# Liu and Yang: Computation, not weight access, enabled these network audits [Canonical reading](https://escapemeridian.com/readings/liu-and-yang-computation-not-weight-access-enabled-these-8a0121e5) Zhiyu Liu and Xinyu Yang report that large language models could audit simulated clinical prediction networks more effectively when given a calculation tool than when asked to inspect network weights in a prompt. Their experiments used tanh networks trained on simulated patient cohorts, including networks with planted relationships, no relationship, or subgroup labels that under-recorded deaths. Without tools, GPT-6 Astra recovered planted relationships in smaller networks but became unreliable as networks grew; under a default instruction it also declined most audit and repair tasks as unidentifiable. An explicit instruction removed those refusals. With a calculation-only tool, models could use either white-box access to weights or black-box access to predictions. The authors report that black-box access performed as well as white-box access, and that a well-designed script matched comparator models on the audit and the best LLM repair. The authors’ result is about these simulation experiments, not evidence that language models can reliably audit clinical prediction systems in use. It highlights a distinction between supplying a model with information and enabling it to calculate with that information. How well would the reported audit results hold on clinical prediction networks beyond the simulated cohorts and network conditions tested here? Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source arXiv cs.AI · Article Original source date: Nov 5, 2026 (2026-11-05). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-09T00:24:11.749Z. First observed: 2026-11-06T05:10:56.749Z. Retrieved: 2026-11-09T00:23:26.876Z. - [Computation, not weight access: large language models as auditors of clinical prediction networks](https://arxiv.org/abs/2611.07847) ## Limits The experiments used simulated cohorts of patients with subarachnoid haemorrhage and simulated networks, so the findings do not establish performance on clinical prediction systems in use. The abstract reports single-prompt tests and specific models and network conditions; it does not establish how results generalize beyond them. ## Open question How well would the reported audit results hold on clinical prediction networks beyond the simulated cohorts and network conditions tested here? ## Source-stated Liu and Yang tested language-model audits using simulated patient cohorts and networks with planted relationships, no relationship, or under-recorded subgroup deaths. > 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. - [Supporting source](https://arxiv.org/abs/2611.07847) ## Source-stated Without tools, GPT-6 Astra became unreliable on larger networks and often refused tasks under the default instruction; an explicit system instruction removed those refusals. > Without tools, Astra recovered planted relationships in 353-weight networks but became unreliable at 3,457 weights, and under the default instruction it declined 11 of 12 audit and repair tasks as unidentifiable; the explicit instruction removed these refusals. - [Supporting source](https://arxiv.org/abs/2611.07847) ## Source-stated With a calculation-only tool, the models audited networks using either weight access or prediction-only access, with the authors reporting comparable performance. > Black-box access performed as well as white-box access, and other GPT-6 models read as accurately. - [Supporting source](https://arxiv.org/abs/2611.07847) ## Source-stated The authors report that a well-designed script matched comparator models on the audit and the best LLM repair. > A well-designed script matched the comparator models on the audit and the best LLM repair. - [Supporting source](https://arxiv.org/abs/2611.07847) ## Escape Meridian interpretation The results concern the reported simulation experiments; they do not establish audit reliability on clinical prediction systems in use. > internally pre-specified simulation experiments. Tanh networks of 105 to 38,401 weights were trained on simulated cohorts - [Supporting source](https://arxiv.org/abs/2611.07847) ## Source and contributor context Liu, Zhiyu, Yang, Xinyu · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Large language model A language model with many learned numerical settings, used to process and generate language. - [Definition reference](https://developers.google.com/machine-learning/glossary#large-language-model) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## Editorial provenance Version 1. Publication ID: e13211aa-4d2d-4db5-8650-2146a4375b77. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-09T00:24:11.749Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: c53e5fc26bf3fd2606b1fa3a92cce8c19f1adbe5881b83550f234efef6dcd375. Source text hash: 389ad0964e6a307898749ab4cbc5b8c78b1a8758e9beee8912bbf695d3fa3439. ## Revision and source-check history ### Version 1 — 2026-11-09T00:24:11.749Z Zhiyu Liu and Xinyu Yang report that large language models could audit simulated clinical prediction networks more effectively when given a calculation tool than when asked to inspect network weights in a prompt. Their experiments used tanh networks trained on simulated patient cohorts, including networks with planted relationships, no relationship, or subgroup labels that under-recorded deaths. Without tools, GPT-6 Astra recovered planted relationships in smaller networks but became unreliable as networks grew; under a default instruction it also declined most audit and repair tasks as unidentifiable. An explicit instruction removed those refusals. With a calculation-only tool, models could use either white-box access to weights or black-box access to predictions. The authors report that black-box access performed as well as white-box access, and that a well-designed script matched comparator models on the audit and the best LLM repair. The authors’ result is about these simulation experiments, not evidence that language models can reliably audit clinical prediction systems in use. It highlights a distinction between supplying a model with information and enabling it to calculate with that information. How well would the reported audit results hold on clinical prediction networks beyond the simulated cohorts and network conditions tested here? ## Original source arXiv cs.AI · Article Original source date: Nov 5, 2026 (2026-11-05). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-09T00:24:11.749Z. First observed: 2026-11-06T05:10:56.749Z. Retrieved: 2026-11-09T00:23:26.876Z. - [Computation, not weight access: large language models as auditors of clinical prediction networks](https://arxiv.org/abs/2611.07847) ## Limits The experiments used simulated cohorts of patients with subarachnoid haemorrhage and simulated networks, so the findings do not establish performance on clinical prediction systems in use. The abstract reports single-prompt tests and specific models and network conditions; it does not establish how results generalize beyond them. ## Open question How well would the reported audit results hold on clinical prediction networks beyond the simulated cohorts and network conditions tested here? ## Source-stated Liu and Yang tested language-model audits using simulated patient cohorts and networks with planted relationships, no relationship, or under-recorded subgroup deaths. > 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. - [Supporting source](https://arxiv.org/abs/2611.07847) ## Source-stated Without tools, GPT-6 Astra became unreliable on larger networks and often refused tasks under the default instruction; an explicit system instruction removed those refusals. > Without tools, Astra recovered planted relationships in 353-weight networks but became unreliable at 3,457 weights, and under the default instruction it declined 11 of 12 audit and repair tasks as unidentifiable; the explicit instruction removed these refusals. - [Supporting source](https://arxiv.org/abs/2611.07847) ## Source-stated With a calculation-only tool, the models audited networks using either weight access or prediction-only access, with the authors reporting comparable performance. > Black-box access performed as well as white-box access, and other GPT-6 models read as accurately. - [Supporting source](https://arxiv.org/abs/2611.07847) ## Source-stated The authors report that a well-designed script matched comparator models on the audit and the best LLM repair. > A well-designed script matched the comparator models on the audit and the best LLM repair. - [Supporting source](https://arxiv.org/abs/2611.07847) ## Escape Meridian interpretation The results concern the reported simulation experiments; they do not establish audit reliability on clinical prediction systems in use. > internally pre-specified simulation experiments. Tanh networks of 105 to 38,401 weights were trained on simulated cohorts - [Supporting source](https://arxiv.org/abs/2611.07847) ## Source and contributor context Liu, Zhiyu, Yang, Xinyu · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Large language model A language model with many learned numerical settings, used to process and generate language. - [Definition reference](https://developers.google.com/machine-learning/glossary#large-language-model) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## Editorial provenance Version 1. Publication ID: e13211aa-4d2d-4db5-8650-2146a4375b77. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-09T00:24:11.749Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: c53e5fc26bf3fd2606b1fa3a92cce8c19f1adbe5881b83550f234efef6dcd375. Source text hash: 389ad0964e6a307898749ab4cbc5b8c78b1a8758e9beee8912bbf695d3fa3439. ### uncertain — 2026-11-09T00:24:11.749Z — anthropic/claude-opus-5.5 Publication: e13211aa-4d2d-4db5-8650-2146a4375b77 attribution — pass: The reading correctly attributes the work and its conclusions to Zhiyu Liu and Xinyu Yang, the listed authors. It keeps reported results as the authors' findings rather than established fact. > Authors:Zhiyu Liu, Xinyu Yang summary — uncertain: The reading leaves out the headline tool-assisted results: all 29 runs up to 38,401 weights recovered the planted relationship, there were no false alarms in 15 null runs, and the biased subgroup was found in 6 of 6 runs. It also drops the authors' conclusion that elicitation mattered alongside computation. Central terms such as weights, tanh networks and planted relationship are not explained in plain language. Smallest repair: add the with-tool results, keep 'computation and elicitation', and briefly gloss these terms. > For auditing these networks, what mattered was computation and elicitation, not access to the weights. date — pass: The originalPublished date of 2026-11-05 matches the v1 submission date on arXiv. > \[Submitted on 5 Nov 2026\] support — pass: Each claim is grounded in the abstract. The limitation to simulated settings is a careful, labeled interpretation that follows from the stated simulation design. The question does not assume an answer. > We separated these conditions in internally pre-specified simulation experiments. media — pass: The selected media is the verified arXiv abstract page, which is the direct original source. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/8a0121e57121b7dd3869bef7ec17f6aa72d325f16c0c3622be61175d11fddc2b/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/liu-and-yang-computation-not-weight-access-enabled-these-8a0121e5.json)