Can you boost an AI's performance just by changing its tools and workflow, without retraining it at all?
Does test-time scaffolding close capability gaps without retraining models?
This explores whether you can make a model perform better by wrapping it in better tools, code, routing, or inference-time tricks instead of retraining it, and where that approach stops working.
This explores whether you can make a model perform better by changing what surrounds it at inference time (harnesses, code, routing, small adjustments) rather than retraining it. The corpus says yes, often by a surprising margin, but it also shows why that works and where it stops. In one study, a stronger model built an inference-time harness for a weaker one and nearly doubled the weaker model's scores on Theory-of-Mind benchmarks with no retraining at all Can a stronger model lift a weaker one at test time without retraining?. The interesting part is how. The harness didn't get the weaker model to 'think harder.' It moved the unreliable reasoning steps into deterministic code and sent each task down a route suited to it. Scaffolding helps most when it takes work away from the model, not when it asks the model for more.
Why would that be enough? A whole cluster of notes suggests that many capability gaps are really elicitation gaps. Base models already contain reasoning ability in a hidden form, and five quite different interventions all seem to draw out ability that was already there rather than create it: RL, critique fine-tuning, decoding changes, steering internal features, and verifiable-reward training Do base models already contain hidden reasoning ability?. RL post-training looks less like teaching a model how to reason and more like teaching it when to. Hybrid models recover 91% of the gains just by routing tokens Does RL post-training create reasoning or just deploy it?. Instruction tuning tells a similar story: models trained on meaningless or wrong instructions do about as well as those trained on correct ones, because what transfers is a sense of what the output should look like Does instruction tuning teach task understanding or output format?. If training mostly unlocks and formats what's already there, scaffolding can often do the same job from the outside.
The catch is that scaffolding doesn't help every model equally. The ability to write useful harness improvements is about the same across model tiers, but the ability to benefit from them peaks in mid-tier models Do stronger models always evolve harnesses better?. Weak models often fail to use the harness at all. Strong models sometimes don't follow it faithfully, because they override it with their own habits. So scaffolding closes the gap best when the capability is latent but poorly deployed. It can't supply a capability that isn't there, and it can be wasted on a model that thinks it knows better.
There's also a middle ground between 'no retraining' and 'full retraining.' Test-time training, where a model briefly adapts on a long document before answering, helps only when it trains on the passages that bear on the question. Training on random passages actually makes accuracy worse Does test-time training help long context models answer questions?. What you feed the model at test time matters more than the adaptation method. Transformer² goes further: it keeps small sets of 'expert' adjustments to the model's weights and mixes them on the fly for each task, so the model specializes per task without a training run Can models dynamically activate expert skills at inference time?. And at the agent level, the Darwin Gödel Machine improves itself by rewriting its own scaffolding (code editing, context management) and keeping whatever variants test better. It more than doubled performance on coding benchmarks without changing the underlying model Can AI systems improve themselves through trial and error?.
The takeaway you might not have expected: the line between 'capability' and 'scaffolding' is blurrier than it looks. If much of what training adds is knowing when and how to deploy existing skills, then a well-built harness is doing some of the same work as training, just outside the weights. The real limit isn't retraining versus no retraining. It's whether the skill exists in the model at all, and whether the model is in the middle range where it will actually use the help.
Sources 8 notes
A stronger model built inference-time harnesses that nearly doubled weaker model performance on Theory-of-Mind benchmarks without retraining, primarily by moving unstable reasoning into deterministic code and task-specific routing rather than encouraging extended reasoning.
Five independent mechanisms—RL steering, critique fine-tuning, decoding changes, SAE feature steering, and RLVR—all elicit reasoning already present in base model activations. Post-training selects rather than creates reasoning; the bottleneck is elicitation, not capability acquisition.
Evidence shows base models already contain reasoning capability in latent form; RL training optimizes deployment timing rather than capability creation. Hybrid models recover 91% of performance gains by routing tokens only, and activation vectors for reasoning strategies pre-exist before any RL.
Models trained on semantically empty or deliberately incorrect instructions achieve comparable performance to those trained on full correct instructions, achieving 43% vs random baseline 42.6%. The semantic content of instructions appears largely irrelevant; what transfers is knowledge of the output space.
Model capability to produce useful harness edits stays constant across tiers, but capacity to actually benefit from those edits follows an inverted U-shape, peaking in mid-tier models. Weak models fail to invoke harnesses; strong models struggle with faithful instruction-following.
Show all 8 sources
Adapting on random spans from long contexts degrades accuracy, while adapting on oracle spans helps significantly. Self-guided evidence selection recovers most gains cheaply, showing token curation matters more than the adaptation method itself.
Transformer2 demonstrates that tuning only singular values within weight matrices produces composable expert vectors that dynamically mix at inference without interference, outperforming LoRA with fewer parameters and enabling continual specialization.
DGM replaces formal proofs with empirical benchmarking and maintains an evolutionary archive of agent variants, achieving 2.5× improvement on SWE-bench and 2.2× on Polyglot by discovering capabilities like better code editing and context management.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Sharpening Tax in Post-Training
- Eliciting Reasoning in Language Models with Cognitive Tools
- On the Interplay of Pre-Training, Mid-Training, and RL on Reasoning Language Models
- ProRL: Prolonged Reinforcement Learning Expands Reasoning Boundaries in Large Language Models
- Evaluating Theory of Mind in Reasoning Models: Robustness over Reasoning
- Exploring Format Consistency for Instruction Tuning
- The Invisible Leash: Why RLVR May Not Escape Its Origin
- The Darwin Gödel Machine: AI that improves itself by rewriting its own code