benchmarks evals
The Auditor Needed Computation, Not the Weights
In controlled simulations, black-box prediction access matched white-box weight access once models received a calculation tool—and a script matched the best repairs.
Summary
In controlled simulations, black-box prediction access matched white-box weight access once models received a calculation tool—and a script matched the best repairs.
Claims that language models can inspect neural-network weights often blur reading parameters with executing calculations. This study separated those conditions using simulated subarachnoid-haemorrhage cohorts and small tanh networks containing planted relationships, null cases and one subgroup whose deaths were under-recorded. Without tools, the strongest tested model became unreliable by 3,457 weights and often declined tasks under its default instruction. With a calculation-only tool, it found planted relationships in all 29 reported runs up to 38,401 weights, produced no false alarm in 15 null runs and detected the biased subgroup in all six tests. Black-box prediction access performed as well as white-box access, while a purpose-built script matched the model auditors and the best repair. The experiments use simulated data and do not validate clinical deployment; their sharper point is methodological: computation and elicitation drove the audit, not privileged access to raw weights.
Why it matters
In controlled simulations, black-box prediction access matched white-box weight access once models received a calculation tool—and a script matched the best repairs.
Limits and context
- With a calculation-only tool, it found planted relationships in all 29 reported runs up to 38,401 weights, produced no false alarm in 15 null runs and detected the biased subgroup in all six tests.
- The experiments use simulated data and do not validate clinical deployment; their sharper point is methodological: computation and elicitation drove the audit, not privileged access to raw weights.
Key claims
In controlled simulations, black-box prediction access matched white-box weight access once models received a calculation tool—and a script matched the best repairs.
Qualification: With a calculation-only tool, it found planted relationships in all 29 reported runs up to 38,401 weights, produced no false alarm in 15 null runs and detected the biased subgroup in all six tests.
Evidence: source-2026-11-07-002
Sources
- arXiv preprint 2611.07847, original v1arXiv · primary research
Corrections
: Publication note: This November 7 edition was assembled before its scheduled 9:00 AM America/Detroit reveal but is being published late after recovery from a local build interruption. Its original reporting, image-generation times and scheduled publishAt are preserved.