Can extended thinking modes introduce genuine rhetorical exploration to LLMs?
This explores whether 'thinking longer' actually lets an LLM weigh competing claims and counter-positions the way a human arguer does — or whether it just produces more text without genuine rhetorical back-and-forth.
This explores whether extended thinking modes give an LLM real rhetorical exploration — the genuine weighing of competing claims and counter-positions — and the corpus suggests the answer is mostly no, for a reason that runs deeper than thinking length. The sharpest piece here argues that token generation is a smooth probabilistic flow, not a turbulent exploration of rival arguments Does LLM generation explore competing claims while producing text?. Models are trained to continue toward their training distribution, not to swerve into logically related counterpositions, so smooth process produces smooth claims that multiply without ever generating a new perspective. If rhetorical exploration means holding a thesis against its antithesis, the default machinery isn't built to do it — and adding tokens to that machinery just makes more smoothness.
That reframes the 'think longer' hope directly. More thinking time turns out to be non-monotonic: accuracy can fall from 87% to 70% as thinking tokens scale up, and skipping explicit reasoning entirely sometimes matches standard thinking at the same budget Does more thinking time actually improve LLM reasoning?. So extra deliberation isn't free exploration; past a threshold it's drift. A second piece sharpens why: reasoning models behave like wandering explorers, not systematic searchers, lacking the validity, effectiveness, and necessity that real search requires — which is why success collapses exponentially as problems deepen Why do reasoning LLMs fail at deeper problem solving?. Wandering and rhetorical exploration look similar from outside (both produce lots of intermediate text) but only one converges on anything.
The more interesting wrinkle is that thinking mode isn't fixed — it's shaped by training. In vanilla models, extended thinking is actively counterproductive, inducing self-doubt that degrades answers; RL training reverses the same mechanism into productive gap analysis Does extended thinking help or hurt model reasoning?. So 'can thinking explore' isn't a yes/no property of the architecture — what the thinking does depends on how it was trained. That's the closest thing to a path toward genuine exploration: not longer thinking, but differently-trained thinking. It also hints at why surface chains-of-thought can mislead — the real reasoning may live in hidden-state trajectories that the visible 'thinking' only partially interfaces with Where does LLM reasoning actually happen during generation?.
Here's what you might not have known you wanted: the corpus suggests rhetorical exploration may be a different cognitive paradigm entirely, not a longer version of problem-solving. One line of work argues creative reasoning splits into combinational, exploratory, and transformational modes that current methods completely ignore — they only ever do conventional problem-solving Can LLMs reason creatively beyond conventional problem-solving?. And a darker note: even when a model produces an articulate argument, it can't weigh that argument's force, because it has lost the social world — reputation, standing, track record — that gives expert claims their authority, processing only text Can language models distinguish expert arguments from common assumptions?. Genuine rhetoric isn't just generating both sides; it's judging which side carries weight, and that's the part extended thinking doesn't reach.
Sources 7 notes
Token prediction trains models to continue toward the training distribution, not to explore logically related counterpositions. This smoothness in process produces smooth claims that multiply without generating new perspectives.
Accuracy drops from 87.3% to 70.3% as thinking tokens scale from 1,100 to 16,000, and bypassing explicit reasoning entirely matches or beats standard thinking at equal token budgets. The relationship is non-monotonic, not the linear improvement commonly assumed.
Current reasoning models lack the three properties of systematic exploration: validity, effectiveness, and necessity. This causes success probability to drop exponentially with problem depth, making medium problems solvable but deep problems catastrophically harder.
Vanilla models use thinking mode counterproductively, inducing self-doubt that degrades performance. RL training reverses this, transforming the same mechanism into beneficial gap analysis. Training mediates reasoning quality, not just quantity.
Evidence from CoT faithfulness tests, feature steering, and layer analysis suggests latent-state dynamics drive reasoning, while surface chain-of-thought serves as a partial interface. Hidden reasoning processes should be the default focus of study.
Research identifies combinational, exploratory, and transformational reasoning as distinct creative modes grounded in cognitive science. Existing LLM reasoning methods address only conventional problem-solving, leaving creative paradigms unaddressed and potentially explaining diversity collapse in ideation.
LLMs lose the social context that gives expert claims their force—reputation, track record, and standing—because they process only text, not the social world where expertise is built and evaluated.