RELAI-VCL compounds agent-optimizer gains via continual learning

A study on Terminal-Bench 2.0 evaluates whether agent-optimization gains compound over time by comparing GEPA, Meta Harness, and RELAI's Verifiable Continual Learning (RELAI-VCL). The research tests if optimized agents can improve on new tasks without eroding previous gains.

GEPA's optimized agent transferred below the unoptimized baseline when new tasks were introduced.
Meta Harness transferred well but failed to improve further with a second optimization budget.
RELAI-VCL was the only method to transfer positively and continue improving, reaching a 76.4% lifelong average pass rate compared to 66.0% for GEPA and 64.6% for Meta Harness.

The authors conclude that optimization gains compound only when regression control is built into the optimization loop to prevent shortcut solutions.

Benchmarks

Benchmark	Model	Score
Terminal-Bench	RELAI-VCL	76.4%
Terminal-Bench	GEPA	66%
Terminal-Bench	Meta Harness	64.6%
Terminal-Bench	baseline agent	58.7%

Benchmark

Model

Score

Terminal-Bench

RELAI-VCL

76.4%

Terminal-Bench

GEPA

66%

Terminal-Bench

Meta Harness

64.6%

Terminal-Bench

baseline agent

58.7%