Tencent's Terminal-RSI raises terminal agent Terminal-Bench average from 20.68% to 50.58%

Tencent researchers present Terminal-RSI, a training method that starts from Qwen3.8-27B and alternates two phases over five stages, according to Elvis Saravia (@omarsar0), who describes the paper on X. One phase fine-tunes on rewritten trajectories that pass a verifier; the other runs PPO with an inherited critic.

The Terminal-Bench four-suite average rises from 20.68% to 50.58%, with 86.52% on Terminal-Bench 2.1 and 79.00% on Terminal-Bench Hard, per the paper as Saravia reports it.

The paper's reward audit, as reported, finds that strict success scoring reduces failed runs' share of reward from 50.09% to 28.72%.