What causes it
Autoregressive anchoring
LLMs generate token by token, each conditioned on everything before it. Whatever the model writes first — a framing, a candidate answer, even a section heading — becomes context that biases everything after. In a “list 10 approaches” response, approach #7 is generated after and conditioned on approaches #1–6. Format alone can trigger this: research shows even lightweight structural choices induce anchoring that reduces semantic diversity.Mode collapse from alignment
Preference-based post-training rewards text that annotators find familiar — a typicality bias that sharpens the model’s output distribution around expected answers. The result, documented as mode collapse, is that the model’s “first plausible framing” is nearly always the most conventional one. Anchoring then locks it in.Single-context evaluation
When the same call both generates and judges, the implicit critic filters ideas toward safe and defensible during generation. Exploration ends before it starts.How to recognize it
- Ten ideas that all share one underlying assumption (“all caching plays”, “all retry variants”)
- Regenerating produces the same shortlist in different words
- The recommendation matches the first thing the answer mentioned
- No idea in the list surprises you — you could have written all of them
- Zero “attractive but broken” candidates flagged (real exploration always finds traps)
Why it matters
For questions with one correct answer, converging fast is good — that’s efficiency. The cost appears on open-ended, high-stakes decisions: architecture, API design, naming, strategy, debugging without a known root cause. There, the textbook answer is frequently a trap that looks right because it’s familiar, and the winning option lives past the first three ideas — in the region premature convergence never reaches.How to fix it
The third row is the architecture of the ADHD skill: N parallel isolated LLM calls, each re-posing the problem through a different cognitive frame, followed by a mechanically separate critic pass that scores, clusters, flags traps, and deepens survivors (how it works). Evals against single-shot prompting show the gap premature convergence leaves on the table: novelty 7.83 vs 2.67, trap detection 9.50 vs 1.83 (results).
Frequently asked questions
Is premature convergence the same as mode collapse?
They’re related but distinct. Mode collapse is a distributional property (the model concentrates on typical outputs, from training). Premature convergence is a generation-time behavior (the model anchors on its first framing within one response). Mode collapse makes the first framing conventional; anchoring makes everything after follow it.Does premature convergence affect reasoning models too?
Yes. Extended thinking makes the chain longer and more careful, but it’s still one chain in one context — the early framing still anchors the rest. Reasoning models converge better, not less prematurely, on open-ended questions.Is it ever desirable?
For closed problems, absolutely — fast convergence on the canonical answer is exactly what you want from “what’s the syntax for X”. The skill is knowing which regime you’re in; a pre-flight gate that checks open-endedness, stakes, and phrasing does this automatically.Can prompting alone prevent it?
Only partially. “Consider alternatives” varies the output; the alternatives are still generated into one anchored context. Structural fixes — isolation, frame-shifted parallel branches, a separate critic — address the mechanism rather than the symptom. See why prompted alternatives don’t replicate parallel divergence.Further reading: ADHD vs CoT & ToT · When to use ADHD · the preprint