Five LLMs fail selective revision test under expert disagreement
Source published ·Modelwire updated
Original coverage: arXiv cs.CL ↗·How Modelwire adds context
The development
A new audit framework reveals that current LLMs fail to selectively revise decisions when challenged by expert disagreement. Researchers tested five models across 1,989 grounded decisions in medicine, science, and law, finding that none met the mathematical threshold for beneficial revision: they must reverse wrong answers at least nine times more often than correct ones at 90% baseline accuracy. The finding exposes a critical gap between apparent responsiveness to authority and actual reasoning improvement, with implications for deployment in high-stakes domains where deference to credentials can mask uncorrected errors.
Modelwire’s AI-generated summary of coverage from arXiv cs.CL.
Modelwire analysis
ExplainerOur AI-generated reading of the wider context and the next developments to watch.
The audit reveals that LLMs fail not just at reasoning, but specifically at the selective part of revision: they can't reliably distinguish when to defer to experts versus when to trust their own outputs. This is distinct from general hallucination or error rates; it's about the decision logic governing when to change course.
This connects directly to two prior findings. The 'Auditing the Judge' work from last week showed that internal model signals outperform confidence scores at detecting errors, but this new study suggests the problem runs deeper: even when models flag uncertainty, they lack the discrimination to revise selectively. Meanwhile, the 'Real Time Error Correction' paper from three days ago proposed catching errors during reasoning generation, but assumes errors are detectable in-flight; this audit implies models may not recognize when expert input should override their trajectory at all. The gap between these approaches suggests the field is treating symptoms (error detection, verification layers) without addressing the core issue: models don't have reliable internal criteria for when external authority should matter.
If the same five models show improvement on this audit after fine-tuning on expert disagreement datasets, that confirms the failure is learnable rather than architectural. If performance remains flat, it suggests the issue is fundamental to how these models weight confidence versus deference, which would require different training approaches entirely. Watch for follow-up work testing whether models trained on explicit reasoning about when to trust experts (versus when to resist) can meet the 9x threshold.
This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error
Coverage behind this analysis
These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.
·arXiv cs.LG
Internal probes catch errors LLM confidence scores miss
Researchers have demonstrated that supervised probes trained on internal model activations can outperform language models' own confidence signals at detecting errors, a critical capability for systems that route outputs to human review. Testing across three 7-12B parameter models on math benchmarks, the approach identified 1.3-22.5 percentage points more errors than the model's native verdict scores…
MentionsLLMs · medicine · science · law
How this coverage is produced
Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.
Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “When Expert Disagreement Hurts: Prestige and Selectivity in LLM Decision Revision”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.