BITE Before You Think: Morphology-Guided Computational Mastication for Language-Model Inputs

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Equal-length inputs need not require the same processing: a separable record, a repetitive log and a tightly coupled dependency chain present different computational workpieces. BITE asks which information should be processed by which specialised “tooth”, and whether matching operations to input morphology can reduce total processing cost while preserving task-relevant information. Drawing on mastication mechanics, we formulate a material–operation–state interaction: CUT separates tractable boundaries, UNTANGLE separates problematic structures, REDUCE consolidates repetition, and INTEGRATE assembles coupled evidence. Selection should depend on the state produced by preceding operations, not length alone. Source tracing and recovery constrain this process; they are not its optimisation objective. We distinguish this programme from the restricted, deterministic implementation actually evaluated. The unchanged prospective study used 18 public structured files and four arms. Its fixed reference gate and ordered planner selected REDUCE on one document and no operations on 17. In the one matched-budget case, BITE reduced 112 input tokens to 91 and retained 1.0 of unique assignment identities, versus 0.8462 for prefix truncation. A separately labelled retrospective correction found reference retention of 1.0 for both methods on one different document. All 72 arm–document cases reconstructed their sources. These findings establish a limited implementation result, not the central morphology-by-operation hypothesis or net savings at matched answer quality. In response to Dream-RSI, we distinguish within-input transformation selection from history-based improvement of exploration policies. This second edition restores computational mastication as the research subject, preserves the original measurements, and specifies the missing tests of tooth–workpiece compatibility. Version 2.0: scope of revision Version 2.0 is a conceptual and interpretive revision of the English and Japanese version-1.0 manuscripts, not a new experimental report. It restores morphology-guided computational mastication as the research subject: which information should be processed by which specialised operation, and how the state produced by an earlier operation should inform the next choice. Source tracing and recovery remain safeguards rather than the optimisation objective. This edition clarifies the material–operation–state formulation and discusses Dream-RSI as related work. The original measurements, historical verdicts and the distinction between prospective observations and retrospective corrections are unchanged. No new benchmark run, model fine-tuning or BITE retuning was performed for this edition. Record contents and evidence boundaries This record contains equivalent English and Japanese manuscripts. The Japanese manuscript is a translation, not an independent replication. The version-2.0 revision companion documents the conceptual revision and does not replace the original version-1.0 empirical supplement (Tables S1–S6), which accompanies the unchanged measurements. This is an unreviewed preprint. The deposited materials do not include executable BITE source code or copies of third-party raw files. The original measurements do not establish the central morphology-by-operation hypothesis, general compression superiority, net savings at matched answer quality or deployment safety.