Frontier Evaluation Exposes Weaknesses in Autonomous Long-Horizon Reasoning

FutureSim, ASI-Bench, and LoopArena continue to show weak real-world forecasting, dependence on human guidance, and low strict task success. Debate over pass@1 versus pass@k and the reliability of self-scored search further favors calibrated, verifier-backed, trajectory-level evaluation over headline capability claims.