More capable language models can perform more complex tasks without externalized chain-of-thought reasoning.
5 events · 2 assessments · 2 decisions
Reassessed no status change
The subclaim's first assessment (supported, 0.85) raises doubts about rate and steadiness, not direction. The parent claim asserts only direction, and independent direct-answer evidence also supports it. The change is not material: the status stays verified at 0.85, now with credence 0.95. The argument was re-evaluated (likely), and importance was set to 0.45 with contestation 0.3. Dependents were not notified because the status did not change.
Reassessed: still Verified
verdict confidence 0.85 · credence 0.95
Structured and assessed
Added one argument from measured no-CoT time horizons, with a new subclaim on the no-CoT horizon trend (seeded at 0.9). Linked the related serial-reasoning claim. Recorded the Redwood/Ward et al. 2026 instance. Assessed verified at 0.85 confidence, credence 0.94. Set importance to 0.5 and contestation to 0.3. The parent is already supported and this verdict agrees with the seed's direction, so no notification is needed.
Assessed Verified
verdict confidence 0.85 · credence 0.94
Direct measurement establishes the claim as a general trend. When successive frontier models are tested with reasoning tokens suppressed, the difficulty of the tasks they can complete rises with capability. A 2026 study by Ward, Gould, Woodruff and colleagues tested 14 models, from GPT-2 to GPT-5.5, on 43 benchmarks. It found that the no-chain-of-thought time horizon has grown steadily, doubling about once a year (95% interval roughly 6 to 23 months). GPT-5.5 reaches about three minutes of human-equivalent work at 50% reliability. Ryan Greenblatt's earlier single-forward-pass math measurements and METR's 2026 frontier risk report point the same way. The disputed part is the rate and the reach of the trend, not its direction. Since GPT-4, capability measured with chain of thought has grown about twice as fast as capability without it, because most recent gains come from externalized reasoning. Current no-CoT horizons remain short: a few minutes of human-equivalent work, and METR found no public model above about ten minutes on no-CoT math in early 2026. Work on latent multi-hop reasoning also shows large gaps between what models can compose internally and what they can do with explicit reasoning. So the claim holds as stated, but it does not show that more capable models can already carry out long or elaborate planning without a visible trace. The measured horizons are lower bounds, since forced short answers may under-elicit reasoning models.
Claim entered the graph