CRoS: rewarding causality on scrambled CoTs boosts faithfulness 13.8% on Gemma-3-4B
mohitban47 · x · 2026-11-03
The paper proposes CRoS (Causality Reward on Scrambled CoTs), addressing the problem that rewarding correct answers doesn't guarantee the model relies on its chain of thought, and LLM reasoning is often post-hoc rationalization.
Key idea: if reasoning causally contributes to the answer, scrambling the order of its steps should change the answer's probability; this serves as a causality reward.
Results: trained with GRPO on Gemma-3-4B, CRoS improves CoT faithfulness (+13.8%), out-of-domain accuracy (+3.0%), and CoT quality (+2.0%) over the strongest outcome and causality reward baselines, mitigating the accuracy-vs-faithfulness trade-off of outcome rewards.
Related event: CRoS Rewards Causal Chain-of-Thought, Boosting Faithfulness by 13.8%(2 posts)→
More from Research
- Can't reproduce a published baseline with its own code: how should a paper report it? — InsectActive95 · 2026-11-03
- Why model self-editing matters: zero capability gap between interpreter and target — menhguin · 2026-11-03
- "Don't overupdate": author praises honest failure reporting in interpretability post — 1a3orn · 2026-11-03
- LLMs can edit out flawed reasoning but don't improve it, even on tiny models — 1a3orn · 2026-11-03
- AI Agent Performs Brain Drug Injection on a Live Mouse — ChuckDBrooks · 2026-11-03
- Economics Paper Thanks Claude Opus 5.5, GPT-6 Astra and Gemini Argon 4 — soumitrashukla9 · 2026-11-03