Newest papers

What the Author Wrote, What the Agent Reads: Silent Cue Loss in Enterprise Document Ingestion

Yikai Zhao, Saurabh Pandey, Pradeep Kumar Misra · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.30026 · cited by 0 works (OpenAlex) · open-access PDF

Portuguese

O agente de corretação corporativa é um agente válido através de documentos de laboratório. Antes que um agente possa seguir uma política, responder a uma pergunta de conformidade ou enviar um bilhete, algumas camadas de ingestão devem transformar um arquivo de Word ou PDF em tokens - e essa camada é onde um agente principal implantado no selvagem silenciosamente perde informações. Mostramos que uma camada é perdida por cada conversor mainstream - o que significa que um autor codifica a forma como o texto parece - e que a perda faz com que os agentes fiquem confiantes em vez de visivelmente incompletos. Nós construímos uma tarefa a partir de 76 documentos de política real: recuperar os requisitos que o autor separou, marcado contra as propriedades OOXML do arquivo por um determinista que combina com nenhum juiz LLM.

Machine translation (Meta M2M100); the English original follows.

Abstract

Enterprise agents act on documents. Before an agent can follow a policy, answer a compliance question or file a ticket, some ingestion layer must turn a Word file or PDF into tokens - and that layer is where an agent deployed in the wild silently loses information. We show that one layer is lost by every mainstream converter - the meaning an author encodes in how text looks - and that the loss makes agents confidently wrong rather than visibly incomplete. We build a task from 76 real policy documents: recover the requirements the author singled out, scored against the file's own OOXML properties by a deterministic matcher with no LLM judge. Formatting-blind ingestion collapses from micro F1 0.690 to 0.150 (docling) and 0.075 (plain text), a gap that replicates across six answering models from two vendors - five Claude models spanning the Opus, Sonnet and Haiku tiers, and OpenAI's GPT-5.6. An agent that cannot see the cue guesses, and docling's false-positive rate on requirements the author did not flag is four times ours. A span-level leave-one-out localises the loss to two cues - highlight (+0.302) and colour (+0.262), both Holm-adjusted p < 1e-4 - while bold, italic, monospace and strikethrough each sit within +/-0.003 of zero. The same asymmetry appears in PDF. Colour is also the cue no mainstream converter preserves and no parsing benchmark scores, and it is common: 32.8% of 6,460 documents carry a deliberate colour or highlight, 10.3% on an obligation. A pre-registered test shows that typing a cue for the agent adds +0.085 beyond merely preserving it (Holm-adjusted p = 0.049). We validate the labels against rendered page images through an independent modality (kappa = 0.72) and bound what wording alone can achieve (AUC 0.664). For agent builders the practical finding is that ingestion, not reasoning, is where this reliability is won or lost, and that the fix is cheap and measurable.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Computer science, Task (project management), Meaning (existential), Layer (electronics), Word (group theory), Mainstream

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange