Pith. sign in

REVIEW 3 major objections 4 minor 56 references

Trie Automata for Constrained Decoding over Large Finite Sets

T0 review · 3 major / 4 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read A trie automaton with precomputed token masks makes constrained decoding over finite sets up to 29× faster in batch serving while producing outputs identical to grammar-based constrained decoding.

desk verdict Solid systems contribution that deserves serious review, but the 'flat per-step cost regardless of set size' claim is a scope condition, not a proven property. read the letter →

arxiv 2608.12574 v1 pith:UI73PMUT submitted 2026-08-12 cs.AI cs.FL

Xingzi Xu , Karim Bouyarmane This is my paper · ORCID
classification cs.AIcs.FL
keywords constraineddecodingtrieautomatonfinite-setconstraintstokenmaskingAho-Corasickmulti-patternmatchingstructuredgenerationlargelanguagemodelsenumcardinality
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 establish that finite-set constraints—selecting one string from a list of thousands of tool names, labels, or codes—should not be compiled through general-purpose grammar finite-state machines, but enforced with a specialized character-level trie whose nodes carry precomputed valid-token masks. It argues that finite sets have exploitable structure (shared prefixes, bounded depth, known cardinality) and that the hard part, aligning byte-pair-encoding tokens that span multiple trie edges with character-level trie nodes, reduces to multi-pattern string matching solvable in near-linear time. If the central claim is right, constrained decoding over sets of 10,000 or more values stops being a slowly degrading bottleneck: per-step valid-token computation becomes a cached lookup around 0.65 microseconds, compilation stays under 100 milliseconds, batch serving throughput rises by up to 29×, and outputs remain identical to grammar-based constrained decoding with 100% validity. The practical consequence is that large tool registries, classification label sets, and entity-link targets can be used as hard constraints rather than capped at a few hundred or a thousand values.

What carries the argument

The central mechanism is a character-level trie built from the finite set of valid strings, with per-node precomputed token masks. To fill those masks, the paper uses an Aho-Corasick multi-pattern matcher—an automaton that scans a text once and reports all occurrences of any vocabulary token—to align BPE tokens with trie paths: a depth-first traversal of the trie carries the matcher state, and each reported token is recorded in the valid set of the node where it starts, provided it does not overshoot a leaf. This turns precomputation from $O(N_{\mathrm{chars}}\cdot V\cdot\ell)$ into $O((N_{\mathrm{chars}}+V)\cdot\ell)$, where $N_{\mathrm{chars}}$ is the total number of characters in the enum, $V$ is the vocabulary size, and $\ell$ is the maximum token length. At decode time masking is a cached lookup costing $O(|\mathrm{valid}[s_t]|)$, and because the tokenizer matcher is reusable across schemas, per-schema compilation reduces to the trie traversal. The automaton's nodes are exactly the minimal DFA states for the finite language, which is what makes output equivalence exact.

What would settle it

Build the trie for $K=1{,}000$ random strings with no shared prefixes (e.g., random UUIDs) using a large tokenizer, and measure the mean size of $\mathrm{valid}[\mathrm{root}]$ and of valid sets at depth 1–2, the per-step masking time, and batch throughput against the grammar backend. If the root valid set contains thousands of tokens and per-step masking time grows with $K$ instead of staying flat, the paper's flat-cost claim fails for unstructured finite sets.

Watch

Extended reading notes

Core claim

The central claim is Proposition 1: for any finite set of allowed strings and any decoding prefix, the constrained token distributions produced by the grammar finite-state machine and by the trie automaton are identical, so greedy and fixed-seed sampling produce exactly the same outputs. The trie automaton is not an approximation; it is the minimal DFA for the finite language, with each node storing the set of vocabulary tokens that are valid continuations. The load-bearing identity is that the BPE-trie alignment problem—knowing which multi-character tokens can start at each trie node—is a multi-pattern string matching problem, and therefore can be solved once per tokenizer in time linear in the trie size plus the vocabulary size, rather than per node times per token. With masks precomputed, decoding becomes a stateless lookup $A(s_t)=\mathrm{valid}[s_t]$, and the paper claims this yields 7× faster per-step masking, 2–6.5× faster compilation at $K\ge 300$, and up to 29× higher end-to-end batch throughput because the stateless path bypasses the guided-decoding pipeline.

Load-bearing premise

The practical promise of flat per-step cost regardless of set size rests on the enum's strings sharing prefixes, so the root node's valid-token set stays small and shrinks rapidly after three or four characters; for large sets of random strings with no shared prefixes, the root mask can approach the full vocabulary and the per-step advantage at shallow depths would shrink or vanish.

Editorial extensions

If this is right

  • Tool registries, classification label sets, and entity-link targets with tens of thousands of entries can be hard constraints instead of being capped at a few hundred or thousand values.
  • Because per-step masking is a cached lookup whose working set fits in L1 cache, CPU masking no longer idles the GPU in batch serving, removing a throughput bottleneck that grows with batch size.
  • Output equivalence with grammar-based constrained decoding means switching backends entails no accuracy or validity tradeoff; 100% of decoded outputs are in the allowed set.
  • Compilation stays under 100 ms for $K$ up to 10,000 across tokenizer vocabularies from 32K to 262K, enabling on-the-fly recompilation when the valid set changes per query.
  • The crossover with token-level tries sits near $K\approx 1{,}000$, exactly where provider enum limits and compile-timeouts previously stopped scaling, so the practical operating range expands by roughly two orders of magnitude.

Reading between the lines

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

  • Editorial inference: the flat-cost guarantee is structural, not universal: for enums whose strings collectively begin with most vocabulary-initial byte sequences (random IDs, hashes, UUIDs), the root valid set can approach $O(V)$, and the advantage over grammar backends at shallow prefix depths would shrink; the paper's benchmarks all use prefix-structured sets.
  • Editorial inference: because the matcher is tokenizer-specific and reusable, a multi-tenant serving layer could cache it once per tokenizer and pay only trie traversal per new schema, making per-query dynamic constraints nearly free; the paper states caching but stops short of quantifying multi-tenant savings.
  • Editorial inference: the dispatch principle generalizes beyond enums: fixed-format strings such as dates or UUIDs can be enforced by character-position masks, and the paper's appendix reports large compilation speedups there; this suggests a broader design in which constraint shape, not schema generality, chooses the enforcement engine.
  • Editorial inference: if serving engines eventually move logit masking onto the GPU, the CPU per-step advantage will matter less; the trie's contribution would then shift to its compact precomputed bitmasks and fast compilation rather than the stateless serving path.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes a trie automaton for constrained decoding from finite enum sets. It builds a character-level trie of the valid strings and uses Aho-Corasick multi-pattern matching over the tokenizer vocabulary to precompute, for every trie node, the set of BPE tokens that are valid continuations. Compilation is claimed to be O((Nchars+V)·ℓ) and per-step masking O(|valid[st]|). Proposition 1 asserts output equivalence with standard FSM-based constrained decoding. Experiments across seven tokenizer families report sub-100ms compilation up to K=10,000, 0.65µs per-step valid-token computation, and up to 29× vLLM batch throughput versus XGrammar, with 100% validity.

Significance. If the practical claims hold, the paper addresses a real production bottleneck: general-purpose grammar compilation for enum constraints becomes prohibitively slow at K in the hundreds to thousands, and the proposed specialized backend is a plausible drop-in replacement. The formal equivalence result, once its proof is repaired, and the complexity analysis are valuable; the controlled comparison in Appendix H.2 is a strong methodological step. The main weakness is that the headline 'flat per-step cost regardless of set size' rests on empirical prefix-structure assumptions that are not tested for unstructured enums. The paper is honest about many of these scope conditions, but the abstract and conclusion state the claim more broadly than the evidence supports.

major comments (3)
  1. [Appendix G.5, Proposition 1; Appendix H.2] The proof's Myhill-Nerode step is incorrect. Distinct prefixes of a finite string union are not necessarily distinct equivalence classes: for E={ab, cb}, the prefixes 'a' and 'c' have the same right language and are merged in the minimal DFA, while the trie keeps them as separate nodes. Therefore 'the trie is isomorphic to the minimal DFA for LE' is false in general, and the H.2 assertion that 'the number of DFA states equals the number of trie nodes (both are the minimal DFA for LE, per Proposition 1)' is not a valid consequence of Proposition 1. The output-equivalence conclusion can be recovered by noting that token validity at a prefix depends only on the prefix's right language, so the proof should be rewritten; as it stands, a load-bearing justification in the paper is wrong.
  2. [Section 4.2, Tables 14 and 15, Appendix G.3, Appendix H.7] The 'flat per-step cost regardless of set size' claim is not established for unstructured enums. The algorithm's per-step cost is O(|valid[st]|), and the paper concedes the worst case is O(V) at the root. The only root-size measurement (Table 15) uses synthetic tool names with prefix-sharing ratio r=0.40; for random strings or UUIDs, which Appendix H.7 itself lists as r≈0.95–1.0, the root's children cover most of the byte alphabet and |valid[root]| can be a substantial fraction of V, so shallow-depth per-step cost approaches the O(V) of the FSM scan the trie is meant to replace. No experiment measures per-step masking on such an unstructured enum, and the G.3 exponential-shrinkage bound is vacuous at depth 0 and assumes uniform independent characters. This does not affect the output-equivalence theorem, but it directly bounds the practical domain of the abstract's headline claim and should be scoped or tested.
  3. [Section 4.2, Table 15] The statement that |valid[st]| 'shrinks exponentially with trie depth' is not supported even by the paper's own data: Table 15 shows the mean valid-set size rising from 3 at depth 2 to 5–10 at depths 3–5 before declining. The non-monotonic behavior reflects real enum structure, but the exponential-shrinkage claim, which is used to justify the effective-constant per-step cost, needs a more careful empirical or analytical statement.
minor comments (4)
  1. [Abstract] The abstract contains a typo: 'acardinality wall' should be 'a cardinality wall'.
  2. [Table 4 and Appendix H.3] The main text reports 0.65µs per-step trie cost while H.3 reports 0.08µs for the raw lookup; the discrepancy is explained, but Table 4's caption should point readers to that reconciliation to avoid apparent inconsistency.
  3. [Section 5.1, Table 3] The statement that the trie 'exceeds unconstrained throughput' is clearly tied to the different token counts (3.2 vs. 8.7 tokens/request), but the sentence could be rephrased to make explicit that this part of the comparison is an output-length effect rather than a decoding-speed effect.
  4. [Appendix H.2] The 196× precomputation slowdown for XGrammar is obtained by instantiating a GrammarMatcher per state; since the paper itself notes that a purpose-built DFA traversal could be faster, the 196× figure should be labeled as an upper-bound illustration for the current API rather than a fundamental algorithmic limit.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the trie construction, complexity bounds, and equivalence proof are self-contained, with external baselines and no fitted parameters renamed as predictions.

full rationale

The paper's derivation chain is self-contained. The trie construction (Section 4.1) and Aho-Corasick mask precomputation (Section 4.2) are standard algorithmic reductions with explicit complexity bounds of O((Nchars+V)·ell) for compilation and O(|valid[st]|) for per-step masking, derived from textbook Aho-Corasick and trie properties rather than from the paper's own conclusions. Proposition 1 (output equivalence) is proved from Myhill-Nerode equivalence classes and transition-function agreement, not assumed, and the proof is independent of the benchmark numbers. The experimental comparisons are against external systems (XGrammar, LLGuidance, GENRE-style token-level tries), and the controlled comparison in Appendix H.2 isolates the precomputation-efficiency claim by measuring an XGrammar-based per-state mask precomputation path, which does not presuppose the trie's advantage. There are no fitted parameters that are later presented as predictions, and no load-bearing self-citations: the reference list contains no work by the present authors, and all cited prior results (Aho-Corasick 1975, Fredkin 1960, Myhill-Nerode, GENRE, XGrammar, LLGuidance) are external and independently checkable. The weakest practical assumption, that |valid[root]| is small for realistic enums, is explicitly conceded in Section 4.2 ('the worst case is O(V) (at the root node)') and is an empirical structural property of the benchmark enums, not a circular derivation; it limits the generality of the 'flat per-step cost' headline but does not contaminate the formal equivalence or the measured speedups.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claims rest on standard automata theory (trie-to-DFA isomorphism, Aho-Corasick guarantees), a domain assumption about byte-level BPE tokenizers, and the standard state-equivalence assumption of constrained decoding. No fitted parameters enter the complexity or equivalence arguments. The only hand-set constant is an illustrative cache-penalty factor in Appendix G.6 that explains, rather than produces, the measured FSM degradation. No new physical or conceptual entities are postulated.

free parameters (1)
  • cache penalty factor alpha = 2-3 (illustrative)
    Appendix G.6 introduces f(|S|) = alpha * |S| * |Sigma| * w / CL2 for FSM transition-table cache misses; alpha is set by hand to a plausible range. It is used to explain observed super-linear FSM degradation, not fitted to data and not load-bearing for the central speedup claims.
assumptions (4)
  • standard math Myhill-Nerode theorem: the minimal DFA for a finite string union is isomorphic to the character trie plus a dead state
    Used in Proposition 1 and Appendix G.5 to prove that trie and FSM admit identical token sets at each state.
  • standard math Aho-Corasick multi-pattern matching correctness and O(L+m) text-processing cost
    Foundation of the precomputation algorithm (Section 4.2); standard result from Aho & Corasick (1975).
  • domain assumption BPE tokenizers decompose into byte-level character strings that correspond to trie paths
    Section 2 treats the alphabet as 256 bytes; valid for the seven cited tokenizer families, but a property of those tokenizers rather than a theorem.
  • domain assumption The set of valid next tokens is a function only of the current character prefix state
    Standard constrained-decoding assumption (Equation 1, Section 2); it licenses replacing runtime FSM simulation with precomputed per-node masks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Trie Automata for Constrained Decoding over Large Finite Sets." pith.science (2026). https://pith.science/paper/UI73PMUT

@misc{pith2026260812574,
  author       = {Pith},
  title        = {Pith review of: Trie Automata for Constrained Decoding over Large Finite Sets},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UI73PMUT}},
  note         = {Machine review of arXiv:2608.12574}
}
read the original abstract

Large language models increasingly need to generate structured outputs that conform to predefined schemas, with one common constraint being selection from a finite set of valid strings. Current constrained decoding systems handle this through general-purpose grammar compilation, which becomes prohibitively slow as the number of valid values grows into the thousands, a cardinality wall. We introduce the trie automaton, a specialized mechanism that exploits finite-set structure (shared prefixes, bounded depth, known cardinality) via Aho-Corasick multi-pattern matching to precompute per-node token masks. The trie achieves 7X faster per-step valid-token computation (0.65 us vs. 5.8 us) compared to XGrammar, one of the primary backends in vLLM and SGLang, and 2--6.5X faster compilation at K >= 300. Because precomputed masks enable a stateless serving path that bypasses the guided decoding pipeline, this advantage compounds in batch serving: end-to-end vLLM throughput reaches 219 req/s vs. XGrammar's 7.5 req/s at batch size 256 (29X). The 29X combines the algorithmic speedup with integration-path savings that only precomputed masks can unlock. Across seven tokenizer families (32K--262K vocabulary), the trie maintains sub-100ms compilation up to K = 10,000 and flat per-step cost regardless of set size, while guaranteeing 100% output validity.

Figures

Figures reproduced from arXiv: 2608.12574 by the authors.

Figure 1
Figure 1. (a) Compilation time vs. enum cardinality (Qwen3-8B, log-log scale). Dotted lines [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The BPE-trie alignment problem. A character-level trie encodes three enum values [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

56 extracted references · 45 canonical work pages

  1. [1]

    2023 , eprint=

    Efficient Guided Generation for Large Language Models , author=. 2023 , eprint=

    2023

  2. [2]

    Gonzalez and Clark Barrett and Ying Sheng , booktitle=

    Lianmin Zheng and Liangsheng Yin and Zhiqiang Xie and Chuyue Sun and Jeff Huang and Cody Hao Yu and Shiyi Cao and Christos Kozyrakis and Ion Stoica and Joseph E. Gonzalez and Clark Barrett and Ying Sheng , booktitle=

    work page

  3. [3]

    NeurIPS , year=

    Grammar-Aligned Decoding , author=. NeurIPS , year=

    work page

  4. [4]

    EMNLP Industry Track , year=

    Let Me Speak Freely? A Study on the Impact of Format Restrictions on Performance of Large Language Models , author=. EMNLP Industry Track , year=

    work page

  5. [5]

    Grammar-Constrained Decoding for Structured

    Saibo Geng and Martin Josifoski and Maxime Peyrard and Robert West , booktitle=. Grammar-Constrained Decoding for Structured. 2023 , note=

    work page 2023

  6. [6]

    Ruan and Yaxing Cai and Ruihang Lai and Ziyi Xu and Yilong Zhao and Tianqi Chen , journal=

    Yixin Dong and Charlie F. Ruan and Yaxing Cai and Ruihang Lai and Ziyi Xu and Yilong Zhao and Tianqi Chen , journal=

    work page

  7. [7]

    ICML , year=

    Guiding LLMs The Right Way: Fast, Non-Invasive Constrained Generation , author=. ICML , year=

    work page

  8. [8]

    SynCode: LLM Generation with Grammar Augmentation , author=. Trans. Mach. Learn. Res. , year=

    work page

Show all 56 references
  1. [9]

    Saibo Geng and Hudson Cooper and Michał Moskal and Samuel Jenkins and Julian Berman and Nathan Ranchin and Robert West and Eric Horvitz and Harsha Nori , journal=

  2. [10]

    NeurIPS , year=

    Automata-based Constraints for Language Model Decoding , author=. NeurIPS , year=

  3. [11]

    Beurer-Kellner, Luca and Fischer, Marc and Vechev, Martin , title =. Proc. ACM Program. Lang. , month = jun, articleno =. 2023 , issue_date =. doi:10.1145/3591300 , abstract =

    2023 doi

  4. [12]

    NAACL , year=

    Constrained Decoding with Speculative Lookaheads , author=. NAACL , year=

  5. [13]

    arXiv preprint arXiv:2507.16768 , year=

    WGRAMMAR: Leverage Prior Knowledge to Accelerate Structured Decoding , author=. arXiv preprint arXiv:2507.16768 , year=

    arXiv

  6. [14]

    2007 , note=

    Regular Expression Matching Can Be Simple And Fast , author=. 2007 , note=

    2007

  7. [15]

    arXiv preprint arXiv:2407.12849 , year=

    Large Language Models are Good Medical Coders, if Provided with Tools , author=. arXiv preprint arXiv:2407.12849 , year=

    arXiv

  8. [16]

    arXiv preprint arXiv:2304.13998 , year=

    Mimic-IV-ICD: A new benchmark for eXtreme MultiLabel Classification , author=. arXiv preprint arXiv:2304.13998 , year=

    arXiv

  9. [17]

    2025 , howpublished=

    2026. 2025 , howpublished=

    2026

  10. [18]

    Introduction to Automata Theory, Languages, and Computation , author=

  11. [19]

    ICML , year=

    Hierarchically Classifying Documents Using Very Few Words , author=. ICML , year=

  12. [20]

    ICML , year=

    Hierarchical Multi-Label Classification Networks , author=. ICML , year=

  13. [21]

    NeurIPS , year=

    Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks , author=. NeurIPS , year=

  14. [22]

    NeurIPS , year=

    Sparse Local Embeddings for Extreme Multi-label Classification , author=. NeurIPS , year=

  15. [23]

    2021 , note=

    Ximing Lu and Peter West and Rowan Zellers and Ronan Le Bras and Chandra Bhagavatula and Yejin Choi , booktitle=. 2021 , note=

    2021

  16. [24]

    NAACL , year=

    NeuroLogic A*esque Decoding: Constrained Text Generation with Lookahead Heuristics , author=. NAACL , year=

  17. [25]

    EMNLP , year=

    PICARD: Parsing Incrementally for Constrained Auto-Regressive Decoding from Language Models , author=. EMNLP , year=

  18. [26]

    NAACL , year=

    Fast Lexically Constrained Decoding with Dynamic Beam Allocation for Neural Machine Translation , author=. NAACL , year=

  19. [27]

    ICML , year=

    Fast Inference from Transformers via Speculative Decoding , author=. ICML , year=

  20. [28]

    EMNLP , year=

    Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks , author=. EMNLP , year=

  21. [29]

    NeurIPS , year=

    Adaptable Logical Control for Large Language Models , author=. NeurIPS , year=

  22. [30]

    arXiv preprint arXiv:2502.09061 , year=

    CRANE: Reasoning with constrained LLM generation , author=. arXiv preprint arXiv:2502.09061 , year=

    arXiv

  23. [31]

    arXiv preprint arXiv:2502.14969 , year=

    Lost in Space: Finding the Right Tokens for Structured Output , author=. arXiv preprint arXiv:2502.14969 , year=

    arXiv

  24. [32]

    arXiv preprint arXiv:2505.04016 , year=

    SLOT: Structuring the Output of Large Language Models , author=. arXiv preprint arXiv:2505.04016 , year=

    arXiv

  25. [33]

    arXiv preprint arXiv:1704.03718 , year=

    Deep Extreme Multi-label Learning , author=. arXiv preprint arXiv:1704.03718 , year=

    arXiv

  26. [34]

    NeurIPS , year=

    A No-Regret Generalization of Hierarchical Softmax to Extreme Multi-Label Classification , author=. NeurIPS , year=

  27. [35]

    Communications of the ACM , volume=

    Trie Memory , author=. Communications of the ACM , volume=

  28. [36]

    EMNLP , year=

    Guided Open Vocabulary Image Captioning with Constrained Beam Search , author=. EMNLP , year=

  29. [37]

    ICLR , year=

    Autoregressive Entity Retrieval , author=. ICLR , year=

  30. [38]

    2024 , note=

    Yujia Qin and Shihao Liang and Yining Ye and Kunlun Zhu and Lan Yan and Yaxi Lu and Yankai Lin and Xin Cong and Xiangru Tang and Bill Qian and Sihan Zhao and Lauren Hong and Runchu Tian and Ruobing Xie and Jie Zhou and Mark Gerstein and Dahai Li and Zhiyuan Liu and Maosong Sun...

    2024

  31. [39]

    Xiang Fei and Xiawu Zheng and Hao Feng , journal=

  32. [40]

    arXiv preprint arXiv:2504.09135 , year=

    Efficient and Asymptotically Unbiased Constrained Decoding for Large Language Models , author=. arXiv preprint arXiv:2504.09135 , year=

    arXiv

  33. [41]

    Transactions of the ACL , year=

    Multilingual Autoregressive Entity Linking , author=. Transactions of the ACL , year=

  34. [42]

    arXiv preprint arXiv:2509.20386 , year=

    Dynamic ReAct: Scalable Tool Selection for Large-Scale MCP Environmentss , author=. arXiv preprint arXiv:2509.20386 , year=

  35. [43]

    Communications of the ACM , volume=

    Efficient String Matching: An Aid to Bibliographic Search , author=. Communications of the ACM , volume=

  36. [44]

    Communications of the ACM , volume=

    Programming Techniques: Regular Expression Search Algorithm , author=. Communications of the ACM , volume=

  37. [45]

    Theory of Machines and Computations , pages=

    An n n Algorithm for Minimizing States in a Finite Automaton , author=. Theory of Machines and Computations , pages=. 1971 , publisher=

    1971

  38. [46]

    2024 , howpublished=

    Structured Model Outputs , author=. 2024 , howpublished=

    2024

  39. [47]

    2024 , howpublished=

    Structured Outputs , author=. 2024 , howpublished=

    2024

  40. [48]

    2025 , howpublished=

    Structured Outputs , author=. 2025 , howpublished=

    2025

  41. [49]

    Vectorizing the Trie: Efficient Constrained Decoding for

    Zhengyang Su and Isay Katsman and Yueqi Wang and Ruining He and Lukasz Heldt and Raghunandan Keshavan and Shao-Chuan Wang and Xinyang Yi and Mingyan Gao and Onkar Dalal and Lichan Hong and Ed Chi and Ningren Han , journal=. Vectorizing the Trie: Efficient Constrained Decoding for

  42. [50]

    COLING , year=

    Learning Question Classifiers , author=. COLING , year=

  43. [51]

    Jack FitzGerald and Christopher Hench and Charith Peris and Scott Mackie and Kay Rottmann and Ana Sanchez and Aaron Nash and Liam Urbach and Vishesh Kakarala and Richa Singh and Swetha Ranganath and Laurie Crist and Misha Britan and Wouter Leeuwis and Gokhan Tur and Prem Natar...

  44. [52]

    Proceedings of the 2nd Workshop on Natural Language Processing for Conversational AI (NLP4ConvAI), ACL , year=

    Efficient Intent Detection with Dual Sentence Encoders , author=. Proceedings of the 2nd Workshop on Natural Language Processing for Conversational AI (NLP4ConvAI), ACL , year=

  45. [53]

    EMNLP-IJCNLP , year=

    An Evaluation Dataset for Intent Classification and Out-of-Scope Prediction , author=. EMNLP-IJCNLP , year=

  46. [54]

    EMNLP , year=

    Efficient Beam Search for Large Language Models Using Trie-Based Decoding , author=. EMNLP , year=

  47. [55]

    Computational Linguistics , volume=

    Tokenization as Finite-State Transduction , author=. Computational Linguistics , volume=

  48. [56]

    APSIPA ASC , year=

    Zero-shot Context Biasing with Trie-based Decoding using Synthetic Multi-Pronunciation , author=. APSIPA ASC , year=

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.