Recent work on distillation transfers the capabilities of large models to smaller ones often by updating the latter's parameters, through teacher forcing, on-policy distillation, and related training-time methods. In this paper, we ask whether such transfer can instead occur at test time. We study strong-to-weak scaffolding: whether a stronger builder model can construct inference-time harnesses that help a weaker target model solve tasks more reliably without any parameter updates. Using four representative Theory-of-Mind benchmarks, each builder model uses 5% of the data as a validation set to iteratively refine its harness over multiple rounds, after which the finalized harness is evaluated on the full test set. Empirically, this form of test-time capability transfer is highly effective, nearly doubling average target-model performance from 0.49 to 0.91. Our analysis shows that the gains come primarily from offloading unstable model reasoning into deterministic code, benchmark-specific routing, and strict answer-format enforcement, rather than from encouraging the target model to reason more extensively or sample more broadly. We further find that builder-model reasoning effort improves harness quality monotonically, platform effects are modest relative to the builder model's own capability, and weaker target models receive the largest gains. These results suggest that inference-time harness design is an important complement to conventional training-time distillation, enabling strong models to transfer cognitive structure to weaker models without retraining.
Alex: Welcome to another episode of ResearchPod. Today, we're looking at a study from Salesforce AI Research on something they call "strong-to-weak scaffolding."
Sam: So the central question is: can you make a smaller AI model perform like a much larger one, without actually retraining it from scratch?
Alex: Exactly. The idea is that a powerful model can build a kind of custom support structure for a weaker one. The weaker model then runs inside that structure and performs significantly better than it would on its own.
Sam: And the core problem isn't that smaller models are ignorant — it's that they get overwhelmed when a task has too many moving parts at once?
Alex: That's the key insight. Think of a student who actually knows the material, but freezes up during an exam because the question is formatted in a confusing way, or because they have to track too many things simultaneously. The knowledge is there — the environment is working against them.
Sam: So rather than teaching the student more content, you redesign the exam to play to their strengths.
Alex: Precisely. The researchers call this redesigned environment a "harness." It's a set of instructions, decision rules, and verification checks that wrap around the smaller model. The stronger model designs the harness; the weaker model just works inside it.
Sam: And how much of a practical difference does that actually make?
Alex: A notable one. In their tests, the best harness nearly doubled the accuracy of a smaller model — bringing it much closer to the performance of a larger, far more expensive system. And crucially, the smaller model's internal parameters weren't changed at all.
Sam: So where do the gains actually come from? What is the harness doing that the model couldn't do on its own?
Alex: The biggest gains come from identifying the steps where the model is most likely to make errors — things like keeping track of output format, or following a precise logical sequence — and then handing those steps off to a rigid, rule-based computer program instead. The model no longer has to juggle everything at once.
Sam: So it's a bit like using a calculator for the arithmetic part of a maths problem. You still have to understand the problem, but you're not wasting mental energy on the sums.
Alex: That's a good analogy. The researchers call this "deterministic offloading." You move the error-prone, mechanical steps out of the model's hands and into a system that executes them perfectly every time. What's left for the model is the part that actually requires understanding — and it handles that better when it's not distracted by everything else.
Sam: So the stronger model is essentially acting as an architect. It studies the task, figures out where the weaker model tends to fail, and builds a workspace that eliminates those failure points.
Alex: That's a precise way to put it. The builder model uses a sample of real task data to test and refine the harness until it performs reliably. It's an automated process — the builder is essentially running experiments on behalf of the weaker model.
Sam: Does the quality of the builder model matter much? Does it need to be extremely capable to produce a useful harness?
Alex: The study finds a consistent pattern: the stronger the builder, the better the resulting scaffold. But even moderately capable builders produce meaningful improvements. The relationship is fairly steady — more builder effort reliably translates to a better harness.
Sam: What kinds of tasks did they actually test this on?
Alex: They focused on social reasoning tasks — problems that require what researchers call "Theory of Mind." That's the ability to understand that other people hold different beliefs, intentions, or knowledge than you do. Imagine tracking a situation like: "She thinks he doesn't know that she moved the object." You have to hold multiple perspectives in your head at once and reason about what each person believes. It's genuinely demanding, even for large models.
Sam: That sounds like mental chess. If you lose track of even one layer, the whole chain of reasoning collapses.
Alex: Exactly. And that's precisely the kind of task where a well-designed harness makes a real difference. By breaking the problem into structured steps and offloading the bookkeeping to reliable code, the scaffold keeps the model from losing its place.
Sam: So the broader shift here isn't really about building smarter models — it's about building smarter environments for the models we already have.
Alex: That's the central argument of the paper. For many practical applications, a carefully designed harness can be just as valuable as scaling up the model itself. It's a different lever — and in some cases, a more accessible one. You don't need to train a new model or spend enormous resources. You need a strong model to study the task and build the right scaffolding.
Sam: That's a meaningful reframing. Instead of asking "how do we make the AI more powerful," you ask "how do we make the task easier for the AI to handle reliably."
Alex: And the evidence here suggests that question is worth taking seriously. Thanks for listening to ResearchPod.