acceptodds
Withdrawn from ICLR 2027

When Does Mixing Candidate Sources Beat Resampling in Diffusion Language Models?

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Masked diffusion language models can spend extra generator calls in three ways: resample the same decoder, decode with a different rule, or query a different model. We ask when drawing from several of these sources beats resampling one. The analysis starts from a simple observation: repeated draws solve the prompts a source already handles, so the prompts left open concentrate on its weaknesses. For conditionally independent draws we derive how this residual set evolves and obtain a switching rule: another source gains ground when it succeeds where the current source fails, and a learned verifier returns only part of what it adds. The rule yields a greedy source order for tasks with an exact checker and a gated escalation rule for learned verifiers, both fixed on development data and tested prospectively. On Countdown and CRUXEval-I, mixing decoders and models outperforms resampling the released dVoting decoder by 16.8 to 21.2 points at the same budget, in the direction predicted, some smaller than predicted, and matches or beats the strongest single model given the same number of draws. With learned verifiers on MATH the gain shrinks to 3.1 to 5.5 points, and on GSM8K, where few prompts remain unsolved, it vanishes, as the analysis predicts.

Primary area:
foundation or frontier models, including LLMs
Keywords:
diffusion language models, test-time scaling, test-time compute allocation, best-of-N with verifiers, decoding rules, resampling, coverage, process reward models, verifier fine-tuning, adaptive inference

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 4 reviews, above 0% of the others.
Reviewer_dDMeOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The submitted manuscript seems to be about test-time compute in diffusion language models (dLLMs). It studies the question if Best-of-N performance can be improved if instead of sampling multiple answers from the same dLLM and sampler (so by just varying the random seed), one varies the dLLM and/or the sampler across generations.

Critical strengths

/

Critical weaknesses
  • The paper to me feels end-to-end AI generated. I tried to understand the abstract but got stuck there already. Why use fancy terms like "candidate sources" or "generator calls" where we have established terms for those concepts already ("samplers" and "NFEs")? Also in abstract they use terms such as "residual set", "gated escalation rule", "dVoting decoder" which are never introduced and I wouldn't say they are that commonly known that one could use them without properly introducing them first?
  • Also concepts like Best-of-N and Pass@k seem to be crucial here since this is a test time compute paper? But they are never used?
  • There is no background section where the notation and the setting would be slowly introduce to aid the reader. In Section 3 they jump straight into problem formulation where the reader is confused even more since they do not bother to connect their presentation to the common concepts in test-time compute literature...
  • By default the paper seems to be using dVoting sampler from [1] . Why use such a "niche" sampler? Why not use something more standard such as Fast-dLLM or EB-sampler? Also dVoting is slow no, since it requires generating samples sequentially, one-by-one?
  • I do not understand how results are presented in Figure 1 and Table 1. Why average across different budgets? Why not present Best-of-N/Pass@k curves for various N as is common in the test-time compute literature?
  • Also for Table 1, why are Two Temperatures [2] and Fast-dLLM [3] baselines? These are samplers, so I thought this could be used within the proposed mechanism of sampling different generations from various samplers?

[1] Feng et al. DVOTING: Fast Voting for dLLMs. arXiv 2026

[2] Olausson et al. A Tale of Two Temperatures: Simple, Efficient, and Diverse Sampling from Diffusion Language Models. COLM 2026

[3] Wu et al. Fast-dLLM: Training-free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decoding. ICLR 2026

Questions for authors

See above

Reviewer_bpK9OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies diffusion LM inference quality with access to multiple models and samplers under a fixed NFE budget. The paper proposes an allocation strategy under a pre-defined verifier. This strategy beats resampling on two reasoning tasks (Countdown and CRUXEval-I) but is less successful for MATH/GSM8K.

Critical strengths

The positive results on Countdown and CRUXEval-I are mildly interesting.

Critical weaknesses

This paper fails entirely to engage with classic literature, and rederives many known results as a consequence. There are obvious connections to Bayesian sequential search, reliability theory, and bandit theory that are completely overlooked. The proposed future work on "a switching rule updated from the failures observed at test time" is an online bandit problem, for which there is extensive prior work.

Proposition 1 is just the standard Bayesian update of a posterior mean after a failure. It follows directly, e.g., from the standard Bayesian update after unsuccessful search.

Equation 6 is Luce’s choice rule, with the Bradley-Terry model. See, e.g., McFadden's work on random utility models. You cannot just re-derive these results without attribution. Connecting to prior work would also clarify the model’s limitations including the failure modes discussed in Section 4.3. Repeated answers are the classic failure of the independence-of-irrelevant-alternatives property. Source-dependent scores indicate model mis-specification. Both are discussed in the literature, along with classic mitigations

Questions for authors
Reviewer_JkkYOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates when mixing candidate sources (different model–decoder configurations) is more effective than repeatedly sampling from a single source in diffusion LLMs. It derives residual success dynamics and source-switching conditions, and proposes greedy source ordering and verifier-gated escalation.

Critical strengths

Careful experimental protocol

Source selection is fixed on development data and evaluated on held-out sealed test sets, with matched generator-call budgets, common verifiers, and confidence intervals.

Useful analysis of source complementarity

The paper quantifies how repeated failures change the relative value of candidate sources and distinguishes candidate coverage from the accuracy achieved by the verifier.

Critical weaknesses

Limited novelty relative to prior work

Heterogeneous sampling and optimized compute allocation have already been studied, notably in OSCA [1]. The proposed framework addresses a closely related problem, but its methodological contributions beyond existing allocation approaches and their application to diffusion LLMs are not sufficiently established. OSCA is neither discussed nor compared experimentally.

Unclear benefit of the proposed source-selection strategy

The substantial gains over resampling appear largely attributable to combining complementary sources rather than the proposed greedy ordering. Appendix B.1 reports nearly identical performance when ordering sources simply by standalone accuracy per call (+0.2, 0.0, and −0.8 percentage points across the exact-checker settings). Comparisons against uniform or random source mixing are also missing. Although residual-based selection improves over the accuracy-based rule on MATH, its additional value is not consistently demonstrated across settings.

Limited applicability of the theoretical analysis

The switching analysis assumes conditionally independent draws, whereas the deployed dVoting sampler uses dependent samples. The covariance-based prediction also fails its prespecified sign criterion (58% versus 70%), although the disagreements occur near zero.


[1] Zhang et al., "Scaling LLM Inference Efficiently with Optimized Sample Compute Allocation", NAACL 2025.

Questions for authors

Novelty relative to OSCA

How does the proposed framework fundamentally differ from OSCA's optimized allocation across heterogeneous sampling configurations [1]? What additional capabilities or guarantees does the residual-based formulation provide?

Simple mixing baselines

How does the proposed policy compare with uniform or random source mixing under the same candidate menu, generator-call budget, and verifier?

Theoretical applicability

To what extent do the proposed switching conditions remain predictive for the dependent sampling process used by dVoting?

Missing baseline evaluations

Why were HEX, Two Temperatures, and Fast-dLLM not evaluated on MATH set 2 and GSM8K? Appendix B.2 confirms that MATH set 2 was not run but does not explain the omission.


[1] Zhang et al., "Scaling LLM Inference Efficiently with Optimized Sample Compute Allocation", NAACL 2025.

Reviewer_zcL3OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper asks how masked diffusion language models (MDLMs) should spend extra forward passes at test time: on more samples from the same sampler and trained model, on another sampler but same model, or on another model altogether. Its theory says that failed samples concentrate around the harder prompts, so the next sample should come from whichever of the three options succeeds most per forward pass. This gives two policies: one for when a program can check answers exactly, and one for when a learned verifier scores them. At equal forward passes on held-out prompts, mixing two LLaDA samplers and a second MDLM, Dream, beats resampling one LLaDA sampler by around 15% with an exact checker on Countdown abd CRUXEval-I, and by around 5% with a learned verifier on MATH, and not at all on GSM8K. The authors present this as confirming their theory.

Critical strengths
  1. The propositions and math all seem correct.
  2. The experimental protocol is careful: budgets and answer selection are held fixed, and all choices are made on "development" (validation) data before the "sealed" (held-out test) sets are used.
Critical weaknesses
  1. The theory is independent of diffusion, elaborates basic ideas, and is not needed for the experiments:
  • Nothing in it uses diffusion, and would apply to any randomized solver.
  • It turns elementary ideas into already familiar policies: best-of-N with a verifier, greedy selection, etc.
  • In particular, none of the results of the theory are needed to motivate, interpret, or execute the experiments.
  1. The experiments have no clear relevant finding, because the gains are exactly what one would expect given the baseline of a single model (LLaDA) with a single sampler. Furthermore, the gains on Countdown seem to come mostly from using the other LLaDA's sampler, and on CRUXEval-I from using the Dream model.
Questions for authors
  1. Which decision in the experiments depends on the theory, rather than on tuning directly on development data?
  2. Section 5.1 says three tasks have exact checkers but names two. Experiments that could change my assessment:
  3. At equal forward passes, how does mixing compare with independent samples of the adaptive sampler, the fixed-step sampler with enough samples, and Dream alone on MATH and GSM8K?

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

Related papers

Discussion (0)

Sign in to comment.