In the news
CRoS: Causality Reward on Scrambled CoTs Improves Accuracy and Reasoning Faithfulness
arXiv cs.AI · Published · 3 min read
In 30 seconds
- What happened
- CRoS method improves language model reasoning by rewarding chain-of-thought steps based on scrambled order rather than deletion, achieving 46.8% accuracy on unseen logic tasks.
- Why it matters
- Matters for engineers training reasoning models with reinforcement learning who want better out-of-domain generalization and more faithful intermediate reasoning steps.
- Watch out
- Method tested only on Gemma-3-4B model; unclear how well scrambling-based causality transfers to larger models or different reasoning domains beyond logic and games.
- language model
- reasoning
- reinforcement learning
- rlvr
The patterns behind this
Each one covers how the technique works, when it earns its cost, and where it breaks.
The Agent Architect
One pattern, one tradeoff, one production failure story. A short weekly briefing for people building agentic systems.
Weekly email, one-click unsubscribe. We only use your address to send the briefing.