aogavrilov/where-quality-breaks-results

Where Quality Breaks in Compressed Short-Text Generation: Staged Bottleneck Localization — reported result summary

This repository contains an author-maintained, machine-readable summary of the key quantitative values reported in Where Quality Breaks in Compressed Short-Text Generation: Staged Bottleneck Localization.

Scope: this is a small table-level result summary. It is not the underlying training corpus, evaluation corpus, model code, checkpoint, benchmark release, or a new experimental run.

Publication

• Canonical publication page: https://aogavrilov.com/publications/where-quality-breaks/
• DOI: https://doi.org/10.23919/FRUCT70069.2026.11506553
• Authors: Alexey Gavrilov, Alan-Barsag Gazzaev, Sergey Muravyov
• Venue: 2026 39th Conference of Open Innovations Association (FRUCT)
• Open preprint record: https://arxiv.org/abs/2607.24176
• Open result section: https://aogavrilov.com/publications/where-quality-breaks/#results
• Hugging Face collection: https://huggingface.co/collections/aogavrilov/publication-result-summaries-6a6b733b3116a87befb37ec2

Reader guides

• English paper guide: https://aogavrilov.com/publications/where-quality-breaks/
• Руководство на русском: https://aogavrilov.com/ru/publications/where-quality-breaks/
• 简体中文论文导读: https://aogavrilov.com/zh/publications/where-quality-breaks/
• Problem-first guide — How to tell whether a compressed text generator fails in the codec or the generator: English · Русский · 简体中文

Files

• results.csv — table shown in the Dataset Viewer.
• results.json — table plus DOI, metric, sample-size, condition, uncertainty, and takeaway metadata.
• results.md — human-readable result summary.
• citation.bib — BibTeX record for the paper.
• manifest.json and SHA256SUMS — source links and integrity metadata for this export.

Experimental scope recorded by the paper

• Source data: TinyStories, as described in the paper.
• Reported sample size: 256 paired reconstruction samples; 251–256 generated samples per mode; four matched geometry settings.
• Conditions: GPT-2 token sequences of length 64 compressed to 16 top-level codes with a hierarchical VQ-VAE-2; all generation modes use the shared external scorer.
• Metrics: External GPT-2 perplexity: mean, median, p95, and maximum; Codebook usage and support size; SBERT, BERTScore, MAUVE, and an LLM-judge summary for geometry runs
• Main reported takeaway: Most of the observed quality loss is introduced before generation; code-space diffusion still reduces median perplexity by 30.9% versus token-space diffusion.

Limitations

• Statistical uncertainty: The reported comparisons are descriptive single runs; confidence intervals and multi-seed significance estimates were not computed.
• Data boundary: The evidence covers short synthetic stories and should not be treated as a benchmark for unrestricted natural-language generation.
• Version boundary: A dataset checksum or immutable TinyStories snapshot identifier is not reported in the paper.
• Reproducibility boundary: A public installation recipe is not yet available; no inactive Code button is shown.

Do not treat absent values as zero, infer functional correctness from structural proxies, or transfer the reported ranking beyond the stated experimental setting.

Rights and provenance

No separate reuse license is asserted for the source paper or upstream data by this export. The paper PDF remains subject to the stated IEEE rights, and the upstream data retain their own terms.

Use of TinyStories remains subject to the dataset's own terms; no dataset files are redistributed by this site.

The authoritative context and current rights statement are maintained at https://aogavrilov.com/publications/where-quality-breaks/#data.

Citation

@inproceedings{Gavrilov2026WhereQuality,
  title      = {Where Quality Breaks in Compressed Short-Text Generation: Staged Bottleneck Localization},
  author     = {Gavrilov, Alexey and Gazzaev, Alan-Barsag and Muravyov, Sergey},
  booktitle  = {2026 39th Conference of Open Innovations Association (FRUCT)},
  publisher  = {IEEE},
  year       = {2026},
  pages      = {69--76},
  doi        = {10.23919/FRUCT70069.2026.11506553},
  url        = {https://doi.org/10.23919/FRUCT70069.2026.11506553},
  isbn       = {978-952-65246-5-8},
}