MINT: meeting-time truncation for multi-step off-policy RL
The key implication is that backup-horizon selection in off-policy RL can be handled automatically by a coupling-time stopping rule instead of hand-tuned trace lengths. RL researchers and practitioners tuning n-step returns or V-trace-style coefficients get a variance-controlled alternative with a single-sample surrogate and no mixing-time estimates. The breadth of benchmarks (MuJoCo, DMC, Atari 100k, RWRL, OGBench) makes the evidence easier to trust than single-domain papers.