EarlyEval introduces early outcome prediction as a way to cut the cost of evaluating LLM agents, training lightweight classifiers to detect when an agent’s final outcome is already evident from its intermediate behavior and halting the run at that point. Tested on SWE-bench Verified, TerminalBench, and Toolathlon, the method eliminates 13-26% of agent steps and up to 44% of input tokens at 89-97% prediction accuracy, while shifting per-agent resolve rates by only one to two percentage points.
