The Burstiness Engine
Closer, but still short
The instruction narrows the gap while human articles remain more dispersed.
NeurIPS 2026 · LP4FM Poster
Burstiness Was Measured Wrong,
and Prompting Cannot Aim It
Can a prompt give generated prose the rhythm of human writing? Across nine model families, the same instruction lands short, near the target, or beyond it.
A study of sentence rhythm, and the difference between changing it and controlling it.
The aim problem
One prompt. Nine landings.
Without an explicit rhythm instruction, the baseline gaps already differ.
- DeepSeek V4 Pro+0.03595% bootstrap interval +0.011 to +0.053. Control +0.035.
- Gemini 3.1 Pro+0.12895% bootstrap interval +0.102 to +0.140. Control +0.128.
- Gemma 3 27B+0.06995% bootstrap interval +0.046 to +0.099. Control +0.069.
- GLM-5.2+0.12695% bootstrap interval +0.102 to +0.150. Control +0.126.
- Grok 4.3+0.02795% bootstrap interval +0.005 to +0.044. Control +0.027.
- Grok 4.5-0.00995% bootstrap interval -0.036 to +0.003. Control -0.009.
- Kimi K2.6+0.09595% bootstrap interval +0.070 to +0.120. Control +0.095.
- Claude Opus 4.8+0.11995% bootstrap interval +0.091 to +0.141. Control +0.119.
- Qwen3.6-27B+0.09795% bootstrap interval +0.071 to +0.116. Control +0.097.
The same instruction narrows some gaps, widens one and produces overshoot.
- DeepSeek V4 Pro+0.00095% bootstrap interval -0.012 to +0.026. Control +0.035.
- Gemini 3.1 Pro+0.07195% bootstrap interval +0.050 to +0.090. Control +0.128.
- Gemma 3 27B+0.04995% bootstrap interval +0.031 to +0.067. Control +0.069.
- GLM-5.2+0.06195% bootstrap interval +0.047 to +0.082. Control +0.126.
- Grok 4.3+0.05095% bootstrap interval +0.030 to +0.068. Control +0.027.
- Grok 4.5-0.02595% bootstrap interval -0.045 to -0.006. Control -0.009.
- Kimi K2.6-0.00695% bootstrap interval -0.034 to +0.020. Control +0.095.
- Claude Opus 4.8-0.03795% bootstrap interval -0.060 to -0.013. Control +0.119.
- Qwen3.6-27B+0.02895% bootstrap interval +0.013 to +0.046. Control +0.097.
With examples, every family remains significantly flatter than its human baseline.
- DeepSeek V4 Pro+0.02995% bootstrap interval +0.009 to +0.049. Control +0.035.
- Gemini 3.1 Pro+0.16995% bootstrap interval +0.148 to +0.183. Control +0.128.
- Gemma 3 27B+0.06395% bootstrap interval +0.047 to +0.100. Control +0.069.
- GLM-5.2+0.14895% bootstrap interval +0.123 to +0.169. Control +0.126.
- Grok 4.3+0.10895% bootstrap interval +0.085 to +0.133. Control +0.027.
- Grok 4.5+0.07795% bootstrap interval +0.058 to +0.112. Control -0.009.
- Kimi K2.6+0.12395% bootstrap interval +0.112 to +0.152. Control +0.095.
- Claude Opus 4.8+0.03995% bootstrap interval +0.011 to +0.058. Control +0.119.
- Qwen3.6-27B+0.11995% bootstrap interval +0.094 to +0.138. Control +0.097.
Planning produces a different pattern of gaps. Near zero alone does not establish equivalence.
- DeepSeek V4 Pro+0.01395% bootstrap interval -0.010 to +0.040. Control +0.035.
- Gemini 3.1 Pro+0.07095% bootstrap interval +0.049 to +0.088. Control +0.128.
- Gemma 3 27B+0.06195% bootstrap interval +0.046 to +0.081. Control +0.069.
- GLM-5.2+0.08095% bootstrap interval +0.062 to +0.099. Control +0.126.
- Grok 4.3+0.06195% bootstrap interval +0.038 to +0.085. Control +0.027.
- Grok 4.5-0.00995% bootstrap interval -0.023 to +0.019. Control -0.009.
- Kimi K2.6+0.02695% bootstrap interval +0.007 to +0.054. Control +0.095.
- Claude Opus 4.8-0.00095% bootstrap interval -0.022 to +0.022. Control +0.119.
- Qwen3.6-27B+0.04595% bootstrap interval +0.023 to +0.065. Control +0.097.
Self-refinement narrows several gaps and pushes others beyond the human reference.
- DeepSeek V4 Pro-0.01895% bootstrap interval -0.032 to +0.003. Control +0.035.
- Gemini 3.1 Pro+0.03995% bootstrap interval +0.016 to +0.053. Control +0.128.
- Gemma 3 27B+0.02995% bootstrap interval +0.009 to +0.048. Control +0.069.
- GLM-5.2+0.01195% bootstrap interval -0.010 to +0.038. Control +0.126.
- Grok 4.3+0.01995% bootstrap interval -0.003 to +0.042. Control +0.027.
- Grok 4.5-0.05995% bootstrap interval -0.085 to -0.024. Control -0.009.
- Kimi K2.6-0.07095% bootstrap interval -0.091 to -0.035. Control +0.095.
- Claude Opus 4.8-0.06595% bootstrap interval -0.088 to -0.045. Control +0.119.
- Qwen3.6-27B+0.03995% bootstrap interval +0.018 to +0.063. Control +0.097.
−0.100+0.20
ΔBₙ← More dispersedFlatter →
Selected mode Control 95% CI
Zero is the human reference. ΔBₙ = human − model. Each row is a paired comparison, not a ranking.
Exact estimates and intervals
| Family | Gap | 95% bootstrap CI | Control |
|---|---|---|---|
| DeepSeek V4 Pro | +0.035 | +0.011 to +0.053 | +0.035 |
| Gemini 3.1 Pro | +0.128 | +0.102 to +0.140 | +0.128 |
| Gemma 3 27B | +0.069 | +0.046 to +0.099 | +0.069 |
| GLM-5.2 | +0.126 | +0.102 to +0.150 | +0.126 |
| Grok 4.3 | +0.027 | +0.005 to +0.044 | +0.027 |
| Grok 4.5 | -0.009 | -0.036 to +0.003 | -0.009 |
| Kimi K2.6 | +0.095 | +0.070 to +0.120 | +0.095 |
| Claude Opus 4.8 | +0.119 | +0.091 to +0.141 | +0.119 |
| Qwen3.6-27B | +0.097 | +0.071 to +0.116 | +0.097 |
| Family | Gap | 95% bootstrap CI | Control |
|---|---|---|---|
| DeepSeek V4 Pro | +0.000 | -0.012 to +0.026 | +0.035 |
| Gemini 3.1 Pro | +0.071 | +0.050 to +0.090 | +0.128 |
| Gemma 3 27B | +0.049 | +0.031 to +0.067 | +0.069 |
| GLM-5.2 | +0.061 | +0.047 to +0.082 | +0.126 |
| Grok 4.3 | +0.050 | +0.030 to +0.068 | +0.027 |
| Grok 4.5 | -0.025 | -0.045 to -0.006 | -0.009 |
| Kimi K2.6 | -0.006 | -0.034 to +0.020 | +0.095 |
| Claude Opus 4.8 | -0.037 | -0.060 to -0.013 | +0.119 |
| Qwen3.6-27B | +0.028 | +0.013 to +0.046 | +0.097 |
| Family | Gap | 95% bootstrap CI | Control |
|---|---|---|---|
| DeepSeek V4 Pro | +0.029 | +0.009 to +0.049 | +0.035 |
| Gemini 3.1 Pro | +0.169 | +0.148 to +0.183 | +0.128 |
| Gemma 3 27B | +0.063 | +0.047 to +0.100 | +0.069 |
| GLM-5.2 | +0.148 | +0.123 to +0.169 | +0.126 |
| Grok 4.3 | +0.108 | +0.085 to +0.133 | +0.027 |
| Grok 4.5 | +0.077 | +0.058 to +0.112 | -0.009 |
| Kimi K2.6 | +0.123 | +0.112 to +0.152 | +0.095 |
| Claude Opus 4.8 | +0.039 | +0.011 to +0.058 | +0.119 |
| Qwen3.6-27B | +0.119 | +0.094 to +0.138 | +0.097 |
| Family | Gap | 95% bootstrap CI | Control |
|---|---|---|---|
| DeepSeek V4 Pro | +0.013 | -0.010 to +0.040 | +0.035 |
| Gemini 3.1 Pro | +0.070 | +0.049 to +0.088 | +0.128 |
| Gemma 3 27B | +0.061 | +0.046 to +0.081 | +0.069 |
| GLM-5.2 | +0.080 | +0.062 to +0.099 | +0.126 |
| Grok 4.3 | +0.061 | +0.038 to +0.085 | +0.027 |
| Grok 4.5 | -0.009 | -0.023 to +0.019 | -0.009 |
| Kimi K2.6 | +0.026 | +0.007 to +0.054 | +0.095 |
| Claude Opus 4.8 | -0.000 | -0.022 to +0.022 | +0.119 |
| Qwen3.6-27B | +0.045 | +0.023 to +0.065 | +0.097 |
| Family | Gap | 95% bootstrap CI | Control |
|---|---|---|---|
| DeepSeek V4 Pro | -0.018 | -0.032 to +0.003 | +0.035 |
| Gemini 3.1 Pro | +0.039 | +0.016 to +0.053 | +0.128 |
| Gemma 3 27B | +0.029 | +0.009 to +0.048 | +0.069 |
| GLM-5.2 | +0.011 | -0.010 to +0.038 | +0.126 |
| Grok 4.3 | +0.019 | -0.003 to +0.042 | +0.027 |
| Grok 4.5 | -0.059 | -0.085 to -0.024 | -0.009 |
| Kimi K2.6 | -0.070 | -0.091 to -0.035 | +0.095 |
| Claude Opus 4.8 | -0.065 | -0.088 to -0.045 | +0.119 |
| Qwen3.6-27B | +0.039 | +0.018 to +0.063 | +0.097 |
- 9
- model families
- 200
- matched headlines
- 5
- prompting modes
- 3
- seeds per prompt
The question
Language has a rhythm. Can we measure what prompts change?
A short sentence. A longer one that gives an idea room to unfold. Variation in sentence length is one part of what we mean by the rhythm of prose.
The challenge is to turn that intuition into a measurable target. Longer sentences are not necessarily more bursty. More variation is not automatically better writing. Before testing a prompt, we need to know exactly what its score describes.
01 / Measurement
First, remove the rhythm that formatting puts there.
Echoed headlines and markdown can change sentence-length statistics before a model has changed the prose at all.
On the audited Qwen leg, an instruction initially appeared to close the gap. After normalizing the output, a significant deficit emerged. The evaluation changed because the measurement became cleaner.
Inside the normalization auditQwen · Instruction mode
Raw output+0.005p = .10
Normalized+0.028p = .00003
Median Bₙ gap: human − model. Raw non-significance did not establish equivalence.
02 / The benchmark
Getting closer is not the same as hitting the target.
We gave nine model families the same 200 news headlines and five ways to approach the writing. A consistent recipe did not produce a consistent landing.
Within a bounded margin
DeepSeek and Kimi meet specified equivalence bounds on Bₙ.
Further away
Grok 4.3 becomes flatter relative to its human baseline.
Beyond the reference
Opus and Grok 4.5 are more dispersed under instruction.
These describe paired estimates against a shared human baseline. Provider decoding differs; between-mode changes are not directly tested contrasts.
Explore all five prompting modes03 / The surprise
Examples do not guarantee a better target.
Few-shot prompting offers a natural intuition: show the model the rhythm you want. Yet in this benchmark, few-shot outputs are significantly flatter than the human articles in all nine families.
They are also the flattest mode by rank-biserial effect size in seven. A technique that helps with many language tasks can move this particular property away from its target.
Inspect the published comparisons7/9
families have their flattest mode under few-shot prompting
By matched rank-biserial effect size.What it means
A target needs more than a persuasive prompt.
The contribution is a way to measure sentence-length dispersion and a benchmark showing how unevenly prompts move it. Reaching the human range is possible. Reaching it reliably is a different problem.
The evidence comes from English news and an offline, descriptive evaluation. Bₙ measures one property of prose, not writing quality or whether readers think a text sounds human.