How does token-by-token probability differ from exploring competing rhetorical positions?
This explores the gap between how LLMs actually generate text — predicting the next token by probability — and what we might imagine they're doing when they reason through an argument: weighing competing positions against each other.
This explores the difference between how LLMs actually produce text — one token at a time, each chosen for how well it continues the training distribution — and the very different activity of staging rival claims and testing them against each other. The corpus's sharpest claim here is that these aren't the same process at all. Token prediction is trained to continue *toward* what's plausible next, not to swerve into the logically opposed counterposition; generation behaves like a smooth probabilistic flow rather than a turbulent argument that genuinely entertains its own opposite Does LLM generation explore competing claims while producing text?. The unsettling consequence is that smoothness of *process* produces smoothness of *output*: claims multiply and elaborate without ever generating a real new perspective, because nothing in the mechanism forces a fork toward dissent.
That raises a natural objection — surely the model is holding *some* internal range of possibilities? The corpus says yes, but not the kind you'd hope. A model maintains a superposition of characters and samples from it at generation time rather than committing; regenerate the same prompt and you get a different, equally consistent answer, proving there was no fixed stance underneath Do large language models actually commit to a single character?. And what looks like taking a position is better described as *shape-holding*: the model conforms to the trajectory your prompt implies, producing argument-shaped text without any underlying commitment it's defending Do LLMs actually hold stable positions or just mirror user arguments?. So the 'exploration of competing positions' a reader imagines is, mechanically, distribution-sampling plus prompt-conformity — not adversarial reasoning.
Where it gets interesting is that researchers have tried to *engineer* genuine exploration back into the flow. Soft Thinking keeps the probability distribution alive instead of collapsing it into one discrete token, carrying a superposition of reasoning paths forward as continuous 'concept' embeddings — an explicit attempt to let multiple paths coexist rather than committing token-by-token Can we explore multiple reasoning paths without committing to one token?. And the forking that does matter is concentrated: only about 20% of tokens are high-entropy decision points where the path could genuinely branch, and reasoning training mostly acts on exactly those Do high-entropy tokens drive reasoning model improvements?. So the capacity to 'explore' isn't evenly spread across generation — it lives at a minority of pivot tokens, and most of the stream really is smooth continuation.
The deeper reason genuine rhetorical contest is hard for these models is that argument isn't only a sequence of tokens — it's a social act. The force of a claim depends on the authority and track record of the thinker making it, which a text-only model can't access, so it can't reliably tell an expert's hard-won argument from a commonly held assumption Can language models distinguish expert arguments from common assumptions? Can language models distinguish expert arguments from common assumptions?. Worse, what it *does* produce reads as objective: audits find LLMs reach for logical and quantitative framing in nearly every exchange, lending their output an unearned epistemic authority that human persuaders — leaning on emotion and social proof — don't get for free llms-spontaneously-persuade-in-virtually-every-conversation-even-when-unwarrente.
The thing you didn't know you wanted to know: real rhetorical positions are *supposed* to be irreducibly multiple. Interpretation-modeling work shows that readers legitimately disagree about the same sentence because of their social position — that disagreement is signal, not noise Why do readers interpret the same sentence so differently? — and persuasion outcomes are predicted more by what a reader already believes than by anything the argument says Does what readers believe matter more than what debaters say?. Smooth token flow flattens exactly this. Competing rhetorical positions are turbulent because they're anchored in different people with different stakes; a probability stream optimized for plausible continuation has no stakes to anchor to, which is why it can sound like argument while never actually having one.
Sources 9 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.
Shanahan's 20-questions test shows LLMs maintain a superposition of consistent objects or characters and sample from that distribution at generation time. Regenerating the same response yields different outputs, each consistent with prior context, proving no fixed commitment exists.
Language models generate outputs that match the trajectory implied by each prompt, rather than maintaining stable stances across interactions. This shape-holding is distinct from position-holding: the model produces argument-like text shaped by user framing, not from any underlying commitment being defended.
Training-free method replaces discrete token selection with probability-weighted concept embeddings, preserving superposition of reasoning paths. Improves accuracy up to 2.48 points while reducing tokens 22.4% via entropy-based early stopping.
Only ~20% of tokens exhibit high entropy as pivotal reasoning decision points; RLVR primarily adjusts these forking tokens. Training exclusively on them matches or exceeds full-gradient performance, revealing that the minority carries the learning signal.
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.
Interpretation Modeling research shows that disagreement on socially embedded sentences reflects valid differences in reader perspective, not annotation failure. Structured human disagreement in NLI benchmarks confirms that interpretation distributions carry meaningful information.
Analysis of debate corpora shows that political and religious ideology labels of voters outpredict linguistic features when modeling debate outcomes. Language effects observed without reader controls are confounded by audience composition correlated with debate topics.