LLM sparsity explains inconsistent coding performance

This paper reframes LLM inconsistency not as random hallucination but as a measurable, structural response to out-of-distribution inputs — which has real implications for how we debug and improve model behavior.

• –The core finding: as inputs move further OOD (harder reasoning, longer context, more choices), the last hidden states concentrate into sparser subspaces — the model is essentially "narrowing focus" under stress
• –This explains the senior-engineer-one-day, syntax-error-the-next phenomenon developers routinely observe — it's not random, it correlates with input difficulty
• –SG-ICL uses sparsity scores to rank and sequence few-shot examples in context, giving a practical handle on a previously opaque failure mode
• –Sparsity scales across multiple difficulty axes: reasoning complexity, context length, and choice count — suggesting it's a general mechanism, not task-specific
• –Opens the door to sparsity-based runtime monitors: detect when a model is about to fail before it does