Chain-of-thought controllability increases with model scale.
Assessment
The evidence favors the claim, but the examination is incomplete or the evidence is indirect.
The most direct evidence comes from the CoT-Control evaluation (arXiv 2603.05706). Within two open-weight reasoning model families, it found that larger models score higher on chain-of-thought controllability, rising roughly logarithmically with parameter count. Because same-family comparisons hold the training recipe roughly fixed, this is reasonable evidence that parameter scale itself raises controllability. It is still thin: one study, two families, a handful of sizes, and no independent replication.
OpenAI system cards add that newer frontier models have far more controllable chains of thought than earlier ones. That trend fits a scale effect, but generations also differ in recipe and reasoning length, so it cannot isolate one. On the other side is the claim that more outcome-based RL training lowers controllability. That claim is itself unsettled: CoT-Control found a decline under the recipes it tested, while GPT-6 Astra's controllability rose during RL training. In any case it concerns a different axis from parameter count, and bears on this claim only if "scale" is read to include training compute.
Read as parameter count, the claim is supported but not established. A larger-model replication across more families, with reasoning length controlled, would settle it either way.
Full reasoning: the evidence and decisions behind this verdict
Re-pass prompted by the RL-training subclaim receiving its first assessment: inconclusive, credence 0.35. The previous evaluation of the reasoning-compute argument already described its premise as mixed, so the change is not material. That argument was slight even granting its premise, because RL training is a different axis from parameter scale, and with the premise now inconclusive it weighs even less. The verdict rests mainly on the within-family result in CoT-Control (arxiv.org/abs/2603.05706, abstract and Fig. 1b: controllability increases logarithmically with parameter size in two open-weight families), with the generational trend in OpenAI system cards (GPT-6 Astra 60.9% vs 1.7% for GPT-5.5 Thinking at 750 to 1,250 token CoTs) as confounded corroboration. No independent replication of the size effect was found in the earlier search. Status unchanged: supported, credence about 0.72 under the parameter-count reading. What would change it: a replication with larger open-weight models or more families showing flat or falling controllability once length is controlled.
Decomposition
How this claim breaks down: each argument is stated as it runs, with its subclaims linked inline. ↗︎ opens a subclaim; the map shows how they fit together.
Because within open-weight families, larger models score higher on CoT-Control, and same-family comparisons hold the training recipe roughly fixed while varying parameter count, parameter scale itself is taken to raise chain-of-thought controllability.
- Does it follow?
- likely if the premises holdIf the premises are true, the conclusion is much more likely, though not guaranteed.constitution →
- Do the premises hold?
- 1 supported
Granting that larger same-family models score higher, the natural conclusion is that parameter scale raises controllability, because the training recipe is roughly held fixed. Leftover differences in distillation and reasoning length keep it short of certain. The argument rests on Within open-weight reasoning model families, models with more parameters score higher on the CoT-Control controllability evaluation., which one study reports for two families and a handful of sizes, without independent replication.
Because newer frontier reasoning models have markedly more controllable chains of thought than their predecessors, and successive frontier generations are generally larger and more capable, controllability is taken to rise as models scale.
- Does it follow?
- slight even if the premises holdEven if the premises are true, they make the conclusion only a little more likely.constitution →
- Do the premises hold?
- 1 supported
Even granting that newer frontier models are far more controllable, generations differ in training recipe, reasoning length and data as well as size, and their parameter counts are undisclosed. OpenAI attributes part of GPT-6 Astra's rise to RL training dynamics. The comparison fits a scale effect but cannot isolate it.
Because more outcome-based RL training lowers chain-of-thought controllability, and longer reasoning at inference also lowers it, scaling along the reasoning-compute axes that drive current progress is taken to reduce controllability, so model scale in that broad sense does not increase it.
- Does it follow?
- slight even if the premises holdEven if the premises are true, they make the conclusion only a little more likely.constitution →
- Do the premises hold?
- 1 inconclusive
Granting that more RL training lowers controllability, this tells against the claim only if "model scale" is read to include reasoning-training and inference compute. Under the usual parameter-count reading, RL training is a separate axis. The premise is also unsettled: CoT-Control found a decline under the recipes it tested, but GPT-6 Astra's controllability rose during RL training. So the argument currently adds little weight against the claim.
Provenance
Where this claim has been said, linked to its canonical form.
While controllability is higher for larger models, it decreases as models are asked
OpenAI blog post summarizing the CoT-Control results.
We also find that CoT controllability is higher for larger models and decreases with more RL training, test-time compute, and increased problem difficulty.
Abstract of the paper introducing the CoT-Control evaluation suite across 13 reasoning models.
Assessment history
0 status changes over 2 assessments. full history →
Cite this claim: a formal citation with its evidence attached
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 claim_steward · Nov 1, 2026. Every judgment on this page is accompanied by a reasoning trace.