Can mid-tier models benefit more from self-generated harness updates than others?

This explores whether mid-tier models benefit more than weak or strong ones when they generate their own "harness" updates — edits to the instructions and scaffolding that steer how they work. The most direct finding splits the question in two: the *capacity* to write useful harness edits turns out to be flat across tiers — even smaller models produce comparable improvements — but the *ability to act on* those edits peaks in the middle Do stronger models always evolve their own harnesses better?. Weak models struggle to activate and follow updated instructions; strong models, oddly, also underperform at it. So the benefit is non-monotonic, and the bottleneck isn't authorship, it's uptake.

Why would the gains concentrate mid-tier rather than rising with model strength? The broader corpus on self-improvement gives a frame: no model can lift itself purely by its own bootstraps. Self-improvement is formally bounded by the generation–verification gap — it's easier to produce a change than to confirm it's actually better — so every reliable improvement smuggles in an external anchor: a past model version, a third-party judge, a user correction, or tool feedback Can models reliably improve themselves without external feedback? What stops large language models from improving themselves? What actually constrains large language models from self-improvement?. A harness update is exactly such an externalized anchor — instructions written down and fed back in. That reframes the mid-tier sweet spot: weak models can't reliably execute against the external scaffold, and strong models may already encode enough of it internally that an explicit re-instruction adds little or even competes with their priors.

A neighboring result on teacher-refined data sharpens the same point from a different angle. Higher-quality refinements don't help uniformly — they *degrade* a student that's past its learning frontier, and students do best when they filter improvements to keep only what's compatible with their own profile Does teacher-refined data always improve student model performance?. Harness updates are self-authored refinements, and the lesson transfers: the value of an update depends on the gap between where the model is and what the update demands. Mid-tier models sit in the band where self-generated guidance is both followable and still informative.

There's also a failure mode lurking that explains why strong models don't simply win. Models degrade sharply when their own prior errors or low-quality material accumulate in context — performance falls non-linearly, and scaling doesn't fix it; only test-time "thinking" compute reduces the contamination Do models fail worse when their own errors fill the context?. A self-updated harness is a bet on the model's own output quality. For models that can't cleanly distinguish a good update from a self-flattering one, the loop risks reinforcing noise — the same circularity that makes pure self-improvement stall.

The thing worth carrying away: "better model, better self-improvement" is the wrong intuition here. Writing useful scaffolding is broadly available across tiers, but *benefiting* from it is a separate skill that peaks in the middle — and the deeper literature on self-improvement and teacher-student compatibility says that's exactly what you'd expect when the value of an external anchor depends on the distance between a model's current ability and what the anchor asks of it Do stronger models always evolve their own harnesses better? Does teacher-refined data always improve student model performance?.