📢 Excited to share our latest work from @SFResearch: Learning Generalizable Behaviors for Terminal Agents 🤔 Q1: How does RL actually help terminal agents generalize? → Not by teaching new skills. It reshapes behavior. • Skills (write the regex, call the tool) are already there from pre-training + SFT • Behaviors (inspect first, verify, stop looping) are what RL shapes, and what predicts success 🔍 Q2: Can you trust your synthetic RL environments? → Often not. The reward can be wrong in both directions. • The cleanest public collection we audited: 35.8% clean • Reward 1 for copying a leaked answer. Reward 0 for a correct solution 🌊 So we built RIVER → Filter defective envs, penalize repetitive loops, then run RL. • River-8B: best among evaluated open RL-trained 8B models on all 4 terminal benchmarks • 2B→27B with <30% of the envs: RL gains +106% on Terminal-Bench-Lite, +30% on Terminal-Bench-v2.1 • Same 3.5K-env budget: filtered 19.4 vs. randomly sampled 17.7 ✅ Takeaway: audit your verifiers before you scale your envs. 🙌 Led by @YYYao45 and @bo_pang0 , with @xuanphinguyen , @zhao__ding and @JotyShafiq 📄 Paper: arxiv.org/abs/2608.22631 🌐 Website: self-evolving-agents.salesfo… 🧵👇