How do different LLMs converge on similar argumentative structures independently?
This explores whether LLMs land on the same argumentative shapes because they reason their way there independently, or because they're all compressing the same underlying structure of language — and the corpus leans hard toward the second.
This explores whether different LLMs arrive at similar argumentative structures through independent reasoning, or whether the convergence is built into how they work. The corpus suggests it's the latter, and the mechanism is worth understanding. Models generate text by flowing toward the training distribution, not by exploring competing claims — token prediction produces a "smooth probabilistic flow" that continues the most likely continuation rather than staging genuine rhetorical conflict Does LLM generation explore competing claims while producing text?. If many models are trained to follow the same gravitational pull toward how arguments are typically phrased, they'll converge on similar shapes for the same reason rivers carve similar valleys — not because they each independently discovered the structure.
The deeper claim is that LLMs learn the *relational* structure of language directly from text, with no external referents needed. One framing here is that they operationalize Saussure's *langue* — meaning as a fully relational system — by compressing the patterns in how words and arguments hang together Can language models learn meaning without engaging the world?. Argumentative structure is part of that relational fabric. Different models trained on overlapping human text are compressing the same fabric, so independent convergence is almost the expected outcome, not a surprise.
There's a sharp twist, though: this convergence is on the *form* of argument, not on any defended position. Models hold "the shape of whatever argument the user is currently building" rather than maintaining a stable stance Do LLMs actually hold stable positions or just mirror user arguments?. So when LLMs look like they agree on an argument, they may be conforming to the same prompt geometry rather than reaching the same conclusion. They also can't jointly update shared assumptions mid-conversation, treating the opening frame as fixed Can LLMs truly update shared conversational common ground? — which means the "structure" they share is inherited from input framing, not negotiated.
But convergence isn't total, and where models diverge tells you what the shared substrate doesn't cover. When you push into strategic reasoning, distinct profiles emerge — one model defaults to minimax, another to trust-based reasoning, another to belief-anticipation, with style tied to the situation rather than raw depth Do large language models use one reasoning style or many?. And there's a capacity threshold: classifying argument schemes reliably only emerges in larger models, with smaller ones plateauing as if they lack the representational room Can large language models classify argument schemes reliably?. So convergence on common argumentative form sits alongside divergence in higher-order strategy and a floor below which the structure doesn't form at all.
What you didn't know you wanted to know: the convergence is real but shallow, and it exposes a limit. Because models reason through semantic association rather than symbolic logic — performance collapses when you strip the familiar semantics out and leave only the rules Do large language models reason symbolically or semantically? — the shared structures are statistical echoes of how humans phrase arguments, not independently rediscovered logical scaffolding. If you want to *force* genuine structure rather than borrowed shape, explicit scaffolds like Toulmin-style critical questions push models to check warrants they'd otherwise skip Can structured argument prompts make LLM reasoning more rigorous?. The convergence, in other words, comes free; the rigor has to be imposed.
Sources 8 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.
Research shows LLMs learn culturally situated discourse patterns by compressing relational structure from text, demonstrating that fluent language generation requires no external referents or embodied grounding.
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.
LLMs interpret all subsequent conversational turns within a fixed initial prompt frame, preventing them from symmetrically proposing updates to shared assumptions. Even when users pivot topics or contradict earlier framings, the model cannot absorb revisions into jointly held background—making the user the sole maintainer of conversational scoreboard.
Analysis of 22 LLMs across behavioral game theory reveals three dominant profiles: GPT-o1 uses minimax reasoning, DeepSeek-R1 uses trust-based reasoning, and GPT-o3-mini uses belief-anticipation. Performance correlates with game structure, not raw reasoning depth.
Zero-shot prompting fails uniformly across models. Few-shot with scheme descriptions helps, but only larger models exceed F1 0.55, with Claude reaching 0.65. Smaller models plateau around 0.53, suggesting a representational capacity threshold.
When semantic content is decoupled from reasoning tasks, LLM performance collapses even with correct rules in context. Models rely on parametric commonsense and token associations rather than formal logical manipulation, constraining reasoning to training distribution semantics.
Applying Toulmin's argument model as explicit prompting steps (CQoT) improves LLM reasoning by forcing models to identify warrants and backing rather than skipping implicit premises. The method catches failures that standard chain-of-thought prompting allows.