Calibration errors vary by target when predicting AI agent outcomes
Testing-Driven Reliability Audit of Trajectory-Based Early Outcome Prediction for LLM Agents: Target-Specific Calibration Transfer Persists Within a Single Benchmark
Artificial Intelligence
Summary
Predicting whether AI agents will succeed or fail early in their tasks helps save time. This paper studies how well these predictions stay accurate when used on new AI agents that the predictor hasn’t seen before. The authors found that in general, the prediction confidence transfers well between agents, but some specific agent and outcome combinations show consistent errors. These errors were not found to be a general problem across different testing environments, suggesting the issue is specific to certain conditions.
What this means in practice
- •For ai evaluation teams: Identify when confidence predictions for early stopping are reliable or prone to errors on new AI agents using trajectory data.
- •For software validation engineers: Use calibrated early outcome predictors to reduce run times by terminating tests once confident predictions are reached, avoiding costly full runs.
Authors
YanZe Cao
Abstract
Predicting early outcomes based on trajectory can decrease the expenses associated with agent evaluation by terminating a run once the outcome becomes sufficiently predictable, assuming that the predictor's confidence is properly calibrated. Calibration is at risk when a predictor is applied to an agent on which it was never trained, but it is not known whether such transfer failures are broad across agent systems or concentrated in specific target agent/head combinations. Using public SWE-bench Verified trajectories and a frozen dual-head early-outcome prediction pipeline, we ran a leave-one-agent-out calibration audit, a shared-predictor leave-two-agents-out control, oracle prior correction, and a robustness battery over training cohorts, task resampling, task halves, jackknife, and thresholds. Fixed-scaffold TerminalBench analysis served as a pre-registered boundary test. Broad same-predictor pairwise heterogeneity was not supported; the median pairwise corrected-gap differences were 0.0180 (SUCCESS head, 45 pairs) and 0.0385 (FAILURE head, 35 pairs), and the pre-registered heterogeneity criterion was not met on either head. Two specific combinations, gpt-5-mini/SUCCESS and claude-opus-4.6/FAILURE, showed persistent calibration-transfer errors (median corrected gaps 0.1377 and 0.1107) without a sign reversal under any frozen control. TerminalBench did not establish cross-benchmark replication: the success target produced zero decisions (INDETERMINATE), and the failure target did not satisfy the pre-registered persistence criterion. Therefore, a strong target-specific calibration-transfer error can exist within one frozen environment, but the evidence does not establish that the error is intrinsic to the model or general across benchmarks.