mlsanityData to 14 Oct 2026

Time-Annealed Perturbation Sampling: Diverse Generation for Diffusion Language Models

arXiv 2601.22629 ↗First posted 30 Jan 2026cs.CL cs.AIcited by 3

On the map: LLM reasoning → Diffusion reasoning and test-time scaling

also in 1 more
Abstract (arXiv)

Diffusion language models (Diffusion-LMs) generate text through iterative denoising, exposing a temporal structure that is largely absent from autoregressive decoding. In this paper, we show that this temporal structure provides a useful control axis for generation diversity: early denoising steps mainly determine high-level semantic trajectories, while later steps refine lexical realization. Motivated by this observation, we propose Time-Annealed Perturbation Sampling (TAPS), a training-free inference strategy that samples nearby conditioning trajectories through time-aware, manifold-constrained perturbations. TAPS encourages semantic branching during early denoising and anneals the perturbation away before refinement, improving exploration while preserving prompt alignment, generation quality, and reasoning ability. Experiments on multiple Diffusion-LM backbones, including non-autoregressive and semi-autoregressive models, show that TAPS consistently improves semantic and lexical diversity across open-ended and instruction-following generation tasks, while preserving reasoning ability on verifiable reasoning benchmarks with negligible overhead.

Key facts. Cited by 3 papers in the map (data through 14 Oct 2026). Placed in Diffusion reasoning and test-time scaling, one of 17 directions in LLM reasoning. Builds on 53 of its 61 references that are in the map. First posted on arXiv 30 Jan 2026.

Builds on

53 of 61 references.

References · 53

  1. 2021Training Verifiers to Solve Math Word Problemscited by 10,109
  2. 2021Measuring Mathematical Problem Solving With the MATH Datasetcited by 5,371
  3. 2024GPT-4o System Cardcited by 4,588
  4. 2019The Curious Case of Neural Text Degenerationcited by 2,916
  5. 2023GPQA: A Graduate-Level Google-Proof Q&A Benchmarkcited by 2,879
  6. 2020Learning to summarize from human feedbackcited by 2,588
  7. 2023Is Your Code Generated by ChatGPT Really Correct? Rigorous Evaluation of Large Language Models for Code Generationcited by 1,391
  8. 2018Hierarchical Neural Story Generationcited by 1,193
  9. 2024OlympiadBench: A Challenging Benchmark for Promoting AGI with Olympiad-Level Bilingual Multimodal Scientific Problemscited by 1,136
  10. 2015A Diversity-Promoting Objective Function for Neural Conversation Modelscited by 1,077
  11. 2022Diffusion-LM Improves Controllable Text Generationcited by 1,027
  12. 2025Large Language Diffusion Modelscited by 912
38 more
  1. 2025AlphaEvolve: A coding agent for scientific and algorithmic discoverycited by 867
  2. 2022Contrastive Decoding: Open-ended Text Generation as Optimizationcited by 609
  3. 2025Beyond the 80/20 Rule: High-Entropy Minority Tokens Drive Effective Reinforcement Learning for LLM Reasoningcited by 559
  4. 2025Dream 7B: Diffusion Large Language Modelscited by 544
  5. 2019Neural Text Generation with Unlikelihood Trainingcited by 503
  6. 2024From Crowdsourced Data to High-Quality Benchmarks: Arena-Hard and BenchBuilder Pipelinecited by 500
  7. 2018Texygen: A Benchmarking Platform for Text Generation Modelscited by 491
  8. 2025Fast-dLLM: Training-free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decodingcited by 420
  9. 2023Understanding the Effects of RLHF on LLM Generalisation and Diversitycited by 396
  10. 2024Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researcherscited by 363
  11. 2024Skywork-Reward: Bag of Tricks for Reward Modeling in LLMscited by 327
  12. 2025Reasoning with Exploration: An Entropy Perspectivecited by 281
  13. 2017Beam Search Strategies for Neural Machine Translationcited by 232
  14. 2023Protein Design with Guided Discrete Diffusioncited by 157
  15. 2024Turning Up the Heat: Min-p Sampling for Creative and Coherent LLM Outputscited by 150
  16. 2023CADS: Unleashing the Diversity of Diffusion Models through Condition-Annealed Samplingcited by 129
  17. 2025SDAR: A Synergistic Diffusion-AutoRegression Paradigm for Scalable Sequence Generationcited by 129
  18. 2022Truncation Sampling as Language Model Desmoothingcited by 123
  19. 2024Is Temperature the Creativity Parameter of Large Language Models?cited by 93
  20. 2025SEED-GRPO: Semantic Entropy Enhanced GRPO for Uncertainty-Aware Policy Optimizationcited by 89
  21. 2025Revolutionizing Reinforcement Learning Framework for Diffusion Large Language Modelscited by 88
  22. 2025Jointly Reinforcing Diversity and Quality in Language Model Generationscited by 80
  23. 2025A Survey on Diffusion Language Modelscited by 78
  24. 2023Particle Guidance: non-I.I.D. Diverse Sampling with Diffusion Modelscited by 71
  25. 2024Preserving Diversity in Supervised Fine-Tuning of Large Language Modelscited by 70
  26. 2025NoveltyBench: Evaluating Language Models for Humanlike Diversitycited by 66
  27. 2025Diverse Preference Optimizationcited by 52
  28. 2024Forcing Diffuse Distributions out of Language Modelscited by 48
  29. 2025Modifying Large Language Model Post-Training for Diverse Creative Writingcited by 40
  30. 2023Hot or Cold? Adaptive Temperature Sampling for Code Generation with Large Language Modelscited by 38
  31. 2025TESS 2: A Large-Scale Generalist Diffusion Language Modelcited by 30
  32. 2025Roll the dice & look before you leap: Going beyond the creative limits of next-token predictioncited by 29
  33. 2022Rethinking and Refining the Distinct Metriccited by 27
  34. 2024Exploration-Driven Policy Optimization in RLHF: Theoretical Insights on Efficient Data Utilizationcited by 25
  35. 2025Creative Preference Optimizationcited by 25
  36. 2025Diverse Preference Learning for Capabilities and Alignmentcited by 25
  37. 2025G2: Guided Generation for Enhanced Output Diversity in LLMscited by 14
  38. 2023KL-Divergence Guided Temperature Samplingcited by 9

50 most cited of 53 All 53 (JSON)

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 3.

3 papers.

  1. 2026Free Lunch for Pass@k? Low Cost Diverse Sampling for Diffusion Language Modelscited by 5
  2. 2026A Tale of Two Temperatures: Simple, Efficient, and Diverse Sampling from Diffusion Language Modelscited by 5
  3. 2026Focus on the Core: Empowering Diffusion Large Language Models by Self-Contrastcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2025Breaking AR's Sampling Bottleneck: Provable Acceleration via Diffusion Language Models
  2. 2026Simple Diffusion Language Models Are More Effective Few-Step Generators Than Reported
  3. 2024Scaling Diffusion Language Models via Adaptation from Autoregressive Models
  4. 2026Diffusion Language Models: An Experimental Analysis
  5. 2026Just on Time: Token-Level Early Stopping for Diffusion Language Models
  6. 2025Decoder-Only LLMs are Better Controllers for Diffusion Models
  7. 2023Can Diffusion Model Achieve Better Performance in Text Generation? Bridging the Gap between Training and Inference!
  8. 2026Temporally-Resolved Token Attribution Reveals the Generation Dynamics of Diffusion Language Models