Pattern informatics and token knockout experiments using the HRA preparation
Year: 2026
Type: preprint
Abstract: We observed that removing a native token frequently failed to remove the subsequent pattern of reasoning. We organize Stages 6–8 of the Hortus Regis Alti (HRA) preparation as pattern informatics: the identification, manipulation and comparison of traceable process patterns, with explicit measurement-failure conditions. Retrospective random-forest screening preceded 216 matched deletion or insertion pairs, comprising 432 continuations and 5,112,618 newly generated native tokens from one frozen DeepSeek-R1-Distill-Qwen-7B preparation. In 87 of 108 deletion pairs, the edited continuation immediately regenerated the deleted token ID; six deletion pairs matched across their entire coordinate-aligned output overlap after regeneration. Deletion restored natural termination in 18 pairs but removed it in 22; strict correct submission improved in three and deteriorated in seven. Across both edit types, all 26 termination recoveries remained strictly incorrect. Subsequent lineage-held-out forests predicted long exact repetition from preceding native-symbol features, yet 234 of 255 eligible trajectories without a detected target interval also admitted a 32-token window whose replacement lowered all three reference scores. These observations expose the ineffectiveness of single-token deletion as a reliable, persistent functional knockout in this preparation. They also motivate a deeper attractor hypothesis: local symbolic removal can be followed by endogenous reconstruction and renewed occupation of a prior pattern. Contextual reconstruction, self-correction, continued self-inference and sampling-path divergence provide competing explanations. HRA makes these alternatives inspectable through preserved trajectories, matched operations and explicit probe calibration. Corresponding authors in order: Yuncheng Jia; Boyu Zhou. Lead contact: Yuncheng Jia ([email protected]). No affiliations stated. Not peer reviewed. Stage 6: screening; Stage 7: 216 matched pairs / 432 arms; Stage 8: returned-data analysis. Beijing time (UTC+08:00); execution and archiving: Chengdu, Sichuan, China.version 2 upload with lab note (chinese)
Authors: Jia, Yuncheng, Zhou, Boyu
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225459489