Supervision Fading: Measuring What LLMs Can Do Once the Scaffold Is Gone

Senior medical residents tell a consistent story about the jump to independent practice. Some of their uncertainty is about missing knowledge, such as an unfamiliar presentation or procedure. A separate kind is about autonomy: they can do each piece of the work when an attending physician frames the problem, but they are unsure they can put the pieces together alone. Residents even rehearse for this deliberately, for example by delaying a consult so they have to "make the call" themselves (Griffith et al.). We propose to test large language models in the same way. For each hard clinical question we build a ladder of prompts. The top rung gives attending-level guidance: the question is broken into sub-questions, with the needed facts supplied. Each lower rung removes some of that guidance, ending with the bare question and then with harmless rewordings of the bare question. The drop in accuracy along this ladder, and where along it the model breaks, separates "the model lacks the knowledge" from "the model has the pieces but cannot assemble them unprompted."

We expect this to explain two findings that are currently reported separately. Medical benchmarks show that scaling helps factual recall but barely helps multi-step crisis management (AnesSuite). Prompt-sensitivity studies show that sensitivity rises with task complexity (ProSA), and that the complex, multi-part dependencies inside the model stay unstable even when simple ones stabilize (IPS). Our hypothesis is that both describe one failure, which we call scaffold dependence. The same measurement also gives a practical signal for when a model should hand a case to a human, analogous to a resident calling the attending.

Core Claim

On top-tier clinical reasoning items, most of a model's sensitivity to prompt wording comes from integration failures. These are items where the model answers every sub-question correctly when asked one at a time, yet fails the bare question. Integration failures should not shrink with scale the way knowledge failures do. A further consequence is that disagreement between a model's bare answer and its answer under a self-built scaffold should predict error better than disagreement across surface paraphrases. Paraphrase disagreement misses errors the model makes the same way every time, which Rethinking Hallucinations calls prompt-agnostic errors. Removing structure probes the specific step that fails.

To be precise, for an item $i$ let $a_i(r)$ be correctness at rung $r$, with $r=0$ the full scaffold and $r=R$ the bare question. The autonomy gap is $\Delta = \mathbb{E}_i[a_i(0) - a_i(R)]$. We split it using a per-item sub-question pass indicator $s_i$, which is 1 if every sub-question is answered correctly in isolation:

$$\Delta = \underbrace{\mathbb{E}_i[(1-s_i)(a_i(0)-a_i(R))]}_{\text{knowledge gap}} + \underbrace{\mathbb{E}_i[s_i(a_i(0)-a_i(R))]}_{\text{integration gap}}.$$

The prediction is that the integration term grows from System 1 to System 2 and stays flat across model sizes, while the knowledge term shrinks with size.

Research Program

The main testbed is AnesBench, with 7,972 bilingual anesthesiology questions split into System 1 (factual recall), System 1.x (formula application) and System 2 (operating-room crisis management) (AnesBench). We will decompose about 400 System 1.x and System 2 items into sub-questions at the lower tiers. A strong model will draft the decompositions, and anesthesiologists will verify and correct them. Each sub-question gets its own gold answer, so $s_i$ can be measured directly. This also addresses the dossier's concern about whether sub-tasks in a hierarchy are independent: we observe propagation directly rather than assuming it away.

The ladder has five rungs. Rung 0 poses the sub-questions with their correct answers filled in. Rung 1 poses the sub-questions without answers. Rung 2 replaces them with a generic, case-independent clinical checklist, in the style of the structured verification in ICT. Rung 3 is the bare item. Rung 4 applies Brittlebench-style meaning-preserving perturbations to the bare item, such as typos, persona framing, newlines and paraphrases (Brittlebench).

Scaffolds must never quote option text. Multiple-choice conversion is known to invite lexical-overlap shortcuts (From Generation to Selection), and scaffolds that echo options would recreate that problem.

Two replications test whether the effect is specific to medicine. ConsisEval supplies naturally paired easy and hard problems, where the easy problem is a strict sub-step of the hard one (ConsisEval). In that setting the rung-0 scaffold is the easy problem together with its solution. ICT supplies cases where the correct "call" is to abstain because information is missing. That lets us test whether removing the scaffold specifically revives the action bias reported by ToolBeHonest and ICT.

The escalation signal is computed without any human scaffold. The model writes its own decomposition, answers each sub-question, and integrates the results. In resident terms this is mental war-gaming. We escalate when that answer disagrees with the bare answer, or when the model's own sub-answers are unstable across two decompositions. Models are Qwen2.5 at 7B, 14B, 32B and 72B, the Morpheus 7B/14B/32B checkpoints released with AnesSuite, Llama-3-70B-Instruct, and one commercial model. Including Morpheus tests whether training on verifiable reasoning traces (AnesR1, trained with GRPO, a reinforcement-learning method) shrinks the integration gap or only the knowledge gap.

Baselines

The naive reference is the flat score: bare-question accuracy per tier, which is how AnesBench is currently reported. For tier-level accuracies we found no reported result on Lacuna; AnesSuite shows them only in figures. We will therefore re-run every model ourselves under one decoding setup. Any claim that the ladder adds information must show that items with identical flat accuracy separate cleanly into knowledge-gap and integration-gap groups, and that these groups behave differently under perturbation.

The strongest existing escalation method is SENSEL (Sensitivity and Accuracy in ICL). It abstains when answers disagree across paraphrases or demonstration orderings, and it reports +4.1 points of area under the F1-coverage curve over max-probability and entropy on GPT-J-6B. A second strong competitor is plain self-consistency: sample $k$ chain-of-thought answers and abstain on disagreement. Both will be given the same number of forward passes as our signal. A self-scaffold uses roughly one decomposition call, $m$ sub-answer calls and one integration call, so each baseline gets that many paraphrases or samples. We will also report max softmax probability and entropy. Every method gets the same held-out threshold-tuning split and is scored on risk-coverage and F1-coverage on System 2 and on ICT. Our method beats them only if it wins at matched compute on System 2 and, especially, catches errors that SENSEL scores as consistent.

For ConsisEval the reference figure is the hard-to-easy consistency score of GPT-4 Turbo, 92.2%. For ICT it is Gemma-3-4B's F1 of 68.57% with structured verification, up from 20.90% zero-shot. That makes the generic checklist rung a real, strong competitor and not a strawman.

Two ablations isolate what matters. The first is a placebo scaffold: structured text of the same length and format that carries no item-relevant content. It separates the effect of content from the effect of format. The second compares a self-generated scaffold with an expert-verified one, which measures how much of the gap a model can close without a human.

Why not just…?

Why not just run paraphrase perturbations on each AnesBench tier? This is the obvious first study, and it is our rung 4. On its own it can show that System 2 is more brittle. It cannot say why, because a paraphrase-induced flip on a knowledge-gap item and one on an integration-gap item look the same. The decisive test is whether paraphrase flips on System 2 concentrate among integration-gap items. If they are spread evenly, the ladder adds nothing to this simpler study and we will say so.

Why not just always deploy with a scaffold, such as divide-and-conquer prompting or a verification checklist? This is a real threat. Divide-and-Conquer shows that splitting a problem into independent parts limits error compounding, and ICT's checklist roughly triples small-model F1. If a self-generated scaffold recovers most of the rung-0 accuracy, the autonomy gap is a prompting artifact with a cheap fix, and that becomes our headline finding. The expert-versus-self-scaffold comparison settles it. ConsisEval shows that even frontier models fail easy sub-steps about 8% of the time, which suggests self-decomposition will inherit some errors. That remains to be measured.

Why not just use a bigger model? AnesSuite already reports a ceiling on System 2 from 7B to 72B, and Brittlebench finds that sensitivity accounts for about 25% of variance even in commercial models. Our size sweep tests the stronger claim that scale shrinks the knowledge term but not the integration term. If both terms shrink together, the scale skeptic wins.

Why not just abstain on low token confidence? ProSA ties sensitivity to flat output distributions, so confidence may carry the same information. However, SENSEL already beats max-probability, and Rethinking Hallucinations finds that entropy-style scores track randomness rather than correctness. Its Med-HALT ambiguity is about 60% for Llama2-13B. Confidence stays in the comparison at matched compute.

Decisive Checks

Four results decide the proposal.

First, the integration share of $\Delta$ should rise from System 1 to System 2 and stay roughly flat from 7B to 72B, while the knowledge share falls. If this fails, the System 2 ceiling is about knowledge, and scaffold dependence is not its explanation.

Second, among System 2 items that are correct at rung 0, paraphrase flips at rung 4 should be significantly more frequent in integration-gap items than in knowledge-gap items. This links brittleness to integration.

Third, on items where SENSEL's paraphrases agree but the answer is wrong, the self-scaffold disagreement signal should flag a meaningful fraction. This tests directly whether removing structure catches the prompt-agnostic errors that paraphrasing cannot.

Fourth, the placebo scaffold should recover little of the gap. If it recovers much, the effect comes from formatting, not from the reasoning the scaffold supplies.

On ICT, unscaffolded rungs should show more fabricated tool calls than the checklist rung, which is the LLM version of a resident freezing or overreaching without a backstop. If the Morpheus models shrink the integration gap more than Qwen models of equal size, that motivates a follow-up we do not claim here: training with scaffolds that fade over the course of training, much as residency is meant to fade supervision.