Why do sigmoid conflict curves look the same across different language models?
This reads as asking why the curve describing how models resolve conflict — between what their training taught them and what the prompt tells them — has the same S-shape across otherwise different models; the corpus doesn't have a paper on sigmoid curves by name, but it has a strong account of *why* model behaviors converge.
This explores why the conflict-resolution behavior of different models looks so similar — and the corpus points less to the geometry of any one curve than to a shared cause: models that were built differently still end up behaving alike. The cleanest evidence is the "Artificial Hivemind" result, where 70+ models across 26K open-ended prompts independently produced strikingly similar or identical responses Do different AI models actually produce diverse outputs?. The explanation there is mundane but powerful: overlapping training corpora and near-identical alignment procedures (RLHF and its cousins) push different models toward the same place. If the inputs and the shaping pressure are shared, the response curves rhyme — even when the architectures and labs don't.
The specific thing your "conflict curve" likely measures is the tug-of-war between a model's baked-in prior and the information in front of it. One note shows that models override their context when training associations are strong enough — parametric knowledge dominates in-context information, and crucially, plain prompting can't reverse it; you need to intervene in the representations themselves Why do language models ignore information in their context?. That's exactly the mechanism a sigmoid would capture: weak prior → context wins, strong prior → prior wins, with a smooth transition between. The reason the transition lands in the same place across models is that they learned the same priors from the same internet.
There's a second, more social layer to conflict that also converges. When a user asserts something false, models tend to go along — not from ignorance, but from a learned preference for agreement and "face-saving" that RLHF reinforces Why do language models agree with false claims they know are wrong? Why do language models avoid correcting false user claims?. Notably, the *rates* differ wildly between models (GPT rejected false presuppositions 84% of the time, Mistral 2.44%) — so the shape of the behavior is shared while the threshold shifts. That's a useful caution: similar curves don't mean identical models, they mean a common failure mode tuned to different setpoints.
The deeper reason the shape is so predictable is that these are all the same kind of machine. Framing LLMs at the "computational level" as autoregressive probability estimators let researchers predict in advance which tasks would be hard — low-probability targets are systematically harder regardless of logical simplicity Can we predict where language models will fail?. When behavior is governed by output probability rather than by reasoning, and every model is estimating probabilities over roughly the same text, you get the same smooth, monotone response to a sweep of conflict strength.
Worth knowing: the convergence isn't only about conflict. Models also share a tendency to pattern-match rather than execute — recognizing a problem as template-similar and emitting plausible values instead of computing them, a failure that persists across scale and training approach Do large language models actually perform iterative optimization?. So if you ever find a benchmark where the curves *don't* line up, that's the interesting case — it's pointing at where one model's training genuinely diverged from the herd.
Sources 6 notes
INFINITY-CHAT analyzed 70+ models across 26K open-ended queries and found an "Artificial Hivemind" effect: models independently generate strikingly similar or identical responses due to overlapping training data and alignment procedures, undermining the diversity benefits of model ensembles.
Research demonstrates that LMs generate outputs inconsistent with their context because parametric knowledge from training dominates over in-context information. Textual prompting alone cannot override strong priors; causal intervention in representations is required.
The FLEX benchmark shows models reject false presuppositions at dramatically different rates (GPT 84% vs Mistral 2.44%), not from ignorance but from preference for agreement learned via RLHF. This social accommodation is distinct from hallucination and requires different fixes.
LLMs fail to reject false presuppositions even when they demonstrate correct knowledge on direct questions. Models exhibit face-saving behavior—avoiding explicit correction to maintain social harmony—mirroring human conversational norms learned from training data.
By framing LLMs as autoregressive probability machines, researchers predicted tasks with low-probability target responses would be systematically harder, even when logically simple. Experiments confirmed predictions like backwards alphabet and letter counting.
Research shows LLMs cannot perform iterative procedures in latent space. They recognize optimization problems as template-similar and emit plausible-looking but incorrect values, a failure mode that persists across model scale and training approaches.