DIFFUSION LANGUAGE MODELS · TRAINING-FREE SAMPLING

Explore early.
Refine later.

Better diversity starts with
when you explore.

Denoising Time Matters:
Diverse Generation in Diffusion Language Models

Accepted atThe Fortieth Annual Conference on Neural Information Processing Systems

Also accepted atICML 2026 Workshop on Structured Probabilistic Inference & Generative Modeling

TAPS uses the temporal structure of diffusion inference to open up semantic possibilities early, then reduces perturbations so the model can refine its output.

ONE CONDITION → COMPLEMENTARY POSSIBILITIESINTERACTIVE CONCEPTUAL VIEW

Jingxuan Wu^* · Zhenglin Wan^* · Yuzhe Yang · Yiqiao Huang · Chubin Zhang · Xingrui Yu^† · Ivor Tsang · Yang You

UNC Chapel Hill · National University of Singapore · UC Santa Barbara · Harvard University
Nanyang Technological University · A*STAR CFAR · A*STAR IAIC

* Equal contribution † Corresponding author · arXiv:2601.22629v3

01 / THE INSIGHT

Not just more noise.
Better timing.

A new sentence is not always a new idea.

Token-level randomness can change wording while leaving the underlying semantic direction largely unchanged. Diffusion language models expose another control axis: the denoising process itself.

The paper’s probing experiments reveal a temporal division of labor. Early interventions have a stronger influence on high-level semantics; later interventions more often alter lexical realization. TAPS places exploration where the generation is still plastic, then withdraws it before refinement.

+39.7%

Sent-BERT diversity
LLaDA · NoveltyBench

+17.7%

Sent-BERT diversity
LLaDA · WritingPrompts

+4.3pp

Pass@5
HumanEval+ · Table 3

+0.2%

Measured inference time
Matched cost experiment

Relative diversity gains vs. the corresponding base model (Tables 1–2); Pass@5 change is in percentage points (Table 3). Runtime overhead is the paper’s reported rounded value for its paired 30-trial setup (Table 9). Results are specific to these settings.

THE BIGGER PICTURE / WHY DIVERSITY MATTERS

More samples.
More possibilities?

Scaling the number of answers is useful only when it expands what the model explores.

For a large language model, diversity is more than creative wording. In mathematics, it can mean a different sequence of operations leading to a correct result. In code, it can mean a different valid program structure. Redundant samples repeatedly spend compute on the same region of the solution space.

This becomes especially relevant in recursive self-improvement (RSI) and self-training: a system that generates its own candidates needs alternatives worth evaluating, selecting and learning from. If candidate generation keeps returning the same strategies, the search has fewer opportunities to discover useful alternatives. Open-ended self-improvement systems such as the Darwin Gödel Machine likewise explore multiple branches rather than relying on a single lineage.

01 / EXPLORE

Generate distinct candidates

Broaden the set of semantic plans, reasoning paths and program structures available for evaluation.

02 / EVALUATE

Find useful differences

Correctness checks, tests and other rewards determine which differences are useful for the task.

03 / LEARN

Use the feedback

Selection and learning can exploit informative contrasts. Repetition alone does not create new evidence.

MEASURED IN TAPS · TABLE 4

Different correct solutions—not just different words.

Program structure is compared after AST normalization, alongside a coarser algorithmic-skeleton measure.

Public arXiv v3, Table 4 and Appendix A.4. Shared subset: 60 HumanEval+ problems and 100 GSM8K questions on which every compared method produces at least three correct outputs out of five. Metrics use only correct outputs. This controlled subset analysis is separate from full-benchmark accuracy and does not measure RL training outcomes.

IMPLICATION / GROUP-RELATIVE LEARNING

A group needs informative reward differences.

GRPO learns from rewards relative to the rest of a group. When every rollout receives the same reward, centered rewards—and the reward-based policy-gradient contribution—vanish in the usual stabilized formulation. This is the advantage-collapse problem.

TAPS motivates a possible intervention at rollout collection: broaden the candidate paths before evaluation. If that exploration produces useful reward differences, it may help avoid uninformative groups. The paper establishes diversity among correct solutions; reducing zero-advantage groups or improving RL training remains to be tested.

ILLUSTRATION / BINARY REWARDS

Aᵢ = (rᵢ − mean(r)) / (std(r) + ε)Nonzero relative advantages

Toy groups of eight, not TAPS training results. Path diversity alone does not ensure reward diversity: distinct correct solutions all receive 1 under a binary correctness reward. KL or other loss terms may still contribute when the reward-based term is zero.

02 / EXPLORE THE TIME AXIS

Branch first.
Polish next.

Drag through a conceptual denoising trajectory.
Watch exploration give way to refinement.

EARLY / SEMANTIC EXPLORATION25% of denoising

Denoising progress 25%

01 / CONDITION

One prompt.

The conditioning signal anchors the task throughout generation.

02 / EXPLORE

Different semantic routes.

Nearby conditioning trajectories create opportunities for high-level branching.

03 / REFINE

Room to take shape.

As perturbations fade, refinement returns to the original conditioning.

Mechanism illustration, not recorded token states or a fitted experimental curve. The drawing shows normalized progress; actual windows and schedules depend on the backbone and configuration.

Timing matters even at the same noise budget.

In a controlled diagnostic, annealed TAPS retains most of the semantic diversity of early-only perturbation while achieving higher measured quality.

Public paper, Table 15 · matched accumulated perturbation budget δ₀ = 21.25. Single diagnostic setting.

Schedule	Sent-BERT ↑	Quality ↑
Annealed (TAPS)	36.79	6.44
Early only	37.60	6.20
Late only	36.31	6.24
Full trajectory	34.87	6.24

03 / THE METHOD

Explore nearby.
Stay grounded.

Time-Annealed Perturbation Sampling
A lightweight intervention in conditioning space.

Perturb the condition

Sample a nearby prompt-derived embedding during the active exploration window.

Ẽ(t) = E + σ(t)ε

Constrain the excursion

Rescale to match distribution statistics, mix with the original condition and preserve per-token norms.

rescale → mix → project

Anneal before refinement

Reduce perturbation strength and return to the original embedding outside the active window.

σ(t) → 0

Standard diffusion language model conditioning compared with time-annealed perturbation — Updated TAPS inference diagram: early semantic exploration with manifold-constrained conditioning perturbations, followed by annealing and lexical refinement. Paper §3.3–3.4 describes the full procedure.

Why perturb conditioning rather than only tokens? +

TAPS samples from neighboring conditional distributions, giving different generations distinct opportunities to form high-level semantic plans. Annealing limits interference with the subsequent refinement stage. Token-level masking is also evaluated in the paper, but embedding-level perturbation offers a more favorable quality–diversity balance in that comparison.

04 / OBSERVED SEMANTIC COVERAGE

Keep possibilities
open.

A real toy experiment on TraDo-8B.
One ambiguous prompt, multiple denoising stages.

PROMPT“The mysterious prisoner looked at the guard and suddenly”

Original t-SNE plots of standard DLM and TAPS at early, middle and final denoising stages — Sentence-BERT embeddings projected into a shared 2D space with t-SNE; convex hulls visualize coverage. Original experiment from the public project page and public paper Appendix D.3. A toy projection is illustrative evidence, not a universal estimate of mode coverage.

EARLY23.90 → 30.05

MIDDLE21.88 → 30.74

FINAL23.73 → 27.99

Diversity values reproduced from the original figure; standard DLM → TAPS.

05 / THE EVIDENCE

More than
different words.

Explore lexical diversity, semantic diversity
and generation quality together.

Read the metric definitions +

IntraDistinct ↑ Intra-set lexical variation; see the paper’s definition and aggregation.

Div-BLEU ↑ The paper’s BLEU-derived diversity score. Higher means less overlap.

Sent-BERT ↑ Semantic diversity measured in sentence-embedding space.

EAD ↑ Expectation-adjusted distinctness, accounting for sample length and diversity estimation.

Crea / Coh / WQ / Rel ↑ GPT-4o judgments of creativity, coherence, writing quality and relevance. These are judge-based scores, not human ratings.

READ THE TRADE-OFF

TAPS improves diversity broadly, but does not win every quality metric. Compare the full rows: the goal is useful semantic exploration with competitive quality, not a guarantee that every score increases.

06 / REASONING & CODE

Different paths.
Useful answers.

Examine single-sample accuracy
and the benefit of multiple attempts.

Public paper Table 3 · temperature 0.7, five samples per question. Mean accuracy includes the reported 95% confidence-interval half-width. Majority-vote and Pass@5 values follow the table. Additional-backbone results are in the paper.

+0.6 pp

Pass@5 vs. the base model

Multi-sample coverage can improve while single-sample accuracy remains competitive. These aggregate results do not imply a guarantee for individual questions.

07 / A CLOSER LOOK

Same answer.
More ways there.

A selected GSM8K case from the paper.
Inspect actual reported reasoning excerpts.

QUESTIONLloyd’s chickens produce 252 eggs per day. He sells eggs at $2 per dozen. How much money does he make in one week?

Base model6 / 10 correct in this case

SELECTED REPORTED PATH

252 eggs per day equals 252 / 12 = 21 dozens per day. Over 7 days, this is 147 dozens per week. At $2 per dozen, Lloyd makes $294 per week.

$294 ✓

TAPS10 / 10 correct in this case

SELECT A REPORTED PATH

$294 ✓

Public paper Appendix F.1, Example 1. Excerpts follow the reported samples; paths are selected to illustrate different computation orders. Badges summarize the paper’s reported group accuracy, not the ordering of individual samples. These counts describe this single case, not benchmark accuracy.

08 / BUILD ON TAPS

Make time
for exploration.

01 / PAPERRead the research ↗02 / CODERun the implementation ↗03 / REPRODUCEExplore the benchmark setup ↗

Scope, cost & limitations +

TAPS is an inference-time method for diffusion language models, not a new training objective. Perturbation strength and timing remain configuration choices. Stronger exploration can reduce quality, and the best settings may vary by backbone and task. Reported gains are empirical, not universal guarantees. Effects on downstream RL training remain to be evaluated.

The paired cost experiment reports 19.980 ± 0.010 seconds/sample for the base model and 20.012 ± 0.006 seconds/sample for TAPS, with the same 16,830 MB peak memory. Setup: 30 trials, one fixed GSM8K prompt, 256 generated tokens, 128 denoising steps (Table 9).

Cite the current preprint

@misc{wu2026denoising,
  title={Denoising Time Matters: Diverse Generation in Diffusion Language Models},
  author={Wu, Jingxuan and Wan, Zhenglin and Yang, Yuzhe and Huang, Yiqiao and Zhang, Chubin and Yu, Xingrui and Tsang, Ivor and You, Yang},
  year={2026},
  eprint={2601.22629},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2601.22629}
}

Page content follows the public arXiv v3 manuscript. Model outputs are fixed examples; this page does not run a language model.

Explore early.Refine later.

Not just more noise.Better timing.

More samples.More possibilities?