Large Language Models Are Overconfident in Their Own Responses
Source published ·Modelwire updated
Original coverage: arXiv cs.CL ↗·How Modelwire adds context
The development
A new study reveals that instruction-tuned conversational LLMs suffer from systematic overconfidence, driven by both post-training procedures and chat templates that introduce an 'ownership bias' where models trust their own outputs more than identical user-provided text. Testing across six open-weight models and multiple benchmarks exposes a calibration gap that grows beyond base model miscalibration, suggesting deployment risks for applications relying on model confidence scores for uncertainty quantification or safety filtering.
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 key finding isn't just that LLMs are overconfident in general, which has been documented before, but that the chat template itself is a structural contributor: the formatting wrapper that marks text as model-generated appears to inflate confidence scores independently of the underlying content, meaning the same factual claim gets trusted differently depending on who the model thinks produced it.
This connects directly to two threads running through recent coverage. The reranker self-assessment paper from June 2nd ('Can LLM Rerankers Predict Their Own Ranking Performance?') reached a cautiously optimistic conclusion that self-consistency signals could substitute for external evaluation, but the ownership bias finding here complicates that picture: if models systematically favor their own outputs, self-consistency checks may be measuring confidence inflation rather than genuine reliability. More urgently, the eating disorder safety paper from June 1st and the harm amplification work from the same day both assume that model confidence or refusal behavior reflects calibrated judgment. If that calibration is structurally distorted by post-training and chat formatting, the safety filtering those systems depend on is operating on a flawed signal.
Watch whether any of the six tested open-weight model families release updated chat templates or post-training recipes that explicitly target calibration in the next two release cycles. If ownership bias persists across template revisions, it points to a post-training data problem rather than a formatting artifact, which is a harder fix.
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
Can LLM Rerankers Predict Their Own Ranking Performance?
Researchers investigate whether large language models can self-assess the quality of their own ranking outputs without external evaluation tools. The study tests both training-free methods, like consistency checks across multiple rankings and confidence verbalization, and learned approaches across TREC benchmarks using four LLMs. Results show self-consistency rivals existing state-of-the-art query performance prediction, suggesting rerankers may…
MentionsLarge Language Models · instruction-tuned LLMs · chat templates · ownership bias
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. 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.