README
Awesome-Parallel-Text-Generation
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
Self-Drafting via Partial Computation
BranchyNet is an early-exit precursor, not itself an LLM self-speculative decoding system.
Self-Drafting via Parallel Prediction
Multi-token prediction is part of drafter design, rather than a separate category alongside draft-and-verify.
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.
Verification Strategies
Efficient Verification
Tree/graph verification and scheduling methods reduce verification or end-to-end draft-and-verify latency.
Accurate Verification
These methods improve acceptance or verification decisions. Relaxed or learned acceptance rules do not automatically preserve the target sampling distribution.
Non-AR-Based
One-Shot Generation
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.
Related Foundations
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.
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.
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.
Structure Planning
SoT expands skeleton points concurrently. WritingPath, Plan-and-Write, and outline-conditioned generation are planning precursors, not evidence of parallel execution.
Process Orchestration
Plan Execution
Branch Reasoning
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.
Related Surveys
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.
Contributors
2 commits
1 commits
Similar repos
README
Awesome-Parallel-Text-Generation
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
Self-Drafting via Partial Computation
BranchyNet is an early-exit precursor, not itself an LLM self-speculative decoding system.
Self-Drafting via Parallel Prediction
Multi-token prediction is part of drafter design, rather than a separate category alongside draft-and-verify.
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.
Verification Strategies
Efficient Verification
Tree/graph verification and scheduling methods reduce verification or end-to-end draft-and-verify latency.
Accurate Verification
These methods improve acceptance or verification decisions. Relaxed or learned acceptance rules do not automatically preserve the target sampling distribution.
Non-AR-Based
One-Shot Generation
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.
Related Foundations
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.
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.
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.
Structure Planning
SoT expands skeleton points concurrently. WritingPath, Plan-and-Write, and outline-conditioned generation are planning precursors, not evidence of parallel execution.
Process Orchestration
Plan Execution
Branch Reasoning
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.
Related Surveys
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.
Similar repos
Contributors
2 commits
1 commits