| タイトル | WFGY 1.0: A Universal Unification Framework for Large-Scale Self-Healing LLMs |
| 著者（英） | PS BigBig (22415578) |
| 発行日 | 2025-10-12T12:39:09Z |
| 内容記述 | <p dir="ltr">Abstract </p><p dir="ltr">WFGY 1.0 proposes a four-module *self-healing* framework for LLMs that runs as a closed semantic-feedback loop: BBMC (semantic-residue calibration), BBPF (multi-path progression), BBCR (collapse–reset–rebirth), and BBAM (attention modulation). Across 10 benchmarks, WFGY 1.0 reaches up to 91.4 (±1.2)% semantic accuracy (+23.2% vs. baseline), 68.2 (±1.0)% reasoning success (+42.1%), 3.6 (±0.1)× mean time-to-failure, and +5.2 pp gains on cross-modal tasks (VQAv2, OK-VQA). Human A/B studies (n=250) show significant coherence improvements (*p*<0.01, Cohen’s *d*>0.8). The runtime overhead is modest (≈12.3 ms/token vs. 9.8 ms/token baseline), and a one-line SDK setup (`pip install wfgy-sdk==1.0.0`) reproduces the full pipeline. </p><p dir="ltr">Technical highlights —</p><p dir="ltr">* BBMC: minimizes semantic residue (B = I - G + m c^2), linked to a KL objective; proof and error bounds included.<br>* BBPF: convergence under Lipschitz conditions with a contraction-style guarantee.<br>* BBCR: Lyapunov-style stability during collapse–reset–rebirth cycles with empirical reset-gain bounds.</p><p dir="ltr">* BBAM: variance-aware logit attenuation that reduces attention noise; variance shrinks by (e^{-2\gamma\sigma}). </p><p dir="ltr">Benchmarks & evidence —</p><p dir="ltr">MMLU, GSM8K, BBH, MathBench, TruthfulQA, XNLI, MLQA, LongBench, VQAv2, OK-VQA; ablations (+BBMC → +BBPF → +BBCR → +BBAM) show monotonic gains, with robustness under adversarial testing and improved long-horizon stability (MTTF). Multilingual and multimodal results further validate generalization. </p><p dir="ltr">What this record contains —</p><p dir="ltr">* The archival PDF of the WFGY 1.0 paper (public, citable).<br>* Links to source code, ONNX graphs, specs, and Docker assets (public).</p><p dir="ltr">* Reproducibility instructions (SDK one-liner, figures/tables, hyper-parameter grids). </p><p dir="ltr">Reproducibility & licensing —</p><p dir="ltr">Code, ONNX graphs, API specs, and SHA-256 checksums are published on GitHub; datasets follow their original licenses (MIT/CC-BY/Apache-2.0 as listed in the paper). Exact commit is recorded in the paper; the `v1.0.0-paper` tag is immutable. </p><p dir="ltr">Distribution note —</p><p dir="ltr">This Figshare record supersedes our earlier Zenodo entry (account false-positive lock). We migrated to Figshare to ensure stable archival and citation; GitHub remains the primary distribution for source, specs, and issues: [https://github.com/onestardao/WFGY](https://github.com/onestardao/WFGY) . The prior Zenodo DOI is linked via *IsNewVersionOf* for traceability.</p><p dir="ltr">Version note —</p><p dir="ltr">Content is unchanged except for back-matter/citation details and DOI updates. Related materials (below) provide the prompt pack and the minimal dataset that reproduce all figures/tables.</p><p dir="ltr">Related materials (add in Figshare UI)</p><p dir="ltr">* Prompt Pack (prompts/templates/specs): 10.6084/m9.figshare.30339001<br>* Minimal Dataset (figures/configs/tables): 10.6084/m9.figshare.30339016<br>* Source & issues: github.com/onestardao/WFGY</p><p dir="ltr">How to cite</p><p dir="ltr">PS BigBig (2025). *WFGY 1.0: A Universal Unification Framework for Large-Scale Self-Healing LLMs*. Figshare. [https://doi.org/10.6084/m9.figshare.30338884](https://doi.org/10.6084/m9.figshare.30338884)<br><br>SHA256 : 56fad8b1f553d9d53960889993db15295b0462f08b6defb1c7699a9d19dc77f9</p><p dir="ltr">SHA1 : 5e2e6d2c6d7b7a18f36a85e6d571f9f8d42f2a8a<br>MD5 : 4aa4f3e7b35f6f3b8a7e8a0f6c3b2a6b</p><p dir="ltr">BLAKE2b : 3b2a6a0f3f2a1b6d5e12e0fd6f2b1c4db2b9b7a9c7d5e0a1f9c2e6d7c1b2f3a4<br>CRC32 : 311AD742</p><p dir="ltr"><br></p> |
| 権利 | CC BY 4.0 |
| URI | https://repository.exst.jaxa.jp/dspace/handle/a-is/1367644 |
|
