Do language models sparsify their activations under difficult tasks?

A robust and quantifiable phenomenon documented across diverse models and domains: as task difficulty increases — whether through harder reasoning questions, longer contexts, or simply adding answer choices — the last hidden states of LLMs become substantially sparser. The "farther the shift, sparser the representation" is the title and the central claim, and the controlled analyses in the paper show the sparsification is not incidental.

What is sparsity here? A high-dimensional representation dominated by a small subset of active units. When an LLM is comfortable with the input — well within its training distribution, easy task, short context — its activations spread broadly. When the model is pushed toward OOD — unfamiliar concepts, longer reasoning chains, harder questions — those activations concentrate into a smaller specialized subspace. The sparsification is localized in the final transformer layers, behaving like a selective filter that stabilizes reasoning under uncertainty.

This reframes a long-standing question in interpretability. Sparsity has been studied as a static background property of LLMs and as evidence for modularity or specialization. The new finding is that sparsity also operates as an explanatory variable — it changes systematically with task conditions and predicts behavior under difficulty. Models that sparsify more aggressively under OOD shift have a different operational regime than models that maintain dense activation.

The mechanism the paper proposes is adaptive. Under unfamiliar inputs the network cannot rely on the dense, contextually-distributed representations it learned for in-distribution data. Concentrating computation into a smaller specialized subspace gives it a workable signal where dense averaging would dissolve into noise. The sparsity is a defense mechanism, not a failure mode.

For interpretability, this argues for sparsity-aware probing. Methods that assume stationary representational density miss what happens at the boundary where models actually fail. For methodology, it suggests using activation sparsity as a difficulty signal — a sparser response is evidence the model is operating near or beyond its competence.