Newest papers

Same Simulation, Different Evidence: Where Retrieved Tool Output Costs LLM Agents Accuracy

Ke Zhang, Sahchit Chundur, Jerry Peprah Owusu, Mohammad Javad Abdolhosseini Qomi, Maziar Raissi · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.26836 · cited by 0 works (OpenAlex) · open-access PDF

French

Les agents LLM sont les agents correcteurs qui calculent la réponse correcte peuvent toujours répondre incorrectement. Les agents utilisent de plus en plus de longues sorties d'outils hors de la fenêtre contextuelle, les enregistrant dans des fichiers que l'agent recherche ou lit sur demande. Nous demandons où cela coûte l'exactitude avec un véritable outil scientifique: PHREEQC, un simulateur de géochimie largement utilisé de l'USGS, dont les rapports atteignent 846.8k caractères et dont le déterminisme permet de suivre chaque valeur imprimée. Les agents l'utilisent pour répondre à 158 éléments de choix multiple vérifiés par l'exécution. Lorsque les agents lisent le rapport du simulateur à travers un indice de section plutôt que dans le contexte, GPT-5.1 et

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

Abstract

LLM agents that compute the right answer can still answer wrongly. Agent harnesses increasingly keep long tool outputs out of the context window, saving them to files that the agent searches or reads on demand. We ask where this costs accuracy with a real scientific tool: PHREEQC, a widely used geochemistry simulator from the USGS, whose reports reach 846.8k characters and whose determinism lets every printed value be traced. Agents use it to answer 158 execution-verified multiple-choice items. When agents read the simulator's report through a section index rather than in context, GPT-5.1 and GPT-5.2 lose 11.3 and 7.9 points of accuracy, while Claude Opus 4.6 and Claude Sonnet 4.6 show no detectable cost. Re-running all 11,007 logged simulator calls shows that both GPT models compute the keyed value about as often as before; what changes is the evidence they answer from. Under retrieval they often skip the report, answer from their own higher-precision output files, and misround to a neighboring option, a failure we call evidence substitution. With the simulation held fixed, retrieval still costs GPT-5.2 9.4 points. For GPT-5.4, more frequent correct simulations mask an evidence-use loss of about ten points, leaving a net change indistinguishable from zero. Naming the authoritative output raises accuracy by 6.3-7.9 points over plain retrieval for the three GPT models; for GPT-5.1 and GPT-5.2, inlining outputs that fit works as well, while asking for correct rounding does not. Context-management choices in agent harnesses should be judged by where accuracy is lost, not only by final accuracy.

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

Topics: Computer science, Rounding, Context (archaeology), Value (mathematics), Section (typography), Ask price

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