Large Language Model Meets Constraint Propagation

Alexandre Bonlarron; Elisabetta De Maria; Florian R\'egin; Jean-Charles R\'egin

REVIEW 4 major objections 6 minor 35 references

Large Language Model Meets Constraint Propagation

T0 review · 4 major / 6 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 The paper claims that previewing future token domains with a masked language model makes GenCP-style constrained text generation find more valid solutions while using fewer autoregressive LLM calls.

desk verdict A plausible incremental extension of GenCP, but the claimed bidirectional benefit is untested and the evaluation is too thin to back the main conclusion. read the letter →

arxiv 2505.24012 v1 pith:DTID57NU submitted 2025-05-29 cs.CL cs.AI

Alexandre Bonlarron , Florian Régin , Elisabetta De Maria , Jean-Charles Régin This is my paper

classification cs.CLcs.AI MSC 68T2068T50

keywords constrainedtextgenerationconstraintprogrammingmaskedlanguagemodelsdomainpreviewmeta-variablesbacktrackingsearchautoregressiveLLMcallsGenCP

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

The paper tries to show that constrained text generation, where every emitted sentence must satisfy hard rules such as exact character counts or banned words, becomes more efficient when a Masked Language Model is allowed to look ahead. The approach, called previewMLM, asks the bidirectional model to propose candidate tokens for positions that have not been generated yet, and feeds those candidates into the constraint solver before the next token is chosen. On the paper's benchmark tasks this yields more valid solutions with fewer autoregressive LLM calls: for the single-sentence task, 10 solutions from 359 LLM calls and 170 backtracks, versus 3 solutions, 444 calls, and 375 backtracks without the preview. The trade-off is a modest number of extra MLM calls and, in one paragraph task, more backtracks, but the authors conclude that the preview pays off exactly where constraints are strict and long-range.

What carries the argument

The load-bearing mechanism is the MLM domain preview operating on meta-variables. A meta-variable bundles one or more token-level decision variables, so a word like "Using" is represented by two variables ("Us", "ing") rather than forcing a one-word-per-variable model that mismatches tokenizers. Given k-1 assigned variables, the method masks the next d positions and asks the masked model for a set of candidate tokens; those candidates become the domains for the upcoming variables, and the CP propagators filter them before the autoregressive model's top-k sample is used to select the next assignment. In short, the preview turns the constraint solver into a bidirectional filter: future domains are known early enough to prune the current decision.

What would settle it

Run the same six tasks with the MLM preview replaced by future tokens sampled from the autoregressive model's own distribution, holding all search parameters fixed; if solution counts and LLM calls remain at previewMLM levels, the preview signal is not the cause. A sharper check is to log the survival rate of MLM-proposed tokens through constraint propagation: if preview candidates survive filtering at no higher rate than arbitrary tokens from the base vocabulary, the preview is not carrying information into the propagator.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that the bottleneck in GenCP is not the search or the constraints but the one-directional flow of domain information: an autoregressive model only knows the left context, so the constraint propagator cannot filter variables that are far ahead. previewMLM removes that bottleneck by treating the next d variables as masked positions, having BERT propose their candidate domains, and running constraint propagation on those domains before deciding the current token. The empirical marker is the jump in solutions on sent-1 (3 to 10) and the appearance of solutions on para-4, where the baseline found none, together with lower LLM call counts. The authors are explicit that this works at look-ahead depth d=2 and concentrates on the last two variables of a sentence, where sum constraints bind the final and penultimate tokens; deeper previews reduced the number of solutions.

Load-bearing premise

The load-bearing premise is that a masked model, looking at both sides of a partially written text, can propose candidate tokens for future positions that are both plausible enough and diverse enough that constraint propagation turns them into useful pruning; if those preview domains are too narrow or too noisy, the extra calls add cost without improving search.

Editorial extensions

If this is right

On sent-1, previewMLM reduces autoregressive LLM calls from 444 to 359 and backtracks from 375 to 170, while raising the number of solutions from 3 to 10.
On para-2, the preview doubles the solution count from 3 to 6 while cutting LLM calls from 446 to 341; it costs 40 MLM calls and 57 extra backtracks.
On para-4, where the metavar baseline found no valid solution in 466 LLM calls and 447 backtracks, previewMLM finds 2 solutions in 422 LLM calls and 258 backtracks.
The benefit is conditional: preview depths d=3 and d=4 decreased the number of solutions, so the useful regime is a short two-position look-ahead near the end of a sentence.
Because the method reduces the expensive autoregressive LLM calls, it shifts cost from the LLM to the cheaper masked model, making strict-constraint generation more affordable.

Reading between the lines

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

Beyond the paper: the mechanism should matter most when constraints link distant positions, such as whole-sentence character sums or banned-word lists, and should matter little for purely local lexical constraints; a researcher could test this by varying constraint distance while holding task difficulty fixed.
Beyond the paper: the paper's depth results suggest the bottleneck is preview quality, not the CP machinery; a stronger or fine-tuned masked scorer may extend the useful look-ahead beyond d=2, which would turn the current two-position preview into a general long-range propagation tool.
Beyond the paper: the same future-domain preview could be applied to other structured decoding settings, such as translation with length constraints or summarization with keyword constraints, by swapping the masked model for any bidirectional scorer that can propose candidates for masked spans.
Beyond the paper: the evaluation metric is LLM call count, a hardware-independent proxy; if MLM calls are substantially cheaper in wall-clock time, the practical speedup is larger than the call counts show, but if they are not, the trade needs to be re-checked in seconds.

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.

Desk Editor's Note

read the letter

The punchline: this is a sensible, clearly written extension of GenCP, but the paper's central claim—that MLM-based domain preview improves GenCP—is not actually established. The reported gains are real in Table 2 for sent-1, para-2, and para-4, but the mechanism billed as bidirectional is never isolated from a simple two-token lookahead.

What's new and good: the meta-variable representation fixes a real mismatch between word-level CP variables and token-level LLM outputs. Using an MLM to propose candidate domains for upcoming variables is a reasonable idea, and the examples look fluent. The authors are honest about where it doesn't work: they only use previewMLM on three tasks, and they admit trying d=3 and d=4 and seeing fewer solutions. That is more than many papers do.

The soft spots, in order of importance. First, the bidirectional story doesn't hold up. The prompt for the penultimate and final variables is '... [MASK] [MASK]' with no tokens after the masks. BERT sees nothing to the right, and the other mask is just a placeholder. So the extra solutions could come entirely from looking two variables ahead rather than from any bidirectional understanding. The paper needs a previewAR baseline—same lookahead with an autoregressive model—to make the advertised claim. Second, Table 2 is single runs, no seeds, no standard deviations. Three tasks is a small basis for 'significantly improves.' Third, the depth d=2 was selected after seeing results for other depths, which is post-hoc selection. Finally, the metric is LLM call counts; the extra MLM calls have a cost that is never measured. The paper itself mentions MLM latency but does not include it in the comparison.

None of this kills the idea. The meta-variable contribution is solid enough on its own, and the preview mechanism might well be an improvement. It's just that the current evidence doesn't separate the interesting part (bidirectionality) from a plain lookahead.

Who should read it: anyone working on CP-guided text generation or constrained decoding. It's a useful reference for the meta-variable trick and a cautionary example of how easy it is to overclaim a mechanism without the right ablation.

My recommendation: send it to peer review, but with the expectation of major revision. The authors should add an autoregressive lookahead baseline, run multiple seeds, and report wall-clock time or a proper cost model. If that comes back clean, this would be a legitimate contribution. As it stands, it's a promising short paper with an unsupported headline.

Referee Report

4 major / 6 minor

Summary. The paper extends the GenCP framework for constrained text generation by using a masked language model (BERT) to propose candidate domains for the next two unassigned variables ("domain preview"), with the goal of giving the CP solver forward information and reducing expensive autoregressive LLM calls. The authors introduce meta-variables to reconcile word-level CP variables with subword tokens, describe the MLM preview procedure, and report experiments on six COLLIE-inspired tasks showing fewer LLM calls and/or more solutions for previewMLM on three of the tasks.

Significance. If the claimed effects hold, the work would be a practical step toward cheaper constrained decoding, since it replaces some autoregressive calls with cheaper MLM calls and improves feasibility on strict length and keyword constraints. The paper is honest about some limitations (Section 4.2 notes that depth d=3 or 4 decreases the number of solutions), uses a fixed and reasonably reproducible protocol (babbage-002, bert-base-cased, top-k=50), and builds on an independently published GenCP baseline. The central contribution, however, is not established by the current experiments because the headline comparison conflates MLM bidirectionality with the addition of any two-variable lookahead, and the results are single runs without variance or statistical tests.

major comments (4)

[§3.3.1 and Table 2] The prompt in §3.3.1 is "{X1=w1,...,X_{k-1}=w_{k-1}, X_k=[M], X_{k+1}=[M]}": both future positions are masked, so BERT has no unmasked right context to exploit. The abstract and §1 claim bidirectional propagation that "leverages both past and future tokens," but the implementation never provides future tokens. Since Table 2 compares previewMLM only against vanilla and metavar (neither performs any lookahead), the observed gains (e.g., sent-1: 359 vs 444 LLM calls, 10 vs 3 solutions) may be due solely to a two-token lookahead, not MLM bidirectionality. Adding a previewAR baseline—using the same autoregressive LLM to generate domains for the next two variables—is necessary to support the central claim.
[Table 2 and §4.2] Each row in Table 2 reports a single run with no seeds, error bars, or statistical tests. The word "significantly" in the abstract and conclusion is therefore unsupported. In addition, previewMLM is evaluated on only three of the six tasks (sent-1, para-2, para-4); para-3, para-5, and para-6 have no preview comparison. Please provide repeated runs and a significance test, or soften the claim to a qualitative observation.
[§4.2 Limitations] The paper itself states that depth d=3 and d=4 decrease the number of solutions and that the preview is applied only to the last two variables of each sentence. This means the claimed benefit is limited to a narrow regime, and the choice d=2 appears to be made after observing the benchmark outcomes. The authors should report the d=3/4 results (or at least the trend) and explicitly limit the conclusion to constraints concentrated in the final variables.
[§4.2 and Table 3] The claim that previewMLM produces "meaningful" outputs is supported only by a few self-selected examples in Table 3. There is no automatic metric (e.g., perplexity, human evaluation) or comparison to the metavar outputs, so the quality component of the contribution is not demonstrated. Either provide a quantitative quality evaluation or drop the quality claim from the abstract.

minor comments (6)

[§2.4] The sentence "Constraint Programming (CP) has been used to formalizes the constrained generation task" is ungrammatical; "formalizes" should be "formalize" or the sentence should be restructured.
[Table 1] In Table 1, "Kindgoms" should be "Kingdoms," and the table caption contains an incomplete sentence fragment that should be rephrased.
[§4.1.1] The sentence "Our benchmark suite consists of several task can be found in Tab. 1" is malformed; please rephrase.
[§5.2] The phrase "when a the domain of a variable is determined" contains a stray article and should be corrected.
[Throughout] The method name is rendered inconsistently as "previewMLM" in Table 2 and "PreviewMLM" in the text; please standardize.
[§3.1.1] The example prompt "The little boy [MASK] [MASK] [MASK]..." shows three masks, whereas the method later uses two; please align the illustration with the d=2 implementation.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the MLM-preview gain is an empirical result measured against a prior-work baseline, not a derivation forced by definitions or self-citation.

full rationale

I walked the claimed derivation chain. The paper's claim is that replacing the autoregressive-only domain generation in GenCP with an MLM domain preview (previewMLM) reduces LLM calls/backtracks and increases feasible solutions. Nothing in Sections 3.1–3.3 defines the success metric in terms of the mechanism: previewMLM is specified algorithmically (masked future variables queried with BERT, domains filtered by CP propagators, assignments made by the autoregressive LLM), and Table 2 reports measured counts. The comparison baseline, GenCP, is cited to the authors' own prior work [11, 24, 25], but that citation is contextual background and baseline implementation, not a proof of the new result; the measured call counts speak independently, and the baseline is externally published. The Section 4.2 limitation—'the depth d to define the preview of the domain is used only on the last two variables of any sentences. We tried further depth d=3 and d=4 and so on but unfortunately the number of solutions decrease for the same amount of running time while keeping the same quality'—is an empirical tuning disclosure that narrows the claim; it is not a fitted input renamed as a prediction, because no held-out prediction is asserted. The strongest criticism, that no autoregressive-lookahead (previewAR) control isolates MLM bidirectionality, is a threat to causal attribution and experimental completeness, not a circularity: an omitted control does not make the reported improvement equal to its input by construction. No uniqueness theorem, ansatz, or known pattern is imported from the authors' prior work to force the choice of MLM preview. I therefore find no circular step that can be quoted, and score the paper 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The empirical result rests on several hand-chosen parameters and domain assumptions, the most fragile being that BERT preview at depth d=2 on the final variables is universally useful. These choices are not derived and are not supported by sensitivity analysis.

free parameters (4)

Domain preview depth d = 2
Section 4.2: tried d=3 and d=4 and observed fewer solutions; d=2 was retained for all reported comparisons.
top-k for domain generation = 50
Section 4.1: each GenD call asks for 50 tokens; no sensitivity analysis.
LLM temperature = 0.8
Section 4.1.2: temperature 0.8 for creative output; no ablation.
Thematic pre-prompt = 'Amidst the crimson glow of a setting sun...'
Hand-authored prompt used to inspire content; its effect on feasibility and quality is not measured.

assumptions (5)

standard math Constraint Programming filtering algorithms (knapsack, arc consistency) are correct and complete for the constraints used.
Used to prune domains and to justify backtracking behavior.
domain assumption Token sequences produced by the LLM tokenizer can be cleanly mapped to CP decision variables via meta-variables.
Section 3.1; if token boundaries are unstable the meta-variable construction fails.
domain assumption BERT's masked token predictions for future positions are informative enough to improve candidate domains after CP filtering.
Core mechanism; only indirectly validated by the small experiments.
domain assumption Fewer autoregressive LLM calls is a fair efficiency measure and MLM calls are comparatively cheap.
Sections 4.1.2 and 4.2; no wall-clock timing is reported.
ad hoc to paper The preview depth d=2 on only the final variables of a sentence generalizes to other constrained generation tasks.
The authors limit preview to the last two variables and found other depths reduce the number of solutions.

invented entities (1)

Meta-variable (Xi composed of one or more decision variables)
purpose: Bridge the mismatch between word-level CP variables and subword LLM tokens.
A modeling abstraction with no falsifiable prediction outside the paper's own generated text.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Large Language Model Meets Constraint Propagation." pith.science (2026). https://pith.science/paper/DTID57NU

@misc{pith2026250524012,
  author       = {Pith},
  title        = {Pith review of: Large Language Model Meets Constraint Propagation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DTID57NU}},
  note         = {Machine review of arXiv:2505.24012}
}

read the original abstract

Large Language Models (LLMs) excel at generating fluent text but struggle to enforce external constraints because they generate tokens sequentially without explicit control mechanisms. GenCP addresses this limitation by combining LLM predictions with Constraint Programming (CP) reasoning, formulating text generation as a Constraint Satisfaction Problem (CSP). In this paper, we improve GenCP by integrating Masked Language Models (MLMs) for domain generation, which allows bidirectional constraint propagation that leverages both past and future tokens. This integration bridges the gap between token-level prediction and structured constraint enforcement, leading to more reliable and constraint-aware text generation. Our evaluation on COLLIE benchmarks demonstrates that incorporating domain preview via MLM calls significantly improves GenCP's performance. Although this approach incurs additional MLM calls and, in some cases, increased backtracking, the overall effect is a more efficient use of LLM inferences and an enhanced ability to generate feasible and meaningful solutions, particularly in tasks with strict content constraints.

Figures

Figures reproduced from arXiv: 2505.24012 by the authors.

**Figure 1.** This diagram positions different methods along the x-axis based on feasibility difficulty for constrained text generation, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 2.** Illustration of GenCP in a toy example where the task is to generate a passage without the letter ”e”. The autoregressive [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 30 canonical work pages

[1]

Hanjie, Runzhe Yang, and Karthik R Narasimhan

Shunyu Yao, Howard Chen, Austin W. Hanjie, Runzhe Yang, and Karthik R Narasimhan. COLLIE: System- atic construction of constrained text generation tasks. In The Twelfth International Conference on Learning Representations, 2024

work page 2024
[2]

Bartoldson, Bhavya Kailkhura, and Ferdi- nando Fioretto

Michael Cardei, Jacob K Christopher, Thomas Hartvigsen, Brian R. Bartoldson, Bhavya Kailkhura, and Ferdi- nando Fioretto. Constrained discrete diffusion, 2025

work page 2025
[3]

Adversarial tokenization, 2025

Renato Lui Geh, Zilei Shao, and Guy Van den Broeck. Adversarial tokenization, 2025

work page 2025
[4]

Chi, Quoc V

Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V . Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. In Sanmi Koyejo, S. Mohamed, A. Agarwal, Danielle Belgrave, K. Cho, and A. Oh, editors, Advances in Neural Information Processing Systems 35: Annual Conference on Neur...

work page 2022
[5]

Griffiths, Yuan Cao, and Karthik Narasimhan

Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: deliberate problem solving with large language models. In Proceedings of the 37th Interna- tional Conference on Neural Information Processing Systems , NIPS ’23, Red Hook, NY , USA, 2023. Curran Associates Inc

work page 2023
[6]

Fast lexically constrained decoding with dynamic beam allocation for neural ma- chine translation

Matt Post and David Vilar. Fast lexically constrained decoding with dynamic beam allocation for neural ma- chine translation. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers) , pages 1314–1324, New Orleans, Louisiana, June 2018. Associa...

work page 2018
[7]

Smith, and Yejin Choi

Ximing Lu, Sean Welleck, Peter West, Liwei Jiang, Jungo Kasai, Daniel Khashabi, Ronan Le Bras, Lianhui Qin, Youngjae Yu, Rowan Zellers, Noah A. Smith, and Yejin Choi. NeuroLogic A*esque decoding: Constrained text generation with lookahead heuristics. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational L...

work page 2022
[8]

Constraints first: A new mdd-based model to generate sentences under constraints

Alexandre Bonlarron, Aur ´elie Calabr`ese, Pierre Kornprobst, and Jean-Charles R ´egin. Constraints first: A new mdd-based model to generate sentences under constraints. In Edith Elkind, editor, Proceedings of the Thirty- Second International Joint Conference on Artificial Intelligence, IJCAI-23, pages 1893–1901, 2023

work page 1901

Show all 35 references

[9]

Intertwining cp and nlp: The generation of unreasonably con- strained sentences

Alexandre Bonlarron and Jean-Charles R ´egin. Intertwining cp and nlp: The generation of unreasonably con- strained sentences. In Kate Larson, editor, Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI-24 , pages 7600–7608. Interna...

2024
[10]

Markov constraint as large language model surrogate

Alexandre Bonlarron and Jean-Charles R ´egin. Markov constraint as large language model surrogate. In Kate Larson, editor, Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI- 24, pages 1844–1852. International Joint Conferences on ...

2024
[11]

Combining Constraint Programming Reason- ing with Large Language Model Predictions

Florian R ´egin, Elisabetta De Maria, and Alexandre Bonlarron. Combining Constraint Programming Reason- ing with Large Language Model Predictions. In 30th International Conference on Principles and Practice of Constraint Programming (CP 2024), 2024

2024
[12]

The curious case of neural text degeneration

Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30,

2020
[13]

Where is the signal in tokenization space? In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 3966–3979, 2024

Renato Geh, Honghua Zhang, Kareem Ahmed, Benjie Wang, and Guy Van Den Broeck. Where is the signal in tokenization space? In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 3966–3979, 2024

2024
[14]

Markov constraints: Steerable generation of markov sequences

Franc ¸ois Pachet and Pierre Roy. Markov constraints: Steerable generation of markov sequences. Constraints, 16(2):148–172, apr 2011

2011
[15]

BERT: pre-training of deep bidirectional transformers for language understanding

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In Jill Burstein, Christy Doran, and Thamar Solorio, editors, Proceed- ings of the 2019 Conference of the North American Chapter ...

2019
[16]

Constraint propagation

Christian Bessiere. Constraint propagation. In Foundations of Artificial Intelligence, volume 2, pages 29–83. Elsevier, 2006

2006
[17]

Markov constraints for generating lyrics with style

Gabriele Barbieri, Franc ¸ois Pachet, Pierre Roy, and Mirko Degli Esposti. Markov constraints for generating lyrics with style. In Proceedings of the 20th European Conference on Artificial Intelligence , ECAI’12, page 115–120, NLD, 2012. IOS Press

2012
[18]

Avoiding plagiarism in markov sequence generation

Alexandre Papadopoulos, Pierre Roy, and Franc ¸ois Pachet. Avoiding plagiarism in markov sequence generation. In Proceedings of the Twenty-Eighth AAAI Conference on Artificial Intelligence , AAAI’14, page 2731–2737. AAAI Press, 2014

2014
[19]

Generating all possible palin- dromes from ngram corpora

Alexandre Papadopoulos, Pierre Roy, Jean-Charles R ´egin, and Franc ¸ois Pachet. Generating all possible palin- dromes from ngram corpora. In Proceedings of the 24th International Conference on Artificial Intelligence , IJCAI’15, page 2489–2495. AAAI Press, 2015

2015
[20]

Assisted lead sheet composition using flowcomposer

Alexandre Papadopoulos, Pierre Roy, and Franc ¸ois Pachet. Assisted lead sheet composition using flowcomposer. In Michel Rueher, editor, Principles and Practice of Constraint Programming , pages 769–785, Cham, 2016. Springer International Publishing

2016
[21]

Expressing musical ideas with constraint programming using a model of tonal harmony

Damien Sprockeels and Peter Van Roy. Expressing musical ideas with constraint programming using a model of tonal harmony. In Kate Larson, editor, Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI-24 , pages 7753–7761. Internationa...

2024
[22]

Lexically constrained decoding for sequence generation using grid beam search

Chris Hokamp and Qun Liu. Lexically constrained decoding for sequence generation using grid beam search. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1535–1546, Vancouver, Canada, July 2017. Associati...

2017
[23]

Neuro- Logic decoding: (un)supervised neural text generation with predicate logic constraints

Ximing Lu, Peter West, Rowan Zellers, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. Neuro- Logic decoding: (un)supervised neural text generation with predicate logic constraints. In Kristina Toutanova, Anna Rumshisky, Luke Zettlemoyer, Dilek Hakkani-Tur, Iz Beltagy, Stev...

2021
[24]

Using on-the-fly model checking to improve constraint programming for dynamic problems

Florian R ´egin and Elisabetta De Maria. Using on-the-fly model checking to improve constraint programming for dynamic problems. In 2023 IEEE 35th International Conference on Tools with Artificial Intelligence (ICTAI) , pages 393–398, 2023

2023
[25]

Generative constraint programming

Florian R ´egin. Generative constraint programming. Theses, Universit ´e Cˆote d’Azur, December 2024

2024
[26]

Why is constrained neural language generation particularly challenging? arXiv preprint arXiv:2206.05395, 2022

Cristina Garbacea and Qiaozhu Mei. Why is constrained neural language generation particularly challenging? arXiv preprint arXiv:2206.05395, 2022

2022 arXiv
[27]

A dynamic programming approach for consistency and propagation for knapsack constraints

Michael Trick. A dynamic programming approach for consistency and propagation for knapsack constraints. Annals of Operations Research, 118:73–84, 2003

2003
[28]

Flashattention: Fast and memory-efficient exact attention with io-awareness

Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher R ´e. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344–16359, 2022

2022
[29]

Mahoney, and Kurt Keutzer

Amir Gholami, Sehoon Kim, Zhen Dong, Zhewei Yao, Michael W. Mahoney, and Kurt Keutzer. A survey of quantization methods for efficient neural network inference. CoRR, abs/2103.13630, 2021

2021 arXiv
[30]

Thinking, fast and slow

Daniel Kahneman. Thinking, fast and slow. Farrar, Straus and Giroux, New York, 2011

2011
[31]

Thinking fast and slow in ai

Grady Booch, Francesco Fabiano, Lior Horesh, Kiran Kate, Jonathan Lenchner, Nick Linck, Andreas Loreggia, Keerthiram Murgesan, Nicholas Mattei, Francesca Rossi, and Biplav Srivastava. Thinking fast and slow in ai. Proceedings of the AAAI Conference on Artificial Intelligence, ...

2021
[32]

Lazy arc consistency

Thomas Schiex, Jean-Charles R ´egin, Christine Gaspin, and Gerard Verfaillie. Lazy arc consistency. In Proceed- ings of the AAAI Conference on Artificial Intelligence, volume 1, pages 216–221, 1996

1996
[33]

Constraint-based scheduling: applying constraint pro- gramming to scheduling problems, volume 39

Philippe Baptiste, Claude Le Pape, and Wim Nuijten. Constraint-based scheduling: applying constraint pro- gramming to scheduling problems, volume 39. Springer Science & Business Media, 2001. 12

2001
[2020]

OpenReview.net, 2020

2020
[2021]

Association for Computational Linguistics

Pith tools

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

[1] [1]

Hanjie, Runzhe Yang, and Karthik R Narasimhan

Shunyu Yao, Howard Chen, Austin W. Hanjie, Runzhe Yang, and Karthik R Narasimhan. COLLIE: System- atic construction of constrained text generation tasks. In The Twelfth International Conference on Learning Representations, 2024

work page 2024

[2] [2]

Bartoldson, Bhavya Kailkhura, and Ferdi- nando Fioretto

Michael Cardei, Jacob K Christopher, Thomas Hartvigsen, Brian R. Bartoldson, Bhavya Kailkhura, and Ferdi- nando Fioretto. Constrained discrete diffusion, 2025

work page 2025

[3] [3]

Adversarial tokenization, 2025

Renato Lui Geh, Zilei Shao, and Guy Van den Broeck. Adversarial tokenization, 2025

work page 2025

[4] [4]

Chi, Quoc V

Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V . Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. In Sanmi Koyejo, S. Mohamed, A. Agarwal, Danielle Belgrave, K. Cho, and A. Oh, editors, Advances in Neural Information Processing Systems 35: Annual Conference on Neur...

work page 2022

[5] [5]

Griffiths, Yuan Cao, and Karthik Narasimhan

Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: deliberate problem solving with large language models. In Proceedings of the 37th Interna- tional Conference on Neural Information Processing Systems , NIPS ’23, Red Hook, NY , USA, 2023. Curran Associates Inc

work page 2023

[6] [6]

Fast lexically constrained decoding with dynamic beam allocation for neural ma- chine translation

Matt Post and David Vilar. Fast lexically constrained decoding with dynamic beam allocation for neural ma- chine translation. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers) , pages 1314–1324, New Orleans, Louisiana, June 2018. Associa...

work page 2018

[7] [7]

Smith, and Yejin Choi

Ximing Lu, Sean Welleck, Peter West, Liwei Jiang, Jungo Kasai, Daniel Khashabi, Ronan Le Bras, Lianhui Qin, Youngjae Yu, Rowan Zellers, Noah A. Smith, and Yejin Choi. NeuroLogic A*esque decoding: Constrained text generation with lookahead heuristics. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational L...

work page 2022

[8] [8]

Constraints first: A new mdd-based model to generate sentences under constraints

Alexandre Bonlarron, Aur ´elie Calabr`ese, Pierre Kornprobst, and Jean-Charles R ´egin. Constraints first: A new mdd-based model to generate sentences under constraints. In Edith Elkind, editor, Proceedings of the Thirty- Second International Joint Conference on Artificial Intelligence, IJCAI-23, pages 1893–1901, 2023

work page 1901