cs.LGNov 18, 2026

Auditing black-box clinical prediction models with a large language model through queries alone: a blinded test with planted flaws

Authors: Zhiyu Liu, Xinyu Yang

Organizations: Tianjin Medical University, Tianjin, China · Tianjin Medical University General Hospital, Tianjin, China

Abstract

Large language models (LLMs) are used to explain clinical prediction models, on the promise that they can say what a model computes, find its faults and even reveal new knowledge in it. We tested the three promises where the answers could be checked. From the International Stroke Trial we trained four kinds of model (logistic regression, explainable boosting machine, gradient-boosted trees, neural network) for two outcomes on patients randomised earlier, planted one flaw in each of 24 copies (an under-predicted subgroup, a reversed risk factor or an implausible step) and mixed them with the 8 clean models. GPT-6 Astra audited each model three times, blind, through a tool that only returned the model's predictions; the judging rules were committed before any run and applied by program. Its surrogate models, with a median of 128 terms, reproduced the audited models' log-odds with a mean R2R^2 of 0.976 (0.959 for an additive surrogate of 42 terms fitted by an analyst). It found the planted flaw in 69 of 72 runs (96%; 95% CI 80% to 99%); a script written for the purpose found all 24. It also reported many more problems: 60% of 1,089 checkable reports failed the pre-specified check against the model's average behaviour; post hoc, most of these were real but small steps, and 15% of the checkable reports were not found in the model. Of 601 unexpected relationships it read out of the clean models, 11% held in patients randomised later (random pairs 6%; a regression scan's shorter list of its strongest candidates, 28%); after a check on the training data, 33% against 30%. With query access alone, an LLM can describe a black-box model faithfully and find nearly every planted flaw, as a purpose-written script also did; its reports of problems must be checked on the model and its discoveries on data.

Figures & tables

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Nov 5, 2026

Computation, not weight access: large language models as auditors of clinical prediction networks

Frontier large language models (LLMs) have been reported to read, edit and write the weights of small neural networks, raising the prospect of auditing clinical prediction models from their files alone. Public demonstrations, however, mixed reading raw weights with running code. We separated these conditions in internally pre-specified simulation experiments. Tanh networks of 105 to 38,401 weights were trained on simulated cohorts of patients with subarachnoid haemorrhage, each with a planted relationship or none, and one network was trained on labels that under-recorded deaths in a subgroup. GPT-6 Astra and four other GPT models answered single prompts under a default or an explicit system instruction, either with the weights in the prompt and no tools, or with a calculation-only tool that exposed the weights (white box) or only the network's predictions (black box). Without tools, Astra recovered planted relationships in 353-weight networks but became unreliable at 3,457 weights, and under the default instruction it declined 17 of 21 audit and repair tasks as unidentifiable; the explicit instruction removed these refusals. With the tool, it recovered the planted relationship in all 39 runs on networks of up to 38,401 weights, raised no false alarm in 21 runs on null networks, and found the biased subgroup in 6 of 6 runs. Black-box access performed as well as white-box access, and other GPT-6 models read as accurately. A well-designed script matched the comparator models on the audit and the best LLM repair. For auditing these networks, what mattered was computation and elicitation, not access to the weights.
Jun 8, 2026

Measurement Under Selection: Decoy-Calibrated Failure Audits for Language Models

Knowing how often a language model fails does not explain where its errors concentrate. When auditors examine many explanations, the strongest observed pattern may arise by chance. We introduce Janus, a procedure for checking proposed error patterns before reporting them. Janus starts with a fixed list of yes/no properties of the examples being evaluated, such as whether the input is long. For each property, it compares the model's error rates on examples with that property and those without it. To see how large a difference can arise by chance, it repeats this calculation after shuffling the yes/no labels across examples without changing the group sizes. These shuffled properties are called decoys. A pattern is reported only if the size of its error difference meets a threshold set using decoys. On separate held-out examples, the same group must still have the higher error rate and the difference must meet a minimum, which was chosen in advance. In a controlled experiment, where the model must find a code in documents containing tables of staff, projects, and renewal codes, Janus confirms five related patterns of higher error rates on tasks requiring more lookups across tables. It also confirms a sixth pattern: lower error rates on examples with the needed information at the ends of the tables. In our samples from the MuSiQue and LongBench v2 public benchmarks, SliceLine finds groups with high error rates, while Janus reports no confirmed error patterns for the example properties we chose to test. For comparison, we use standard tests that shuffle errors and account for testing many candidates. With the same holdout check, they confirm two to six controlled patterns, depending on the test and threshold, and none on either benchmark. In simulations with no real error patterns, Janus reports false patterns more often than Benjamini-Hochberg, depending on the decoy count.
Jun 6, 2026

Stress-testing medical large language models reveals latent safety pathology beyond benchmark accuracy

Large language models (LLMs) are entering clinical practice based on benchmark accuracy that may fail to detect safety-relevant failure modes. Here we present AI-MASLD, a stress-audit framework that adapts the logic of metabolic stress testing from hepatology to the evaluation of clinical LLMs. Using 240 clinical cases across six narrative perturbation probes, we subjected seven models to double-stress testing and quantified performance through three indices: metabolic index (MI), perturbation flip rate (PFR), and counterfactual fairness index (CFI). Under clean baseline conditions, all models performed uniformly well. Under realistic narrative stress, performance diverged sharply, revealing two distinct stress-response phenotypes. Quantized models exhibited pseudonormalization, in which low flip rates hid functional collapse. Medical supervised fine-tuning systematically degraded logical stability, fairness, and information extraction. An open-weight model matched or exceeded proprietary alternatives on every safety dimension. These findings establish narrative stress auditing as a necessary complement to accuracy-based evaluation.