Different phrasings of the same idea seem to collapse into a single shared pattern inside a model's middle layers — why?

This explores why prompts that mean similar things seem to collapse toward shared internal representations partway through a transformer's layers. The collection doesn't have a paper that names 'attractor states' directly, but several notes triangulate the phenomenon from different angles — and together they suggest the convergence isn't mysterious so much as a consequence of how models turn surface wording into reusable internal patterns.

The strongest handle is the idea that reasoning lives in hidden-state trajectories, not the visible text. One line of work argues LLM reasoning should be studied as latent-state trajectory formation, where surface chain-of-thought is only a partial readout of what the middle layers are actually doing Where does LLM reasoning actually happen during generation?. If two differently-worded prompts share meaning, they get routed toward the same region of that trajectory space — which is exactly what an 'attractor' describes. Layer-lens analysis reinforces the timing: models compute their real answer in early-to-middle layers and only later reshape it into format-compliant surface tokens Do transformers hide reasoning before producing filler tokens?. The middle is where semantics consolidate before the model commits to output form.

Why consolidation, specifically? Two notes point at the same mechanism from opposite directions. Models fit instance-based patterns rather than general algorithms, so anything resembling a familiar instance snaps to the pattern learned for it — novelty, not complexity, is what breaks this Do language models fail at reasoning due to complexity or novelty?. And when the training signal has low variance, policies collapse toward input-agnostic templates, actively erasing prompt-specific differences Why do language models collapse into generic templates?. Both describe a basin: distinct inputs pulled toward a shared learned response. Attractor convergence is the healthy version of the same dynamic that, pushed too far, becomes template collapse.

There's a useful caveat about what converges. Whether a prompt's meaning aggregates into a stable internal state depends on the information actually flowing through the prompt first — saliency analysis shows chain-of-thought only works when question semantics propagate into the prompt structure before reasoning begins, and for simple questions a direct route works better Why do some questions perform better without step-by-step reasoning?. So convergence isn't automatic; it requires the semantic content to reach the layers where consolidation happens. And the ceiling on all of this: prompting only reorganizes and activates knowledge already in the model's distribution — it can't inject anything new Can prompt optimization teach models knowledge they lack?. That reframes the whole picture — the attractor a prompt falls into isn't created by the prompt, it's a stable configuration already carved into the weights that the prompt merely selects.

The thing worth carrying away: the same middle-layer basin that lets a model recognize 'these two prompts mean the same thing' is also the machinery that, under weak or uniform training pressure, flattens genuinely different prompts into one generic answer. Semantic convergence and semantic collapse are the same phenomenon at different intensities — which is why work on latent reasoning treats hidden-state dynamics, not the visible tokens, as the real object of study Can models reason without generating visible thinking tokens?.