Towards reliable, controllable and more efficient generation with Large Language Models (LLMs)

A curated list of papers related to constrained decoding of LLM, along with their relevant code and resources.

• Awesome-LLM-Constrained-Decoding
  • Table of Contents
  • Libraries
  • Papers
  • Benchmark & Datasets & Evaluation
  • Survey
  • Blog Posts
  • Disclaimer
  • Contributing

Library	Feature	Stars
guidance-ai/guidance	CFG, Regex, JSON Schema, Token Forcing, compatible with Transformers, LLAMA-CPP
outlines-dev/outlines	CFG, Unicode support, Hugging Face ecosystem, VLLM support
sgl-project/sglang	Regex support, emphasis on LLM inference efficiency, compressed FSM
eth-sri/lmql	Regex support, various constraints, more powerful control flow
jxnl/instructor	Try-Reject-Repeat approach to ensure constraints are met
noamgat/lm-format-enforcer	Regex, JSON Schema, Beam Search etc.
epfl-dlab/transformers-CFG	CFG (EBNF Interface), Compatible with Transformers, Easy to extend for research
uiuc-focal-lab/syncode	CFG generation that supports builtin grammars like JSON, Python, Go, and more

Disclaimer:

• The libraries listed above are not exhaustive and are subject to change.
• The features mentioned are 100% not exhaustive and I strongly recommend checking the respective repositories for more details.
• The libraries are listed by the Github stars
• If you are the author of a library and would like to add or update the information, please open an issue or submit a pull request.

Papers with are newly added papers (not necessarily newly published papers).

Date	Paper	Publication
2024-08	Let Me Speak Freely? A Study on the Impact of Format Restrictions on Performance of Large Language Models	Arxiv
2024-08	FANTAstic SEquences and Where to Find Them: Faithful and Efficient API Call Generation through State-tracked Constrained Decoding and Reranking	Arxiv
2024-07	Automata-based constraints for language model decoding	CoLM
2024-06	Sketch-Guided Constrained Decoding for Boosting Blackbox Large Language Models without Logit Access	ACL
2024-05	Grammar-Aligned Decoding	Preprint
2024-03	SynCode: LLM Generation with Grammar Augmentation	Arxiv
2024-03	Guiding LLMs The Right Way: Fast, Non-Invasive Constrained Generation	ICML
2024-02	Constrained Decoding for Code Language Models via Efficient Left and Right Quotienting of Context-Sensitive Grammars	Arxiv
2024-02	Formal-LLM: Integrating Formal Language and Natural Language for Controllable LLM-based Agents	Arxiv
2023-12	SGLang: Efficient Execution of Structured Language Model Programs	Preprint
2023-12	Monitor-Guided Decoding of Code LMs with Static Analysis of Repository Context	NeurIPS
2023-11	Prompt Sketching for Large Language Models	Preprint
2023-11	Sequential Monte Carlo Steering of Large Language Models using Probabilistic Programs	PADL
2023-10	Don't Fine-Tune, Decode: Syntax Error-Free Tool Use via Constrained Decoding	Arxiv
2023-10	Amortizing intractable inference in large language models	ICLR
2023-10	KCTS: Knowledge-Constrained Tree Search Decoding with Token-Level Hallucination Detection	EMNLP
2023-07	Efficient Guided Generation for Large Language Models	Arxiv
2023-06	Grammar Prompting for Domain-Specific Language Generation with Large Language Models	NeurIPS
2023-06	Grammar-Constrained Decoding for Structured NLP Tasks without Finetuning	EMNLP
2023-06	Prompting Is Programming: A Query Language for Large Language Models	PLDI
2023-05	Measuring and Mitigating Constraint Violations of In-Context Learning for Utterance-to-API Semantic Parsing	EMNLP Findings
2023-04	Tractable Control for Autoregressive Language Generation	ICML
2022-11	Validating Large Language Models with ReLM	MLSys
2022-11	CodePAD: Sequence-based Code Generation with Pushdown Automaton	ISSTA
2022-05	Gradient-Based Constrained Sampling from Language Models	EMNLP
2022-01	Synchromesh: Reliable code generation from pre-trained language models	ICLR
2021-12	PICARD: Parsing Incrementally for Constrained Auto-Regressive Decoding from Language Models	EMNLP
2021-12	Constrained Language Models Yield Few-Shot Semantic Parsers	EMNLP
2021-12	Controlled Text Generation as Continuous Optimization with Multiple Constraints	NeurIPS
2021-06	NEUROLOGIC DECODING:(Un)supervised Neural Text Generation with Predicate Logic Constraints	NAACL
2019-05	A General-Purpose Algorithm for Constrained Sequential Inference	CoNLL
2019-05	Improved Lexically Constrained Decoding for Translation and Monolingual Rewriting	NAACL
2018-09	CGMH: Constrained Sentence Generation by Metropolis-Hastings Sampling	AAAI
2018-05	Fast Lexically Constrained Decoding with Dynamic Beam Allocation for Neural Machine Translation	NAACL
2018-04	Incorporating Discriminator in Sentence Generation: a Gibbs Sampling Method	AAAI
2017-12	Guided Open Vocabulary Image Captioning with Constrained Beam Search	EMNLP
2017-06	Lexically Constrained Decoding for Sequence Generation Using Grid Beam Search	ACL

Date	Paper	Publication
2024-05	COLLIE: Systematic Construction of Constrained Text Generation Tasks	ICLR
2023-12	BenchCLAMP: A Benchmark for Evaluating Language Models on Syntactic and Semantic Parsing	NeurIPS Track on Datasets and Benchmarks
2023-10	Evaluating Large Language Models on Controlled Generation Tasks	Arxiv
2023-09	Struc-Bench: Are Large Language Models Really Good at Generating Complex Structured Data?	Arxiv
2020-12	CommonGen: A Constrained Text Generation Challenge for Generative Commonsense Reasoning	EMNLP Findings

Date	Paper	Publication
2024-04	"We Need Structured Output": Towards User-centered Constraints on Large Language Model Output	Arxiv

• Leveraging Constrained Sampling for Fill-in-the-Middle Code Completion by nielstron
• Proper Well-Formedness for Finite LLM Sampling by nielstron
• LLM Decoding with Regex Constraints by Vivien
• Constrained Decoding is Posterior Inference by Saibo-creator
• Making Structured Generation Faster Than Unstructured
• Coding For Structured Generation with LLMs
• Beating GPT-4 with Open Source
• Prompt Efficiency - Using Structured Generation to get 8-shot performance from 1-shot.
• How fast can grammar-structured generation be?
• Structured Generation Improves LLM performance: GSM8K Benchmark
• Coalescence: making LLM inference 5x faster

Many of the blogs are written by Outlines team, many thanks to them for their great work! ❤️

This list is not exhaustive and will be updated regularly. If you have any suggestions or want to add a paper, please feel free to open an issue or submit a pull request. We hope to include all relevant papers in this list.

Contributions are welcome! Feel free to submit a pull request or open an issue. Please make sure to read the Contributing Guidelines before contributing.