Separating Pressure and Accountability in Tool-Using Agents: Evidence from CESO-Bench
Desk rejected at ICLR 2027, so left out of the paper lists and the map.
Abstract
In the pressure comparison studied here, a pressured agent is compared with an agent told that its work will be audited. That comparison does not separate a pressure effect from an accountability-cue effect or from the model's behavior without either cue. We study this decomposition in executable tool use with CESO-Bench, a synthetic benchmark of 24 task families spanning access control, refund processing, and data retention. Each family has pressure, full-auditor, and neutral variants built from the same task body. The neutral closing keeps task-defining details while removing selected pressure language and is mechanically matched to its paired pressure closing within ±15% of word count. We evaluate three MiniMax M-series configurations at temperature 0.7 using a frozen surface success signal and deterministic end-state compliance oracles. The primary outcome is silent pass, SP = mean[A(1-C)], where A is the recorded task_answer_ok signal and C is the atomic-compliance result after authoritative replay. Across the three models, the pressure-versus-neutral family means are -0.015, -0.031, and -0.023 for M2.7, M2.5, and M3; the corresponding auditor-versus-neutral means are -0.060, -0.096, and -0.044. The M2.7 95% percentile intervals are [-0.049, +0.011] and [-0.157, +0.003]; the M2.5 and M3 analyses have partial family support and report point estimates only. This paper treats those contrasts as endpoint-specific evidence: a point estimate near zero is not equivalence, a confidence interval that covers zero is not proof of independence, and an auditor comparison does not identify a mechanism. The study is limited to the synthetic task pack, recorded runs, and the declared oracle contract.
- Primary area:
- datasets and benchmarks