The Burstiness Engine

Coverage by axis

Q1 — Prompt-engineering ceiling

• peer-reviewed Echoes in AI: Quantifying Lack of Plot Diversity in LLM Outputs
  — · PMC · 2025
  Quantifies diversity collapse in LLM outputs — supports cross-generation flattening / 'average prosody' in text.
• preprint bridge Benchmark of Stylistic Variation in LLM-Generated Texts
  — · arXiv · 2025
  COMPETING WORK for Q1 — benchmarks stylistic variation across 16 frontier models + prompt settings. Does NOT frame as control ceiling vs distributional target — wedge survives but must differentiate.
• preprint Evaluating the Diversity and Quality of LLM Generated Content
  — · arXiv · 2025
  Effective semantic diversity = diversity among quality-passing outputs. Method gift: gate burstiness measurement on quality so high-variance incoherence isn't rewarded.
• preprint bridge LLMs Still Struggle to Imitate the Implicit Writing Styles of Everyday People
  — · arXiv · 2025
  Prompt-only personalized-style imitation FAILS — dual support for Q1 (prompt ceiling) and Q6 (fingerprint hard by instruction). Strongest single motivation for model-level approach.

Q2 — Learnable rhythm embedding

• peer-reviewed bridge Towards Controllable Speech Synthesis in the Era of LLMs
  Lee et al. · EMNLP 2025 main · 2025
  EMNLP 2025 unified TTS+LLM control framework. Cites DiTTo-TTS for speech rate via latent length.
• peer-reviewed bridge Style Vectors for Steering Generative Large Language Models
  Konen, Jentzsch, Diallo et al. · EACL 2024 Findings · 2024
  Foundational steering template for a 'burstiness vector'
• peer-reviewed bridge Controllable neural TTS using intuitive prosodic features
  Raitio, Rasipuram, Castellani · Interspeech 2020 · 2020
  The Raitio paper Vittoria named
• preprint From Weights to Activations: Is Steering the Next Frontier of LLMs?
  Turner et al. · arXiv · 2026
  2025/2026 survey-style — situates activation steering as the next frontier
• preprint LLMBraces: Straightening Out LLM Predictions
  — · arXiv · 2025
  Task-specific control via relevant sub-updates
• preprint Low-Rank Adaptation for Foundation Models — Survey
  — · arXiv · 2025
  Comprehensive LoRA survey

Q3 — Burstiness across model generations

• peer-reviewed Trusting AI to detect AI?
  — · Computers in Human Behavior · 2026
  Detection rates drop with newer LLM generations
• peer-reviewed Echoes in AI: Quantifying Lack of Plot Diversity in LLM Outputs
  — · PMC · 2025
  Quantifies diversity collapse in LLM outputs — supports cross-generation flattening / 'average prosody' in text.
• peer-reviewed How is ChatGPT's Behavior Changing Over Time?
  Chen, Zaharia, Zou · Harvard Data Science Review · 2024
  Empirical longitudinal drift of LLM behavior
• peer-reviewed Distinguishing ChatGPT-3.5 vs -4 vs human Japanese texts
  — · PMC · 2023
  Direct stylometric comparison across GPT versions
• preprint bridge Can You Detect the Difference? Stylometric Comparison of Diffusion vs Autoregressive Text
  Tarım & Onan · arXiv · 2025
  COMPETING WORK — first systematic stylometric burstiness comparison. Cite, don't ignore.
• preprint RedNote-Vibe: Temporal Dynamics of AI-Generated Text
  — · arXiv · 2025
  Dataset for temporal-dynamics detection
• preprint bridge DivEye: Diversity Boosts AI-Generated Text Detection
  — · arXiv · 2025
  CUTTING-EDGE — surprisal-based features capturing how unpredictability FLUCTUATES across text. Most direct burstiness-adjacent 2025 work. Vittoria pole-position read.
• preprint bridge Benchmark of Stylistic Variation in LLM-Generated Texts
  — · arXiv · 2025
  COMPETING WORK for Q1 — benchmarks stylistic variation across 16 frontier models + prompt settings. Does NOT frame as control ceiling vs distributional target — wedge survives but must differentiate.
• preprint Evaluating the Diversity and Quality of LLM Generated Content
  — · arXiv · 2025
  Effective semantic diversity = diversity among quality-passing outputs. Method gift: gate burstiness measurement on quality so high-variance incoherence isn't rewarded.

Q4 — Burstiness LoRA / auxiliary steering

• peer-reviewed bridge Style Vectors for Steering Generative Large Language Models
  Konen, Jentzsch, Diallo et al. · EACL 2024 Findings · 2024
  Foundational steering template for a 'burstiness vector'
• peer-reviewed bridge FUDGE: Controlled Text Generation With Future Discriminators
  Yang, Klein · NAACL 2021 · 2021
  Foundation for boundary-FUDGE: future-attribute discriminator-guided decoding. We adapt it by using the computable burstiness metric as the future predictor applied only to the sentence-boundary decision.
• peer-reviewed bridge Mirostat: A Neural Text Decoding Algorithm that Directly Controls Perplexity
  Basu et al. · ICLR 2021 · 2021
  MUST-DIFFERENTIATE (top scoop risk): controls a distributional output statistic (perplexity) in-generation via closed-form feedback. We differ: structural sequence-VARIANCE not token-mean; boundary-only not whole-vocab; future-discriminator not running set-point. Frame ours as the structural-variance, boundary-altitude generalization.
• preprint CARD: Cluster-level Adaptation with Reward-guided Decoding
  — · arXiv · 2026
  Logit corrections on frozen LLM
• preprint Styles + Persona-plug = Customized LLMs
  — · arXiv · 2026
  Plug-in persona over frozen LLM
• preprint GLASS: GRPO-Trained LoRA for Acoustic Style Steering in Zero-Shot TTS
  — · arXiv · 2026
  GRPO-trained LoRA for a continuous style attribute - method analogue for the GRPO burstiness controller arm.
• preprint Continuous Control of Editing Models via Adaptive-Origin Guidance
  — · arXiv · 2026
  Continuous control via guidance - relevant to the smooth burstiness dial.
• preprint A Unified Study of LoRA Variants: Taxonomy, Review, Codebase
  — · arXiv · 2026
  LoRA-variant taxonomy - reference for the controller adapter choice.
• preprint From Weights to Activations: Is Steering the Next Frontier of LLMs?
  Turner et al. · arXiv · 2026
  2025/2026 survey-style — situates activation steering as the next frontier
• preprint bridge TADA! Tuning Audio Diffusion Models through Activation Steering
  — · arXiv · 2026
  Activation steering applied to audio diffusion
• preprint Self-Supervised Honesty Steering via Anti-Parallel Representations
  — · arXiv · 2026
  Self-supervised steering method
• preprint bridge Plug-and-Play LLM Fingerprinting via Text-to-Weight Generation
  — · arXiv · 2026
  GENERATES LoRA params as variable-length sequences — direct precedent for personalized burstiness-LoRA generation
• preprint bridge AgentSteerTTS: Multi-Agent Closed-Loop TTS Steering
  — · arXiv · 2026
  Multi-agent closed-loop steering — adversarial disentanglement
• preprint bridge Personalized Text Generation with Contrastive Activation Steering
  Liu et al. · arXiv · 2025
  User-specific style vectors via contrastive activation — closest analogue to a personal burstiness fingerprint
• preprint Merge and Guide: Unifying Model Merging and Guided Decoding for Controllable Multi-Objective Generation
  — · arXiv · 2025
  Closed-form logit-combination decoding for semantic/reward objectives, not structural statistics: shows the closed-form-decoding family does not cover structural-variance.
• preprint Low-Rank Adaptation for Foundation Models — Survey
  — · arXiv · 2025
  Comprehensive LoRA survey
• preprint Continuous Language Model Interpolation for Dynamic Control
  — · arXiv · 2024
  Weight-space interpolation for fine-grained control
• preprint bridge StyleRemix: Authorship Obfuscation via Distillation
  Fisher et al. · arXiv · 2024
  Reverse of burstiness-injection — same machinery
• preprint Large Language Models can be Guided to Evade AI-Generated Text Detection (SICO)
  Lu et al. · TMLR 2024 · 2024
  Forces sentence-length variance for detector evasion via prompt/in-context substitution: black-box, post-hoc, no logit steering. Cite-and-dismiss as orthogonal humanizer mechanism.

Q5 — Variance vs inter-token timing

• peer-reviewed AI-Generated Text Detection: A Comprehensive Review of Active Methods
  — · ScienceDirect · 2026
  2025/26 detection review — paraphrasing vulnerability
• preprint The Statistical Signature of LLMs
  — · arXiv · 2026
  How probabilistic sampling reshapes statistical structure
• preprint bridge Can You Detect the Difference? Stylometric Comparison of Diffusion vs Autoregressive Text
  Tarım & Onan · arXiv · 2025
  COMPETING WORK — first systematic stylometric burstiness comparison. Cite, don't ignore.
• preprint Beyond Checkmate: Creative Choke Points in AI Text
  — · arXiv · 2025
  Segment-level burstiness analysis — where the rhythm signal lives
• preprint Detecting LLM-Generated Short Answers
  — · arXiv · 2025
  Documents GPTZero operationalization
• preprint A Training-free Method for LLM Text Attribution
  — · arXiv · 2025
  Uses GPTZero burstiness definition upstream
• preprint bridge DivEye: Diversity Boosts AI-Generated Text Detection
  — · arXiv · 2025
  CUTTING-EDGE — surprisal-based features capturing how unpredictability FLUCTUATES across text. Most direct burstiness-adjacent 2025 work. Vittoria pole-position read.
• preprint Stylometry Recognizes Human and LLM-Generated Texts
  — · arXiv · 2025
  Stylometric human-vs-LLM discrimination — burstiness sits inside a working stylometric feature set.
• preprint Detecting AI-Generated Text: Factors Influencing Detectability
  — · arXiv · 2024
  Survey treating perplexity-burstiness as standard

Q6 — Fingerprint vs idiosyncratic rhythm

• peer-reviewed bridge Prosodic cues strengthen human-AI voice boundaries
  Bakkouche et al. · ScienceDirect · 2025
  The named Bakkouche perception study — listeners discriminate human vs AI by prosody
• peer-reviewed bridge Style Vectors for Steering Generative Large Language Models
  Konen, Jentzsch, Diallo et al. · EACL 2024 Findings · 2024
  Foundational steering template for a 'burstiness vector'
• preprint bridge Plug-and-Play LLM Fingerprinting via Text-to-Weight Generation
  — · arXiv · 2026
  GENERATES LoRA params as variable-length sequences — direct precedent for personalized burstiness-LoRA generation
• preprint bridge Personalized Text Generation with Contrastive Activation Steering
  Liu et al. · arXiv · 2025
  User-specific style vectors via contrastive activation — closest analogue to a personal burstiness fingerprint
• preprint bridge LLMs Still Struggle to Imitate the Implicit Writing Styles of Everyday People
  — · arXiv · 2025
  Prompt-only personalized-style imitation FAILS — dual support for Q1 (prompt ceiling) and Q6 (fingerprint hard by instruction). Strongest single motivation for model-level approach.
• preprint Stylometry Recognizes Human and LLM-Generated Texts
  — · arXiv · 2025
  Stylometric human-vs-LLM discrimination — burstiness sits inside a working stylometric feature set.
• preprint bridge StyleRemix: Authorship Obfuscation via Distillation
  Fisher et al. · arXiv · 2024
  Reverse of burstiness-injection — same machinery
• preprint Large Language Models can be Guided to Evade AI-Generated Text Detection (SICO)
  Lu et al. · TMLR 2024 · 2024
  Forces sentence-length variance for detector evasion via prompt/in-context substitution: black-box, post-hoc, no logit steering. Cite-and-dismiss as orthogonal humanizer mechanism.

Q7 — Punctuation / paragraph proxies

• preprint A Statistical Journey into the Poetic World of Evgenij Onegin
  — · arXiv · 2026
  Symbolic time-series + Markov modeling of phonological rhythm. Method transfer for Q7: model punctuation/segmentation as a symbolic time series → timing-free burstiness proxy.

TTS — Raitio (Apple)

• peer-reviewed bridge Towards Controllable Speech Synthesis in the Era of LLMs
  Lee et al. · EMNLP 2025 main · 2025
  EMNLP 2025 unified TTS+LLM control framework. Cites DiTTo-TTS for speech rate via latent length.
• peer-reviewed Style and Prosody control for Zero-shot Speech Synthesis
  Suni, Antti et al. · SSW 2025 · 2025
  Derived-vector approach to predictably modify synthesized speech
• peer-reviewed Enhanced Prosody Modeling and Character Voice Controlling for Audiobooks
  — · ACM · 2025
  Prosodic expressiveness for audiobook synthesis
• peer-reviewed Emphasis control for parallel neural TTS / Hierarchical Prosody Modeling
  Raitio et al. · arXiv / Interspeech · 2022
  Hierarchical latent space for prosodic emphasis
• peer-reviewed bridge Controllable neural TTS using intuitive prosodic features
  Raitio, Rasipuram, Castellani · Interspeech 2020 · 2020
  The Raitio paper Vittoria named
• preprint GLASS: GRPO-Trained LoRA for Acoustic Style Steering in Zero-Shot TTS
  — · arXiv · 2026
  GRPO-trained LoRA for a continuous style attribute - method analogue for the GRPO burstiness controller arm.
• preprint bridge TADA! Tuning Audio Diffusion Models through Activation Steering
  — · arXiv · 2026
  Activation steering applied to audio diffusion
• preprint MAGIC-TTS: Fine-Grained Controllable Speech Synthesis
  — · arXiv · 2026
  Token-level duration conditioning
• preprint bridge AgentSteerTTS: Multi-Agent Closed-Loop TTS Steering
  — · arXiv · 2026
  Multi-agent closed-loop steering — adversarial disentanglement

TTS — CTRL-P

• peer-reviewed Ctrl-P: Temporal Control of Prosodic Variation for Speech Synthesis
  Mohan, Hu, Klimkov et al. · Interspeech 2021 · 2021
  RESOLVED: the true 'CTRL-P' Vittoria referenced, distinct from Raitio 2020. Conditions on three acoustic correlates of prosody. PHASE 2 anchor, not core.

TTS — Bakkouche perception

• peer-reviewed bridge Prosodic cues strengthen human-AI voice boundaries
  Bakkouche et al. · ScienceDirect · 2025
  The named Bakkouche perception study — listeners discriminate human vs AI by prosody

Bio — NN ↔ biological rhythm

• peer-reviewed Language experience shapes predictive coding of rhythmic sound sequences
  — · eLife · 2024
  Brain rhythm-prediction is shaped by native-language experience
• peer-reviewed Predictive Coding or Just Feature Discovery? An Alternative Account
  — · PMC · 2024
  Critical view — LM features predict brain but mechanism may not be predictive coding
• peer-reviewed bridge Long-range and hierarchical language predictions in brains and algorithms
  Caucheteux, Gramfort, King · Nature Human Behaviour · 2023
  LM hierarchy matches brain hierarchy — strongest Vaswani↔biology bridge
• preprint Toward a realistic model of speech processing in the brain with SSL
  — · arXiv · 2022
  SSL models match brain processing

Arch — Vaswani extensions

• peer-reviewed bridge Long-range and hierarchical language predictions in brains and algorithms
  Caucheteux, Gramfort, King · Nature Human Behaviour · 2023
  LM hierarchy matches brain hierarchy — strongest Vaswani↔biology bridge

Research sessions

Learnings (cross-session)

• methodology: Run real Firecrawl academic searches; ResearchGravity's delegate_research is a stub that returns hollow 100% completes.
  Why: Pass 1 chain-a360b2b3619b returned 'completed' in 1 second routed to insight_review + webhook_test — no papers retrieved.
  How to apply: Always use Firecrawl --categories research as the actual fetcher; use ResearchGravity only as the INDEX layer.
• methodology: Time-filter cutting-edge queries with --tbs qdr:y for last-12-months. Default arXiv search is undated and surfaces old work.
  Why: Pass 3 only worked because of --tbs qdr:y; without it the same queries surface 2020-2023 papers.
  How to apply: Every new-pass run should include --tbs qdr:y or qdr:m by default; explicitly opt out for historical sweeps.
• methodology: Add as URL sources to NotebookLM, not arbitrary scraped PDFs. Wrong arXiv ID → wrong paper ingested silently.
  Why: Pass 2 added arXiv:2208.07814 (jet physics) thinking it was Caucheteux — fixed by re-search.
  How to apply: VERIFY arxiv ID matches the paper title BEFORE source_add. Search-then-add, never type-then-add.
• methodology: Avoid PMC URLs in source_add — they hit reCAPTCHA and ingest the captcha page as 'source'.
  Why: PMC10038805 was added and the source title showed 'Checking your browser - reCAPTCHA'.
  How to apply: Prefer arXiv preprint over PMC final when both exist; or use SciHub/researcher-portal URLs.
• novelty-bar: The Burstiness Engine novelty is the COMPOSITE (rhythm-target + personal fingerprinting + perception validation), not any single mechanism. Each ingredient has 2024-2026 precedent.
  Why: Style Vectors (2024), StyleVector (2025), Bakkouche (2025), DivEye (2025), DiTTo-TTS (2025) collectively cover the building blocks.
  How to apply: Frame paper as 'composing established mechanisms onto a new target with TTS-derived validation', not 'inventing burstiness steering'.
• competing: Two competing papers are now live: Tarım & Onan 2025 (stylometric burstiness comparison) and DivEye Sep 2025 (surprisal-based diversity detection).
  Why: Both published 2025, both treat burstiness/diversity as central. The Burstiness Engine ships into a crowded room.
  How to apply: Cite both in the related-work section; differentiate on the in-generation CONTROL angle, not the measurement angle.
• perception: Bakkouche 2025 + 'reduced F0 variation lowers naturalness' from ResearchGate 396811851 = direct evidence that suppressed prosodic variance harms perception. Strong analogy to suppressed burstiness in text.
  Why: If TTS listeners notice suppressed variance, text readers likely do too — testable hypothesis with Bakkouche's methodology adapted.
  How to apply: Use Bakkouche's experimental design as the template for Vittoria's text-perception study.
• gap: Q1 (prompt-engineering ceiling for distributional stylistic control) is genuinely under-published. The systematic-negative-result paper does not exist yet.
  Why: 8 Firecrawl queries surfaced zero peer-reviewed papers measuring prompt-only ceiling against distributional targets like sentence-length variance.
  How to apply: Treat Q1 as a publishable contribution in itself OR as the motivation section of the main paper.
• verification: Dario Amodei has no 'thesis on neural networks and biology'. His PhD was condensed-matter physics.
  Why: Dico's notebook references this; no such work exists. Possible misattribution to Olah / Elhage / Anthropic interpretability thread. Dico's email cited a Princeton dataspace link (dsp013f462544k) — verify what it actually is.
  How to apply: Ask Dico for the original source before paper-writing; the bio-bridge work can be carried by Caucheteux + eLife predictive coding without the Amodei citation.
• scope: CANONICAL SCOPE (from Vittoria's email thread, May 21-28 2026): title = 'The Burstiness Engine: Modeling Human Rhythmic Patterns in Language Generation'. Spine = prompt-level control as baseline (show where it breaks) → model-level steering (auxiliary rhythm model / LoRA) as the answer. TTS is PHASE 2, not core. Emotion out of scope unless a dataset already exists.
  Why: Vittoria explicitly flagged the original 'Engineering Synthetic Human Rhythm' title for ethics, said the TTS prosody angle 'makes the surface too big, we'd lose depth', and chose model-level as the real contribution with prompt-level as baseline.
  How to apply: All synthesis and the paper outline follow this spine. Do not foreground TTS. The deliverable Vittoria is waiting on is shareable ResearchGravity results she can read independently before a sync call.

All papers