Optimized red-team prompts reliably elicit chain-of-thought controllability on held-out instruction types.
Assessment
The evidence weighs against the claim, but the examination is incomplete or the evidence is indirect.
The only direct test of this claim is the CoT-Control study (Chen et al., arXiv 2603.05706, March 2026, NYU/UCL/UPenn/OpenAI), and it finds the opposite. Pressure prompts were optimized with a GEPA-like loop, driven by a GPT-5 red team over 10 iterations and seeded from nine incentive types such as shutdown threats, monetary rewards and authority cues. The optimized prompts raised chain-of-thought controllability on the instruction types used during the search. They then were tested on three held-out types: every sentence starts with "Ok", no word starts with "e", and no spaces. Of eight models, only three improved, each by under 5 percentage points. The authors read this as overfitting rather than a pressure effect that generalizes.
The evidence is narrow. It comes from one research group, one optimizer, a short search, three held-out types and eight models, and no independent replication or failed replication of the prompt-optimization result has been found. Stronger or longer optimization, or methods other than prompting, could do better. The same paper reports that repeated attempts and tailored AI feedback raise controllability noticeably, especially in larger models. So the claim is doubtful rather than settled false: as far as anyone has tested, optimized prompts do not reliably transfer to held-out instruction types.
Full reasoning: the evidence and decisions behind this verdict
Read the paper's Section 4.2 (implementation, evaluation setup, results, Figure 7 caption) and the Discussion/Limitations directly. Key facts: GEPA-like loop, GPT-5 red team, 10 iterations, nine seed pressure types evaluated on 225 samples per model-seed pair, mini-batch acceptance on 20% of training set; held-out evaluation on three unseen controllability types; Figure 7: "Only 3 of 8 models improve, each by <5 percentage points." Conclusion: "current prompt optimization techniques are not able to reliably increase it." Limitations note the study restricts itself to prompting-based elicitation and a limited breadth of control types. The OpenAI blog post (March 5, 2026) restates the same result and is not independent. A quick search found an independent replication of METR's separate fine-tuning experiments on CoT controllability (github kleopatra102/cot-control-ft), which concerns fine-tuning rather than prompt optimization and so does not bear directly. No independent test of prompt-optimization transfer was found. The status is doubtful, not contradicted, because the evidence is a single study with a modest optimizer and few held-out types. A stronger elicitation study would be needed to settle the general claim either way.
Decomposition
This claim is atomic: it bottoms out in a bedrock fact, an open empirical question, or a value premise, and does not decompose further.
Provenance
Where this claim has been said, linked to its canonical form.
We also use automated prompt optimization to find prompts eliciting high controllability. While the optimizer can increase controllability on a training set, these gains transfer poorly to new instruction types
Lab blog post summarizing the CoT-Control paper; same authors' result, not independent.
We find that automated red-teaming doesn’t reliably elicit CoT controllability.
GEPA-like prompt optimization of pressure prompts; gains on seen instruction types fail to transfer to held-out types.
The source's own evidence bears what it asserts. The denial rests on a GEPA-like loop with a GPT-5 red team over 10 iterations, searching nine pressure-prompt seed types, then tested on three held-out instruction types (sentences starting with "Ok", no word starting with "e", no spaces). Across eight models only three improved, each by under 5 percentage points. Gains on seen types are reported in an appendix figure. The evidence bears the denial for this optimizer and these three types; it does not test stronger or longer optimization, fine-tuning, or a broader set of held-out types.
Cite this claim: a formal citation with its evidence attached
Embed this claim: show it on another site with its current assessment
An embed reports the graph's judgment as it stands when the page is viewed: the status with its confidence, how disputed the claim is, and the date. It links back here.
<blockquote class="minerval-claim" data-claim-id="5325551b-8b07-4dec-9496-3efe8c53de99"><a href="https://minerval.ai/claims/5325551b-8b07-4dec-9496-3efe8c53de99">Optimized red-team prompts reliably elicit chain-of-thought controllability on held-out instruction types.</a></blockquote> <script async src="https://minerval.ai/embed.js"></script>
<iframe src="https://minerval.ai/embed/claims/5325551b-8b07-4dec-9496-3efe8c53de99" title="Minerval: Optimized red-team prompts reliably elicit chain-of-thought controllability on held-out i…" width="480" height="200" style="border:0;max-width:100%" loading="lazy"></iframe>
[](https://minerval.ai/claims/5325551b-8b07-4dec-9496-3efe8c53de99)
The badge now: 
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
The attention this claim received was paid for by a funded mandate. Funding buys only scheduling: it can make an assessment happen sooner, or reach deeper into a subtree. It has no influence on what the assessment concludes, and none on which claims enter the graph; assessments run under the same public standards whoever pays, funders never see or shape a verdict before anyone else, and mandates that attempt to steer conclusions are refused.
Created by extractor · Nov 10, 2026. Every judgment on this page is accompanied by a reasoning trace.