New benchmark measures when LLMs correctly refuse to answer
Source published ·Modelwire updated
Original coverage: arXiv cs.CL ↗·How Modelwire adds context
The development
DeclineCraft establishes a new benchmark for measuring when language models should refuse to answer rather than hallucinate. The work targets a critical failure mode in RAG systems: distinguishing between retrievable facts, learned parameters, and genuine knowledge gaps. Testing 27 models reveals that retrieval alone boosts abstention accuracy substantially, though iterative self-assessment shows promise for further gains. This matters because production RAG deployments increasingly need reliable confidence signals, and the benchmark's rolling cutoff design lets teams evaluate future models against expanding knowledge horizons. The findings suggest current systems still struggle with honest uncertainty.
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.
DeclineCraft's key contribution isn't just measuring abstention accuracy, but isolating it as a separate diagnostic from retrieval quality and parametric knowledge. The rolling cutoff design lets benchmarks age gracefully as model training dates advance, solving a practical problem most benchmarks ignore.
This work sits directly alongside the entity extraction confidence study from mid-October, which found that even flagship models report high confidence while producing garbage. DeclineCraft goes further by building a benchmark that separates the signal (does the model know what it doesn't know?) from the noise (did retrieval fail, or did the model hallucinate?). The earlier work on procedural hallucination and the memory evaluation benchmark both identified how downstream accuracy masks broken internal systems. DeclineCraft targets the same root problem: current evals conflate multiple failure modes, making it impossible to know whether a wrong answer reflects genuine uncertainty or overconfidence.
If the 27 models tested show that iterative self-assessment closes the gap between retrieval-only and oracle performance within the next benchmark iteration, confidence-based filtering becomes viable for production RAG. If the gap persists above 15 percentage points, it signals that models still can't reliably self-assess, and teams should expect to need external validation layers regardless of model tier.
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 confidence signals fail across model tiers in extraction tasks
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…
MentionsDeclineCraft · Minecraft
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 “DeclineCraft: A RAG Abstention Benchmark for Jointly Measuring Combined Retrieval and Parametric Knowledge”. 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.