Pith. sign in

REVIEW 4 major objections 5 minor 51 references

By planning slot-filling order with Monte Carlo Tree Search at inference time, this paper claims, masked diffusion language models can beat greedy plan-and-infill decoding and autoregressive baselines on average, with the largest gains in c

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 23:44 UTC pith:HMV6ST7L

load-bearing objection A promising training-free MCTS slot-planning idea for masked diffusion LMs, but the reported gains are confounded by simultaneous decoding-parameter changes and an inconsistent hyperparameter report; the central empirical claim is not yet supported. the 4 major comments →

arxiv 2602.12586 v2 pith:HMV6ST7L submitted 2026-02-13 cs.AI

Can I Have Your Order? Monte-Carlo Tree Search for Slot Filling Ordering in Diffusion Language Models

Joshua Ong Jun Leang , Yu Zhao , Mihaela C\u{a}t\u{a}lina Stoian , Wenda Li , Shay B. Cohen , Eleonora Giunchiglia This is my paper
classification cs.AI
keywords masked diffusion modelsMonte Carlo Tree Searchslot-filling orderingplan-and-infill decodinginference-time planningnon-autoregressive generationcode generationreasoning benchmarks
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Masked diffusion language models generate text by infilling masked slots, and the order in which those slots are filled strongly affects output quality. This paper claims that slot order can be planned in advance, at inference time and without retraining, by framing slot selection as a sequential decision problem and solving it with Monte Carlo Tree Search: the planner simulates full completion trajectories, using the model's own token-confidence scores as rewards, before committing to the next slot. On six reasoning and code benchmarks, the resulting orderings beat the greedy plan-and-infill baseline by 8.0% on average and autoregressive models by 3.2%, with a 19.45% jump on MBPP. Two analyses carry the argument: rare non-sequential orderings, not the mostly left-to-right default, drive a disproportionate share of the gains; and a large exploration constant, rather than more simulations, is what lets the search find them.

Core claim

On its own terms, the paper claims that slot selection in plan-and-infill masked diffusion models is a solvable decision problem, not a heuristic to be tuned once. MCDIFFUSE treats the partially filled sequence as a state and the choice of which masked slot to fill next as an action; slot-level token confidence serves as both the immediate reward and the prior that guides tree search. Before committing to any slot, Monte Carlo simulations complete the remaining slots stochastically, and the average confidence of those trajectories is backpropagated; after a fixed budget, the planner commits to the most-visited action. Across six reasoning and code benchmarks, this planner reports the highest

What carries the argument

The carrying mechanism is slot-level confidence — the model's mean token probability for a slot — used two ways. As a reward R(s,a) it scores the immediate quality of filling a slot; normalised across unfilled slots it becomes the prior P(a|s) in a PUCT selection rule, biasing early search toward confident actions while visit counts eventually override that bias. The look-ahead term is a stochastic confidence rollout: a temperature-scaled softmax over remaining slots' confidences samples a completion path, and the average confidence along it estimates long-term coherence. The hybrid value V = λ·R + (1−λ)·G (λ = 0.3) is backpropagated up the tree. This lets the search override locally confide

Load-bearing premise

The load-bearing premise is that the experiments isolate the planner: MCDIFFUSE is run with a different slot size, serial-block count, and confidence thresholds than the plan-and-infill baseline it is compared against (section C.1), and the autoregressive baselines are kept at a 512-token cap that the paper's own extended-context comparison (Table 6) shows is decisive — given longer budgets, autoregressive models reach 63.4% on MATH500 versus 54.6% for MCDIFFUSE.

What would settle it

Re-run MCDIFFUSE on top of the plan-and-infill baseline with every hyperparameter held at the baseline's values — slot size 8, 2 serial blocks, confidence thresholds 0.9 — so that only the slot-selection rule changes. If the average gain over the baseline collapses or reverses, the reported improvement is not caused by the search. Separately, give the autoregressive baselines the same token budget as MCDIFFUSE on each benchmark; if the head-to-head advantage shrinks, part of the result is a token-budget artefact, not planning.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Slot-filling order is a decision variable worth optimising: treating it as a sequential decision problem and searching over it at inference time closes, and on five of six benchmarks reverses, the accuracy gap between masked diffusion models and autoregressive models.
  • Gains concentrate on structured tasks: the paper reports 19.45% and 16.32% absolute accuracy jumps on MBPP and HumanEval, consistent with code's strict ordering dependencies such as declarations before use, function headers, and indentation.
  • Exploration breadth beats simulation depth: larger exploration constants reliably improve accuracy, while increasing simulations under low exploration can actively hurt by entrenching locally confident but globally myopic orderings.
  • Rare non-sequential decisions carry the gains: the planner follows left-to-right order for roughly 91–94% of decisions, but among samples where it succeeds and sequential generation fails, 60.7% involve a non-sequential deviation — a purely sequential planner would forfeit most of the improvement.
  • Planned infilling is token-efficient: on reasoning-heavy benchmarks it produces an average 64.77% fewer tokens than the autoregressive baseline while improving accuracy, with reductions reported as statistically significant (p < 0.001).

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: given that slot size, serial-block count, and confidence thresholds change alongside the planner in the reported configuration (section C.1), and tuning was done on a single benchmark (section C.2), the 8.0% average gain is best read as 'search plus its tuned configuration,' not as the isolated effect of search; a matched planner on/off ablation would separate the two.
  • Editorial inference: the token-budget asymmetry cuts the other way too — the paper's own extended-context results (Table 6) show autoregressive models reaching 63.4% on MATH500 versus 54.6% for MCDIFFUSE, so the headline parity claim holds inside a fixed budget; the durable advantage may be token efficiency rather than raw ceiling accuracy.
  • Testable extension: because the reward signal is purely the model's own confidence, the same planner could order spans in other confidence-scored decoders — for instance, deciding which proof step or code block to elaborate next in an autoregressive draft-then-verify loop — turning slot ordering into a general test-time planning knob.
  • Testable extension: the exploration-width-over-depth result suggests a cheap proxy — a prior learned from low-budget searches, or parallel independent searches with a large exploration constant — could capture most of the gain at a fraction of the 256-simulation wall-clock cost.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes MCDIFFUSE, a training-free framework that uses Monte Carlo Tree Search (MCTS) to select the order in which masked slots are infilled by a masked diffusion language model (MDM). Slot selection is formulated as an MDP: states are partial slot fillings, actions are choices of the next slot, and rewards are slot-level confidence scores from the generative model. MCTS with PUCT, temperature-scaled rollouts, and a mixing coefficient between immediate confidence and rollout value selects the next slot. The authors evaluate on six reasoning and code benchmarks, reporting an average gain of 3.2% over autoregressive baselines and 8.0% over the plan-and-infill baseline ReFusion, with particularly large gains on MBPP and MATH500. They further analyze the sequentiality of the selected orderings and the interaction between exploration constant and simulation budget.

Significance. If the empirical claims were supported, MCDIFFUSE would be a useful contribution: a training-free way to improve diffusion LM decoding via lookahead slot planning, with an interesting and falsifiable analysis showing that rare non-sequential decisions drive accuracy and that exploration breadth matters more than simulation depth. The paper gives a complete algorithm description, a worked numerical example, and extensive benchmark tables. However, the experimental design as presented does not isolate the effect of MCTS from simultaneous changes in decoding hyperparameters, and the comparison to autoregressive baselines is sensitive to the token budget. These issues must be resolved before the central claim can be accepted.

major comments (4)
  1. [§C.1, Table 1] The headline comparison against ReFusion is confounded. §C.1 states that MCDIFFUSE uses slot size 4, 32 serial blocks, slot threshold 0.5, and token threshold 0.6, while ReFusion uses slot size 8, 2 serial blocks, slot threshold 0.9, and token threshold 0.9. These are not minor details: slot size changes the conditional-independence granularity, and thresholds change how many tokens are unmasked per step. The reported gains (e.g., +19.45 on MBPP, +16.32 on HumanEval) therefore cannot be attributed to the MCTS ordering policy without a control that fixes all decoding hyperparameters and varies only the ordering mechanism. In addition, §4 reports c=50 and Nsim=256, but §C.1 reports 30 simulations and c=10.0; §C.2 uses cpuct=50 with Nsim=256. It is unclear which configuration produced Table 1.
  2. [§D.3, Table 6] The claim that MCDIFFUSE 'matches or exceeds' autoregressive performance is budget-dependent. §C.1 caps all models at 512 tokens (1024 for MBPP). Table 6 shows that with extended context (Qwen2.5/Qwen3 at 32,768 tokens), both ARMs outperform MCDIFFUSE on MATH500 (63.40% and 68.00% vs 54.60%). The abstract's 3.2% average improvement over AR baselines is therefore an artifact of the 512-token cap. The authors should either re-evaluate AR baselines without truncation or restrict the claim to 'under identical token budgets' and discuss the trade-off explicitly.
  3. [Algorithm 5, Eq. (5)] The backpropagation update appears inconsistent with the value definition. Eq. (5) defines V(s_t,a) as the value of taking action a at s_t, combining the one-step reward and a rollout estimate from the successor. However, Algorithm 5 backpropagates the same V to all ancestors without adding the rewards of intermediate edges along the path. In standard MCTS, the Q-value of an edge at depth d should be updated with the cumulative return from that depth. As written, Q(s,a) in Eq. (10) for non-root nodes is an average of leaf-value estimates rather than the value of the corresponding action, which biases the PUCT criterion in Eq. (2) and the final robust-child selection in Eq. (11). Please correct the pseudocode and verify the implementation.
  4. [Eq. (3)–(4)] The reward and the prior are both derived from the model's own confidence scores, so the search optimizes the model's self-assessment. This is not logically circular because final accuracy is measured externally, but it is a correctness risk: if confidence is miscalibrated, the rollout values in Eq. (7) may not track generation quality. The paper cites evidence that confidence correlates with accuracy, but no calibration analysis is provided for the specific ReFusion model. At minimum, the hyperparameter-controlled ablation requested above with a confidence-greedy ordering policy would test whether MCTS adds value beyond confidence-based ranking.
minor comments (5)
  1. [§D.1, Table 4] The stated average token reduction (64.77%) and average accuracy improvement (3.38 percentage points) are inconsistent with the table: the simple averages of the reported per-dataset values are approximately 71.3% and 6.3 points. Please correct the arithmetic or clarify the weighting.
  2. [§H, final paragraph] The text refers to 'Figure 7b ... in Figure 7a' for MATH500, but the figure caption identifies (a) as MBPP and (b) as MATH500. The cross-reference is inconsistent.
  3. [§F.2] The worked example uses c=1.4, while the main configuration in §4 is c=50. Please state that the walkthrough uses illustrative values.
  4. [§C.1, Table 6] §C.1 says 'All models use an evaluation length of 512 tokens', but Table 6 lists MCDIFFUSE at 1024 tokens. Please clarify the exact budget used for each model in each table.
  5. [References] The citation 'Granter et al., 2017' for AlphaGo is incorrect; that paper is a commentary in pathology, not the AlphaGo system paper. Please cite the original Nature paper.

Circularity Check

1 steps flagged

MATH500 headline is grid-selected on the same benchmark; the MCTS objective is self-referential, but the central empirical claim retains independent content.

specific steps
  1. fitted input called prediction [Section C.2 (Hyperparameter Tuning, Table 3); results in Table 1; claimed in Abstract]
    "We further perform an additional hyperparameter tuning on the temperature, τ, and λ. ... We perform the hyperparameter tuning on MATH-500 dataset. ... Based on the table, we can observe that the best hyperparameter tuning lies on τ = 0.5, and λ= 0.3 ."

    The paper selects λ and τ by maximizing MATH500 accuracy in Table 3 and then reports MCDIFFUSE's MATH500 accuracy in Table 1 (47.80%) and advertises a 4.9% gain in the abstract. The reported MATH500 number is therefore the grid-selected optimum (48.20% in Table 3) of the same benchmark: the hyperparameters are fit to the very result that is presented as an outcome, so that particular 'prediction' is statistically forced by the selection rather than being an independent evaluation.

full rationale

The MCTS objective is defined entirely through the model's own confidence: the prior (Eq. 4) is the normalized slot-level confidence R (Eq. 3), the rollout return G (Eq. 7) averages the same R, and the leaf value (Eq. 5) mixes R and G. So the planner re-ranks slots by the model's self-assigned probabilities rather than by any external correctness signal. This is self-referential, but not logically circular: final Pass@1 accuracy is measured against external benchmarks, so the empirical claim that MCTS slot planning improves generation quality is falsifiable and is not entailed by the reward definition. The one fitted-called-prediction step is the MATH500 hyperparameter selection: λ, τ (and apparently c, Nsim) are chosen by grid search on MATH500, and then MATH500 appears in Table 1 and the abstract as a 4.9% gain; that number is a grid-selected value, not an independent prediction. The remaining five benchmarks and the qualitative ordering (MCTS > ReFusion > Random/Sequential) are not fitted to the same test sets, so the central claim still has independent content. No load-bearing self-citation or imported uniqueness theorem was found: the confidence-correlation premise cites external work in addition to the authors' own. The C.1/C.2 hyperparameter discrepancy is a reproducibility/confound issue rather than a circularity issue.

Axiom & Free-Parameter Ledger

8 free parameters · 5 axioms · 0 invented entities

The method rests on the assumption that a model's own token confidence is a usable reward for planning, on MCTS/PUCT convergence, and on several hyperparameters tuned on MATH500. The largest free parameters are the decoding settings that differ between MCDIFFUSE and the ReFusion baseline; these confound the headline comparison.

free parameters (8)
  • lambda (mixing coefficient) = 0.3
    Chosen by grid search on MATH500 (Table 3); balances immediate reward vs rollout estimate in Eq. 5.
  • cpuct exploration constant = 50 (main text), 10 (Appendix C.1)
    Exploration constant in PUCT; tuned on MATH500; inconsistent values reported across sections.
  • Nsim simulation budget = 256 (main text), 30 (Appendix C.1)
    Number of MCTS simulations; tuned; inconsistent values reported across sections.
  • tau rollout temperature = 0.5
    Temperature in rollout softmax (Eq. 6); tuned on MATH500.
  • slot size = 4 (MCDIFFUSE) vs 8 (ReFusion)
    Changed between method and baseline; not isolated from MCTS effect.
  • serial blocks = 32 (MCDIFFUSE) vs 2 (ReFusion)
    Changed between method and baseline; affects model parallelism and decoding behavior.
  • slot threshold = 0.5 (MCDIFFUSE) vs 0.9 (ReFusion)
    Changed between method and baseline.
  • token threshold = 0.6 (MCDIFFUSE) vs 0.9 (ReFusion)
    Changed between method and baseline.
axioms (5)
  • domain assumption Slot-level confidence P_theta correlates with downstream correctness.
    Reward in Eq. 3 uses confidence as a quality signal; authors cite prior work (Leang et al. 2025b; Nie et al. 2025), but it remains an assumption about the base model.
  • ad hoc to paper The rollout average G (Eq. 7) is a sufficient statistic for long-term trajectory value.
    No proof is offered; hyperparameters lambda and tau are tuned to make this heuristic work.
  • domain assumption Conditional independence among masked tokens in MDMs permits arbitrary slot ordering.
    Standard MDM assumption, invoked in Sections 1 and 2.
  • standard math PUCT/MCTS converges to a good action in the simulation budget used.
    Background theorem (Kocsis & Szepesvari 2006); the finite-budget behavior is empirical and tuned.
  • domain assumption Final-answer extraction and Pass@1 scoring are fair across all compared models.
    Not detailed in the paper; assumes the metric does not systematically favor MCDIFFUSE's shorter outputs.

pith-pipeline@v1.3.0-alltime-deepseek · 22655 in / 12911 out tokens · 103611 ms · 2026-08-02T23:44:59.502818+00:00 · methodology

0 comments
read the original abstract

While plan-and-infill decoding in Masked Diffusion Models (MDMs) shows promise for mathematical and code reasoning, performance remains highly sensitive to slot infilling order, often yielding substantial output variance. We introduce McDiffuSE, a framework that formulates slot selection as decision making and optimises infilling orders through Monte Carlo Tree Search (MCTS). McDiffuSE uses look-ahead simulations to evaluate partial completions before commitment, systematically exploring the combinatorial space of generation orders. Experiments show an average improvement of 3.2% over autoregressive baselines and 8.0% over baseline plan-and-infill, with notable gains of 19.5% on MBPP and 4.9% on MATH500. Our analysis reveals that while McDiffuSE predominantly follows sequential ordering, incorporating non-sequential generation is essential for maximising performance. We observe that larger exploration constants, rather than increased simulations, are necessary to overcome model confidence biases and discover effective orderings. These findings establish MCTS-based planning as an effective approach for enhancing generation quality in MDMs.

Figures

Figures reproduced from arXiv: 2602.12586 by Eleonora Giunchiglia, Joshua Ong Jun Leang, Mihaela C\u{a}t\u{a}lina Stoian, Shay B. Cohen, Wenda Li, Yu Zhao.

Figure 1
Figure 1. Figure 1: Overview of MCDIFFUSE. We formulate slot selection as a sequential decision-making process optimised via Monte Carlo Tree Search. As illustrated in the Statistics box, the model’s greedy prior (P(a = 1 | s0) = 0.37) favours immediately generating the function definition (i.e., slot 2: “def get max length(words):”). However, through look-ahead simulations, the search algorithm discovers that starting with t… view at source ↗
Figure 3
Figure 3. Figure 3: Impact of exploration constant (c) and simulation budget (Nsim) on task performance. Config Mean H (bits) Std H Median H Concentration Low Exploration (c = 2) N = 30 1.4062 0.3785 1.4824 0.4958 N = 270 1.1842 0.4043 1.1492 0.6118 High Exploration (c = 100) N = 30 1.4176 0.3771 1.5058 0.4954 N = 270 1.4340 0.3768 1.5211 0.4879 [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of ReFusion and MCDIFFUSE on a coding prompt from MBPP. Superscripts denote the infilling slot order and colours indicate the specific generation step. Regarding computational overhead, generation time scales approximately linearly with Nsim while varying c causes negligible cost as it only affects selection without requir￾ing additional forward passes. We observe an accuracy￾efficiency trade-of… view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of token reduction across models. We observe that MCDIFFUSE significantly reduces tokens while generating responses, demonstrating the compactness and coherence of using MCDIFFUSE. D.2. Task-Specific Improvements To investigate whether MCTS-based slot planning benefits different task types differentially, we analyse the absolute accuracy improvements from ReFusion to MCDIFFUSE across our evaluat… view at source ↗
Figure 6
Figure 6. Figure 6: demonstrates a comparative analysis of sequential slot selection behaviour and its relationship to accuracy on the MBPP code generation benchmark for both ReFusion (baseline) and MCDIFFUSE. The baseline ReFusion method (left panel, 58.81% accuracy) shows a weak positive correlation between sequential ordering and task success. In contrast, MCDIFFUSE (right panel, 72.43% accuracy) demonstrates a more strate… view at source ↗
Figure 7
Figure 7. Figure 7: Accuracy versus generation time across different hyperparameter settings on coding and mathematical reasoning tasks. over Nsim = 30 is marginal (typically < 2%), whereas the computational cost increases ninefold. This supports our earlier finding that simulation budgets suffice when exploration is adequate. Although MCTS introduces additional overhead compared to the baseline ReFusion inference, a high-exp… view at source ↗
Figure 8
Figure 8. Figure 8: Impact of exploration constant (cpuct) and simulation budget on task performance on MATH500 mathematical reasoning tasks. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

51 extracted references · 1 canonical work pages

  1. [1]

    S., Gokaslan, A., Yang, Z., Qi, Z., Han, J., Chiu, J

    Arriola, M., Sahoo, S. S., Gokaslan, A., Yang, Z., Qi, Z., Han, J., Chiu, J. T., and Kuleshov, V. Block diffusion: Interpolating between autoregressive and diffusion language models. In International Conference on Learning Representations, 2025

    2025

  2. [2]

    Program synthesis with large language models

    Austin, J., Odena, A., Nye, M., Bosma, M., Michalewski, H., Dohan, D., Jiang, E., Cai, C., Terry, M., Le, Q., et al. Program synthesis with large language models. ArXiv preprint, abs/2108.07732, 2021

    Pith/arXiv arXiv 2021

  3. [3]

    A markovian decision process

    Bellman, R. A markovian decision process. Indiana University Mathematics Journal, 6: 0 679--684, 1957

    1957

  4. [4]

    Accelerated sampling from masked diffusion models via entropy bounded unmasking

    Ben-Hamu, H., Gat, I., Severo, D., Nolte, N., and Karrer, B. Accelerated sampling from masked diffusion models via entropy bounded unmasking. ArXiv preprint, abs/2505.24857, 2025

    Pith/arXiv arXiv 2025

  5. [5]

    Bie, T., Cao, M., Chen, K., Du, L., Gong, M., Gong, Z., Gu, Y., Hu, J., Huang, Z., Lan, Z., et al. Llada2. 0: Scaling up diffusion language models to 100b. ArXiv preprint, abs/2512.15745, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    B., Powley, E., Whitehouse, D., Lucas, S

    Browne, C. B., Powley, E., Whitehouse, D., Lucas, S. M., Cowling, P. I., Rohlfshagen, P., Tavener, S., Perez, D., Samothrakis, S., and Colton, S. A survey of monte carlo tree search methods. IEEE Transactions on Computational Intelligence and AI in games, 4 0 (1): 0 1--43, 2012

    2012

  7. [7]

    M.-B., Winands, M

    Chaslot, G. M.-B., Winands, M. H., and van Den Herik, H. J. Parallel monte-carlo tree search. In International Conference on Computers and Games, pp.\ 60--71. Springer, 2008

    2008

  8. [8]

    Evaluating large language models trained on code

    Chen, M. Evaluating large language models trained on code. ArXiv preprint, abs/2107.03374, 2021

    Pith/arXiv arXiv 2021

  9. [9]

    Think you have solved question answering? try arc, the ai2 reasoning challenge

    Clark, P., Cowhey, I., Etzioni, O., Khot, T., Sabharwal, A., Schoenick, C., and Tafjord, O. Think you have solved question answering? try arc, the ai2 reasoning challenge. ArXiv preprint, abs/1803.05457, 2018

    Pith/arXiv arXiv 2018

  10. [10]

    Training verifiers to solve math word problems

    Cobbe, K., Kosaraju, V., Bavarian, M., Chen, M., Jun, H., Kaiser, L., Plappert, M., Tworek, J., Hilton, J., Nakano, R., Hesse, C., and Schulman, J. Training verifiers to solve math word problems. ArXiv preprint, abs/2110.14168, 2021

    Pith/arXiv arXiv 2021

  11. [11]

    Efficient selectivity and backup operators in monte-carlo tree search

    Coulom, R. Efficient selectivity and backup operators in monte-carlo tree search. In International conference on computers and games, pp.\ 72--83. Springer, 2006

    2006

  12. [12]

    From bits to rounds: Parallel decoding with exploration for diffusion language models

    Fu, H., Huang, B., Adams, V., Wang, C., Srinivasan, V., and Jiao, J. From bits to rounds: Parallel decoding with exploration for diffusion language models. ArXiv preprint, abs/2511.21103, 2025

    arXiv 2025

  13. [13]

    Interpretable contrastive monte carlo tree search reasoning

    Gao, Z., Niu, B., He, X., Xu, H., Liu, H., Liu, A., Hu, X., and Wen, L. Interpretable contrastive monte carlo tree search reasoning. ArXiv preprint, abs/2410.01707, 2024

    Pith/arXiv arXiv 2024

  14. [14]

    Scaling diffusion language models via adaptation from autoregressive models

    Gong, S., Agarwal, S., Zhang, Y., Ye, J., Zheng, L., Li, M., An, C., Zhao, P., Bi, W., Han, J., et al. Scaling diffusion language models via adaptation from autoregressive models. ArXiv preprint, abs/2410.17891, 2024

    Pith/arXiv arXiv 2024

  15. [15]

    Diffucoder: Understanding and improving masked diffusion models for code generation

    Gong, S., Zhang, R., Zheng, H., Gu, J., Jaitly, N., Kong, L., and Zhang, Y. Diffucoder: Understanding and improving masked diffusion models for code generation. ArXiv preprint, abs/2506.20639, 2025

    Pith/arXiv arXiv 2025

  16. [16]

    R., Beck, A

    Granter, S. R., Beck, A. H., and Papke Jr, D. J. Alphago, deep learning, and the future of the human microscopist. Archives of pathology & laboratory medicine, 141 0 (5): 0 619--621, 2017

    2017

  17. [17]

    L., Liu, Y., Shang, N., Sun, Y., Zhu, Y., Yang, F., and Yang, M

    Guan, X., Zhang, L. L., Liu, Y., Shang, N., Sun, Y., Zhu, Y., Yang, F., and Yang, M. rstar-math: Small llms can master math reasoning with self-evolved deep thinking. ArXiv preprint, abs/2501.04519, 2025

    Pith/arXiv arXiv 2025

  18. [18]

    Measuring mathematical problem solving with the math dataset

    Hendrycks, D., Burns, C., Kadavath, S., Arora, A., Basart, S., Tang, E., Song, D., and Steinhardt, J. Measuring mathematical problem solving with the math dataset. NeurIPS, 2021

    2021

  19. [19]

    Denoising diffusion probabilistic models

    Ho, J., Jain, A., and Abbeel, P. Denoising diffusion probabilistic models. In Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., and Lin, H. (eds.), Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020

    2020

  20. [20]

    Train for the worst, plan for the best: Understanding token ordering in masked diffusions

    Kim, J., Shah, K., Kontonis, V., Kakade, S., and Chen, S. Train for the worst, plan for the best: Understanding token ordering in masked diffusions. ArXiv preprint, abs/2502.06768, 2025

    Pith/arXiv arXiv 2025

  21. [21]

    and Szepesv \'a ri, C

    Kocsis, L. and Szepesv \'a ri, C. Bandit based monte-carlo planning. In European conference on machine learning, pp.\ 282--293. Springer, 2006

    2006

  22. [22]

    S., Reid, M., Matsuo, Y., and Iwasawa, Y

    Kojima, T., Gu, S. S., Reid, M., Matsuo, Y., and Iwasawa, Y. Large language models are zero-shot reasoners. In Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., and Oh, A. (eds.), Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28...

    2022

  23. [23]

    Leang, J. O. J., Gema, A. P., and Cohen, S. B. Comat: Chain of mathematically annotated thought improves mathematical reasoning. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pp.\ 20256--20285, 2025 a

    2025

  24. [24]

    Leang, J. O. J., Zhao, Z., Gema, A. P., Yang, S., Kwan, W.-C., He, X., Li, W., Minervini, P., Giunchiglia, E., and Cohen, S. B. Picsar: Probabilistic confidence selection and ranking for reasoning chains. ArXiv preprint, abs/2508.21787, 2025 b

    Pith/arXiv arXiv 2025

  25. [25]

    Refusion: A diffusion large language model with parallel autoregressive decoding

    Li, J.-N., Guan, J., Wu, W., and Li, C. Refusion: A diffusion large language model with parallel autoregressive decoding. ArXiv preprint, abs/2512.13586, 2025

    arXiv 2025

  26. [26]

    X., and Wen, J

    Li, Y., Zhou, K., Zhao, W. X., and Wen, J. Diffusion models for non-autoregressive text generation: A survey. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI 2023, 19th-25th August 2023, Macao, SAR, China , pp.\ 6692--6701. ijcai.org, 2023. doi:10.24963/IJCAI.2023/750

    work page doi:10.24963/ijcai.2023/750 2023

  27. [27]

    Deepseek-v3 technical report

    Liu, A., Feng, B., Xue, B., Wang, B., Wu, B., Lu, C., Zhao, C., Deng, C., Zhang, C., Ruan, C., et al. Deepseek-v3 technical report. ArXiv preprint, abs/2412.19437, 2024

    Pith/arXiv arXiv 2024

  28. [28]

    Faithful chain-of-thought reasoning

    Lyu, Q., Havaldar, S., Stein, A., Zhang, L., Rao, D., Wong, E., Apidianaki, M., and Callison-Burch, C. Faithful chain-of-thought reasoning. In Park, J. C., Arase, Y., Hu, B., Lu, W., Wijaya, D., Purwarianti, A., and Krisnadhi, A. A. (eds.), Proceedings of the 13th International Joint Conference on Natural Language Processing and the 3rd Conference of the ...

    2023

  29. [29]

    Large language diffusion models

    Nie, S., Zhu, F., You, Z., Zhang, X., Ou, J., Hu, J., Zhou, J., Lin, Y., Wen, J.-R., and Li, C. Large language diffusion models. ArXiv preprint, abs/2502.09992, 2025

    Pith/arXiv arXiv 2025

  30. [30]

    Maximizing confidence alone improves reasoning

    Prabhudesai, M., Chen, L., Ippoliti, A., Fragkiadaki, K., Liu, H., and Pathak, D. Maximizing confidence alone improves reasoning. ArXiv preprint, abs/2505.22660, 2025

    Pith/arXiv arXiv 2025

  31. [31]

    and Mardani, M

    Ramesh, V. and Mardani, M. Test-time scaling of diffusion models via noise trajectory search. ArXiv preprint, abs/2506.03164, 2025

    Pith/arXiv arXiv 2025

  32. [32]

    L., Stickland, A

    Rein, D., Hou, B. L., Stickland, A. C., Petty, J., Pang, R. Y., Dirani, J., Michael, J., and Bowman, S. R. Gpqa: A graduate-level google-proof q&a benchmark. In First Conference on Language Modeling, 2024

    2024

  33. [33]

    S., Yang, Z., Akhauri, Y., Liu, J., Singh, D., Cheng, Z., Liu, Z., Xing, E., Thickstun, J., and Vahdat, A

    Sahoo, S. S., Yang, Z., Akhauri, Y., Liu, J., Singh, D., Cheng, Z., Liu, Z., Xing, E., Thickstun, J., and Vahdat, A. Esoteric language models. ArXiv preprint, abs/2506.01928, 2025

    Pith/arXiv arXiv 2025

  34. [34]

    Mastering atari, go, chess and shogi by planning with a learned model

    Schrittwieser, J., Antonoglou, I., Hubert, T., Simonyan, K., Sifre, L., Schmitt, S., Guez, A., Lockhart, E., Hassabis, D., Graepel, T., et al. Mastering atari, go, chess and shogi by planning with a learned model. Nature, 588 0 (7839): 0 604--609, 2020

    2020

  35. [35]

    Confident adaptive language modeling

    Schuster, T., Fisch, A., Gupta, J., Dehghani, M., Bahri, D., Tran, V., Tay, Y., and Metzler, D. Confident adaptive language modeling. In Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., and Oh, A. (eds.), Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orl...

    2022

  36. [36]

    Mastering the game of Go without human knowledge

    Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou, I., Huang, A., Guez, A., Hubert, T., Baker, L., Lai, M., Bolton, A., Chen, Y., Lillicrap, T., Hui, F., Sifre, L., van den Driessche, G., Graepel, T., and Hassabis, D. Mastering the game of Go without human knowledge. Nature, 550 0 (7676): 0 354--359, 2017

    2017

  37. [37]

    P., Kumar, A., Ermon, S., and Poole, B

    Song, Y., Sohl - Dickstein, J., Kingma, D. P., Kumar, A., Ermon, S., and Poole, B. Score-based generative modeling through stochastic differential equations. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021 . OpenReview.net, 2021

    2021

  38. [38]

    Team, Q. et al. Qwen2 technical report. ArXiv preprint, abs/2407.10671, 2024

    Pith/arXiv arXiv 2024

  39. [39]

    Neurosymbolic diffusion models

    van Krieken, E., Minervini, P., Ponti, E., and Vergari, A. Neurosymbolic diffusion models. ArXiv preprint, abs/2505.13138, 2025

    arXiv 2025

  40. [40]

    Mcts-judge: Test-time scaling in llm-as-a-judge for code correctness evaluation

    Wang, Y., Ji, P., Yang, C., Li, K., Hu, M., Li, J., and Sartoretti, G. Mcts-judge: Test-time scaling in llm-as-a-judge for code correctness evaluation. ArXiv preprint, abs/2502.12468, 2025

    Pith/arXiv arXiv 2025

  41. [41]

    H., Le, Q

    Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E. H., Le, Q. V., and Zhou, D. Chain-of-thought prompting elicits reasoning in large language models. In Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., and Oh, A. (eds.), Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Proces...

    2022

  42. [42]

    Beyond examples: High-level automated reasoning paradigm in in-context learning via mcts

    Wu, J., Feng, M., Zhang, S., Che, F., Wen, Z., Liao, C., and Tao, J. Beyond examples: High-level automated reasoning paradigm in in-context learning via mcts. ArXiv preprint, abs/2411.18478, 2024

    Pith/arXiv arXiv 2024

  43. [43]

    P., Kawaguchi, K., and Shieh, M

    Xie, Y., Goyal, A., Zheng, W., Kan, M.-Y., Lillicrap, T. P., Kawaguchi, K., and Shieh, M. Monte carlo tree search boosts reasoning via iterative preference learning. ArXiv preprint, abs/2405.00451, 2024

    Pith/arXiv arXiv 2024

  44. [44]

    Qwen3 technical report

    Yang, A., Li, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Gao, C., Huang, C., Lv, C., et al. Qwen3 technical report. ArXiv preprint, abs/2505.09388, 2025 a

    Pith/arXiv arXiv 2025

  45. [45]

    Markov chain of thought for efficient mathematical reasoning

    Yang, W., Liao, M., and Fan, K. Markov chain of thought for efficient mathematical reasoning. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pp.\ 7132--7157, 2025 b

    2025

  46. [46]

    Dream 7b: Diffusion large language models

    Ye, J., Xie, Z., Zheng, L., Gao, J., Wu, Z., Jiang, X., Li, Z., and Kong, L. Dream 7b: Diffusion large language models. ArXiv preprint, abs/2508.15487, 2025

    Pith/arXiv arXiv 2025

  47. [47]

    Monte carlo tree diffusion for system 2 planning

    Yoon, J., Cho, H., Baek, D., Bengio, Y., and Ahn, S. Monte carlo tree diffusion for system 2 planning. ArXiv preprint, abs/2502.07202, 2025 a

    arXiv 2025

  48. [48]

    Fast monte carlo tree diffusion: 100x speedup via parallel sparse planning

    Yoon, J., Cho, H., Bengio, Y., and Ahn, S. Fast monte carlo tree diffusion: 100x speedup via parallel sparse planning. ArXiv preprint, abs/2506.09498, 2025 b

    arXiv 2025

  49. [49]

    d1: Scaling reasoning in diffusion large language models via reinforcement learning

    Zhao, S., Gupta, D., Zheng, Q., and Grover, A. d1: Scaling reasoning in diffusion large language models via reinforcement learning. ArXiv preprint, abs/2504.12216, 2025

    Pith/arXiv arXiv 2025

  50. [50]

    Llada 1.5: Variance-reduced preference optimization for large language diffusion models

    Zhu, F., Wang, R., Nie, S., Zhang, X., Wu, C., Hu, J., Zhou, J., Chen, J., Lin, Y., Wen, J.-R., et al. Llada 1.5: Variance-reduced preference optimization for large language diffusion models. ArXiv preprint, abs/2505.19223, 2025

    Pith/arXiv arXiv 2025

  51. [51]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...