constrained-generation

Nedimo presents an evidence-grounded, multilingual pipeline for generating low-arousal AI sleep stories, separating deterministic planning from LLM generation with attention tapering, validation guardrails, and TTS loudness normalization, while positioning the system as a relaxation tool without clinical efficacy claims.

提出 Primal-Dual Flow Matching (PDFM)，一种通过原始-对偶优化实现样本级约束满足的生成方法，仅需二元成员预言即可直接控制约束满足率，并在保持分布保真度的同时优于标准流匹配基线。

The paper introduces function-symbol grounding in LTN-GANs to handle hard structural constraints in generative models, showing it learns margin distributions more faithfully than existing methods, as validated on high-resolution datasets.

This paper introduces a tractable method to control autoregressive large language model generation to satisfy LR(k) context-free grammars in polynomial time, improving over exponential previous methods. It demonstrates that current LLMs often fail to generate sequences satisfying simple nested constraints, motivating the need for efficient constrained generation.

This paper proposes a constrained, verifiable agent framework for open-web data collection that shifts LLM output from free-form code to typed JSON collector configurations, achieving zero execution-stage LLM tokens and low latency on 80 tasks.

This paper evaluates how LLMs perform in the game Taboo under lexical constraints, showing trade-offs between compliance and communicative effectiveness, and finding that models are weaker guessers than humans.

Introduces Constrained Tabular Diffusion for Finance (CTDF), a method that integrates feasibility operations into diffusion sampling to enforce hard constraints for financial synthetic data generation.

This paper proposes a two-stage fine-tuning pipeline combining domain-adaptive fine-tuning and reinforcement learning to generate protein sequences that match a desired amino-acid composition profile while maintaining sequence quality.

LatticeBridge proposes a twisted sequential Monte Carlo decoder for structured sequence generation that improves constraint satisfaction by treating the problem as rare-event inference, outperforming greedy and beam baselines on CommonGen, E2E NLG, and WikiBio.

This paper presents a system for constrained humor generation that uses a generate-many select-best strategy with a preference model learned from human comparisons. It achieved top ranks in English and Chinese subtasks and second in Spanish at SemEval-2026 Task 1.