Does upgrading an AI's tools and instructions help weak and strong models equally? Surprisingly, the biggest winners are in the middle.

This explores whether an upgrade to the scaffolding around a model (its prompts, tools, and checking code) helps a weak model as much as a strong one. Not equally, and the pattern isn't simply 'more capable, more gain.' One study found that the ability to write useful harness edits is flat across model tiers, but the ability to benefit from them follows an inverted U that peaks in mid-tier models Do stronger models always evolve harnesses better?.

The two ends fail for opposite reasons. Weak models often don't invoke the harness at all, so the update sits unused. Strong models struggle with faithful instruction-following, so they drift from the script the harness gives them. Mid-tier models are capable enough to use the tools and compliant enough to stay on them. Writing an update and gaining from one are also separate skills, so the model that authors a patch needn't be the one that profits from it.

The weak end isn't hopeless, because it depends on how much the harness asks of the model. When a stronger model built inference-time harnesses for weaker ones on Theory-of-Mind benchmarks, performance nearly doubled. The gains came mostly from moving shaky reasoning into deterministic code and routing by task type, not from asking the weak model to think longer Can a stronger model lift a weaker one at test time without retraining?. Giving weaker planners a map from runtime behavior to code raised win rates by 10–19 points and let them match stronger models at finding the right code Can explicit behavior maps help weaker planners compete with stronger models?. A committee of weak model calls can match strong models too, but only when something external like tests, proofs, or type checks picks the right answer, because sampling alone can't When can weak models match strong model performance?. Reading across these, updates that take work away from the model seem to help weak models more than updates that depend on the model following them well. That is an inference from the notes, not something they test directly.

Gains also don't carry between models automatically. One runbook lifted several frozen models on Terminal-Bench 2.1 and transferred to newer ones unchanged, including a 5.4-point gain for DeepSeek-V4 Flash Can execution harnesses lift model performance without retuning weights?. But a safety harness strict enough for one model over-blocks another Should safety harnesses be customized for each deployment?, and an update that worked once shouldn't be reused after the base model changes without checking that it still applies Should past update success guide future model changes?. Who writes the update matters as well. A 9B model trained on whether its patches actually worked lifted a frozen agent by 9.3 points and beat prompted frontier models, because it reran patches to verify their impact instead of optimizing for plausibility Does training editors on real outcomes beat prompting larger models?.

The practical upshot is to tune a harness for each model tier and test it on the weak model before assuming the gain carries over. The collection doesn't yet say how the inverted U changes when a harness is mostly deterministic code versus mostly instructions, and that is where the answer for weak models probably lies.