In-generation control · EMNLP / ACL / arXiv
Burstiness is the variation in sentence length and complexity that makes writing feel human. We make it a reproducible target and steer it during generation.
METAVENTIONS AI · Research preprint in progress
Same content, two rhythms. We make the top distribution a target and steer it during generation.
Abstract
Burstiness is the most reliable signal that text was produced by a language model, yet it is defined only by a detector heuristic and controlled only by post-hoc rewriting. We treat it instead as a property that can be specified as a distributional target and steered during generation.
We frame the work as modeling human rhythmic patterns, not as evading detection, and we report detector behavior across the full control range so the contribution serves measurement and model analysis as well as generation. The novelty is the composition: applying established steering machinery to a rhythm-specific target, with a reusable definition and perceptual validation.
The wedge
Demand for human rhythm in AI text is everywhere: detectors, humanizer tools, prompt tricks. The literature splits cleanly, and nobody has crossed the gap.
A strong, recent line of work quantifies burstiness after the fact to catch machine text. None of it controls the property.
Activation and LoRA methods steer frozen models toward style targets: sentiment, persona, authorship. None of them target rhythm.
A rhythm-specific distributional target, steered in-generation, validated by human perception. The composite is the novelty, not any single mechanism.
A reproducible definition
The canonical operational definition of burstiness in detection practice is a blog post. The field runs on an un-formalized metric. We replace it with a decomposable, reproducible target.
Variance and kurtosis of sentence length, surprisal fluctuation under a fixed reference model, and punctuation entropy. Our ablation shows the field's implicit metric, standard deviation of surprisal, is the weakest discriminator. Local jumpiness and mean surprisal separate human from machine far better.
Four contributions
Burstiness as a decomposable distributional target, replacing the single-scalar detector heuristic.
A clean measurement of how prompt-only control degrades against the target over output length. A standalone negative result.
Biasing only the sentence-boundary decision yields a monotonic, coherence-preserving length dial. On distilgpt2: Spearman ρ 1.0, Cohen's d 0.93.
A pre-registered human study, adapted from a method validated in speech synthesis, linking controlled rhythm to perceived humanness.
The lab
The Burstiness Engine is built inside METAVENTIONS AI as a self-rebuilding research system: a curated literature corpus, an experiment harness, and a paper that regenerates from a single source of truth. Two people, one rhythm.
Co-author
Owns the research spine and the perception protocol: prompt-level control is the baseline, model-level steering is the answer.
Burstiness Engine
The research questions
| # | Question | Status |
|---|---|---|
| Q1 | The prompt-engineering ceiling for rhythmic control | Gap, publishable |
| Q2 | Burstiness as a learnable rhythm embedding, no base retrain | Mechanism exists |
| Q3 | Is AI text getting more bursty across generations? | Empirically yes |
| Q4 | Small auxiliary rhythm model, a burstiness LoRA | Controller in progress |
| Q5 | Sentence variance vs inter-token timing | Variance formalized |
| Q6 | Burstiness fingerprint vs idiosyncratic style | Stylometry strong |
| Q7 | Punctuation and paragraph structure as proxies | Genuine gap |
A deeper walk through the four research passes and the methodology lives on the research page.
The framing, definition, ablation, and negative result are done. A working controller with quantitative wins is the open frontier, with GRPO runs in progress. Explore the live corpus dashboard or the results and figures.