The Burstiness Engine

Live corpus dashboard

Coverage, sessions, and the full corpus.

Engineering human rhythm in synthetic text · Vittoria Lanzo and Dico Angelo

64 papers 29 peer-reviewed 23 bridge papers 9 research sessions Regenerated 2026-07-12 03:30 UTC

Coverage

Coverage by axis

AxisPapersBridgeStatus
Q1 · Prompt-engineering ceiling129Gap Publishable on its own
Q2 · Learnable rhythm embedding63Strong Mechanism established for other style axes
Q3 · Burstiness across model generations93Strong Empirically confirmed in 2024-2026
Q4 · Burstiness LoRA / auxiliary steering198Strong Direct analogue: StyleVector contrastive
Q5 · Variance vs inter-token timing92Partial Variance OK, timing TTS-only
Q6 · Fingerprint vs idiosyncratic rhythm86Strong Stylometry + Bakkouche perception bridge
Q7 · Punctuation / paragraph proxies10Partial Targeted query needed
TTS · Raitio (Apple)94Confirmed Raitio + Suni + DiTTo-TTS
TTS · CTRL-P10Verify Likely = Raitio 2020 itself
TTS · Bakkouche perception11Located Bakkouche 2025
Bio · NN ↔ biological rhythm41Strong Caucheteux + eLife predictive coding
Arch · Vaswani extensions11Implicit Covered via Q2 / Q4 LM descendants

Q1 · Prompt-engineering ceiling

  • peer-reviewed Echoes in AI: Quantifying Lack of Plot Diversity in LLM Outputs
    Weijia Xu and Nebojsa Jojic and Sudha Rao and Chris Brockett · PNAS · 2025
    Quantifies diversity collapse in LLM outputs — supports cross-generation flattening / 'average prosody' in text.
  • peer-reviewed Large Language Models Cannot Self-Correct Reasoning Yet
    Jie Huang and Xinyun Chen and Swaroop Mishra and Huaixiu Steven Zheng and Adams Wei Yu and Xinying Song and Denny Zhou · ICLR 2024 · 2024
    Caveat that intrinsic self-correction fails for reasoning; motivates treating M4 style/rhythm self-refine as the friendlier case and the self-leniency guard.
  • peer-reviewed bridge Lost in the Middle: How Language Models Use Long Contexts
    Nelson F. Liu and Kevin Lin and John Hewitt and Ashwin Paranjape and Michele Bevilacqua and Fabio Petroni and Percy Liang · TACL 2024 · 2024
    U-shaped primacy/recency; motivates the M1 end re-anchor and the M2 exemplar ordering (strongest first and last).
  • peer-reviewed bridge Plan-and-Solve Prompting: Improving Zero-Shot Chain-of-Thought Reasoning by Large Language Models
    Lei Wang and Wanyu Xu and Yihuai Lan and Zhiqiang Hu and Yunshi Lan and Roy Ka-Wei Lee and Ee-Peng Lim · ACL 2023 · 2023
    Plan-then-write / decomposition method behind the M3 condition; no news-specific plan template exists, so M3 adapts it to journalism's inverted-pyramid (5 W's) schema.
  • peer-reviewed bridge Self-Refine: Iterative Refinement with Self-Feedback
    Aman Madaan and Niket Tandon and Prakhar Gupta and others · NeurIPS 2023 · 2023
    The init/feedback/refine structure of the M4 condition; adopted as structure with qualitative (non-numeric) news feedback, deliberately dropping the paper's numeric rubric.
  • peer-reviewed bridge Rethinking the Role of Demonstrations: What Makes In-Context Learning Work?
    Sewon Min and Xinxi Lyu and Ari Holtzman and Mikel Artetxe and Mike Lewis and Hannaneh Hajishirzi and Luke Zettlemoyer · EMNLP 2022 · 2022
    Demonstration format/distribution drives ICL as much as content; motivates the clean XML-delimited, register-matched M2 exemplars.
  • peer-reviewed bridge Language Models are Few-Shot Learners
    Tom B. Brown and Benjamin Mann and Nick Ryder and others · NeurIPS 2020 · 2020
    Few-shot / in-context learning foundation for the M2 (show) condition.
  • preprint bridge Benchmark of Stylistic Variation in LLM-Generated Texts
    Jiří Milička and Anna Marklová and Václav Cvrček · 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
    Alexander Shypula and Shuo Li and Botong Zhang and Vishakh Padmakumar and Kayo Yin and Osbert Bastani · 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
    Zhengxiang Wang and Nafis Irtiza Tripto and Solha Park and Zhenzhen Li and Jiawei Zhou · 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 bridge The Prompt Report: A Systematic Survey of Prompt Engineering Techniques
    Sander Schulhoff and Michael Ilie and Nishant Balepur and others · arXiv · 2024
    Taxonomy anchor: its 6 families define the non-overlapping mechanism coverage (tell/show/plan/fix/select) for the prompt-ceiling conditions; the fairness guarantee is coverage at the mechanism level, not technique names.
  • preprint bridge A Prompt Pattern Catalog to Enhance Prompt Engineering with ChatGPT
    Jules White and Quchen Fu and Sam Hays and Michael Sandborn and Carlos Olea and Henry Gilbert and Ashraf Elnashar and Jesse Spencer-Smith and Douglas C. Schmidt · arXiv · 2023
    The Persona Pattern ('Act as persona X') is the canonical, metric-free role template adopted verbatim for the M1 instruction condition.

Q2 · Learnable rhythm embedding

Q3 · Burstiness across model generations

Q4 · Burstiness LoRA / auxiliary steering

Q5 · Variance vs inter-token timing

Q6 · Fingerprint vs idiosyncratic rhythm

Q7 · Punctuation / paragraph proxies

TTS · Raitio (Apple)

TTS · CTRL-P

  • peer-reviewed Ctrl-P: Temporal Control of Prosodic Variation for Speech Synthesis
    Devang S Ram Mohan and Vivian Hu and Tian Huey Teh and Alexandra Torresquintero and Christopher G. R. Wallis and Marlene Staib and Lorenzo Foglianti and Jiameng Gao and Simon King · 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

Bio · NN ↔ biological rhythm

Arch · Vaswani extensions

Research sessions

pass-1-2-baseline

Started 2026-06-28 17:03:26 · isolation: no-personal-carryover

Initial 8-query Firecrawl sweep of academic literature mapped to Vittoria's 7 questions + TTS + Bio anchors.

Queries (8)

  • Q2,Q4 · 10 results
    controllable text generation LoRA style adapter frozen LLM stylistic steering arXiv
  • TTS-A · 10 results
    Raitio Apple controllable prosody TTS conditioning CTRL-P prosodic speech synthesis
  • Q5,Q7 · 10 results
    burstiness sentence length variance perplexity AI generated text detection GPTZero stylometry
  • Q6 · 10 results
    stylometric fingerprint authorship attribution sentence rhythm English prose Mendenhall
  • Q3 · 10 results
    AI generated text burstiness longitudinal GPT-3 GPT-4 stylistic evolution detection across model generations
  • Q1,Q7 · 10 results
    prompt engineering ceiling stylistic distributional control LLM negative results long output drift punctuation paragraph rhythm
  • TTS-C · 9 results
    Bakkouche prosody perception Interspeech 2025 TTS speech synthesis listener
  • Bio,Arch · 10 results
    neural oscillations language predictive coding speech rhythm thought chunking transformer attention prosody biology Pickering Garrod

Findings (7)

  • Q4 novelty-bar The 'burstiness LoRA' steering mechanism already exists for OTHER style axes (sentiment, persona, authorship). Novelty bar = apply to rhythm + personal fingerprinting + perception validation, NOT invent the mechanism. cite · cite · cite
  • Q5 competing-work Tarım & Onan 2025 published the first systematic stylometric burstiness comparison (diffusion vs autoregressive). Must cite and position against, not ignore. cite
  • Q5 gap GPTZero's burstiness operationalization (Tian 2023) is a blog post, not peer-reviewed. This IS the formalization gap the Burstiness Engine identifies.
  • Q3 general Empirically confirmed: AI text style measurably shifts across model generations, detection rates drop with newer models. cite · cite
  • TTS-C perception Bakkouche 2025 perception study: listeners do NOT easily perceive humans and AI clones as the same person — prosody is the discriminator. Direct precedent for Vittoria's perception hypothesis.
  • Bio,Arch methodology Caucheteux et al. (Nature Human Behaviour 2023) show brain uses long-range hierarchical predictions matching LM architecture — the strongest single bridge between Vaswani and biology. cite
  • Q1 gap Prompt-engineering ceiling for distributional stylistic control is UNDER-PUBLISHED. The negative-result paper Vittoria proposes is itself a publishable contribution.

pass-3-cutting-edge

Started 2026-06-28 17:03:26 · isolation: no-personal-carryover

Last-12-months arXiv sweep for progression vs baseline. A/B against pass 1+2.

Queries (4)

  • Q2,Q4 [qdr:y] · 10 results
    controllable text generation LoRA style adapter 2025 frozen LLM activation steering
  • Q3,Q5 [qdr:y] · 4 results
    burstiness LLM text generation 2025 2026 stylometric variance peer-reviewed EMNLP ACL NAACL
  • TTS-A [qdr:y] · 10 results
    prosody control TTS 2025 NaturalSpeech VALL-E neural speech synthesis variance adaptor latent
  • Q3,Q5 [qdr:y] · 4 results
    AI generated text detection 2025 2026 burstiness perplexity DetectGPT Binoculars latest

Findings (5)

  • Q5,Q3 competing-work DivEye (arXiv:2509.18880, Sep 2025) captures how unpredictability FLUCTUATES across text via surprisal-based features. This is the most direct 2025 burstiness-adjacent detection paper. Must be in Vittoria's prior-art section. cite
  • TTS-A,Q2 methodology EMNLP 2025 main 'Towards Controllable Speech Synthesis in the Era of LLMs' (Lee et al.) is the unified TTS+LLM control framework. DiTTo-TTS controls speech rate via latent-length prediction — direct precedent for in-generation rhythm conditioning.
  • Q4,Q6 methodology Plug-and-Play LLM Fingerprinting (arXiv:2605.18474) generates LoRA params as variable-length sequences. This is the precedent for GENERATING personalized burstiness LoRAs from a user sample, not just training one. cite
  • Q2,Q4 methodology 'From Weights to Activations' (arXiv:2604.14090) positions activation steering as the 2025/26 frontier. The Burstiness Engine should explicitly choose between weight-space (LoRA) and activation-space (steering vectors) and justify. cite
  • Q4,TTS-A methodology AgentSteerTTS (arXiv:2605.17583) multi-agent closed-loop steering shows the field moving toward composed-controller systems. Burstiness Engine could be one controller in such a system.

pass-4-gap-closing

Started 2026-06-28 17:03:26 · isolation: no-personal-carryover

Targeted Firecrawl sweep closing Q1 prompt-ceiling, Q7 punctuation-proxy, and CTRL-P verification. Re-anchored to Vittoria's canonical scope (prompt-baseline → model-level; TTS Phase 2).

Queries (4)

  • Q1 [qdr:y] · 10 results
    prompt-only stylistic control distributional ceiling negative results LLM sentence length variance kurtosis
  • Q7 [qdr:y]
    punctuation paragraph rhythm proxy text generation LLM burstiness sentence segmentation
  • Q7 · 10 results
    punctuation as proxy for sentence rhythm stylometry text segmentation author attribution
  • TTS-B · 10 results
    Wagner Klimkov CTRL-P prosodic boundaries phrasing prominence ICASSP text-to-speech

Findings (5)

  • TTS-B resolution RESOLVED: CTRL-P is a distinct paper (arXiv:2106.08352, Ctrl-P: Temporal Control of Prosodic Variation, Interspeech 2021), NOT Raitio 2020. Closes the Pass-3 open action. Per Vittoria's scope it anchors Phase 2 (voice personalization), not the core paper. cite
  • Q1 competing-work Q1 is no longer empty: 'Benchmark of Stylistic Variation in LLM-Generated Texts' (2509.10179) measures variation across 16 frontier models + prompts. It does NOT frame this as a control ceiling vs a distributional target, so the Q1 wedge survives — but must now cite and differentiate. cite
  • Q1,Q6 methodology 'LLMs Still Struggle to Imitate Implicit Writing Styles' (2509.14543) shows prompt-only personalized-style imitation fails. This is the strongest single motivation for the model-level (LoRA/steering) approach over prompting. cite
  • Q1 methodology Measure burstiness only among quality-passing generations (effective semantic diversity, 2504.12522) so the Q1 experiment does not reward high-variance incoherence. cite
  • Q7 gap Q7 remains a genuine gap — no dedicated paper treats punctuation/paragraph as a burstiness proxy. Onegin time-series (2604.20221) gives a method: model segmentation as a symbolic time series for a timing-free proxy. Like Q1, a targeted standalone contribution. cite

pass-1-2-baseline

Started 2026-07-11 12:42:49 · isolation: no-personal-carryover

Initial 8-query Firecrawl sweep of academic literature mapped to Vittoria's 7 questions + TTS + Bio anchors.

Queries (8)

  • Q2,Q4 · 10 results
    controllable text generation LoRA style adapter frozen LLM stylistic steering arXiv
  • TTS-A · 10 results
    Raitio Apple controllable prosody TTS conditioning CTRL-P prosodic speech synthesis
  • Q5,Q7 · 10 results
    burstiness sentence length variance perplexity AI generated text detection GPTZero stylometry
  • Q6 · 10 results
    stylometric fingerprint authorship attribution sentence rhythm English prose Mendenhall
  • Q3 · 10 results
    AI generated text burstiness longitudinal GPT-3 GPT-4 stylistic evolution detection across model generations
  • Q1,Q7 · 10 results
    prompt engineering ceiling stylistic distributional control LLM negative results long output drift punctuation paragraph rhythm
  • TTS-C · 9 results
    Bakkouche prosody perception Interspeech 2025 TTS speech synthesis listener
  • Bio,Arch · 10 results
    neural oscillations language predictive coding speech rhythm thought chunking transformer attention prosody biology Pickering Garrod

Findings (7)

  • Q4 novelty-bar The 'burstiness LoRA' steering mechanism already exists for OTHER style axes (sentiment, persona, authorship). Novelty bar = apply to rhythm + personal fingerprinting + perception validation, NOT invent the mechanism. cite · cite · cite
  • Q5 competing-work Tarım & Onan 2025 published the first systematic stylometric burstiness comparison (diffusion vs autoregressive). Must cite and position against, not ignore. cite
  • Q5 gap GPTZero's burstiness operationalization (Tian 2023) is a blog post, not peer-reviewed. This IS the formalization gap the Burstiness Engine identifies.
  • Q3 general Empirically confirmed: AI text style measurably shifts across model generations, detection rates drop with newer models. cite · cite
  • TTS-C perception Bakkouche 2025 perception study: listeners do NOT easily perceive humans and AI clones as the same person — prosody is the discriminator. Direct precedent for Vittoria's perception hypothesis.
  • Bio,Arch methodology Caucheteux et al. (Nature Human Behaviour 2023) show brain uses long-range hierarchical predictions matching LM architecture — the strongest single bridge between Vaswani and biology. cite
  • Q1 gap Prompt-engineering ceiling for distributional stylistic control is UNDER-PUBLISHED. The negative-result paper Vittoria proposes is itself a publishable contribution.

pass-3-cutting-edge

Started 2026-07-11 12:42:49 · isolation: no-personal-carryover

Last-12-months arXiv sweep for progression vs baseline. A/B against pass 1+2.

Queries (4)

  • Q2,Q4 [qdr:y] · 10 results
    controllable text generation LoRA style adapter 2025 frozen LLM activation steering
  • Q3,Q5 [qdr:y] · 4 results
    burstiness LLM text generation 2025 2026 stylometric variance peer-reviewed EMNLP ACL NAACL
  • TTS-A [qdr:y] · 10 results
    prosody control TTS 2025 NaturalSpeech VALL-E neural speech synthesis variance adaptor latent
  • Q3,Q5 [qdr:y] · 4 results
    AI generated text detection 2025 2026 burstiness perplexity DetectGPT Binoculars latest

Findings (5)

  • Q5,Q3 competing-work DivEye (arXiv:2509.18880, Sep 2025) captures how unpredictability FLUCTUATES across text via surprisal-based features. This is the most direct 2025 burstiness-adjacent detection paper. Must be in Vittoria's prior-art section. cite
  • TTS-A,Q2 methodology EMNLP 2025 main 'Towards Controllable Speech Synthesis in the Era of LLMs' (Lee et al.) is the unified TTS+LLM control framework. DiTTo-TTS controls speech rate via latent-length prediction — direct precedent for in-generation rhythm conditioning.
  • Q4,Q6 methodology Plug-and-Play LLM Fingerprinting (arXiv:2605.18474) generates LoRA params as variable-length sequences. This is the precedent for GENERATING personalized burstiness LoRAs from a user sample, not just training one. cite
  • Q2,Q4 methodology 'From Weights to Activations' (arXiv:2604.14090) positions activation steering as the 2025/26 frontier. The Burstiness Engine should explicitly choose between weight-space (LoRA) and activation-space (steering vectors) and justify. cite
  • Q4,TTS-A methodology AgentSteerTTS (arXiv:2605.17583) multi-agent closed-loop steering shows the field moving toward composed-controller systems. Burstiness Engine could be one controller in such a system.

pass-4-gap-closing

Started 2026-07-11 12:42:49 · isolation: no-personal-carryover

Targeted Firecrawl sweep closing Q1 prompt-ceiling, Q7 punctuation-proxy, and CTRL-P verification. Re-anchored to Vittoria's canonical scope (prompt-baseline → model-level; TTS Phase 2).

Queries (4)

  • Q1 [qdr:y] · 10 results
    prompt-only stylistic control distributional ceiling negative results LLM sentence length variance kurtosis
  • Q7 [qdr:y]
    punctuation paragraph rhythm proxy text generation LLM burstiness sentence segmentation
  • Q7 · 10 results
    punctuation as proxy for sentence rhythm stylometry text segmentation author attribution
  • TTS-B · 10 results
    Wagner Klimkov CTRL-P prosodic boundaries phrasing prominence ICASSP text-to-speech

Findings (5)

  • TTS-B resolution RESOLVED: CTRL-P is a distinct paper (arXiv:2106.08352, Ctrl-P: Temporal Control of Prosodic Variation, Interspeech 2021), NOT Raitio 2020. Closes the Pass-3 open action. Per Vittoria's scope it anchors Phase 2 (voice personalization), not the core paper. cite
  • Q1 competing-work Q1 is no longer empty: 'Benchmark of Stylistic Variation in LLM-Generated Texts' (2509.10179) measures variation across 16 frontier models + prompts. It does NOT frame this as a control ceiling vs a distributional target, so the Q1 wedge survives — but must now cite and differentiate. cite
  • Q1,Q6 methodology 'LLMs Still Struggle to Imitate Implicit Writing Styles' (2509.14543) shows prompt-only personalized-style imitation fails. This is the strongest single motivation for the model-level (LoRA/steering) approach over prompting. cite
  • Q1 methodology Measure burstiness only among quality-passing generations (effective semantic diversity, 2504.12522) so the Q1 experiment does not reward high-variance incoherence. cite
  • Q7 gap Q7 remains a genuine gap — no dedicated paper treats punctuation/paragraph as a burstiness proxy. Onegin time-series (2604.20221) gives a method: model segmentation as a symbolic time series for a timing-free proxy. Like Q1, a targeted standalone contribution. cite

pass-1-2-baseline

Started 2026-07-12 03:30:36 · isolation: no-personal-carryover

Initial 8-query Firecrawl sweep of academic literature mapped to Vittoria's 7 questions + TTS + Bio anchors.

Queries (8)

  • Q2,Q4 · 10 results
    controllable text generation LoRA style adapter frozen LLM stylistic steering arXiv
  • TTS-A · 10 results
    Raitio Apple controllable prosody TTS conditioning CTRL-P prosodic speech synthesis
  • Q5,Q7 · 10 results
    burstiness sentence length variance perplexity AI generated text detection GPTZero stylometry
  • Q6 · 10 results
    stylometric fingerprint authorship attribution sentence rhythm English prose Mendenhall
  • Q3 · 10 results
    AI generated text burstiness longitudinal GPT-3 GPT-4 stylistic evolution detection across model generations
  • Q1,Q7 · 10 results
    prompt engineering ceiling stylistic distributional control LLM negative results long output drift punctuation paragraph rhythm
  • TTS-C · 9 results
    Bakkouche prosody perception Interspeech 2025 TTS speech synthesis listener
  • Bio,Arch · 10 results
    neural oscillations language predictive coding speech rhythm thought chunking transformer attention prosody biology Pickering Garrod

Findings (7)

  • Q4 novelty-bar The 'burstiness LoRA' steering mechanism already exists for OTHER style axes (sentiment, persona, authorship). Novelty bar = apply to rhythm + personal fingerprinting + perception validation, NOT invent the mechanism. cite · cite · cite
  • Q5 competing-work Tarım & Onan 2025 published the first systematic stylometric burstiness comparison (diffusion vs autoregressive). Must cite and position against, not ignore. cite
  • Q5 gap GPTZero's burstiness operationalization (Tian 2023) is a blog post, not peer-reviewed. This IS the formalization gap the Burstiness Engine identifies.
  • Q3 general Empirically confirmed: AI text style measurably shifts across model generations, detection rates drop with newer models. cite · cite
  • TTS-C perception Bakkouche 2025 perception study: listeners do NOT easily perceive humans and AI clones as the same person — prosody is the discriminator. Direct precedent for Vittoria's perception hypothesis.
  • Bio,Arch methodology Caucheteux et al. (Nature Human Behaviour 2023) show brain uses long-range hierarchical predictions matching LM architecture — the strongest single bridge between Vaswani and biology. cite
  • Q1 gap Prompt-engineering ceiling for distributional stylistic control is UNDER-PUBLISHED. The negative-result paper Vittoria proposes is itself a publishable contribution.

pass-3-cutting-edge

Started 2026-07-12 03:30:36 · isolation: no-personal-carryover

Last-12-months arXiv sweep for progression vs baseline. A/B against pass 1+2.

Queries (4)

  • Q2,Q4 [qdr:y] · 10 results
    controllable text generation LoRA style adapter 2025 frozen LLM activation steering
  • Q3,Q5 [qdr:y] · 4 results
    burstiness LLM text generation 2025 2026 stylometric variance peer-reviewed EMNLP ACL NAACL
  • TTS-A [qdr:y] · 10 results
    prosody control TTS 2025 NaturalSpeech VALL-E neural speech synthesis variance adaptor latent
  • Q3,Q5 [qdr:y] · 4 results
    AI generated text detection 2025 2026 burstiness perplexity DetectGPT Binoculars latest

Findings (5)

  • Q5,Q3 competing-work DivEye (arXiv:2509.18880, Sep 2025) captures how unpredictability FLUCTUATES across text via surprisal-based features. This is the most direct 2025 burstiness-adjacent detection paper. Must be in Vittoria's prior-art section. cite
  • TTS-A,Q2 methodology EMNLP 2025 main 'Towards Controllable Speech Synthesis in the Era of LLMs' (Lee et al.) is the unified TTS+LLM control framework. DiTTo-TTS controls speech rate via latent-length prediction — direct precedent for in-generation rhythm conditioning.
  • Q4,Q6 methodology Plug-and-Play LLM Fingerprinting (arXiv:2605.18474) generates LoRA params as variable-length sequences. This is the precedent for GENERATING personalized burstiness LoRAs from a user sample, not just training one. cite
  • Q2,Q4 methodology 'From Weights to Activations' (arXiv:2604.14090) positions activation steering as the 2025/26 frontier. The Burstiness Engine should explicitly choose between weight-space (LoRA) and activation-space (steering vectors) and justify. cite
  • Q4,TTS-A methodology AgentSteerTTS (arXiv:2605.17583) multi-agent closed-loop steering shows the field moving toward composed-controller systems. Burstiness Engine could be one controller in such a system.

pass-4-gap-closing

Started 2026-07-12 03:30:36 · isolation: no-personal-carryover

Targeted Firecrawl sweep closing Q1 prompt-ceiling, Q7 punctuation-proxy, and CTRL-P verification. Re-anchored to Vittoria's canonical scope (prompt-baseline → model-level; TTS Phase 2).

Queries (4)

  • Q1 [qdr:y] · 10 results
    prompt-only stylistic control distributional ceiling negative results LLM sentence length variance kurtosis
  • Q7 [qdr:y]
    punctuation paragraph rhythm proxy text generation LLM burstiness sentence segmentation
  • Q7 · 10 results
    punctuation as proxy for sentence rhythm stylometry text segmentation author attribution
  • TTS-B · 10 results
    Wagner Klimkov CTRL-P prosodic boundaries phrasing prominence ICASSP text-to-speech

Findings (5)

  • TTS-B resolution RESOLVED: CTRL-P is a distinct paper (arXiv:2106.08352, Ctrl-P: Temporal Control of Prosodic Variation, Interspeech 2021), NOT Raitio 2020. Closes the Pass-3 open action. Per Vittoria's scope it anchors Phase 2 (voice personalization), not the core paper. cite
  • Q1 competing-work Q1 is no longer empty: 'Benchmark of Stylistic Variation in LLM-Generated Texts' (2509.10179) measures variation across 16 frontier models + prompts. It does NOT frame this as a control ceiling vs a distributional target, so the Q1 wedge survives — but must now cite and differentiate. cite
  • Q1,Q6 methodology 'LLMs Still Struggle to Imitate Implicit Writing Styles' (2509.14543) shows prompt-only personalized-style imitation fails. This is the strongest single motivation for the model-level (LoRA/steering) approach over prompting. cite
  • Q1 methodology Measure burstiness only among quality-passing generations (effective semantic diversity, 2504.12522) so the Q1 experiment does not reward high-variance incoherence. cite
  • Q7 gap Q7 remains a genuine gap — no dedicated paper treats punctuation/paragraph as a burstiness proxy. Onegin time-series (2604.20221) gives a method: model segmentation as a symbolic time series for a timing-free proxy. Like Q1, a targeted standalone contribution. cite

Learnings (cross-session)

All papers

TitleAuthorsVenueYearPeer?
Trusting AI to detect AI?Yicheng Sun and Yihan Liao and Xiaoxue MaComputers & Education2026
AI-Generated Text Detection: A Comprehensive Review of Active and Passive ApproachesLingyun Xiang and Nian Li and Yuling Liu and Jiayong HuComputers, Materials & Continua2025
Echoes in AI: Quantifying Lack of Plot Diversity in LLM OutputsWeijia Xu and Nebojsa Jojic and Sudha Rao and Chris BrockettPNAS2025
Enhanced Prosody Modeling and Character Voice Controlling for Audiobook Speech SynthesisNing-Qian Wu and Zhen-Hua LingACM Transactions on Asian and Low-Resource Language Information Processing2025
Finding the Human Voice in AI: Insights on the Perception of AI-Voice Clones from Naturalness and Similarity RatingsLinda Bakkouche and Charles McGhee and Emily Lau and Stephanie Cooper and Xinbing Luo and Madeleine Rees and Kai Alter and Brechtje Post and Julia SchwarzInterspeech 20252025
Style and Prosody control for Zero-shot Speech SynthesisSuni, Antti et al.SSW 20252025
Towards Controllable Speech Synthesis in the Era of LLMsLee et al.EMNLP 2025 main2025
How is ChatGPT's Behavior Changing Over Time?Lingjiao Chen and Matei Zaharia and James ZouHarvard Data Science Review2024
Language experience shapes predictive coding of rhythmic sound sequencesPiermatteo Morucci and Sanjeev Nara and Mikel Lizarazu and Clara Martin and Nicola MolinaroeLife2024
Large Language Models Cannot Self-Correct Reasoning YetJie Huang and Xinyun Chen and Swaroop Mishra and Huaixiu Steven Zheng and Adams Wei Yu and Xinying Song and Denny ZhouICLR 20242024
Lost in the Middle: How Language Models Use Long ContextsNelson F. Liu and Kevin Lin and John Hewitt and Ashwin Paranjape and Michele Bevilacqua and Fabio Petroni and Percy LiangTACL 20242024
Style Vectors for Steering Generative Large Language ModelsKai Konen and Sophie Jentzsch and Diaoulé Diallo and Peer Schütt and Oliver Bensch and Roxanne El Baff and Dominik Opitz and Tobias HeckingEACL 2024 Findings2024
Distinguishing ChatGPT(-3.5, -4)-generated and human-written papers through Japanese stylometric analysisWataru Zaitsu and Mingzhe JinPLoS One2023
Long-range and hierarchical language predictions in brains and algorithmsCharlotte Caucheteux and Alexandre Gramfort and Jean-Remi KingNature Human Behaviour2023
Plan-and-Solve Prompting: Improving Zero-Shot Chain-of-Thought Reasoning by Large Language ModelsLei Wang and Wanyu Xu and Yihuai Lan and Zhiqiang Hu and Yunshi Lan and Roy Ka-Wei Lee and Ee-Peng LimACL 20232023
Predictive Coding or Just Feature Discovery? An Alternative AccountRichard Antonello and Alexander HuthNeurobiology of Language2023
Self-Refine: Iterative Refinement with Self-FeedbackAman Madaan and Niket Tandon and Prakhar Gupta and othersNeurIPS 20232023
Emphasis control for parallel neural TTSShreyas Seshadri and Tuomo Raitio and Dan Castellani and Jiangchuan LiInterspeech 20222022
Rethinking the Role of Demonstrations: What Makes In-Context Learning Work?Sewon Min and Xinxi Lyu and Ari Holtzman and Mikel Artetxe and Mike Lewis and Hannaneh Hajishirzi and Luke ZettlemoyerEMNLP 20222022
Ctrl-P: Temporal Control of Prosodic Variation for Speech SynthesisDevang S Ram Mohan and Vivian Hu and Tian Huey Teh and Alexandra Torresquintero and Christopher G. R. Wallis and Marlene Staib and Lorenzo Foglianti and Jiameng Gao and Simon KingInterspeech 20212021
FUDGE: Controlled Text Generation With Future DiscriminatorsKevin Yang and Dan KleinNAACL 20212021
Mirostat: A Neural Text Decoding Algorithm that Directly Controls PerplexitySourya Basu and Govardana Sachitanandam Ramachandran and Nitish Shirish Keskar and Lav R. VarshneyICLR 20212021
Controllable neural text-to-speech synthesis using intuitive prosodic featuresTuomo Raitio and Ramya Rasipuram and Dan CastellaniInterspeech 20202020
Language Models are Few-Shot LearnersTom B. Brown and Benjamin Mann and Nick Ryder and othersNeurIPS 20202020
Measuring burstiness for finite event sequencesKim, Eun-Kyeong and Jo, Hang-HyunPhysical Review E2016
Burstiness and memory in complex systemsGoh, Kwang-Il and Barabási, Albert-LászlóEPL (Europhysics Letters)2008
A note on the statistical analysis of sentence-length as a criterion of literary styleWilliams, C. B.Biometrika1940
On sentence-length as a statistical characteristic of style in proseYule, G. UdnyBiometrika1939
Variabilità e MutabilitàGini, Corradobook:C. Cuppini, Bologna1912
A Statistical Journey into the Poetic World of Evgenij OneginAngelo Maria SabatiniarXiv2026·
A Unified Study of LoRA Variants: Taxonomy, Review, CodebaseHaonan He and Jingqi Ye and Minglei Li and Zhengbo Wang and Tao Chen and Lei Bai and Peng YearXiv2026·
AgentSteerTTS: Multi-Agent Closed-Loop TTS SteeringBin Kang and Shaoguo Wen and Yang Fan and Shunlong Wu and Junjie Wang and Yulin Li and Junzhi Zhao and Junle Wang and Zhuotao TianarXiv2026·
CARD: Cluster-level Adaptation with Reward-guided Decoding for Personalized Text GenerationYutong Song and Jiang Wu and Weijia Zhang and Chengze Shen and Shaofan Yuan and Weitao Lu and Jian Wang and Yu Wang and Nikil Dutt and Amir M. RahmaniarXiv2026·
Continuous Control of Editing Models via Adaptive-Origin GuidanceAlon Wolf and Chen Katzir and Kfir Aberman and Or PatashnikarXiv2026·
From Weights to Activations: Is Steering the Next Frontier of Adaptation?Simon Ostermann and Daniil Gurgurov and Tanja Baeumel and Michael A. Hedderich and Sebastian Lapuschkin and Wojciech Samek and Vera SchmittarXiv2026·
GLASS: GRPO-Trained LoRA for Acoustic Style Steering in Zero-Shot TTSJaehoon Kang and Yejin Lee and Kyuhong ShimarXiv2026·
MAGIC-TTS: Fine-Grained Controllable Speech SynthesisJialong Mai and Xiaofen Xing and Xiangmin XuarXiv2026·
Prompt2Fingerprint: Plug-and-Play LLM Fingerprinting via Text-to-Weight GenerationSixu Chen and Xiang Chen and Hongyao Yu and Jiaxin Hong and Hao Fang and Shuoyang Sun and Bin Chen and Shu-Tao XiaarXiv2026·
Self-Supervised Honesty Steering via Anti-Parallel RepresentationsMichael J. ClarkarXiv2026·
Styles + Persona-plug = Customized LLMsYutong Song and Jiang Wu and Shaofan Yuan and Chengze Shen and Jian Wang and Amir Rahmani and Nikil Dutt and Yu WangarXiv2026·
TADA! Tuning Audio Diffusion Models through Activation SteeringŁukasz Staniszewski and Katarzyna Zaleska and Mateusz Modrzejewski and Kamil DejaarXiv2026·
The Statistical Signature of LLMsOrtal Hadad and Edoardo Loru and Jacopo Nudo and Niccolò Di Marco and Matteo Cinelli and Walter QuattrociocchiarXiv2026·
A Training-free Method for LLM Text AttributionTara Radvand and Mojtaba Abdolmaleki and Mohamed Mostagir and Ambuj TewariarXiv2025·
Benchmark of Stylistic Variation in LLM-Generated TextsJiří Milička and Anna Marklová and Václav CvrčekarXiv2025·
Beyond Checkmate: Creative Choke Points in AI TextNafis Irtiza Tripto and Saranya Venkatraman and Mahjabin Nahar and Dongwon LeearXiv2025·
Can You Detect the Difference?Tarım, İsmail and Onan, AytuğarXiv2025·
Detecting LLM-Generated Short Answers and Effects on Learner PerformanceShambhavi Bhushan and Danielle R Thomas and Conrad Borchers and Isha Raghuvanshi and Ralph Abboud and Erin Gatz and Shivang Gupta and Kenneth KoedingerarXiv2025·
Diversity Boosts AI-Generated Text DetectionAdvik Raj Basani and Pin-Yu ChenarXiv2025·
Evaluating the Diversity and Quality of LLM Generated ContentAlexander Shypula and Shuo Li and Botong Zhang and Vishakh Padmakumar and Kayo Yin and Osbert BastaniarXiv2025·
LLM Braces: Straightening Out LLM Predictions with Relevant Sub-UpdatesYing Shen and Lifu HuangarXiv2025·
LLMs Still Struggle to Imitate the Implicit Writing Styles of Everyday PeopleZhengxiang Wang and Nafis Irtiza Tripto and Solha Park and Zhenzhen Li and Jiawei ZhouarXiv2025·
Merge and Guide: Unifying Model Merging and Guided Decoding for Controllable Multi-Objective GenerationGuofu Xie and Chen Zhang and Xiao Zhang and Yunsheng Shi and Ting Yao and Jun XuarXiv2025·
Personalized Text Generation with Contrastive Activation SteeringJinghao Zhang and Yuting Liu and Wenjie Wang and Qiang Liu and Shu Wu and Liang Wang and Tat-Seng ChuaarXiv2025·
RedNote-Vibe: Temporal Dynamics of AI-Generated TextYudong Li and Yufei Sun and Peiru Yang and Yuhan Yao and Wanyue Li and Jiajun Zou and Haoyang Yang and Haotian Gan and Linlin Shen and Yongfeng HuangarXiv2025·
Stylometry Recognizes Human and LLM-Generated TextsKarol Przystalski and Jan K. Argasiński and Iwona Grabska-Gradzińska and Jeremi K. OchabarXiv2025·
Continuous Language Model Interpolation for Dynamic and Controllable Text GenerationSara Kangaslahti and David Alvarez-MelisarXiv2024·
Detecting AI-Generated Text: Factors Influencing Detectability with Current MethodsKathleen C. Fraser and Hillary Dawkins and Svetlana KiritchenkoarXiv2024·
Large Language Models can be Guided to Evade AI-Generated Text DetectionNing Lu and Shengcai Liu and Rui He and Qi Wang and Yew-Soon Ong and Ke TangTMLR 20242024·
Low-Rank Adaptation for Foundation Models: A Comprehensive ReviewMenglin Yang and Jialin Chen and Jinkai Tao and Yifei Zhang and Jiahong Liu and Jiasheng Zhang and Qiyao Ma and Harshit Verma and Regina Zhang and Min Zhou and othersarXiv2024·
StyleRemix: Interpretable Authorship Obfuscation via Distillation and Perturbation of Style ElementsJillian Fisher and Skyler Hallinan and Ximing Lu and Mitchell Gordon and Zaid Harchaoui and Yejin ChoiarXiv2024·
The Prompt Report: A Systematic Survey of Prompt Engineering TechniquesSander Schulhoff and Michael Ilie and Nishant Balepur and othersarXiv2024·
A Prompt Pattern Catalog to Enhance Prompt Engineering with ChatGPTJules White and Quchen Fu and Sam Hays and Michael Sandborn and Carlos Olea and Henry Gilbert and Ashraf Elnashar and Jesse Spencer-Smith and Douglas C. SchmidtarXiv2023·
Toward a realistic model of speech processing in the brain with SSLJuliette Millet and Charlotte Caucheteux and Pierre Orhan and Yves Boubenec and Alexandre Gramfort and Ewan Dunbar and Christophe Pallier and Jean-Remi KingarXiv2022·
Wikinews: free-content news source (English, 2016-2021)Wikinews·2021·