LLM confidence signals fail across model tiers in extraction tasks
Source published ·Modelwire updated
Original coverage: arXiv cs.CL ↗·How Modelwire adds context
The development
A preregistered evaluation of extraction reliability across LLM tiers reveals that confidence calibration remains severely broken even in flagship models. Testing DeepSeek-V4-Flash, GLM-5.3-Flash, and GPT-6.1-Sol on cyber threat intelligence extraction shows micro precision ranging from 0.03 to 0.34 and calibration error between 0.53 and 0.94. The finding matters because it exposes a fundamental gap: LLMs can report high confidence while producing garbage, making their verbalized uncertainty signals unreliable as filters for production use. This challenges the assumption that cheaper models can be salvaged through confidence thresholding, forcing teams to either accept low precision or invest in post-processing pipelines.
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 study isolates a specific failure mode: LLMs don't just produce wrong answers, they produce wrong answers while reporting high confidence. This means confidence thresholding (the standard cost-cutting workaround) doesn't work as a filter.
This connects directly to the pattern exposed in 'All Verdicts are Not Equal' from early October, which found that LLM outputs shift unpredictably even under controlled conditions. That work challenged whether LLMs could be trusted as evaluators; this one goes deeper by showing that even when models attempt to signal uncertainty, those signals are decoupled from actual accuracy. The cyber threat intelligence domain also echoes 'Measuring Artifact Reliance in Machine-Generated Text Detectors' (October 9), which found that systems appear reliable on benchmarks but fail on surface-level artifacts in deployment. Here, the artifact is the confidence score itself.
If the authors release a follow-up showing that post-hoc calibration methods (temperature scaling, Platt scaling) can bring calibration error below 0.30 on the same CTI extraction task, that would suggest the problem is fixable in production. If calibration error remains above 0.50 even after recalibration, it signals that confidence is fundamentally unreliable for this class of task, forcing teams to abandon thresholding entirely.
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.CL
LLM judges fail reliability audit, undermining evaluation standard
A systematic audit of LLM-as-a-Judge evaluation exposes fundamental brittleness in a practice now treated as industry standard. Testing six frontier models across multiple conditions reveals that verdicts shift even at zero temperature, position bias flips majority outcomes on hard tasks, and the most consistent judge achieves reliability by repeating errors while matching ground truth only…
MentionsDeepSeek-V4-Flash · GLM-5.3-Flash · GPT-6.1-Sol
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 “From Unreliable to Usable: Decomposing Failure Modes and Confidence Signals in LLM Entity Extraction - Evidence from Cyber Threat Intelligence”. 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.