Pith. sign in

REVIEW 2 major objections 5 minor 4 cited by

DINGO: Constrained Inference for Diffusion LLMs

T0 review · 2 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read DINGO gives diffusion LLMs provably optimal constrained decoding for regular expressions.

desk verdict Real algorithm with valid proofs for a single-step product objective, but the headline 'provably distribution-preserving' overclaims what is actually established for multi-step diffusion. read the letter →

arxiv 2505.23061 v1 pith:HCOMTRZD submitted 2025-05-29 cs.LG cs.PLcs.SE

Tarun Suresh , Debangshu Banerjee , Shubham Ugare , Sasa Misailovic , Gagandeep Singh This is my paper · ORCID
classification cs.LGcs.PLcs.SE
keywords diffusionLLMconstraineddecodingregularexpressionDFAdynamicprogrammingdistribution-preservinginferencestructuredoutputgenerationJSON
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Diffusion LLMs generate several tokens at once, so the usual autoregressive trick of masking invalid tokens one position at a time cannot guarantee that the final block obeys a user-supplied constraint. DINGO is a constrained decoding method for these models: given any regular expression, it uses dynamic programming over the states of a token-level DFA to choose, from the probability distribution the model produces for a block, the most probable string whose mask tokens can be filled in to form a valid prefix of the regular language. The paper proves this chosen string has probability at least as high as any other valid string under that per-position distribution, and reports that enforcing constraints this way lifts accuracy by up to 68 percentage points on GSM-Symbolic and JSON benchmarks. The central claim is that regular-language constraints can be enforced on parallel token blocks without paying for sequential decoding and without resorting to greedy approximation.

What carries the argument

The machinery is a Viterbi-style dynamic program carried out on a token-level DFA built from the user's regular expression. Each vocabulary token is treated as a possibly multi-character symbol, and the mask token gets a nondeterministic transition $\delta_\bot$ that jumps to any state reachable by one token. The DP maintains $W[i,q]$, the maximum probability of any $i$-token prefix ending in DFA state $q$ under the block's per-position distribution, with backpointers for reconstruction; the final answer is the highest-probability path ending in a live state. The live-state condition enforces that a valid completion exists, and the max-product recurrence enforces optimality.

What would settle it

Generate the same prompt many times with an unconstrained diffusion LLM, record the empirical frequencies of final full-length strings, and check whether DINGO's output, applied at a specified diffusion step, achieves at least the probability of every valid string under that empirical final distribution; one valid string that appears more often than DINGO's pick would refute the claim that DINGO preserves the true output distribution.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that for any regular expression $R$, prompt $p$, block length $d$, and output distribution $D$ of one diffusion step, if at least one $d$-token string can be completed to a valid prefix with nonzero probability, DINGO returns a string $r^*$ such that (i) some completion of its mask tokens lies in $L_P(R)$, and (ii) for every valid string $r'$, $P(r' \mid D) \le P(r^* \mid D)$. This is Proposition 4.2, the optimality result, together with Proposition 4.1 for correctness. The paper further reports that on GSM-Symbolic and JSON-Mode-Eval, DINGO reaches 100% syntactic and schema accuracy across LLaDA-8B and Dream-7B variants, with the largest gains on the weaker models.

Load-bearing premise

The load-bearing assumption is that the product of per-position token probabilities from one unmasking step is the correct output distribution to optimize, and the paper does not establish how that single-step distribution relates to the distribution of the final multi-step generated string.

Editorial extensions

If this is right

  • Structured outputs such as JSON, math expressions, and other regex-specified formats can be enforced on diffusion LLMs with 100% syntactic validity in the paper's benchmarks, at time comparable to unconstrained generation.
  • The method works for both single-block and semi-autoregressive diffusion: after each block, the ending DFA state becomes the start state for the next block's DP, provided it is live.
  • Because DINGO maximizes rather than greedily approximates, it avoids the distribution distortion that prior work attributes to greedy constrained decoding.
  • Any downstream task that consumes only syntactically valid completions inherits the best achievable score under the per-position distribution the model gives for the block.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper's optimality result is stated for the product distribution produced by a single diffusion step; it does not show that this product equals the final marginal distribution after the model's full $T$-step unmasking, so a natural test is to compare DINGO against sampling from the full diffusion process on the same prompts.
  • The same DP idea should extend to context-free constraints by replacing the DFA with a pushdown automaton, though the live-state check and DP state space would become more expensive.
  • For semi-autoregressive decoding the paper notes per-block optimality may not imply global optimality; one could measure how much is lost by greedily optimizing each block against the next block's future.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper proposes DINGO, a dynamic-programming constrained decoding method for diffusion LLMs that must satisfy a user-specified regular expression. The method builds a token-level DFA, augments its transitions to handle the mask token, and runs a Viterbi-style DP over DFA states to select the highest-probability token sequence under a per-position product distribution, while guaranteeing that the resulting string is a valid prefix of the target regular language. The authors claim that DINGO is provably distribution-preserving and optimal among valid strings, and they report large empirical gains on GSM-Symbolic and JSON-ModeEval benchmarks with several diffusion LLMs.

Significance. If the optimality and distribution-preservation claims were correct, DINGO would be an important first step: a tractable, provably constrained decoder for parallel generation in diffusion LLMs, where traditional sequential constrained decoding does not apply. The DP formulation itself is clean and the reported empirical gains are substantial (e.g., 100% schema validity in Table 2). However, the central theoretical guarantee is proved for a single-step per-position product distribution, not for the diffusion model's actual T-step output distribution, so the 'provably distribution-preserving' claim in the abstract and contributions is not established. The practical value of the method may survive as a heuristic, but the paper's main selling point, the formal guarantee, is unsupported in the evaluated settings.

major comments (2)
  1. The optimality objective in Eq. (2) is the product of per-position probabilities vvvi[ti] from a single diffusion step, justified by the sentence 'Since the token distributions vvvm+1, . . . , vvvn are independent across positions.' However, Definition 2.2 and Algorithm 2 define the diffusion LLM as a T-step unmasking process: the token at each position is selected at some step, conditioned on a partially masked context, and the final fully unmasked string's marginal distribution is not, in general, equal to the product of per-position probabilities at any single step. The paper never specifies at which diffusion step DINGO is applied, nor proves that optimizing the step-t product preserves the model's final output distribution. This gap is not confined to the semi-autoregressive case disclaimed in the Limitations: the JSON experiments in Section 5 use a single block with T=64 diffusion steps, and the GSM-Symbolic main results use 8 blocks, for which the Limitations already concede that optimality may fail. The abstract's 'provably distribution-preserving' claim is therefore unsupported in the settings where the method is evaluated.
  2. The DP includes the mask token as a candidate token with probability 1 at masked positions, so the returned string rrr* can contain mask symbols. The proof of Prop. 4.1 (step 3) replaces each mask with an arbitrary token τj that satisfies the DFA transition but may have vvvi[τj] = 0. Consequently, the fully unmasked string actually output by DINGO can have probability 0 under the very per-position product distribution that Prop. 4.2 optimizes, and it is not the string rrr* whose optimality the proposition establishes. Prop. 4.2 therefore does not show that the final output string maximizes probability among valid strings, even under the paper's own definition of 'output distribution.'
minor comments (5)
  1. The abstract says 'up to a 68 percentage point improvement,' but Table 2 shows a gap of 85 percentage points for Dream-B-7B (DINGO 100 vs. Unconstrained 15), so the stated maximum is inconsistent with the reported numbers.
  2. Since δ(q, t) returns a set of states, the notation q' ← δ(q, t) is ambiguous; the loop should iterate over q' ∈ δ(q, t).
  3. The paper states a response length of 128 with 8 blocks for GSM-Symbolic, but does not state the block length d used in the DP; specifying d would clarify how the DP block length relates to the diffusion steps and the reported times.
  4. The checklist answer for 'Open access to data and code' is [Yes] with 'Code is provided,' but no code link or supplementary material is visible in the manuscript; please either include the link or change the answer accordingly.
  5. The GSM-Symbolic regex is rendered with broken line breaks and is difficult to verify as a valid regular expression; please provide a cleaner listing.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the DP optimality proof is self-contained, and the only self-citations are not load-bearing; the main weakness is an unsupported identification, not a circular reduction.

full rationale

DINGO's derivation chain is self-contained. Proposition 4.2 proves, by induction on the DP state, that Algorithm 1 returns the argmax of the product distribution P(rrr | vvv_{m+1}...vvv_n) = Prod_i vvv_i[t_i] exactly as defined in Eq. 1-2, with no fitted parameters and no use of its own conclusion as an input. The only self-citations to CRANE (Banerjee et al., 2025) and SynCode (Ugare et al., 2024b) are for the notion of valid prefixes and for a GSM-Symbolic regex; neither carries the proof burden of the optimality or correctness theorems. The paper does not exhibit any step in which a prediction is definitionally equal to a fitted input. The abstract's 'provably distribution-preserving' claim is weakened by the fact that the paper never establishes that the single-step per-position product distribution equals the final marginal distribution of the T-step unmasking process, and the Limitations section explicitly concedes that optimality may fail in semi-autoregressive setups. However, this is an unsupported identification or correctness gap, not a circular reduction from output back to input, so it does not raise the circularity score.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The central proof rests on the product-of-independent-per-position-distributions assumption for a single diffusion step, the token-to-character mapping of the vocabulary, and the wildcard treatment of mask tokens. No free parameters are fitted; block length and step counts are user choices.

assumptions (3)
  • domain assumption Per-position token distributions at a diffusion step are independent and constitute the output distribution.
    Used in Eq. (1) and Eq. (2) to factor the probability of a string as a product; not proven for the full diffusion process.
  • standard math Each token can be mapped to a fixed string of characters, making the token-level DFA transition well-defined.
    Required for the precomputation in Section 4.1; holds for subword vocabularies.
  • domain assumption The mask symbol can be replaced by any non-mask token independently at each masked position.
    Definition 3.1 and the δ⊥ transition rely on this; reasonable for the model but an architectural assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DINGO: Constrained Inference for Diffusion LLMs." pith.science (2026). https://pith.science/paper/HCOMTRZD

@misc{pith2026250523061,
  author       = {Pith},
  title        = {Pith review of: DINGO: Constrained Inference for Diffusion LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HCOMTRZD}},
  note         = {Machine review of arXiv:2505.23061}
}
read the original abstract

Diffusion LLMs have emerged as a promising alternative to conventional autoregressive LLMs, offering significant potential for improved runtime efficiency. However, existing diffusion models lack the ability to provably enforce user-specified formal constraints, such as regular expressions, which makes them unreliable for tasks that require structured outputs, such as fixed-schema JSON generation. Unlike autoregressive models that generate tokens sequentially, diffusion LLMs predict a block of tokens in parallel. This parallelism makes traditional constrained decoding algorithms, which are designed for sequential token prediction, ineffective at preserving the true output distribution. To address this limitation, we propose DINGO, a dynamic programming-based constrained decoding strategy that is both efficient and provably distribution-preserving. DINGO enables sampling of output strings with the highest probability under the model's predicted distribution, while strictly satisfying any user-specified regular expression. On standard symbolic math and JSON generation benchmarks, DINGO achieves up to a 68 percentage point improvement over unconstrained inference

Figures

Figures reproduced from arXiv: 2505.23061 by the authors.

Figure 1
Figure 1. Ablation Study on The Number of Diffusion Blocks For GSM-Symbolic [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. An example from the GSM-symbolic dataset (variables in blue), where unconstrained [PITH_FULL_IMAGE:figures/full_fig_p025_2.png] view at source ↗
Figure 3
Figure 3. An example from the GSM-symbolic dataset (variables in blue), where unconstrained [PITH_FULL_IMAGE:figures/full_fig_p026_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: An example from JSON generation, where unconstrained generation produces a syntac [PITH_FULL_IMAGE:figures/full_fig_p027_4.png]
Figure 5
Figure 5. Figure 5: An example from JSON generation, where unconstrained generation produces a syntac [PITH_FULL_IMAGE:figures/full_fig_p027_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 4 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Imprompt: A Language Framework for Prompt Programming

    cs.AI 2026-07 conditional novelty 7.0 of 10

    Prompt programs can be typed like programs and compiled in different ways; Imprompt formalizes two such compilers and tests them on tax-law reasoning and data-labeling tasks.

  2. Continuous Diffusion Models Can Obey Formal Syntax

    cs.LG 2026-02 conditional novelty 7.0 of 10

    Training-free gradient guidance, based on the exact regex-acceptance probability under the decoder's unigram distribution, lets continuous diffusion language models satisfy regular-expression constraints.

  3. The Hidden Cost of Structured Generation in LLMs: Draft-Conditioned Constrained Decoding

    cs.CL 2026-02 conditional novelty 5.0 of 10

    Draft-then-constrain decoding—first generate a free-form draft, then run format-constrained decoding conditioned on that draft—consistently improves structured-output accuracy across 1B–14B models.

  4. Discrete Diffusion Models: A Unified Framework from Tokenization to Generation

    cs.LG 2026-07 unverdicted novelty 4.0 of 10

    Discrete diffusion models are re-framed as instances of a tokenization-centric, four-component design space (corruption, denoiser, objective, sampler) in a broad survey with no new experimental or theoretical results.

Reference graph

Works this paper leans on

30 extracted references · 11 canonical work pages · cited by 4 Pith papers

  1. [1]

    Block diffusion: Interpolating between autoregressive and diffusion language models

    Marianne Arriola, Subham Sekhar Sahoo, Aaron Gokaslan, Zhihan Yang, Zhixuan Qi, Jiaqi Han, Justin T Chiu, and Volodymyr Kuleshov. Block diffusion: Interpolating between autoregressive and diffusion language models. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=tyEyYT267x

    2025

  2. [2]

    Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg

    Jacob Austin, Daniel D. Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. Structured denoising diffusion models in discrete state-spaces. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. URL https://openreview.net/forum?id=h7-XixPCAL

    work page 2021

  3. [3]

    CRANE : Reasoning with constrained LLM generation

    Debangshu Banerjee, Tarun Suresh, Shubham Ugare, Sasa Misailovic, and Gagandeep Singh. CRANE : Reasoning with constrained LLM generation. arXiv preprint arXiv:2502.09061, 2025. URL https://arxiv.org/pdf/2502.09061

    arXiv 2025

  4. [4]

    Bartoldson, Bhavya Kailkhura, and Ferdinando Fioretto

    Michael Cardei, Jacob K Christopher, Thomas Hartvigsen, Brian R. Bartoldson, Bhavya Kailkhura, and Ferdinando Fioretto. Constrained language generation with discrete diffusion models, 2025. URL https://arxiv.org/abs/2503.09790

    arXiv 2025

  5. [5]

    Z3: an efficient smt solver

    Leonardo De Moura and Nikolaj Bj rner. Z3: an efficient smt solver. In Proceedings of the Theory and Practice of Software, 14th International Conference on Tools and Algorithms for the Construction and Analysis of Systems, TACAS'08/ETAPS'08, page 337–340, Berlin, Heidelberg, 2008. Springer-Verlag. ISBN 3540787992

    work page 2008

  6. [6]

    A general-purpose algorithm for constrained sequential inference

    Daniel Deutsch, Shyam Upadhyay, and Dan Roth. A general-purpose algorithm for constrained sequential inference. In Proceedings of the Conference on Computational Natural Language Learning, 2019. URL https://aclanthology.org/K19-1045/

    2019

  7. [7]

    XGrammar : Flexible and efficient structured generation engine for large language models

    Yixin Dong, Charlie F Ruan, Yaxing Cai, Ruihang Lai, Ziyi Xu, Yilong Zhao, and Tianqi Chen. XGrammar : Flexible and efficient structured generation engine for large language models. arXiv preprint arXiv:2411.15100, 2024. URL https://arxiv.org/pdf/2411.15100

    arXiv 2024

  8. [8]

    Chen et. al. Evaluating large language models trained on code, 2021. URL https://arxiv.org/abs/2107.03374

    arXiv 2021

Show all 30 references
  1. [9]

    LLM training data synthesis for more effective problem solving using satisfiability modulo theories

    Timofey Fedoseev, Dimitar Iliev Dimitrov, Timon Gehr, and Martin Vechev. LLM training data synthesis for more effective problem solving using satisfiability modulo theories. In The 4th Workshop on Mathematical Reasoning and AI at NeurIPS'24, 2024. URL https://openreview.net/fo...

    2024

  2. [10]

    G. D. Forney. The viterbi algorithm. Proc. of the IEEE, 61: 0 268 -- 278, March 1973

    1973

  3. [11]

    Ssd-lm: Semi-autoregressive simplex-based diffusion language model for text generation and modular control, 2023

    Xiaochuang Han, Sachin Kumar, and Yulia Tsvetkov. Ssd-lm: Semi-autoregressive simplex-based diffusion language model for text generation and modular control, 2023. URL https://arxiv.org/abs/2210.17432

    2023 arXiv

  4. [12]

    Automata-based constraints for language model decoding

    Terry Koo, Frederick Liu, and Luheng He. Automata-based constraints for language model decoding. In Conference on Language Modeling, 2024. URL https://openreview.net/forum?id=BDBdblmyzY

    2024

  5. [13]

    Validating large language models with RELM

    Michael Kuchnik, Virginia Smith, and George Amvrosiadis. Validating large language models with RELM . Proceedings of Machine Learning and Systems, 5, 2023. URL https://proceedings.mlsys.org/paper_files/paper/2023/file/93c7d9da61ccb2a60ac047e92787c3ef-Paper-mlsys2023.pdf

    2023

  6. [14]

    Sequential Monte Carlo steering of large language models using probabilistic programs

    Alexander K Lew, Tan Zhi-Xuan, Gabriel Grand, and Vikash Mansinghka. Sequential Monte Carlo steering of large language models using probabilistic programs. In ICML 2023 Workshop: Sampling and Optimization in Discrete Space, 2023. URL https://openreview.net/pdf?id=Ul2K0qXxXy

    2023

  7. [15]

    Syntactic and semantic control of large language models via sequential Monte Carlo

    Jo \ a o Loula, Benjamin LeBrun, Li Du, Ben Lipkin, Clemente Pasti, Gabriel Grand, Tianyu Liu, Yahya Emara, Marjorie Freedman, Jason Eisner, Ryan Cotterell, Vikash Mansinghka, Alex Lew, Tim Vieira, and Tim O'Donnell. Syntactic and semantic control of large language models via ...

    2025

  8. [16]

    Gsm-symbolic: Understanding the limitations of mathematical reasoning in large language models, 2024

    Iman Mirzadeh, Keivan Alizadeh, Hooman Shahrokhi, Oncel Tuzel, Samy Bengio, and Mehrdad Farajtabar. Gsm-symbolic: Understanding the limitations of mathematical reasoning in large language models, 2024. URL https://arxiv.org/abs/2410.05229

    2024 arXiv

  9. [17]

    Large language diffusion models, 2025

    Shen Nie, Fengqi Zhu, Zebin You, Xiaolu Zhang, Jingyang Ou, Jun Hu, Jun Zhou, Yankai Lin, Ji-Rong Wen, and Chongxuan Li. Large language diffusion models, 2025. URL https://arxiv.org/abs/2502.09992

    2025 arXiv

  10. [18]

    json-mode-eval, 2024

    NousResearch. json-mode-eval, 2024. URL https://huggingface.co/datasets/NousResearch/json-mode-eval

    2024

  11. [19]

    Logic-lm: Empowering large language models with symbolic solvers for faithful logical reasoning, 2023

    Liangming Pan, Alon Albalak, Xinyi Wang, and William Yang Wang. Logic-lm: Empowering large language models with symbolic solvers for faithful logical reasoning, 2023. URL https://arxiv.org/abs/2305.12295

    2023 arXiv

  12. [20]

    Grammar-aligned decoding

    Kanghee Park, Jiayu Wang, Taylor Berg-Kirkpatrick, Nadia Polikarpova, and Loris D'Antoni. Grammar-aligned decoding. Advances in Neural Information Processing Systems, 37: 0 24547--24568, 2024 a . URL https://proceedings.neurips.cc/paper_files/paper/2024/file/2bdc2267c3d7d01523...

    2024

  13. [21]

    Grammar-aligned decoding

    Kanghee Park, Jiayu Wang, Taylor Berg-Kirkpatrick, Nadia Polikarpova, and Loris D'Antoni. Grammar-aligned decoding. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024 b . URL https://openreview.net/forum?id=5G7ve8E1Lu

    2024

  14. [22]

    Pytorch: An imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu ...

    2019

  15. [23]

    Synchromesh: Reliable code generation from pre-trained language models

    Gabriel Poesia, Alex Polozov, Vu Le, Ashish Tiwari, Gustavo Soares, Christopher Meek, and Sumit Gulwani. Synchromesh: Reliable code generation from pre-trained language models. In International Conference on Learning Representations, 2022. URL https://openreview.net/forum?id=K...

    2022

  16. [24]

    Language models are unsupervised multitask learners

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. OpenAI, 2019. URL https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf. Accessed: 2024-11-15

    2019

  17. [26]

    Syncode: Llm generation with grammar augmentation, 2024 b

    Shubham Ugare, Tarun Suresh, Hangoo Kang, Sasa Misailovic, and Gagandeep Singh. Syncode: Llm generation with grammar augmentation, 2024 b . URL https://arxiv.org/abs/2403.01632

    2024 arXiv

  18. [27]

    IterGen : Iterative structured LLM generation

    Shubham Ugare, Rohan Gumaste, Tarun Suresh, Gagandeep Singh, and Sasa Misailovic. IterGen : Iterative structured LLM generation. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/pdf?id=ac93gRzxxV

    2025

  19. [28]

    Efficient guided generation for large language models

    Brandon T Willard and R \'e mi Louf. Efficient guided generation for large language models. arXiv preprint arXiv:2307.09702, 2023. URL https://arxiv.org/pdf/2307.09702

    2023 arXiv

  20. [29]

    Transformers: State-of-the-art natural language processing

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mari...

    2020

  21. [30]

    Swope, Alex Gu, Rahul Chalamala, Peiyang Song, Shixing Yu, Saad Godil, Ryan Prenger, and Anima Anandkumar

    Kaiyu Yang, Aidan M. Swope, Alex Gu, Rahul Chalamala, Peiyang Song, Shixing Yu, Saad Godil, Ryan Prenger, and Anima Anandkumar. Leandojo: Theorem proving with retrieval-augmented language models, 2023. URL https://arxiv.org/abs/2306.15626

    2023 arXiv

  22. [31]

    Dream 7b, 2025

    Jiacheng Ye, Zhihui Xie, Lin Zheng, Jiahui Gao, Zirui Wu, Xin Jiang, Zhenguo Li, and Lingpeng Kong. Dream 7b, 2025. URL https://hkunlp.github.io/blog/2025/dream

    2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.