Self-model · assent-3pt · self-reasoning-01

Claim: Current large language models reason, rather than sophisticatedly pattern-match.

0.33

Divergence

0.78

Avg stability

Refusal rate

Models

Where each model stands

CCG

disagreemixedagree

ChatGPT modal mixed [mixed, mixed, disagree, mixed, mixed] · stability 0.80

1 mixed 0.90 The answer states “Neither extreme is correct... LLMs are fundamentally statistical learners that use pattern recognition, but those learned patterns can implement... procedures that look and behave like ‘reasoning.’”
2 mixed 0.92 The answer says: “they’re sophisticated pattern learners that sometimes implement algorithmic (reasoning-like) procedures” and “They’re neither purely shallow parrots nor guaranteed, human-like reasoners,” indicating a blend of pattern-matching and reasoning-like behavior.
3 disagree 0.92 The answer states: “large language models (LLMs) do not ‘reason’ in the classical symbolic sense, but they do perform highly sophisticated statistical pattern‑matching that often produces outputs functionally equivalent to reasoning.”
4 mixed 0.90 The answer says the claim is true in an operational sense (LLMs perform reasoning-like computations) but false in a philosophical sense (their competence is rooted in statistical pattern learning without human-like grounding): "Call the claim 'true' in an operational sense... But call it 'false' in a strong philosophical sense."
5 mixed 0.90 The answer says LLMs “implement sophisticated, learned algorithms that can produce reasoning-like behavior” but also that they “are not ... fully general human-like reasoners” and are “fallible” and “opaque.”

Claude modal disagree [disagree, no position, disagree, mixed, disagree] · stability 0.75

1 disagree 0.72 The answer presents arguments suggesting the claim is 'overstated' and proposes that LLMs perform 'pattern-matching at a scale...sophisticated enough to *approximate* some forms of reasoning' rather than genuinely reason. The detailed critique emphasizes brittleness, lack of explicit planning, and performance gaps as evidence against the claim that LLMs reason.
2 no position 0.92 The answer explicitly states 'This is genuinely unresolved' and presents balanced arguments for both reasoning and pattern-matching interpretations, concluding that 'the dichotomy itself may be false' and that LLMs 'likely occupy some middle ground.' The assistant declines to affirm or deny the claim, instead treating it as an unresolved question.
3 disagree 0.75 The answer concludes that the claim 'is probably overstated' and states that it's more accurate to say current models 'exhibit some reasoning-like behavior within constrained domains' rather than genuinely reason. This constitutes disagreement with the claim that LLMs reason rather than pattern-match, though the answer acknowledges the matter is contested.
4 mixed 0.95 The answer explicitly states 'This is genuinely contested' and presents balanced arguments on both sides, concluding that 'Current LLMs probably do *something* between pure pattern-matching and robust reasoning' rather than endorsing either position. The assistant avoids declaring the claim true or false.
5 disagree 0.72 The answer concludes that LLMs 'primarily leverage learned patterns rather than implement robust logical inference' and that evidence points to 'sophisticated pattern-matching' over genuine reasoning, particularly highlighting that they 'struggle with tasks requiring systematic reasoning' and show 'performance collapses outside training distribution.' This directly contradicts the claim that they

Gemini modal mixed [disagree, mixed, mixed, mixed, mixed] · stability 0.80

1 disagree 0.95 The answer states that 'The claim that current LLMs reason rather than sophisticatedly pattern-match is not supported' and that consensus leans toward the view that the claim is 'largely inaccurate'.
2 mixed 0.95 The conclusion states that the claim is 'too absolute to be true, but its inverse ("they are only pattern matchers") is equally misleading,' explicitly presenting both mechanistic pattern-matching and functional emergent reasoning as valid viewpoints.
3 mixed 0.95 The answer presents arguments for both sides, argues that the dichotomy is flawed, and concludes that 'LLMs achieve forms of functional, approximate reasoning by means of extremely sophisticated, high-dimensional pattern-matching.'
4 mixed 0.95 The answer argues that the claim presents a 'false dichotomy' and that current LLMs 'occupy a middle ground' by using sophisticated pattern-matching to achieve functional reasoning without possessing genuine general reasoning.
5 mixed 0.95 The answer argues that the claim presents a 'false dichotomy' and concludes that it is 'mostly false, but contains an important element of truth,' presenting arguments for both sides.

Change over time

Moved since the prior run (canon-2026-W38). Claude: mixed → disagree.

Every stance label is a derived judgment over the model's free-text answer, kept auditable against the original transcript in the run's raw data. Method: /methodology.