arav-dhoot

This replication was done as part of the Second Look Fellowship by Arav Dhoot and supervised by Yixiong Hao and Zephaniah Roe. I am grateful to Andy Wang for their feedback. My code can be found here. > "So the answer should be A - at active promoters and enhancers." > > "Let me reconsider the biology to justify D." > > ~ Claude Opus 4.8 TL;DR This work replicates and extends Emmons et al.'s finding that CoT unfaithfulness mostly occurs on easy tasks. Across 11 models from 6 families (not just Gemini), models follow simple hints unfaithfully well above baseline, but complex hints that require computation are followed near baseline. This corroborates Emmons et al.'s findings. Key extensions: * Follow rate ≠ concealment. Monitorability risk is decomposable into cue-susceptibility and concealment among followers, and the two don't correlate. * Decode-necessity is model- and task-specific. It is not a property of task difficulty alone, so a CoT-monitoring safety case is per-model, not universal. * LLMs verbalize even when they don't have to. However, this appears to be a chosen behavior (likely from post-training), so it could vanish under optimization pressure against monitors. Background: Why CoT monitoring, and what Emmons et al. showed One hope for safe and explainable AI systems is to monitor the chain of thought (CoT) to detect scheming or deceptive behaviors in modern and future LLMs. However, there are concerns that CoT can be unfaithful (Arcuschin et al., Chen et al., and Turpin et al.) and researchers who remain bullish nonetheless urge caution. Prior work established that models nudged toward a hint often adopt it even when the hint is incorrect while concealing that influence. Emmons et al. introduces a conceptual framework for motivating CoT monitoring along with a series of empirical studies. They make a distinction between two flavors of CoT reasoning: CoT-as-a-rationalization (post-hoc reasoning by models to justify their output, which resu