

On MATH-500, the measured gain is largest at the smallest tested budget, and at N=8N=8 LP-SMC is higher in all eight model-benchmark settings. We will add both tables to the appendix.

Regarding Weakness 1:

We agree that particle decoding uses roughly NN times the decoder compute of single-sample decoding. This cost is shared by SR-SMC and best-of-NN, and LP-SMC adds no measurable overhead to SMC (Appendix B.6). The MATH-500 table above gives the trade-off within SMC: LP-SMC with 6 particles reaches the accuracy of SR-SMC with 16 (49.2).

We also ran best-of-NN controls: independent complete rollouts of the same decoder, with the same final RM selector. We report the control on MBPP without PRM guidance and on MATH-500 with PRM guidance. PRM guidance adds one reward-model call per particle at each block boundary; each call is a single forward pass over the current prefix and generates no tokens. We give independent sampling 18 complete rollouts on MATH-500 as an allowance for the intermediate PRM evaluations. LP-SMC uses N=16N=16 particles. On MBPP, where no intermediate PRM is used, both methods use N=16N=16.

Setting (LLaDA-1.5)LP-SMC (N=16N=16)Best-of-NNDifference (95% confidence interval)
MATH-500, PRM-guided55.454.0 (best-of-18)+1.4+1.4 (−1.8,+4.6)(-1.8,+4.6)
MBPP, without PRM guidance50.650.2 (best-of-16)+0.4+0.4 (−2.3,+3.1)(-2.3,+3.1)

The point estimates favour LP-SMC in both settings, with the intervals given in the table. With majority voting as the final selector, which requires no additional reward-model call at selection, PRM-guided LP-SMC is 5.2 points above best-of-18 on MATH-500, with a 95% confidence interval of +2.2+2.2 to +8.4+8.4. Resampling raises the share of correct candidates from 38.1% to 45.7%. Under decoder-confidence guidance the rise is 1.3 points, so the benefit over best-of-NN grows with how informative the guide is. We will add the best-of-NN baseline and the full comparison, including systematic + floor, to the revised paper.

We thank the reviewer again and are glad to clarify anything further.

Thank you for the additional experiments and clarifications. I appreciate the authors' efforts to address my concerns. The new results demonstrate that LP-SMC consistently achieves better solution quality than SR-SMC across several settings. However, when compared with the simpler Best-of-N baseline, the observed improvements appear relatively modest. Overall, I am satisfied with the current assessment and will keep my score unchanged.

AuthorsOpenReview ↗

Best-of-N comparison under the four selection rules

Thank you for reading the new results and for the summary.

We agree that under the reward-model selector the difference from best-of-NN is modest. However, we think the other selection rules are worth considering too. The table gives the same two comparisons under the four selection rules of the paper (Section 5.3), on the same saved candidates (LLaDA-1.5, LP-SMC with N=16N=16). Accuracy in %.

SelectorMATH-500: LP-SMCMATH-500: best-of-18MBPP: LP-SMCMBPP: best-of-16
Reward model55.454.050.650.2
Confidence43.440.640.941.1
Majority vote52.647.441.338.8
Distinct-output vote50.247.641.340.5

LP-SMC is higher in seven of the eight comparisons. On MATH-500 the difference is larger under each of the other three rules than under the reward model, and on MBPP it is larger under the two voting rules. Under majority voting the difference is +5.2+5.2 on MATH-500, with 95% confidence interval (+2.2,+8.4)(+2.2,+8.4), and +2.5+2.5 on MBPP, with interval (0.0,+5.0)(0.0,+5.0). On MATH-500, LP-SMC is guided by the PRM during decoding, so we compare it with best-of-18 to allow for the extra computation.

So the gain over best-of-NN is modest with a reward-model selector, as you say, and larger under most of the other selection rules.

Best-of-NN is also not a single baseline in this literature. PG-DLM and FK-Steering select the sample by reward, the November revision of SR-SMC (our baseline) selects it by log-probability, and S$^3$ defines its best-of-KK baseline as independent samples with a majority vote; the SR-SMC revision lists majority voting as a further compute-matched baseline (its Table 11). We therefore report the comparison under both the reward model and the vote. Under the vote, LP-SMC is higher in both settings:

Setting (majority vote, %)LP-SMCIndependent samplingDifference (95% CI)
LLaDA-1.5, MATH-500, PRM-guided52.647.4+5.2+5.2 (+2.2,+8.4)(+2.2,+8.4)
LLaDA-1.5, MBPP, no PRM guidance41.338.8+2.5+2.5 (0.0,+5.0)(0.0,+5.0)

To our knowledge, earlier SMC decoders for diffusion language models each report one final selector; we report four on the same candidates so that this comparison can be read under each.

We will add these tables to the revision as an additional ablation and state this scope there. If these results address your concerns, we would be grateful if you would consider raising your score. Thank you again for your time and for a review that improved the paper.

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

This paper addresses a highly practical issue in sequential Monte Carlo (SMC) decoding for block-masked diffusion language models: while standard multinomial resampling (as used by the baseline SR-SMC) can increase the effective sample size (ESS), it tends to prematurely eliminate candidate branches (lineages) with lower early weights that could later lead to the correct answer. To this end, the authors propose LP-SMC, which combines Chopthin resampling with a non-equal-weight input weight floor, allowing low-weight potential paths a chance to survive. The paper theoretically distinguishes between ESS and lineage diversity, derives a lower bound on ESS under weight-ratio constraints, and discusses the effects of the weight floor on the resampling trigger conditions and the weighted measure.

Experiments are conducted on two diffusion language models across four math and code benchmarks, with both internal confidence guidance and external PRM guidance, and all answers are selected by the final reward model (RM). Under the same model, guidance signal, and final selector, LP-SMC improves coverage (pass@k) on all 12 settings and final accuracy on 11 settings. The paper also includes component ablation, lineage and output diversity analyses, and paired significance tests to verify the contributions of each component.

Critical strengths
  1. The controlled experimental design is solid, and the reporting on statistical significance is very objective. The main experiments strictly matched the base model, guidance signal, final RM selector, and test set (using the original method SR-SMC's default multinomial resampling as the controlled baseline), eliminating the confounding effect of "boosted scores from stronger selectors." Across all 12 experimental settings, the point estimates of LP-SMC's coverage all increased (with 11 settings' paired bootstrap confidence intervals strictly above zero, indicating statistical significance), and the final accuracy also had 6 settings with confidence intervals excluding zero. The authors were very honest in their reporting, clearly distinguishing the strength of evidence between "improved coverage" and "improved final accuracy," and not overclaiming the modest point estimate increases as overwhelming superiority. This demonstrates that the method indeed addresses a real pain point in diffusion decoding.

  2. The component ablation is clear, the changes are lightweight, and the additional inference cost is negligible. LP-SMC requires no changes to model architecture or retraining; it is entirely a plug-and-play improvement at the inference-stage population management level. The step-by-step ablation in Table 2 and the complete 3×2 crossed ablation in Table 3 (three resamplers × with/without floor) clearly show the independent gains from replacing the resampler and introducing the weight floor. In particular, on the mechanism diagnostic indicators such as surviving lineages (10.04 vs. 7.62) and number of particles with no offspring per step (3.61 vs. 6.97), LP-SMC shows clear advantages. Meanwhile, the measured computational cost of LLaDA-1.5 on GSM8K is nearly identical to the baseline (mean NFE 708.0 vs. 701.6, wall time 119.6 s vs. 120.8 s), making it a practical, low-cost decoding improvement.

Critical weaknesses
  1. The "resampling-free independent sampling" baseline is missing, and the comparison with the strongest alternative (Systematic + Floor) is still not sufficient.
  • Lack of pure independent sampling (Vanilla Best-of-NN) comparison: The so-called best-of-NN in the paper is actually RM selecting the best from the NN particles left after SMC, not the baseline of truly independently generating NN complete trajectories. Without a pure independent sampling comparison under the same computational budget and same RM selector, it is difficult to judge whether LP-SMC is enjoying the bonus of SMC's dynamic compute allocation while preserving diversity, or merely compensating for the diversity loss caused by resampling itself (and it is even unclear how much net benefit it has compared to no resampling at all).
  • Incomplete verification scope and indicators for Systematic + Floor: Although the 3×2 crossed experiments in Table 3 show that under GSM8K (internal confidence guidance), LP-SMC's lineage retention indicators are clearly better than those of systematic + floor, and coverage is slightly higher (94.2% vs. 93.6%), this most critical strong baseline comparison is only given under a single model-task setting. More importantly, Table 3 only reports diagnostic indicators and coverage, without reporting the final accuracy after RM selection, nor providing confidence intervals for the differences. Whether better lineage diversity (10.04 vs. 7.62) can be stably translated into higher coverage and final accuracy on other tasks (such as code generation and external PRM guidance) currently lacks empirical support across settings.
  1. The resampling logic at the Final Block boundary is unclear, and the boundary of incremental theoretical contributions needs to be clarified.
  • Resampling questions at the endpoint: Section 4 mentions that weight floor and ESS-gated resampling are performed at each block boundary, but it does not state whether they are performed as usual after the final block is fully generated. If the entire sequence has been fully generated and no new tokens will be produced afterwards, then doing resampling at this point (duplicating high-weight particles and removing low-weight particles), apart from needlessly discarding candidate answers and lowering or maintaining coverage, has no positive effect whatsoever. The authors need to clarify whether the final candidate set being evaluated was saved before or after the endpoint resampling.
  • Original boundary of the theoretical framework: Resampling causing ancestor lineage degeneration is a classic problem of SMC, and Chopthin has already been used for power sampling of autoregressive models. Section 6 lists "lineage-based formulation" alongside input floor and diffusion decoding as core contributions, but the paper has not yet clearly defined: which mathematical properties belong to known consequences of general SMC resampling, and which are unique new conclusions after block diffusion decoding combined with weight floor (furthermore, the theoretical lower bound on ESS itself is not directly equivalent to preserving correct lineages).
Questions for authors
  1. Handling logic at the final block boundary: After the last block is fully generated, does the code trigger another resampling? Is the final candidate set sent for coverage evaluation and RM selector saved before or after the last resampling? If the baseline methods did indeed perform this "only removing samples, no new words generated" resampling at the endpoint, then after turning off all methods' endpoint resampling, how much of LP-SMC's advantage remains?

  2. Supplement cross-setting and final accuracy comparisons between independent sampling and Systematic + Floor: Can LP-SMC, pure independent sampling (without SMC resampling), and systematic + floor be compared under the same budget on representative math and code settings (no need to re-run the full matrix), with fixed generation configuration and final RM selector? In particular, supplement the final accuracy and paired confidence intervals of LP-SMC vs. systematic + floor under Table 3 and other representative settings, to prove that the lineage diversity advantage observed in Table 3 can indeed be stably translated into final performance.

  3. Clarify the original boundary of "lineage-based formulation": Please clearly state, compared to the prior work cited in Section 6, what specific new definitions, propositions, or mechanism insights this paper has added in the theoretical analysis section? Which conclusions belong to general resampling properties, and which are highly dependent on block diffusion decoding?

4 comments
AuthorsOpenReview ↗

Response to Reviewer wNeR (1/2): Questions 1 and 2

We thank the reviewer for the detailed and constructive review. The review also made us realize that the scope of the paper should be presented more clearly, and we will revise the manuscript accordingly.

Regarding Question 1 and the first point under Weakness 2:

Yes. The submitted experiments inherit the baseline implementation: the gate is evaluated after the final block, and the population is saved after that step. Because no decoding follows, endpoint resampling cannot increase oracle coverage, although it can change selected accuracy by removing candidates. This point led us to check the final gate directly: we reran both SR-SMC and LP-SMC with it disabled on representative math and code settings.

BenchmarkFinal gateSR-SMCLP-SMC
HumanEvalon48.254.9
HumanEvaloff48.855.5
MATH-500on49.252.8
MATH-500off49.052.8

Both settings use decoder-confidence guidance with LLaDA-1.5 and N=16N=16. Accuracy is measured after the final RM selector, using the code RM for HumanEval and the math RM for MATH-500. All results in this response are scored with the same evaluation pipeline. Minor grading differences from the submitted tables change the results by at most two solved problems on HumanEval and MATH-500; we will document the exact conventions in the appendix.

Disabling the final gate increases or preserves coverage. On MATH-500, coverage changes from 54.4% to 55.8% for SR-SMC and from 63.8% to 64.4% for LP-SMC. Thus, the coverage advantage of LP-SMC is +9.4+9.4 points with the gate and +8.6+8.6 points without it. Selected accuracy changes by at most 0.6 points. The LP-SMC advantage is +6.7+6.7 points on HumanEval with or without the gate and +3.6+3.6 versus +3.8+3.8 points on MATH-500. We will clarify the inherited endpoint behavior in Section 4 and note that final-boundary resampling is unnecessary when no further decoding follows. These comparisons are therefore not driven by endpoint resampling.

Regarding Question 2 and Weakness 1:

We compare LP-SMC, systematic + floor, and independent sampling on representative math and code settings with the same decoder configuration and final RM selector. MATH-500 uses PRM guidance, while MBPP uses decoder-confidence guidance because no suitable step-level code PRM meeting our criteria is available. LP-SMC and systematic + floor use N=16N=16. Independent sampling receives 18 complete rollouts on MATH-500 as an allowance for the intermediate PRM evaluations of the two guided methods; we do not claim exact wall-time equivalence. On MBPP, where no method uses an intermediate PRM, all methods use N=16N=16. The table reports final RM-selected accuracy (%).

SettingSystematic + floorLP-SMCIndependent sampling
MATH-500, PRM-guided54.655.454.0
MBPP, confidence-guided50.650.650.2

Paired differences with 95% confidence intervals are:

SettingLP-SMC minus systematic + floorLP-SMC minus independent sampling
MATH-500, PRM-guided+0.8+0.8 (−2.2,+3.8)(-2.2,+3.8)+1.4+1.4 (−1.8,+4.6)(-1.8,+4.6)
MBPP, confidence-guided0.00.0 (−2.5,+2.5)(-2.5,+2.5)+0.4+0.4 (−2.3,+3.1)(-2.3,+3.1)

In the Table 3 setting, GSM8K with decoder-confidence guidance, LP-SMC is +0.5+0.5 points above systematic + floor in final RM-selected accuracy, with a 95% confidence interval of (−0.6,+1.7)(-0.6,+1.7).

LP-SMC also retains more lineages than systematic + floor in both new settings: 6.7 versus 4.2 out of 16 on MATH-500 and 13.5 versus 12.0 on MBPP, consistent with the result reported in Table 3.

Against independent sampling, this estimates favour LP-SMC in both settings, with the intervals given in the table. Under majority voting, which requires no reward model at selection, LP-SMC is +5.2+5.2 points higher on PRM-guided MATH-500 (52.6% versus 47.4%; 95% confidence interval (+2.2,+8.4)(+2.2,+8.4)).

We will add the comparison with independent sampling and these tables to the revised paper.

Our response to Question 3 continues in the next comment.

AuthorsOpenReview ↗

Response to Reviewer wNeR (2/2): Question 3

Regarding Question 3 and the second point of Weakness 2:

Known general SMC properties. Repeated resampling reduces ancestral diversity; multinomial lineage survival follows a general occupancy calculation (the count in Proposition 1(i)); and Chopthin bounds the ratio of its output weights. A weight-ratio bound also implies an ESS bound, although our Lemma 2 sharpens the previously stated constant.

New but mathematically general results in this paper. We do not claim that the multinomial occupancy calculation in Proposition 1(i) is itself new. By “lineage-based formulation,” we mean using root labels and the surviving-lineage count to formalize the ESS-blindness mismatch in Proposition 1 and Remark 1, and connecting that mismatch to population management in block diffusion decoding.

We introduce the input weight floor and set its default using the particle count, κ=ln⁡N\kappa=\ln N, without tuning it to accuracy. We derive the gate-compatibility threshold, κ>ln⁡(3+8)\kappa > \ln(3+\sqrt{8}), show why κ=ln⁡N\kappa=\ln N is compatible for N≥6N\ge6, and bound the weight mass that the floor can move. We also show that the floor and Chopthin act on complementary sides of a resampling event: the floor bounds the weights entering the resampler, while Chopthin bounds the weights it returns.

In the revision, we will add a bound on the maximum expected offspring count of any input particle. With the floor at κ=ln⁡N\kappa=\ln N, the expected offspring count of every particle under Chopthin is bounded by 2N22N+η(N−1),\frac{2N^2}{2N+\eta(N-1)}, where η\eta is Chopthin’s ratio bound. This is 4.29 of 16, regardless of the input weights. With the same floor, a proportional resampler such as systematic resampling permits an expected offspring count of up to 8.26. Across 40,470 recorded events, the largest realized count under LP-SMC was 5.

Specific to block diffusion decoding. We integrate these components at block boundaries and evaluate their effects under both accumulated decoder-confidence and external PRM guidance. We also separate population construction from final selection by reporting selector-independent coverage and applying four selection rules to the same saved populations.

The mathematical bounds limit weight and offspring concentration; they do not determine which lineages are correct. Correct-lineage retention is therefore an empirical question. Relative to multinomial resampling, coverage rises in all twelve settings, and we will revise the manuscript to state this theoretical and empirical boundary more clearly.

We will add these results to the revised paper. We thank the reviewer again and are glad to clarify anything further.

Reviewer_wNeROpenReview ↗

Feedback

Thank you for running the controls I asked for. I find the diagnosis of lineage loss convincing and the evaluation against SR-SMC persuasive.

Two concerns remain. First, the additional benefit of Chopthin for final accuracy is uncertain. With the proposed input floor held fixed, replacing systematic resampling with Chopthin changes RM-selected accuracy by +0.8 percentage points on MATH-500, 0.0 on MBPP, and +0.5 on GSM8K. All three paired 95% confidence intervals include zero. The separate comparison against independent sampling evaluates LP-SMC as a whole. Its RM-selected accuracy gains are +1.4 points on MATH-500 and +0.4 on MBPP, with both intervals including zero. The 5.2-point gain over independent sampling under majority voting on PRM-guided MATH-500 is encouraging. Given the prior work on Chopthin, I regard the input floor, its analysis, and its integration into block diffusion decoding as a useful but modest methodological advance.

Second, the paper’s clearest empirical gain over SR-SMC is in oracle coverage. I would welcome the corresponding comparison on MBPP, where both LP-SMC and independent sampling produce 16 candidates per problem. Could the authors report oracle coverage (pass@16) for the independent rollouts and a paired confidence interval for the coverage difference relative to LP-SMC?

On balance, the rebuttal has resolved my main questions, and I will raise my rating to 3 (weak accept).

We sincerely thank the reviewer for their thoughtful and constructive feedback. We are pleased that our responses have addressed your main concerns, and we greatly appreciate your reconsideration of the rating. We have ensured that the points raised in our earlier discussion are reflected in the revised manuscript. For your latest suggestions, we have incorporated additional supporting experiments and clarifications wherever feasible within the limited time available. We hope these additions will further address your concerns and would be grateful if you could take a look at the revised manuscript. Thank you again for your valuable feedback, which has helped us strengthen the paper.

Reviewer_Au4LOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes lineage-preserving sequential Monte Carlo decoding for masked diffusion language models. It combines Chopthin resampling with an input weight floor to reduce the loss of potentially useful generation paths.

Critical strengths
  • Consistent coverage gains: LP-SMC increases oracle coverage in all twelve main experimental settings, across mathematics and code generation.

  • Low computational overhead: LP-SMC modifies population management without adding direct decoder or reward-model evaluations.

Critical weaknesses
  • Limited methodological novelty: Prior work already applies Chopthin to LLM decoding. The main additions here are the input weight floor and its integration into diffusion decoding, making the methodological contribution relatively incremental.

  • Selector-dependent accuracy gains: Accuracy improves in eleven of the twelve settings under reward-model selection, but the gains are less consistent and can reverse with confidence-based selection.

Questions for authors

Could the authors compare the proposed input weight floor with alternative weight-flattening strategies while keeping the resampler and final selector fixed?

4 comments
AuthorsOpenReview ↗

Response to Reviewer Au4L (1/2): Weaknesses 1 and 2

We thank the reviewer for recognizing the consistent coverage gains, low computational overhead, and soundness of the study. We appreciate the question about alternative weight-flattening strategies. It led us to run a controlled comparison that clarified which properties are specific to the floor and showed that Chopthin can also be paired with other flattening rules. We will add this comparison to the revised ablation section, with the full calibration and event-level analysis in the appendix.

Regarding Weakness 1:

We agree that Chopthin is prior work (Gandy & Lau, 2016) and that Ahmadi et al. (2026) applied it to autoregressive decoding, as Section 6 acknowledges. This paper adapts it to block-wise diffusion language-model decoding and pairs it with a floor on the input weights.

Proposition 1 shows that ESS cannot detect lineage loss: even with uniform weights, multinomial resampling removes about one third of the lineages in expectation while resetting ESS to its maximum.

The two components address different parts of the resampling step. Chopthin bounds the weights that a resampling event returns, while the floor bounds the weights that it receives. On the Dream-7B settings in Table 2, Chopthin alone changes coverage by at most 3 points, while adding the floor contributes a further 8 to 23 points. No single component reaches the combination (Table 3).

The two components together also give a new guarantee that neither provides alone; we will add the proof to the appendix in the revision. With the floor and Chopthin, the expected number of offspring of any particle is at most 4.29 of the 16 slots, regardless of the input weights; the largest realized count across 40,470 events was 5. Under Chopthin alone, the expected count can approach 16, while the floor with multinomial resampling allows an expected count of up to 8.26.

Our evaluation separates building the candidate set from choosing among its candidates. We report oracle coverage and apply four selection rules to the same saved populations under two guidance signals, the decoder’s confidence and an external PRM. Coverage rises in all twelve settings, including the four PRM-guided settings, so the method is not tied to a single guidance signal.

In the revision, we will distinguish the derived condition, κ>ln⁡(3+8)\kappa > \ln(3+\sqrt{8}), from our accuracy-independent default, κ=ln⁡N\kappa=\ln N.

Regarding Weakness 2:

We agree that final accuracy depends on the selector. LP-SMC changes the candidate set, while the selector determines how much of the additional coverage becomes accuracy. Confidence selection is the rule that most often favours the baseline. It relies on the model’s own confidence, which the floor keeps from dominating early, so it can pass over a correct lineage that started with low confidence and that LP-SMC preserved.

The gain is not limited to the reward-model selector. Majority voting and distinct-output voting require no reward model, and they are higher or tied under LP-SMC in 10 and 9 of the 12 settings, respectively (point estimates, Tables 4 and 5). We will add these counts to Section 5.3 and state the limitation more clearly there.

AuthorsOpenReview ↗

Response to Reviewer Au4L (2/2): Alternative flattening strategies

Regarding the question:

We kept the Chopthin resampler and final reward-model selector fixed and replaced the floor with two alternatives. The first is power tempering, w~i∝wiα\tilde w_i\propto w_i^\alpha with α<1\alpha < 1, which is related to the entropic annealing of Giannone et al. (2025). The second is a cap on the largest weights, w~i∝min⁡(wi,c wˉ)\tilde w_i\propto\min(w_i,c\,\bar w), following truncated importance sampling (Ionides, 2008).

We did not tune either strength on evaluation accuracy. Using weights recorded from runs without flattening, we selected α\alpha and cc so that the expected number of input particles surviving a Chopthin resampling event matched that of the floor. We evaluated LLaDA-1.5 with N=16N=16 on MATH-500 under PRM guidance and HumanEval under the decoder’s confidence.

Accuracy (%)Floor (ours)TemperingCap
MATH-50055.455.654.8
HumanEval54.951.255.5

The Table 1 floor population and both alternatives were scored with the same evaluation pipeline. Minor grading differences from Table 1 change the floor result by two solved problems on each benchmark; we will document the exact conventions in the appendix. Relative to the floor, the paired differences and 95% confidence intervals are +0.2+0.2 (−2.8,+3.4)(-2.8,+3.4) on MATH-500 and −3.7-3.7 (−7.3,−0.6)(-7.3,-0.6) on HumanEval for tempering, and −0.6-0.6 (−3.4,+2.2)(-3.4,+2.2) and +0.6+0.6 (−3.0,+4.3)(-3.0,+4.3) for the cap. The cap intervals include zero on both benchmarks, while tempering is lower on HumanEval. We do not claim that the floor dominates, and we will report the cap as a viable alternative.

We retain the floor as the default for two reasons that do not depend on accuracy. First, it uses the same rule, κ=ln⁡N\kappa=\ln N, in every setting, whereas the strength required to match it changes across settings: α\alpha ranges from 0.68 to 0.80 and cc from 1.49 to 2.25. Second, only the floor bounds the ratio of the weights entering the resampler, so only it provides the guarantee above. Under tempering and the cap, a single particle received as many as 15 and 13 of the 16 slots, respectively, compared with an observed maximum of 5 under the floor.

We will add these results to the revised paper. We thank the reviewer again and are glad to clarify anything further.

Thank you for the new experiments comparing the weight floor with tempering and capping while keeping Chopthin and the final selector fixed. These directly address my question about alternative weight-flattening strategies. The default rule and the additional bound on expected offspring counts also provide a clearer rationale for the floor. However, capping achieves similar accuracy in the two tested settings, and I still view the methodological contribution as incremental. I remain positive about the paper and am keeping my rating at 3.

Thank you for the follow up and for the positive assessment.

One distinction between the two is worth noting: the cap requires a strength parameter whose value depends on the setting and calibration procedure, whereas the floor uses a single rule determined by the number of particles, κ=ln⁡N\kappa=\ln N. The two approaches are also not mutually exclusive: a cap or tempering could be applied on top of the floor, which we leave to future work.

Thank you again for the helpful suggestion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.