When a language model works alongside a person, a small routing component usually decides which of them answers each question. This is called learning to defer: a model is trained to hand a case to the human whenever the AI is likely to be wrong. The router depends on one thing, the model's estimate of its own correctness. Benchmark contamination, meaning test questions that leaked into training data, corrupts that estimate in a specific way. A model that has seen a problem template becomes confident whenever it recognizes the template. If a deployment case is a small variant of a memorized item, with different numbers, reordered premises, or a changed answer, the model stays confident while its accuracy drops. The router then keeps the case, and the human never sees the one problem where their judgment was needed.
This note proposes to measure that failure under controlled conditions and to fix it with a small, separately trained router. The router is a low-rank adapter on the language model, a few trainable matrices added beside frozen weights. It is trained on pairs of memorized items and changed variants, so that it learns to separate "I recognize this" from "I will get this right." The goal is a better trade-off between team accuracy and human workload. A naive confidence router, a router that uses a contamination-detection score, and a linear probe trained on the same pairs are the competitors. The design specifies what result would mean the adapter is unnecessary.
Core Claim
Contamination hurts human–AI teams less through inflated leaderboard numbers than through a familiarity gap: the drop in accuracy between memorized items and their answer-changing variants at the same confidence level. For a set of template families, let $\mathrm{acc}_{\text{orig}}(p)$ and $\mathrm{acc}_{\text{var}}(p)$ be the model's accuracy on original and variant items to which it assigns confidence $p$. The gap is
$$ G(p) = \mathrm{acc}_{\text{orig}}(p) - \mathrm{acc}_{\text{var}}(p). $$
For a well-calibrated model, $G(p) \approx 0$ at every $p$. The claim has two parts. First, $G$ grows with contamination density and is not removed by temperature scaling, because a single global temperature rescales both populations together. Second, a router trained on original–variant contrast pairs can drive the router's own familiarity gap toward zero. Contamination-detection scores cannot do this, because they flag both the exact duplicates (where the AI is right and deferring wastes human effort) and the variants (where it is wrong).
The motivating evidence comes from the file. Calibrated Learning to Defer with One-vs-All Classifiers shows that deferral only works when the router's correctness estimates are calibrated. It replaces the usual softmax deferral loss with independent per-option ("one-vs-all") classifiers and reduces router calibration error from 7.58% to 3.01% on CIFAR-10. That work uses image classifiers and never contaminates anything. Data Contamination Calibration for Black-box LLMs introduces PAC, a detector that perturbs text and compares the model's most- and least-certain token probabilities. PAC shows memorized items produce local certainty spikes and estimates a 91.4% GSM8K contamination rate for davinci-002, but it only audits a fixed model and never feeds the signal into a decision.
Other work shows that variants break memorizing models:
- DyVal procedurally generates reasoning problems from graphs and reports a 13.6% GPT-4 accuracy drop on reachability when the graph is described in reverse order.
- Can Large Language Models Reason? A Characterization via 3-SAT shows accuracy collapsing where surface patterns stop predicting satisfiability.
No paper in the file connects these variant failures to routing decisions.
Research Program
A controlled contamination testbed. Natural contamination is unobservable for most models, so the first step is to inject it deliberately.
- Base models: Pythia, whose Pile pretraining data is public and which On the Calibration of Large Language Models and Alignment already profiled for calibration, and LLaMA-2 13B, the base used in Platypus.
- Items: DyVal-generated arithmetic, logic and reachability problems at depths D1–D4, plus GSM8K test items.
- Splits: Items are split by template family, not by instance.
- Injection: A chosen subset of original items is mixed into an instruction-tuning set at repetition densities of zero, one and ten.
- Variants: For every injected original we generate a changed variant. Variants use new numbers, reversed premise order, or a perturbed graph, and in each case the correct answer differs from the memorized one. DyVal's generator makes these changes exact and checkable.
The rank question. The same testbed also answers the file's open question about whether low rank limits memorization. We fine-tune the contaminated mixture with LoRA at ranks 4, 16 and 64 and with full-parameter tuning, then measure three things:
- verbatim accuracy on injected originals;
- variant accuracy;
- $G(p)$.
Xu reports that full fine-tuning on Alpaca pushes MMLU calibration error above 0.20, from below 0.1 for pretrained Pythia, and that LoRA limits that drift. Whether the same protection holds when the tuning data carries leaked items is untested, and this experiment tests it directly.
The router. On each contaminated task model we train a deferral router using the one-vs-all loss of Verma and Nalisnick. It outputs two independent probabilities:
- that the AI's answer is correct;
- that the human's answer is correct.
The router is implemented as its own low-rank adapter plus a small head, and the task model is kept frozen. It is trained on contrast pairs drawn from template families disjoint from the test families. Each pair is a contaminated original the task model gets right and a variant it gets wrong. This teaches the router to attend to where the input departs from the recognized template, rather than to recognition itself.
The human partner. The human side follows standard deferral practice.
- Simulated experts (main experiments): Each expert's accuracy depends on problem depth and type. The profiles are anchored to the 82-person human comparison in DyVal, where humans fall behind GPT-4 on long, tedious computation but hold up better on abductive logic.
- Real participants (smaller study): A small study with real participants on DyVal-style items checks that the simulated results transfer. The collaboration protocol follows HAI-Eval, which reports 31.11% collaborative pass rate against 18.89% for unaided humans and 0.67% for the model alone.
Where this sits among the three directions. Complementarity is the goal being optimized. Contamination is the controlled stressor. Low-rank adaptation is both the proposed router parameterization and an experimental variable, and the design allows that its role may turn out to be minor.
Baselines
No deferral and fixed-rate random deferral. The first reference points are the contaminated task model answering everything, and random deferral at each workload level. They bound how much any router adds.
Confidence thresholding with temperature scaling. This is the default in practice: defer whenever the task model's top probability falls below a threshold. The temperature is fitted on a clean validation split drawn from the training template families. It is the strongest naive competitor because it costs nothing and is what practitioners deploy.
The strongest existing method: a one-vs-all router with a contamination feature. This combines the Verma and Nalisnick router with a PAC polarized-distance score, and, as a second variant, with the IRT guessing factor from Efficiently Measuring the Cognitive Ability of LLMs: An Adaptive Testing Perspective. That paper uses item response theory, the statistics of standardized testing, to flag items where a weak model succeeds on hard questions. This combination is the most direct way to "make deferral contamination-aware" using existing tools.
- Reported figures: 3.01% router calibration error on CIFAR-10 for one-vs-all, from Verma and Nalisnick; detection AUC gains of 4.5% on WikiMIA and 5.9% on StackMIA for PAC, from Ye et al.
- Missing figures: No reported result found on Lacuna for any deferral router on language tasks under contamination.
Fairness conditions. Every router receives the same training template families, the same contrast pairs where applicable, the same hyperparameter search budget (an equal number of trials) and the same GPU hours.
What counts as winning. The proposed router beats a baseline if two things hold on held-out template families:
- the area under its team-accuracy-versus-deferral-rate curve is higher, with confidence intervals over five seeds excluding zero difference;
- its familiarity gap is smaller at matched workload.
Ablations that isolate the key ingredients. Two factors are crossed:
- training data: contrast pairs versus the same number of unpaired items;
- router parameterization: linear probe on frozen hidden states, low-rank adapter, or full fine-tuning.
This separates the contribution of the contrast data from the contribution of the adapter.
Why not just…?
Why not just decontaminate the training data, as Platypus did? Platypus filters fine-tuning data by keywords and embedding similarity before LoRA tuning. That helps the fine-tuning set but cannot reach pretraining corpora of the size of FineWeb, which users of most models cannot inspect. Ye et al. also argue that paraphrased leakage evades string and embedding matching. Our testbed checks this directly: we run Platypus-style filtering on the injected mixture with paraphrased injections and report how much of $G$ survives.
Why not just add a membership-inference score to the router? This is the real threat. If the PAC-feature router matches ours on the accuracy–workload curve, the contrast-trained adapter is unnecessary, and we will report that. We expect it to lose on workload rather than accuracy: a membership score flags exact duplicates, which the AI answers correctly, as readily as changed variants, so it over-defers. PAC also needs token probabilities, which the file notes commercial APIs are withdrawing. The per-population deferral rates (duplicates versus variants) will show which explanation holds.
Why not just train a linear probe on the contrast pairs instead of an adapter? This may well work. It would mean the useful ingredient is the contrast data, not low-rank adaptation. Our hypothesis is that detecting a departure from a memorized template requires computation over the whole input, not just over a final hidden state, so an adapter that can reshape attention should do better. The data-by-parameterization ablation settles the question. If the probe matches the adapter, the conclusion becomes "train routers on variant contrasts," and the adapter drops out of the recommendation.
Why not just prompt the model to be careful with familiar-looking problems? SCALAR tested careful, skeptical and expert personas for an agent and found no stable effect on reasoning quality. We include a "check whether this problem differs from ones you have seen" prompt as an additional zero-training baseline.
Proposed Experiments and Decisive Checks
Experiment one: does contamination create a familiarity gap, and does rank change it? We measure $G(p)$ across injection densities, ranks and full fine-tuning on DyVal and GSM8K items. The decisive outcome has two parts:
- $G$ rises with density and survives temperature scaling;
- whether it is smaller at low rank than under full tuning.
A null result on the first part would undercut the whole proposal, so this experiment comes first.
Experiment two: team performance. We compare team accuracy against deferral rate for all baselines and ablations with simulated experts, on held-out template families and at every contamination density. Reliability diagrams are drawn separately for the original, variant and clean populations. The decisive outcome is whether the contrast-trained router lowers the variant-population calibration error without raising deferral on exact duplicates.
Experiment three: transfer beyond the testbed.
- Unseen perturbation type: train on number changes and test on reversed descriptions, the DyVal failure mode.
- Model with unknown training data: use PAC-flagged GSM8K items as proxy contamination labels.
- Real participants: a small human study on DyVal-style items tests whether gains measured with simulated experts hold up. Its metrics are team accuracy and the rate at which participants accept wrong AI answers. The second metric targets the over-reliance that Accounting for Human Engagement Behavior to Enhance AI-Assisted Decision Making documents, where teams underperform the model alone because human trust is poorly matched to AI accuracy.
References
- Calibrated Learning to Defer with One-vs-All Classifiers
- Data Contamination Calibration for Black-box LLMs
- DyVal: Dynamic Evaluation of Large Language Models for Reasoning Tasks
- Can Large Language Models Reason? A Characterization via 3-SAT
- On the Calibration of Large Language Models and Alignment
- Platypus: Quick, Cheap, and Powerful Refinement of LLMs
- HAI-Eval: Measuring Human-AI Synergy in Collaborative Coding
- Efficiently Measuring the Cognitive Ability of LLMs: An Adaptive Testing Perspective
- When Does Critique Improve AI-Assisted Theoretical Physics? SCALAR: Structured Critic--Actor Loop for Agentic Reasoning
- Accounting for Human Engagement Behavior to Enhance AI-Assisted Decision Making