acceptodds
Under review as a conference paper at ICLR 2027

SONAR: Spectral Diversity for Lost Reasoning Modes in Masked Diffusion LMs

Abstract

Effective exploration for complex reasoning problems remains a critical challenge in masked diffusion language models (MDLMs), where standard sampling frequently collapses into redundant failure modes. Raising the sampling temperature can increase diversity but also degrade solution quality, leaving the full reasoning capabilities of these models largely underutilized. To unlock these capabilities, existing approaches are primarily concerned with algorithms for token-reveal decisions, or promoting exploration over the vocabulary space. These methods, though effective to varying degrees, generally overlook the rich representation structure of the MDLMs. In this paper, we thus propose Sonar, a sampling method that first scans for overlap via a spectral objective in RKHS, subsequently transmitting the gradient of this objective throughout the model. Sonar then utilizes this gradient feedback to steer samples to systematically uncover lost reasoning modes of the MDLM. Across HumanEval (coding) and GSM8K (mathematics), Sonar improves fixed-budget reasoning success for LLaDA-8B-Instruct and Dream-7B-Instruct. It also expands solution coverage and reduces algorithmic collapse by uncovering distinct valid solution strategies missed by ordinary sampling, providing empirical evidence for a capability–sampling gap in MDLMs.

Primary area:
generative models
Keywords:
diversity, diverse sampling, diffusion language models

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_g17KOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Sonar, a training-free sampling method for masked diffusion language models. It aims to improve reasoning diversity by measuring spectral overlap among candidates’ contextual token representations and using this signal to steer generation.

Critical strengths
  • Distinctive design: Sonar uses the spectral geometry of contextual token representations to guide candidate diversity. It measures diversity at the final layer while applying steering at an intermediate layer.

  • Training-free approach: The method keeps pretrained model parameters fixed and requires no additional model training.

Critical weaknesses
  • Task- and temperature-dependent gains: Sonar’s advantage over ODD varies across tasks and temperatures. Although it achieves substantial gains in some low-temperature settings, its advantage on GSM8K is inconsistent, and for Dream it shrinks or reverses at higher temperatures.

  • Unclear computational trade-off: Sonar requires backpropagation through transformer blocks and additional forward evaluations. The paper does not quantify this overhead relative to ODD or provide compute-matched comparisons, leaving unclear whether the performance gains justify the additional computation.

Questions for authors
  1. Could the authors explain why Sonar’s advantage over ODD shrinks or reverses at higher temperatures for Dream?

  2. Could the authors report success-versus-compute curves for Sonar, ODD, and ordinary sampling, allowing the baselines to generate additional candidates within the same compute budget?

5 comments
AuthorsOpenReview ↗

Summary and (w1,q1)

We thank reviewer g17K for their constructive feedback and for recognizing SONAR's distinctive design and training-free approach to representation-space exploration.

We address the reviewer's concerns through additional experiments and analyses, summarized below:

  1. (W1, Q1) Task- and temperature-dependent gains for GSM8K. We explain why different diversity methods need not outperform one another at every temperature. Their effectiveness depends on how their exploration biases interact with sampling stochasticity and the representation structure learned by each model. Following the reviewer's observation, we also conducted additional experiments exploring a broader range of temperatures and diversity strengths, leading to improved SONAR results on GSM8K.
  2. (W2, Q2) Computational efficiency. We evaluate SONAR, ODD, and conventional sampling under comparable GPU-time budgets, combining multiple settings to better explore each method's capabilities. SONAR solves 143 HumanEval problems, compared with 133 for ODD and 129 for LLaDA, using less compute. Moreover, SONAR discovers 9 problems missed by both baselines combined, demonstrating its ability to uncover overlooked reasoning modes.

W1 / Q1. Task- and temperature-dependent gains

The reviewer observes that SONAR's gains over ODD vary across temperatures and models. We believe that a diversity method need not outperform other methods at every sampling temperature. The goal is to improve the method's overall ability to discover alternative reasoning modes under suitable exploration settings. Performance at other temperatures reflects the interaction between sampling stochasticity and the method's exploration bias.

Temperature introduces stochasticity into token sampling, whereas SONAR explicitly guides exploration through a geometric bias defined by its diversity metric. Our hypothesis is that this bias can exploit the structure learned in the model's representation space to discover overlooked reasoning modes. How effectively this bias interacts with temperature depends on the representation structure learned by each model. Intuitively, lower temperatures may complement stronger geometric diversification, but this need not hold across all models and tasks.

Our initial experiments explored a limited range of settings. Following the reviewer's observation, we expanded our hyperparameter search over sampling temperature and diversity strength. This identified more effective configurations for GSM8K (e.g., T=0.25 for Dream and T=0.5 for LLaDA). The revised results strengthen SONAR's performance, particularly at lower temperatures.

Table R9. Updated results on GSM8K. Values are reported across low, medium, and high sampling-temperature settings (low / medium / high).

ModelMethodEmpirical Pass@16 (low / med / high)Pass@8 (low / med / high)
LLaDABase76.2 / 81.5 / 82.672.2 / 76.7 / 76.7
ODD87.6 / 87.4 / 88.282.1 / 82.0 / 82.3
SONAR88.8 / 88.3 / 86.183.2 / 82.9 / 80.0
DreamBase57.1 / 70.1 / 74.155.2 / 64.8 / 65.3
ODD84.8 / 86.2 / 89.876.9 / 79.2 / 82.1
SONAR90.6 / 85.8 / 84.285.4 / 80.1 / 77.7

On GSM8K, both methods benefit from stronger diversification. ODD achieves its best results with larger diversity step sizes (64 and 128), similar to SONAR's improvement with stronger diversity settings. The key difference is that ODD benefits from higher sampling temperatures, whereas SONAR performs better at lower temperatures. Neither behavior is inherently better or worse. Rather, this reflects how the different exploration biases of the two methods interact with sampling stochasticity and the representation structure learned by each model. Please refer to Table 1 and Figure 4 in the revised manuscript for the updated results.

W2 / Q2. Computational efficiency under comparable inference budgets

We agree that computational efficiency should be evaluated under comparable inference budgets, allowing baselines to generate more samples. As also emphasized by Reviewer B9mR, evaluating multiple generation settings provides a more comprehensive picture of each method's exploration potential. Different temperatures and diversity strengths may discover complementary solutions, so relying on a single configuration could underestimate a method's capabilities.

Following this, we selected mixtures of settings for LLaDA, ODD, and SONAR through independent hyperparameter searches on a small subset of problems, maximizing their combined problem coverage on this subset. We allow the baselines to generate substantially more candidates to account for SONAR's additional overhead.

To reduce SONAR's overhead, we use interventions closer to the output. Our measurements show that 1-block and 4-block interventions require 1.85× and 2.12× the runtime of ordinary LLaDA, respectively, compared with 3.97× for the original 16-block intervention used in the submitted manuscript. (See table R4 and appendix F2)

Experimental settings.

  • LLaDA: 6 settings combining temperatures and sample allocations, totaling 144 samples per seed.
  • ODD: 9 settings combining temperature and diversity step size, totaling 144 samples per seed.
  • SONAR: 4 settings (two 1-block and two 4-block intervention profiles), totaling 64 samples per seed.

The resulting relative GPU-time budgets are 1.00× for LLaDA, 1.08× for ODD, and 0.88× for SONAR. Thus, SONAR uses less compute than either baseline.

We measure exploration performance using Coverage@M (Sec. 5.2, Fig. 4), which reports the expected number of unique problems solved when combining (M) independent seed trials.

Table R10. Coverage@M under mixed settings and comparable computational budgets.

MethodM=1M=2M=4M=8Relative GPU budget
LLaDA (144 samples/seed)102.875 ± 1.053112.036 ± 1.569121.514 ± 1.8031291.00×
ODD (144 samples/seed)111.750 ± 2.634119.607 ± 2.241127.200 ± 1.5641331.08×
SONAR (64 samples/seed)109.500 ± 3.240121.179 ± 3.163132.029 ± 2.2681430.88×

Although ODD performs better for a single seed, SONAR achieves higher coverage as exploration extends across multiple seeds. At (M=8), SONAR solves 143 problems, compared with 133 for ODD and 129 for LLaDA, while using less GPU time and generating only 512 samples versus 1,152 for each baseline.

Furthermore, SONAR discovers solutions to 14 problems missed by extensive LLaDA sampling, compared with 8 for ODD. Even combining both baselines (2,304 samples), SONAR discovers 9 problems that neither baseline solves, despite using approximately 2.4× less GPU time.

These results demonstrate that SONAR can achieve greater problem coverage under a smaller computational budget, while also uncovering reasoning modes missed by extensive conventional sampling.

For a detailed computational analysis, including inference latency, GPU memory, and forward/backward operations, please refer to Table R4 in our response to Reviewer B9mR and Appendix F.2–F.3.


Thank you again for your constructive review. If you feel our clarifications have successfully addressed your questions and concerns, we kindly ask that you consider raising your score.

Thank you again for your constructive review and feedback. Following your observations, we expanded our GSM8K experiments across sampling temperatures and diversity strengths, obtaining improved results on both LLaDA and Dream. We also conducted additional matched-compute comparisons against LLaDA and ODD. We hope these results address your concerns and would welcome any further questions or feedback before the discussion period ends.

Reviewer_B9mROpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SONAR to encourage candidates' diversity sampled by diffusion-LM. SONAR encourages candidates to explore different solutions by reducing spectral overlap in their mid-layer hidden contextual representations. It gives the downstream layers a role in shaping the edit, and the paper provides both a local geometric explanation and experiments to motivate that choice.

Critical strengths
  1. The choice of intervention layer is well motivated. Section 3.2 and Appendix B.1 explain how an earlier-layer edit changes the local search geometry at the final representation through the downstream Jacobian. Figure 2 and Appendix C.2 add mechanistic and behavioral evidence, which makes this design choice more convincing.

  2. The coding gains are good. With the same number of candidates, SONAR improves on ODD's best HumanEval empirical pass@16 and pass@8 for both models (Table 1).

Critical weaknesses

The most important missing comparison is at equal compute. SONAR can intervene at steps 4–24 of a 32-step denoising run, with spectral calculations, repeated backward passes, and additional forward proposals (Appendix A). Freezing the weights does not remove the cost of differentiating through the later blocks. No latency, peak memory, or forward/backward counts are reported, and the actual backward budget in Algorithm 1 is not specified in Table 2. The current fixed-candidate comparison leaves open whether ordinary sampling or ODD could match or exceed SONAR by generating more candidates in the same time, including candidates drawn with different sampling settings (e.g., top-p, top-k, temperature, etc.). The results establish an advantage for a fixed number of candidates, but not for a fixed inference budget.

Questions for authors

I'm curious whether the same runtime would be better spent drawing more candidates with ordinary sampling or ODD. Could you compare solved-problem coverage against measured GPU-time on the same hardware with comparable batching, allowing the baselines to generate as many candidates as the budget permits? A particularly useful baseline would combine different temperatures and, where supported by the decoder, top-p/top-k settings, with the configurations and mixture chosen on held-out development data. Please also report latency, peak memory and forward/backward counts.

4 comments
AuthorsOpenReview ↗

Summary and (w1a,q1b)

Summary

We thank reviewer B9mR for their positive feedback on our paper. We are glad that the reviewer finds our choice of intervention layer well motivated (S1) and highlights our strong coding gains (S2).

We address the reviewer's concerns through additional experiments and analyses, summarized below:

  1. (W1a, Q1b) Computational analysis. We provide detailed measurements of inference latency, GPU memory, and forward/backward operations on an NVIDIA H200. Our analysis shows that the main overhead comes from transformer forward/backward passes rather than spectral calculations. Reducing backpropagation depth from 16 to 1 block decreases SONAR's overhead from 3.97× to 1.85× over base sampling.
  2. (W1b, Q1a) Comparison under comparable computational budgets. We conduct extensive experiments combining multiple generation settings for LLaDA, ODD, and SONAR. With a smaller GPU-time budget, SONAR solves 143 problems, compared with 133 for ODD and 129 for LLaDA. Even combining both baselines, with approximately 2.4× SONAR's computational budget, leaves 9 problems unsolved that SONAR discovers solutions to. These results support SONAR's computationally efficient exploration and its ability to uncover reasoning modes missed by extensive conventional sampling.

W1a / Q1b. Computational overhead and forward/backward counts

The dominant factor in SONAR's computational overhead is the number of transformer blocks involved in backpropagation. While our main experiments intervene at layer 16, SONAR also supports interventions closer to the output, substantially reducing computational cost.

We report inference latency and GPU memory in Table R4, a detailed computational breakdown in Table R5, and average forward/backward counts in Table R6. All measurements were performed on an NVIDIA H200 GPU.

Table R4. Wall time and GPU memory comparison of LLaDA, ODD, and SONAR with different backpropagation depths.

MethodWall time (s)MultiplierGPU memory: mean peak / maximum (GiB)
LLaDA4.461.00×16.91 / 17.76
ODD4.921.08×18.82 / 18.91
SONAR, 1 block8.261.85×16.91 / 17.76
SONAR, 4 blocks9.462.12×17.72 / 18.70
SONAR, 16 blocks (used in manuscript)17.713.97×23.38 / 26.94

Table R5. Detailed computational cost breakdown of SONAR (ms).

Component1 block4 blocks16 blocks
Tail forward3.2210.5039.45
Geometry + loss (including EVP)2.933.013.11
EVP alone (included above)0.430.440.43
Local U gradient0.991.011.02
Local V gradient0.580.620.64
Gradient surgery0.190.210.21
Transformer backward3.1912.5249.92

Table R6. Average forward and backward counts per denoising step.

BlocksFull fwd.Partial-block fwd.Partial-block bwd.
11.662.560.66
41.662.580.65
161.663.421.38

These measurements show that spectral calculations contribute relatively little to SONAR's overhead. The primary cost comes from partial forward and backward passes through the transformer, starting at the intervention layer.

Additional details are provided in Appendix F.2.

AuthorsOpenReview ↗

(W1b , Q1a)

W1b / Q1a. Exploration performance under comparable computational budgets and mixed diversity settings

We thank the reviewer for highlighting the importance of comparing methods under equal inference budgets. We agree that combining different settings provides a more comprehensive evaluation of each method's exploration potential.

We therefore conduct a broader comparison using mixed settings for LLaDA, ODD, and SONAR, allowing the baselines to generate substantially more candidates to account for SONAR's measured overhead.

Experimental settings.

  • LLaDA: We use 6 settings combining different temperatures and sample allocations, generating a total of 144 samples per seed (relative GPU budget = 1.00).
  • ODD: We use 9 settings combining temperature and step size α\alpha. Each generates 16 samples, for a total of 144 samples per seed (relative GPU budget = 1.08).
  • SONAR: We use 4 settings (two 1-block and two 4-block interventions). Each generates 16 samples, for a total of 64 samples per seed (relative GPU budget = 0.88, based on Table R4).

The settings for each method are selected through a hyperparameter search to maximize coverage on a small subset of problems.

We evaluate the methods using Coverage@M (Sec. 5.2, Fig. 4), measuring the expected number of unique problems solved when combining MM seed trials.

Table R7. Coverage@M under mixed generation settings and comparable computational budgets.

MethodM=1M=2M=4M=8Relative budget
LLaDA (144 samples/seed)102.875 ± 1.053112.036 ± 1.569121.514 ± 1.8031291.00×
ODD (144 samples/seed)111.750 ± 2.634119.607 ± 2.241127.200 ± 1.5641331.08×
SONAR (64 samples/seed)109.500 ± 3.240121.179 ± 3.163132.029 ± 2.2681430.88×

Although ODD achieves higher expected coverage for a single seed, SONAR discovers substantially more unique problems as exploration is extended across multiple seeds. At M=8M=8, SONAR solves 143 problems, compared with 133 for ODD and 129 for LLaDA, while using less GPU time (see Appendix F.3 for the plot).

This corresponds to 14 additional solved problems over LLaDA and 10 over ODD, using only 512 samples across eight seeds, compared with 1,152 samples for each baseline.

To further examine SONAR's ability to uncover reasoning modes missed by conventional sampling, we restrict the evaluation to problems that remain unsolved by LLaDA.

Table R8. Coverage@M restricted to problems unsolved by LLaDA.

MethodM=1M=2M=4M=8
ODD2.000 ± 1.4143.607 ± 1.5435.857 ± 1.2578
SONAR2.125 ± 0.9274.143 ± 1.1877.857 ± 1.32314

SONAR discovers valid solutions to 14 problems that LLaDA fails to solve, compared with 8 for ODD.

Combined-baseline comparison. We further combine all discoveries from LLaDA and ODD, corresponding to 2,304 generated samples across eight seeds. Together, these baselines solve 137 problems, whereas SONAR alone solves 143, using only 512 samples and approximately 2.4× less GPU time. Importantly, SONAR discovers solutions to 9 problems that remain unsolved by both baselines combined, while the combined baselines solve 3 problems missed by SONAR.

Overall, these experiments support our central hypothesis that exploiting learned representation structure for diversity can reveal lost reasoning modes that remain undiscovered by conventional sampling under a tractable computational budget.


We truly appreciate the time you have taken to help improve our paper. If our response addresses the concerns and questions you raised, we hope you will consider raising your score.

Thank you again for your thoughtful feedback and suggestions. In response to your computational-efficiency concerns, we conducted additional experiments with mixed generation settings under comparable GPU-time budgets, along with detailed runtime and memory measurements. SONAR achieves higher problem coverage than both baselines while using less compute. We would appreciate any further feedback on whether these results address your concerns.

Thank you again for your thoughtful feedback. With the discussion period coming to an end, we wanted to ask whether our additional computational analysis and matched-budget experiments have addressed your questions. We would be happy to provide any further clarification.

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

This paper introduces SONAR, a sampling method for improving reasoning diversity in MDLMs. To quantify diversity, it introduces a spectral overlap measure based on RKHS, which compares the structures of contextual hidden representations across candidate responses.

Critical strengths
  • SONAR separates the diversity measurement and the intervention, which is a reasonable design choice. It measures diversity at the final hidden layer while applying gradient-based interventions at an earlier layer, which allows the perturbations to propagate through subsequent blocks.
  • The experimental results are promising, where SONAR demonstrates improvements over standard sampling and ODD across two MDLMs.
Critical weaknesses
  • The evaluation does not effectively demonstrate that the improvements arise from the proposed RKHS. More comparisons are needed to isolate the contribution.
  • SONAR requires additional computations. But the paper does not provide a computational analysis, making it difficult to assess whether the improvements justify the additional inference overhead.
  • A higher algorithmic novelty score does not imply that the model discovers more effective reasoning strategies. The current results do not show if SONAR increases the diversity of correct solutions.
Questions for authors
  • Could you provide comparisons with other diversity methods, such as cosine similarity, embedding repulsion, and Maximum Mean Discrepancy?
  • Does SONAR still outperform the baselines under matched computational budgets?
  • Could you compute algorithmic novelty using only correct solutions and report the number of distinct correct solutions per problem?
4 comments
AuthorsOpenReview ↗

Summary and (W1,Q1)

Summary

We thank the reviewer for their constructive feedback. We are glad that the reviewer finds our separation of diversity measurement and intervention a reasonable design choice (S1), and our experimental improvements over the baselines promising (S2).

We address the reviewer's concerns through additional experiments and analyses, summarized below:

  1. (W1, Q1) Contribution of holistic spectral comparison. We implement and compare two additional diversity methods, MMD and Cosine-Mean-DPP. We investigate the role of nonlinear RKHS representations and whether retaining richer spectral structure helps uncover overlooked reasoning modes. SONAR achieves the highest overall coverage and discovers more problems missed by conventional sampling.
  2. (W2, Q2) Computational efficiency. We provide detailed runtime and GPU-memory measurements and demonstrate that SONAR outperforms conventional sampling and ODD under comparable computational budgets, while its overhead can be controlled by changing the intervention depth.
  3. (Q3) Correct Algorithmic Novelty. We revisit our algorithmic novelty analysis using only correct solutions and introduce the Expected Fraction of Correct Algorithms metric. SONAR discovers the most correct algorithms (199), achieves the highest expected fraction (90.32%), and discovers 82.88% of the correct algorithms missed by the baseline, averaged per problem.

W1. The evaluation does not effectively demonstrate that the improvements arise from the proposed RKHS. More comparisons are needed to isolate the contribution.

Q1. Could you provide comparisons with other diversity methods, such as cosine similarity, embedding repulsion, and Maximum Mean Discrepancy?

To our knowledge, our baseline ODD is the closest existing method explicitly designed to diversify MDLM samples. Regarding the suggested alternatives, we clarify that SONAR works via embedding repulsion, while cosine similarity and MMD are general comparison measures rather than established MDLM exploration methods. Nevertheless, we agree that investigating these alternatives within the same exploration framework is valuable.

SONAR’s Connection to MMD: MMD is also an RKHS-based approach and is closely related to the Gaussian bump interpretation presented in our paper (see Fig. 3 and L240). Under an RBF kernel, each token representation defines a Gaussian bump, and linear combinations of these bumps span a rich family of functions. MMD compares responses by averaging (equal weight) these bumps, represented by a single vector in RKHS. In contrast, SONAR uses the spectral structure of the RKHS subspaces spanned by these functions to measure how much one response can explain another.

This distinction motivates our second contribution: investigating whether comparing richer subspace structures, rather than collapsing responses into mean vectors, helps discover reasoning modes missed by conventional sampling.

We investigate this by implementing two alternative diversity losses : (i) MMD and (ii) Cosine-Mean-DPP. Comparison with these alternatives isolates two aspects of our design:

  • Cosine-Mean-DPP vs. MMD² : Whether using an RKHS representation provides an advantage over cosine-based repulsion of mean embeddings.
  • MMD² vs. SONAR: Whether retaining the richer spectral subspace structure provides any benefits over comparing a single RKHS mean vector.

Settings. We conduct an independent hyperparameter search per method on a small subset of problems, restricting intervention to just 1 block to accelerate both search and evaluation. We combine 4 diversity settings per method that maximize their problem coverage on the subset. As also noted by reviewer B9mR, combining different settings shows a more robust picture of the exploration potential of a method.

In Table R1, for LLaDA, we generate 144 samples per seed with 6 combinations of temperature and sample count. For MMD, Cosine, and SONAR, we select 4 settings with 16 samples each (total 64 samples per seed). SONAR with 1-block backprop require ~1.9x compute over the base LLaDA (see Table R2 in response to W2, Q2) which makes its relative budget around 0.84 (1.9*64/144).

Table R1. Comparison of SONAR with base model and additional diversity objectives (over 8 seeds).

MethodAvg solved/seed ± SDUnique Problems Solved / 164Unique solved among 35 problems LLaDa fails onRelative Budget
LLaDA102.875 ± 1.05312901x
MMD105.875 ± 2.84813470.84x
Cosine100.750 ± 4.46512640.84x
SONAR108.875 ± 2.26013790.84x

These results support our central design choice. SONAR discovers more solutions to problems missed by extensive conventional sampling (9 vs. 7 for MMD² and 4 for Cosine-Mean-DPP), at comparable or less budget. Please refer to App F.1 for more details.

AuthorsOpenReview ↗

(w2,q2) and q3

W2 / Q2. Computational efficiency under comparable inference budgets

The dominant factor in SONAR's computational overhead is the number of transformer blocks involved in backpropagation. While our main experiments intervene at layer 16, SONAR allows interventions closer to the output, substantially reducing computational cost. As shown in Table R2, reducing backpropagation from 16 blocks to 4 or 1 block decreases the runtime overhead.

Table R2. Runtime and GPU memory comparison on NVIDIA H200.

MethodWall time (s)MultiplierGPU memory: mean peak / maximum (GiB)
LLaDA4.461.00×16.91 / 17.76
ODD4.921.08×18.82 / 18.91
SONAR, 1 block8.261.85×16.91 / 17.76
SONAR, 4 blocks9.462.12×17.72 / 18.70
SONAR, 16 blocks17.713.97×23.38 / 26.94

Moreover, different intervention depths also change the geometry through which the diversity objective is optimized, as discussed in section 3.2 of the manuscript. Therefore, combining different intervention depths therefore offers an opportunity to explore different geometry bias.

Following the mixed-setting evaluation discussed in W1/Q1 and emphasized by reviewer B9mR, we combine multiple configurations to better evaluate each method's exploration potential. We select configurations through a hyperparameter search on a small subset of problems:

  • LLaDA: 6 settings combining temperatures and sample allocations, totaling 144 samples per seed.
  • ODD: 9 settings combining temperature and diversity step size, totaling 144 samples per seed.
  • SONAR: 4 settings (two 1-block and two 4-block intervention profiles), totaling 64 samples per seed.

Based on the measured inference costs, the relative GPU-time budgets are 1.00× for LLaDA, 1.08× for ODD, and 0.88× for SONAR.

We evaluate exploration using Coverage@M (Sec. 5.2, Fig. 4), which measures the expected number of unique problems solved when combining (M) independent seed trials.

Table R3. Coverage@M under mixed settings and comparable computational budgets.

MethodM=1M=2M=4M=8Relative budget
LLaDA (144 samples/seed)102.875 ± 1.053112.036 ± 1.569121.514 ± 1.8031291.00×
ODD (144 samples/seed)111.750 ± 2.634119.607 ± 2.241127.200 ± 1.5641331.08×
SONAR (64 samples/seed)109.500 ± 3.240121.179 ± 3.163132.029 ± 2.2681430.88×

Although ODD achieves higher coverage for a single seed, SONAR achieves higher coverage as exploration extends across multiple seeds. At (M=8), SONAR solves 143 problems, compared with 133 for ODD and 129 for LLaDA, while using less GPU time and generating only 512 samples versus 1,152 for each baseline.

Furthermore, SONAR discovers solutions to 14 problems missed by extensive LLaDA sampling, compared with 8 for ODD. Even combining both baselines (LLaDA and ODD for a total of 2,304 samples), SONAR discovers 9 problems that neither baseline solves, despite using approximately 2.4× less GPU time. For further details on computational cost, please refer to our response to reviewer B9mR and Appendix (F2 - F3).

These results demonstrate that SONAR can achieve greater problem coverage under a smaller computational budget.

Q3. algorithmic novelty using only correct solutions and number of distinct correct solutions per problem

We measure the correct unique problem-algorithm pairs in Table R4, which shows the total number of unique algorithms with correct solutions, discovered across all problems over 8 seeds. SONAR discovers 68 more correct algorithms over the base model, while ODD discovers 25.

Further, in line with the reviewer’s question, we ask, among all the correct algorithms discovered per problem, what fraction is discovered by SONAR?

To measure this, we introduce the Expected Fraction of Correct Algorithm metric. For each problem, we normalize each method’s correct algorithms by the total correct algorithm discovered by the union of all methods. Then we take the expectation over all problems in HumanEval.

Table R4. Correct Novel Algorithms and Expected Fraction of Correct Algorithms (among union of all methods).

MethodUnique Correct AlgorithmExpected Fraction of Correct AlgorithmExpected Fraction of Correct Algorithm Missed by Base
Base LLaDA13160.45 %0.0%
ODD15670.38 %47.72%
SONAR19990.32 %82.88%

So in addition to discovering more novel algorithms, SONAR also discovers more correct algorithms. These results support our motivation of discovering hidden reasoning in MDLMS. Additional figures provided in Appendix F4.


Thank you again for your thoughtful feedback. If our clarifications have successfully addressed your concerns and questions, we would be very grateful if you consider raising your score.

Thank you again for your constructive feedback and for taking the time to review our paper. Following your suggestions, we implemented two additional diversity methods (MMD and Cosine-Mean-DPP), conducted matched-compute experiments, and extended our algorithmic novelty analysis to correct solutions. We hope these additional results address your concerns and would welcome any further questions or feedback during the remaining discussion period.

Thank you again for your time and feedback. As today is the last day of the discussion period, we wanted to check whether our additional experiments and analyses have addressed your concerns. We would be happy to clarify anything further if you have any remaining questions.

Comments on the paper

1 comment

Summary of Additional Experiments and Rebuttal

We thank all reviewers and the Area Chair for their time and feedback. With about one day left in the discussion period, we wanted to briefly summarize the additional experiments we conducted in response to the reviews.

  1. Additional diversity methods: We implemented MMD and Cosine-Mean-DPP within the same exploration framework. SONAR achieves higher problem coverage (137 vs. 134 for MMD and 126 for Cosine-Mean-DPP) and discovers more problems missed by conventional sampling (9 vs. 7 and 4). These results support our hypothesis that retaining richer spectral structure, rather than reducing each response to a single mean vector, helps discover overlooked reasoning modes.

  2. Computational analysis: We measured inference latency, GPU memory, and forward/backward operations on an NVIDIA H200. Most of SONAR's overhead comes from transformer forward/backward passes rather than spectral computations. Reducing backpropagation from 16 blocks to 1 lowers the runtime overhead from 3.97× to 1.85×.

  3. Comparison under comparable compute: We combined multiple settings for each method and allowed the baselines to generate more samples. SONAR solves 143 HumanEval problems, compared with 133 for ODD and 129 for LLaDA, while using less GPU time. Even after combining both baselines (2,304 samples per problem), SONAR discovers 9 problems missed by both while generating only 512 samples per problem.

  4. Correct algorithmic novelty: We revisited our original algorithmic novelty analysis, considering only correct solutions. SONAR discovers 199 unique correct algorithms, compared with 156 for ODD and 131 for LLaDA. It also achieves the highest expected fraction of correct algorithms discovered per problem (90.32%).

  5. Updated GSM8K results: We explored a broader range of sampling temperatures and diversity strengths, improving SONAR's results on both LLaDA and Dream.

We have included the additional results, plots, and analyses in our responses and the revised manuscript (Appendix F).

We hope these results address the reviewers' concerns. We would be happy to discuss any remaining questions before the discussion period ends.

Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.