Awesome-Parallel-Text-Generation

September 22, 2026 · View on GitHub

Our Survey

A Survey on Parallel Text Generation: From Parallel Decoding to Diffusion Language Models · PDF

A systematic survey of token-level and task-level parallel text generation. This index follows the revised manuscript's taxonomy; the linked arXiv version may precede the manuscript revision.

Taxonomy

Token-level and task-level parallelism are composable: concurrent tasks can each use AR-based, Non-AR-based, or hybrid decoding. Categories organize mechanisms, not mutually exclusive complete systems. Drafting and verification can also be combined, and masked generation overlaps with masked-edit refinement. Papers are indexed by the mechanism discussed in the survey. A dash in the Code column means that no code link has been verified for that entry.

Token-Level Parallelism

AR-Based

Draft-and-verify accelerates a target autoregressive decoder through multi-token proposals and parallel verification. Greedy verification and target-distribution-preserving speculative sampling have different correctness guarantees.

Drafter Design

External Drafter
PaperVenueCode
Speculative Decoding with Big Little DecoderNeurIPS 2023Github
Cascade speculative drafting for even faster llm inferenceNeurIPS 2023Github
Distillspec: Improving speculative decoding via knowledge distillationICLR 2024-
Fast Inference from Transformers via Speculative DecodingICML 2023Github
Online Speculative DecodingICML 2024Github
Speculative Decoding: Exploiting Speculative Execution for Accelerating Seq2seq GenerationEMNLP 2023Github
Accelerating Large Language Model Decoding with Speculative Samplingarxiv 2023Github
Accelerating LLM Inference with Staged Speculative DecodingarXiv 2023
Self-Drafting via Partial Computation

BranchyNet is an early-exit precursor, not itself an LLM self-speculative decoding system.

PaperVenueCode
Draft & Verify: Lossless Large Language Model Acceleration via Self-Speculative DecodingACL 2024Github
Speculative Decoding via Early-Exiting for Faster LLM Inference with Thompson Sampling Control MechanismACL 2024-
Kangaroo: Lossless self-speculative decoding via double early exitingNeurIPS 2024Github
Layer-skip: Enabling early-exit inference and self-speculative decodingACL 2024Github
Speed: speculative pipelined execution for efficient decodingNeurIPS-ENLSP 2023-
Swift: On-the-fly self-speculative decoding for llm inference accelerationICLR 2025Github
CAS-Spec: Cascade Adaptive Self-Speculative Decoding for On-the-Fly Lossless Inference Acceleration of LLMsNeurIPS 2025-
BranchyNet: Fast Inference via Early Exiting from Deep Neural NetworksICPR 2016
Self-Drafting via Parallel Prediction

Multi-token prediction is part of drafter design, rather than a separate category alongside draft-and-verify.

PaperVenueCode
EAGLE: Speculative Sampling Requires Rethinking Feature UncertaintyICML 2024Github
Falcon: Faster and Parallel Inference of Large Language Models through Enhanced Semi-Autoregressive Drafting and Custom-designed Decoding TreeAAAI 2025Github
Learning Harmonized Representations for Speculative SamplingICLR 2025Github
Hydra: Sequentially-dependent draft heads for medusa decodingCOLM 2024Github
Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding HeadsICML 2024Github
Mixture of Attentions for Speculative DecodingICLR 2025Github
Pass: Parallel speculative samplingNeurIPS-ENLSP 2023-
L-MTP: Leap Multi-Token Prediction Beyond Adjacent Context for Large Language Modelsarxiv 2025-
On multi-token prediction for efficient LLM inferencearxiv 2025-
Multi-Token Prediction Needs Registersarxiv 2025Github
Blockwise Parallel Decoding for Deep Autoregressive ModelsNeurIPS 2018-
Your LLM Knows the Future: Uncovering Its Multi-Token Prediction Potentialarxiv 2025-
ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-trainingEMNLP 2020Github
Better & faster large language models via multi-token predictionICML 2024-
Deepseek-v3 technical reportarxiv 2024Github
MiMo: Unlocking the Reasoning Potential of Language Model--From Pretraining to Posttrainingarxiv 2025Github
EAGLE-3: Scaling up Inference Acceleration of Large Language Models via Training-Time TestNeurIPS 2025Github

Drafting controls and complementary speculative extensions include adaptive draft lengths, alignment, diffusion proposals, and reasoning-level speculation; they are not additional mutually exclusive drafter types.

PaperVenueCode
A Drop-in Solution for On-the-fly Adaptation of Speculative Decoding in Large Language ModelsACL 2025-
Specdec++: Boosting speculative decoding via adaptive candidate lengthsCOLM 2025Github
SpecReason: Fast and Accurate Inference-Time Compute via Speculative ReasoningNeurIPS 2025Github
Fail Fast, Win Big: Rethinking the Drafting Strategy in Speculative Decoding via Diffusion LLMsarXiv 2025Github
Scaling Speculative Decoding with Lookahead ReasoningNeurIPS 2025Github
Griffin: Effective token alignment for faster speculative decodingNeurIPS 2025Github
Speculative diffusion decoding: Accelerating language generation through diffusionNAACL 2025

Verification Strategies

Efficient Verification

Tree/graph verification and scheduling methods reduce verification or end-to-end draft-and-verify latency.

PaperVenueCode
Adaptive Draft-Verification for Efficient Large Language Model DecodingAAAI 2025Github
Dynamic Depth Decoding: Faster Speculative Decoding for LLMsarxiv 2024-
DySpec: Faster Speculative Decoding with Dynamic Token Tree StructureWWW 2025-
Eagle-2: Faster Inference of Language Models with Dynamic Draft TreesEMNLP 2024Github
Graph-Structured Speculative DecodingACL 2024Github
OPT-Tree: Speculative Decoding with Adaptive Draft Tree StructureTACL 2025Github
Ouroboros: Generating Longer Drafts Phrase by Phrase for Faster Speculative DecodingEMNLP 2024Github
Parallel Speculative Decoding with Adaptive Draft LengthICLR 2025Github
PipeInfer: Accelerating LLM Inference using Asynchronous Pipelined SpeculationSC 2024Github
Predictive Pipelined Decoding: A Compute-Latency Trade-off for Exact LLM DecodingTMLR 2024-
Sequoia: Scalable, robust, and hardware-aware speculative decodingarxiv 2024Github
Generation meets verification: Accelerating large language model inference with smart parallel auto-correct decodingACL 2024Github
Specinfer: Accelerating generative large language model serving with tree-based speculative inference and verificationASPLOS 2024Github
STree: Speculative Tree Decoding for Hybrid State-Space ModelsNeurIPS 2025Github
Accurate Verification

These methods improve acceptance or verification decisions. Relaxed or learned acceptance rules do not automatically preserve the target sampling distribution.

PaperVenueCode
Block Verification Accelerates Speculative DecodingICLR 2025-
Dynamic-Width Speculative Beam Decoding for Efficient LLM InferenceAAAI 2025Github
Judge Decoding: Faster Speculative Sampling Requires Going Beyond Model AlignmentICLR 2025-
Optimized multi-token joint decoding with auxiliary model for llm inferenceICLR 2025Github
ProPD: Dynamic Token Tree Pruning and Generation for LLM Parallel DecodingICCAD 2024-
REST: Retrieval-based speculative decodingNAACL 2024Github
Recursive Speculative Decoding: Accelerating LLM Inference via Sampling without ReplacementICLR-LLMA 2024-
SpecTr: Fast Speculative Decoding via Optimal TransportNeurIPS 2023-
Towards Optimal Multi-draft Speculative DecodingICLR 2025
Multi-Draft Speculative Sampling: Canonical Decomposition and Theoretical LimitsarXiv 2024
A Theoretical Perspective for Speculative Decoding AlgorithmNeurIPS 2024

Non-AR-Based

One-Shot Generation

PaperVenueCode
Non-autoregressive neural machine translationICLR 2018Github
End-to-end non-autoregressive neural machine translation with connectionist temporal classificationEMNLP 2018
Lava nat: A non-autoregressive translation model with look-around decoding and vocabulary attentionarxiv 2025-
AligNART: Non-autoregressive neural machine translation by jointly learning to estimate alignment and translateEMNLP 2021-
Guiding non-autoregressive neural machine translation decoding with reordering informationAAAI 2021Github
Non-monotonic latent alignments for ctc-based non-autoregressive machine translationNeurIPS 2022Github
DePA: Improving Non-autoregressive Translation with Dependency-Aware DecoderIWSLT 2023Github
Directed acyclic transformer for non-autoregressive machine translationICML 2022Github
Viterbi decoding of directed acyclic transformer for non-autoregressive machine translationEMNLP 2022Github
Fully Non-autoregressive Neural Machine Translation: Tricks of the TradeACL-IJCNLP 2021-
Aligned cross entropy for non-autoregressive machine translationICML 2020Github
ngram-OAXE: Phrase-based order-agnostic cross entropy for non-autoregressive machine translationCOLING 2022Github
Multi-granularity optimization for non-autoregressive translationEMNLP 2022Github
Multilingual Non-Autoregressive Machine Translation without Knowledge DistillationIJCNLP-AACL 2023-
Tree-Structured Non-Autoregressive Decoding for Sequence-to-Sequence Text GenerationEMNLP 2025Github
Enriching Non-Autoregressive Transformer with Syntactic and Semantic Structures for Neural Machine TranslationarXiv 2021
Non-autoregressive text generation with pre-trained language modelsEACL 2021
One Reference Is Not Enough: Diverse Distillation with Reference Selection for Non-Autoregressive TranslationNAACL 2022Github
What have we achieved on non-autoregressive translation?Findings ACL 2024
Non-Autoregressive Machine Translation with Latent AlignmentsEMNLP 2020
Glancing transformer for non-autoregressive neural machine translationACL-IJCNLP 2021Github

Masked Generation

This category includes masked-token generation, diffusion formulations, and their training and inference optimizations, as well as related theoretical and evaluation studies. Diffusion-LM and Plaid are included as continuous-diffusion precursors, not as masked diffusion models.

PaperVenueCode
Masked diffusion models are secretly time-agnostic masked models and exploit inaccurate categorical samplingICLR 2025-
A continuous time framework for discrete denoising modelsNeurIPS 2022Github
Discrete diffusion modeling by estimating the ratios of the data distributionICML 2024Github
Simplified and generalized masked diffusion for discrete dataNeurIPS 2024google-deepmind/md4 · darioShar/pytorch-md4
Seed Diffusion: A Large-Scale Diffusion Language Model with High-Speed Inferencearxiv 2025-
Target concrete score matching: A holistic framework for discrete diffusionICML 2025-
Score-based continuous-time discrete diffusion modelsICLR 2023-
Fast-dllm: Training-free acceleration of diffusion llm by enabling kv cache and parallel decodingarxiv 2025Github
Large language diffusion modelsICLR 2025Github
Beyond autoregression: Discrete diffusion for complex reasoning and planningICLR 2025Github
A reparameterized discrete diffusion model for text generationCOLM 2024Github
Train for the Worst, Plan for the Best: Understanding Token Ordering in Masked DiffusionsICML 2025-
Accelerated Sampling from Masked Diffusion Models via Entropy Bounded Unmaskingarxiv 2025-
Accelerating Diffusion Large Language Models with SlowFast: The Three Golden Principlesarxiv 2025Github
Remasking discrete diffusion models with inference-time scalingICLR 2025Github
Path planning for masked diffusion model samplingarxiv 2025Github
Think while you generate: Discrete diffusion with planned denoisingICLR 2025Github
Accelerating Diffusion LLMs via Adaptive Parallel Decodingarxiv 2025-
Reviving any-subset autoregressive models with principled parallel sampling and speculative decodingarxiv 2025Github
dkv-cache: The cache for diffusion language modelsarxiv 2025Github
Accelerating diffusion language model inference via efficient kv caching and guided diffusionarxiv 2025-
Esoteric Language Modelsarxiv 2025Github
Beyond Autoregression: Fast LLMs via Self-Distillation Through TimeICLR 2025-
Cllms: Consistency large language modelsICML 2024Github
The diffusion dualityICML 2025Github
d1: Scaling reasoning in diffusion large language models via reinforcement learningarxiv 2025Github
LLaDA 1.5: Variance-Reduced Preference Optimization for Large Language Diffusion ModelsAAAI 2025Github
DiffuCoder: Understanding and Improving Masked Diffusion Models for Code Generationarxiv 2025Github
Scaling diffusion language models via adaptation from autoregressive modelsICLR 2025Github
Dream 7B: Diffusion Large Language Modelsarxiv 2025Github
DiffPO: Diffusion-styled Preference Optimization for Inference Time Alignment of Large Language ModelsACL 2025Github
Wide-In, Narrow-Out: Revokable Decoding for Efficient and Effective DLLMsarxiv 2025Github
Dream-Coder 7B: An Open Diffusion Language Model for Codearxiv 2025Github
Spg: Sandwiched policy gradient for masked diffusion language modelsarxiv 2025Github
Revolutionizing reinforcement learning framework for diffusion large language modelsICLR 2026Github
Diffusion llms can do faster-than-ar inference via discrete diffusion forcingICLR 2026Github
WeDLM: Reconciling Diffusion Language Models with Standard Causal Attention for Fast Inferencearxiv 2025Github
d-TreeRPO: Towards More Reliable Policy Optimization for Diffusion Language Modelsarxiv 2025Github
d2: Improved Techniques for Training Reasoning Diffusion Language Modelsarxiv 2025-
wd1: Weighted Policy Optimization for Reasoning in Diffusion Language Modelsarxiv 2025Github
Step-Aware Policy Optimization for Reasoning in Diffusion Large Language Modelsarxiv 2025Github
Improving reasoning for diffusion language models via group diffusion policy optimizationarxiv 2025Github
The Flexibility Trap: Why Arbitrary Order Limits Reasoning Potential in Diffusion Language Modelsarxiv 2026Github
Principled rl for diffusion llms emerges from a sequence-level perspectiveICLR 2026Github
DreamOn: Diffusion Language Models For Code Infilling Beyond Fixed-size Canvasarxiv 2026Github
SDAR: A Synergistic Diffusion-AutoRegression Paradigm for Scalable Sequence Generationarxiv 2025Github
Lopa: Scaling dllm inference via lookahead parallel decodingarxiv 2025Github
FAST-dLLM V2: Efficient Block-Diffusion LLMICLR 2026Github
d3LLM: Ultra-Fast Diffusion LLM using Pseudo-Trajectory Distillationarxiv 2026Github
dParallel: Learnable Parallel Decoding for dLLMsICLR 2026Github
Diffusion language models know the answer before decodingICLR 2026Github
Creditdecoding: Accelerating parallel decoding in diffusion large language models with trace creditsarxiv 2025-
Block Diffusion: Interpolating Between Autoregressive and Diffusion Language ModelsICLR 2025Github
Set Block Decoding is a Language Model Inference Acceleratorarxiv 2025-
LLaDA-MoE: A Sparse MoE Diffusion Language Modelarxiv 2025Github
dInfer: An Efficient Inference Framework for Diffusion Language Modelsarxiv 2025Github
ParallelBench: Understanding the Trade-offs of Parallel Decoding in Diffusion LLMsICLR 2026Github
Mask-Predict: Parallel Decoding of Conditional Masked Language ModelsEMNLP-IJCNLP 2019Github
Likelihood-Based Diffusion Language ModelsNeurIPS 2023Github
Diffusion-LM Improves Controllable Text GenerationNeurIPS 2022Github
Gemini DiffusionTechnical release 2025
Mercury: Ultra-Fast Language Models Based on DiffusionarXiv 2025
LLaDA2.0: Scaling Up Diffusion Language Models to 100BarXiv 2025Github
LLaDA2.1: Speeding Up Text Diffusion via Token EditingarXiv 2026Github
DiffusionGemma Model CardModel card 2026
Diffusion Beats Autoregressive in Data-Constrained SettingsarXiv 2025
Fast-decoding diffusion language models via progress-aware confidence schedulesFindings ACL 2026
From Bits to Rounds: Parallel Decoding with Exploration for Diffusion Language ModelsarXiv 2025
Learning to Parallel: Accelerating Diffusion Large Language Models via Learnable Parallel DecodingarXiv 2025
Introspective Diffusion Language ModelsarXiv 2026
ReFusion: A Diffusion Large Language Model with Parallel Autoregressive DecodingICLR 2026Github
LLaDA-V: Large Language Diffusion Models with Visual Instruction TuningarXiv 2025
Dimple: Discrete Diffusion Multimodal Large Language Model with Parallel DecodingarXiv 2025Github
Diffusion Language Models Can Perform Many Tasks with Scaling and Instruction-FinetuningarXiv 2023
Your Absorbing Discrete Diffusion Secretly Models the Conditional Distributions of Clean DataICLR 2025Github
Simple and Effective Masked Diffusion Language ModelsNeurIPS 2024Github
Scaling up Masked Diffusion Models on TextICLR 2025Github
Diffusion Language Models are Super Data LearnersarXiv 2025
Locally Confident, Globally Stuck: The Quality-Exploration Dilemma in Diffusion Language ModelsarXiv 2026
Structured Denoising Diffusion Models in Discrete State-SpacesNeurIPS 2021
Accelerating Transformer Inference for Translation via Parallel DecodingarXiv 2023
CtrlDiff: Boosting Large Diffusion Language Models with Dynamic Block Prediction and Controllable GenerationarXiv 2025
dLLM-Cache: Accelerating Diffusion Large Language Models with Adaptive CachingarXiv 2025Github
Attention Is All You Need for KV Cache in Diffusion LLMsICLR 2026Github
Generation Order and Parallel Decoding in Masked Diffusion Models: An Information-Theoretic PerspectivearXiv 2026
Theoretical Benefit and Limitation of Diffusion Language ModelarXiv 2025
Is Your Diffusion Sampler Actually Correct? A Sampler-Centric Evaluation of Discrete Diffusion Language ModelsarXiv 2026
Parallelism and Generation Order in Masked Diffusion Language Models: Limits Today, Potential TomorrowarXiv 2026

The following works underpin the formulations, decoding strategies, or training objectives discussed in this section and the appendices. Image-generation methods and general training objectives are background sources, not additional evidence of parallel text-generation speedups.

PaperVenueRole
MaskGIT: Masked Generative Image TransformerCVPR 2022Masked image-generation precursor
Improved Masked Image Generation with Token-CriticECCV 2022Token-selection precursor
Consistency ModelsICML 2023Step-reduction foundation
Improved Techniques for Training Consistency ModelsarXiv 2023Step-reduction foundation
Simplifying, Stabilizing and Scaling Continuous-Time Consistency ModelsarXiv 2024Step-reduction foundation
Denoising Diffusion Probabilistic ModelsNeurIPS 2020Diffusion formulation
Deep unsupervised learning using nonequilibrium thermodynamicsICML 2015Diffusion formulation
Score-Based Generative Modeling through Stochastic Differential EquationsICLR 2021Continuous-time formulation
Approximate accelerated stochastic simulation of chemically reacting systemsJ. Chem. Phys. 2001Tau-leaping foundation
Concrete Score Matching: Generalized Score Matching for Discrete DataNeurIPS 2022Discrete score matching
A deep and tractable density estimatorICML 2014Order-agnostic modeling
Training and Inference on Any-Order Autoregressive Models the Right WayNeurIPS 2022Any-order modeling
Autoregressive Diffusion ModelsarXiv 2021Any-order / diffusion connection
XLNet: Generalized Autoregressive Pretraining for Language UnderstandingNeurIPS 2019Permutation language modeling
Direct Preference Optimization: Your Language Model is Secretly a Reward ModelNeurIPS 2023Preference-optimization foundation
DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language ModelsarXiv 2024GRPO training foundation

Edit-Based Refinement

Following the survey, this section covers discrete editing, continuous optimization, hybrid refinement, and masked-edit refinement, together with supporting training methods. Hybrid self-correction may use AR generators internally; iterative refinement alone does not establish a parallel speedup.

PaperVenueCode
Insertion transformer: Flexible sequence generation via insertion operationsICML 2019-
Levenshtein transformerNeurIPS 2019Github
EDITOR: An edit-based transformer with repositioning for neural machine translation with soft lexical constraintsTACL 2021Github
FELIX: Flexible Text Editing Through Tagging and InsertionEMNLP 2020-
Levenshtein OCRECCV 2022Github
FastCorrect: Fast Error Correction with Edit Alignment for Automatic Speech RecognitionNeurIPS 2021Github
Non-autoregressive Text Editing with Copy-aware Latent AlignmentsEMNLP 2023Github
Reinforcement Learning for Edit-Based Non-Autoregressive Neural Machine TranslationNAACL-SRW 2024-
Summarizing Like Human: Edit-Based Text Summarization with KeywordsICANN 2024-
Deterministic non-autoregressive neural sequence modeling by iterative refinementEMNLP 2018Github
Flowseq: Non-autoregressive conditional sequence generation with generative flowEMNLP 2019Github
Latent-variable non-autoregressive neural machine translation with deterministic inference using a delta posteriorAAAI 2020Github
Iterative Refinement in the Continuous Space for Non-Autoregressive Neural Machine TranslationEMNLP 2020Github
Non-autoregressive machine translation with auxiliary regularizationAAAI 2019-
Imitation learning for non-autoregressive neural machine translationACL 2019-
An imitation learning curriculum for text editing with non-autoregressive modelsACL 2022Github
Fast structured decoding for sequence modelsNeurIPS 2019Github
An EM approach to non-autoregressive conditional sequence generationICML 2020-
Imputer: Sequence modelling via imputation and dynamic programmingICML 2020Github
Align-Refine: Non-Autoregressive Speech Recognition via Iterative RealignmentNAACL 2021Github
Learning to rewrite for non-autoregressive neural machine translationEMNLP 2021Github
RenewNAT: renewing potential translation for non-autoregressive transformerAAAI 2023-
Learning to recover from multi-modality errors for non-autoregressive neural machine translationACL 2020Github
Hybrid-regressive neural machine translationICLR 2023-
Iterative Translation Refinement with Large Language ModelsEAMT 2024-
IterGen: Iterative Semantic-aware Structured LLM Generation with BacktrackingICLR 2025Github
Rejuvenating low-frequency words: Making the most of parallel data in non-autoregressive translationACL 2021Github
Understanding and Improving Lexical Choice in Non-Autoregressive TranslationICLR 2021Github
SlotRefine: A fast non-autoregressive model for joint intent detection and slot fillingEMNLP 2020Github
Non-autoregressive dialog state trackingICLR 2020Github
Unleashing the True Potential of LLMs: A Feedback-Triggered Self-Correction with Long-Term Multipath Decodingarxiv 2025-
ProRefine: Inference-Time Prompt Refinement with Textual Feedbackarxiv 2025Github
A Probabilistic Inference Scaling Theory for LLM Self-CorrectionEMNLP 2025Github
Self-Refine: Iterative Refinement with Self-FeedbackNeurIPS 2023Github
Saber: Efficient Sampling with Adaptive Acceleration and Backtracking Enhanced Remasking for Diffusion Language Model in Code GenerationACL 2026
Edit-Based Refinement for Parallel Masked Diffusion Language ModelsarXiv 2026
An Empirical Study of Iterative Refinements for Non-autoregressive TranslationACL 2025
Towards More Efficient Insertion Transformer with Fractional Positional EncodingEACL 2023

Task-Level Parallelism

Task-level methods expose independent content units, subplans, or reasoning branches. Evaluation should distinguish end-to-end latency, aggregate computation, concurrent resources, and synchronization/merging overhead.

Content Decomposition

Query Decomposition

Decomposed Prompting and Least-to-Most Prompting are decomposition precursors; dependent subtasks remain sequential unless their dependencies permit parallel execution.

PaperVenueCode
PARALLELPROMPT: Extracting Parallelism from Large Language Model Queriesarxiv 2025-
Decomposed Prompting: A Modular Approach for Solving Complex TasksICLR 2023Github
Least-to-Most Prompting Enables Complex Reasoning in Large Language ModelsICLR 2023

Structure Planning

SoT expands skeleton points concurrently. WritingPath, Plan-and-Write, and outline-conditioned generation are planning precursors, not evidence of parallel execution.

PaperVenueCode
Navigating the Path of Writing: Outline-guided Text Generation with Large Language ModelsNAACL 2025-
Skeleton-of-thought: Prompting llms for efficient parallel generationICLR 2024Github
Plan-And-Write: Towards Better Automatic StorytellingAAAI 2019
Advancing Precise Outline-Conditioned Text Generation with Task Duality and Explicit Outline ControlEACL 2024

Process Orchestration

Plan Execution

PaperVenueCode
SPRINT: Enabling Interleaved Planning and Parallelized Execution in Reasoning Modelsarxiv 2025-
Decompose, Plan in Parallel, and Merge: A Novel Paradigm for Large Language Models based Planning with Multiple ConstraintsarXiv 2025
Optimizing Sequential Multi-Step Tasks with Parallel LLM AgentsarXiv 2025
AgentSPEX: An Agent SPecification and EXecution LanguagearXiv 2026

Branch Reasoning

PaperVenueCode
Learning to Keep a Promise: Scaling Language Model Decoding Parallelism with Learned Asynchronous Decoding (PASTA)ICML 2025
APAR: LLMs Can Do Auto-Parallel Auto-Regressive DecodingarXiv 2024
Learning Adaptive Parallel Reasoning with Language ModelsarXiv 2025
Accelerate Parallelizable Reasoning via Parallel Decoding within One SequenceEMNLP 2025
Training Large Language Models To Reason In Parallel With Global Forking TokensICLR 2026
Towards Direct Latent-Space Synthesis for Parallel Branches in LLM-Agent WorkflowsarXiv 2026

Complementary Acceleration

These techniques are discussed in the manuscript's composability appendix, not treated as additional parallel-generation paradigms. AWQ and GPTQ concern model compression, FlashAttention-2 concerns attention kernels, and LazyMAR is a visual-generation caching example rather than a text-decoding result.

PaperVenueCode
AWQ: Activation-aware Weight Quantization for LLM Compression and AccelerationMLSys 2024Github
GPTQ: Accurate Post-Training Quantization for Generative Pre-trained TransformersICLR 2023Github
LazyMAR: Accelerating Masked Autoregressive Models via Feature CachingICCV 2025Github
FlashAttention-2: Faster Attention with Better Parallelism and Work PartitioningICLR 2024Github

The entries and main scopes below match Table 1 of the revised manuscript. Scopes describe each review's emphasis and may overlap. Publication years refer to the cited versions.

ReferenceYearSurveyMain scope
Xia et al.2024Unlocking Efficiency in Large Language Model Inference: A Comprehensive Survey of Speculative DecodingSpeculative decoding: drafter selection and verification
Zhang et al.2024Beyond the Speculative Game: A Survey of Speculative Execution in Large Language ModelsSpeculative execution in LLM inference
Ryu and Kim2024Closer Look at Efficient Inference Methods: A Survey of Speculative DecodingSpeculative decoding: draft- and model-centric methods
Hu et al.2025Mitigating Sequential Dependencies: A Survey of Algorithms and Systems for Generation-Refinement Frameworks in Autoregressive ModelsGeneration–refinement frameworks: algorithms and systems
Khoshnoodi et al.2024A Comprehensive Survey of Accelerated Generation Techniques in Large Language ModelsAccelerated generation: speculation, early exiting, and NAR methods
Wan et al.2024Efficient Large Language Models: A SurveyLLM efficiency: model-, data-, and framework-centric methods
Li et al.2023Diffusion Models for Non-autoregressive Text Generation: A SurveyDiffusion models for non-autoregressive text generation
Yi et al.2024Diffusion Models in Text Generation: A SurveyDiffusion-based text generation across task settings
Yu et al.2025Discrete Diffusion in Large Language and Multimodal Models: A SurveyDiscrete diffusion language and multimodal models
Li et al.2025A Survey on Diffusion Language ModelsDiffusion language models: training, inference, and applications