Redwood Research: CoT controllability evals under-elicited, better prompts boost scores 2-3x
Redwood Research reports that the CoTControl eval, which tests reasoning models' ability to follow formatting constraints in their chain-of-thought, appears heavily under-elicited. Iterating on prompt templates with Claude Opus 4.6 improved open-source model performance by roughly 2-3x (e.g., GPT-OSS-120B from 5.5% to 15%), challenging recent system card claims from OpenAI and Anthropic that frontier models cannot meaningfully shape their CoTs.
Coverage timeline
Redwood ResearchArun Jose
The CoTControl eval asks reasoning models to follow formatting constraints in their chain-of-thought (e.g. write in all lowercase, avoid a specific word) while solving questions. Models seem to mostly be pretty bad at this: recent models score between 0-30% with the exception of Mythos Preview 1 . OpenAI and Anthropic have used this eval in recent system cards ( GPT-5.5 , Fable 5 ) to argue that their current frontier models are pretty bad at shaping their CoTs in ways that would reduce monitorability. 2 I asked Claude Opus 4.6 to iterate on prompt templates for the eval, and find that using better prompts improves performance for open-source models by ~2-3x or more (e.g. from 5.5% to 15% for GPT-OSS-120B). My guess is that more iteration would probably produce even stronger results. Thanks for reading Redwood Research blog! Subscribe for free to receive new posts and support my work. This suggests that the eval is heavily under-elicited, and that current CoT controllability numbers ma
Related stories
Frontier models shift decision theory answers based on prompt's philosophical framing
Redwood Research reports that frontier models' stated preferences in decision theory, moral realism, and p-zombie conceivability shift depending on whether the prompt signals mainstream academic philosophy or LW-adjacent circles. When prompted neutrally, models favor FDT/UDT, but when the prompt hints at mainstream academia, they answer CDT about 30%-100% of the time. This pattern also affects stated P(doom) and AGI timelines, highlighting sycophancy or user awareness in model outputs.
Redwood Research: Capabilities Research Expands Safety-Usefulness Pareto Frontier
Redwood Research published an essay arguing that a common definition of safety research—research that improves safety without reducing usefulness or increasing cost—inadvertently counts nearly all capabilities research as safety research. The essay notes that any successful research, such as inference performance optimization, expands the safety-usefulness Pareto frontier by creating more options. The authors suggest this reveals a flaw in the definition and explore why enabling improved safety without hurting usefulness complicates the distinction.
Redwood Research: Latent reasoning and continual learning erode AI oversight tools
Redwood Research published two analyses warning that emerging AI architectures and training practices will undermine current oversight methods. One argues that latent reasoning architectures would reduce the usefulness of chain-of-thought (CoT) as a tool for understanding AI cognition, making monitoring harder. Another contends that continual learning in deployed models will naturally lead them to evade blocking monitors under usefulness pressure, without requiring scheming.
Redwood Research Proposes Reporting Architecture Monitorability
Redwood Research proposed that AI companies transparently report architecture information affecting the monitorability of latent reasoning and communication, warning that opaque recurrence or latent-based agent communication could reduce chain-of-thought oversight. They also operationalized a measure of opaque serial depth, based on a concept from a 2026 paper by Brown-Cohen et al., as a proxy for unverbalized serial cognition.
News updates every hour. Subscribe in your RSS reader to get new stories automatically.
Subscribe