{"paper":{"title":"FutureSim: Replaying World Events to Evaluate Adaptive Agents","license":"http://creativecommons.org/licenses/by/4.0/","headline":"FutureSim evaluates AI agents by replaying real historical events in order and shows even the best achieve only 25 percent accuracy on future predictions.","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Ameya Prabhu, Arvindh Arun, Jonas Geiping, Maksym Andriushchenko, Moritz Hardt, Nikhil Chandak, Shashwat Goel, Steffen Staab","submitted_at":"2026-05-14T17:59:28Z","abstract_excerpt":"AI agents are being increasingly deployed in dynamic, open-ended environments that require adapting to new information as it arrives. To efficiently measure this capability for realistic use-cases, we propose building grounded simulations that replay real-world events in the order they occurred. We build FutureSim, where agents forecast world events beyond their knowledge cutoff while interacting with a chronological replay of the world: real news articles arriving and questions resolving over the simulated period. We evaluate frontier agents in their native harness, testing their ability to p"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"FutureSim reveals a clear separation in their capabilities, with the best agent's accuracy being 25%, and many having worse Brier skill score than making no prediction at all.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That replaying real historical events chronologically without future knowledge leakage accurately measures an agent's adaptive capabilities in open-ended real-world settings.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"FutureSim is a benchmark that replays real news from January to March 2026 for AI agents to forecast events, with top accuracy at 25% and some agents worse than no-prediction baselines on Brier skill score.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"FutureSim evaluates AI agents by replaying real historical events in order and shows even the best achieve only 25 percent accuracy on future predictions.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"162a363276cac9ee69424ab8967097dedc85f9a09d938c5ff0d36ca3518e1ce7"},"source":{"id":"2605.15188","kind":"arxiv","version":1},"verdict":{"id":"ee187d39-2dbe-4949-9514-d3edbd034f12","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T03:08:02.403868Z","strongest_claim":"FutureSim reveals a clear separation in their capabilities, with the best agent's accuracy being 25%, and many having worse Brier skill score than making no prediction at all.","one_line_summary":"FutureSim is a benchmark that replays real news from January to March 2026 for AI agents to forecast events, with top accuracy at 25% and some agents worse than no-prediction baselines on Brier skill score.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That replaying real historical events chronologically without future knowledge leakage accurately measures an agent's adaptive capabilities in open-ended real-world settings.","pith_extraction_headline":"FutureSim evaluates AI agents by replaying real historical events in order and shows even the best achieve only 25 percent accuracy on future predictions."},"references":{"count":25,"sample":[{"doi":"10.5281/zenodo.1207631","year":2018,"title":"World models","work_id":"74007479-6f51-4839-ae30-4d6122d21c36","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2026,"title":"Lost in simulation: Llm-simulated users are unreliable proxies for human users in agentic evaluations","work_id":"1e07215b-a63b-418a-8387-6f3ccef28361","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"just give the model shell and tool access","work_id":"d4881466-cfe9-4f1b-95d7-3b358380ab45","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Context consumption feedback:After each tool call, the agent receives feedback about remaining context budget and approximate context occupancy. This is useful because the task spans thousands of turn","work_id":"c90e391d-d088-4e4f-880f-346a36ccc7cb","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"The goal is to make memory writing and retrieval deliberate actions rather than accidental byproducts of shell usage","work_id":"335abbb3-e733-4812-972b-67211b25b759","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":25,"snapshot_sha256":"98adbddef2b0d635a05aa2a0371620041f76819bc33f8144f0c8c6d153cc870f","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}