Metadata timestamps hijack LLM temporal reasoning across five models
Source published ·Modelwire updated
Original coverage: arXiv cs.CL ↗·How Modelwire adds context
The development
Researchers have exposed a critical vulnerability in how large language models handle temporal reasoning. By manipulating collection-date metadata on Wikipedia revisions, they induced 37-51 percentage point drops in exact-match accuracy across five open-weight LLMs, even when the underlying evidence remained unchanged. The effect persists across prompt variations designed to mitigate it, suggesting models conflate metadata timestamps with factual grounding rather than treating them as auxiliary signals. Extended reasoning partially recovers performance in Qwen3.5, hinting that chain-of-thought approaches may offer partial mitigation. This finding has immediate implications for retrieval-augmented generation systems and highlights a fundamental brittleness in how production LLMs anchor answers to temporal context.
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 vulnerability isn't just that models struggle with temporal reasoning (we knew that). It's that models treat collection-date metadata as a factual signal rather than auxiliary information, meaning the *source* of the timestamp corrupts the answer more than the actual facts do.
This extends the temporal reasoning fragility documented two days ago in 'Keeping Time vs. Understanding Time', which showed models fail to transfer temporal skills across tasks. That work suggested conceptual confusion about how models learn time; this paper identifies a specific mechanism: models anchor to metadata timestamps instead of reasoning through evidence. It also parallels the metadata-sensitivity pattern seen in 'Prompt Sensitivity in Tool-Augmented Model Comparisons' (same week), where auxiliary signals (tool calls) amplified rather than stabilized model behavior. Both suggest that adding structure to model inputs can backfire if models treat that structure as ground truth rather than context.
If the same 37-51 point accuracy drops replicate on a held-out Wikipedia snapshot collected after October 2026 (ensuring no training contamination), this confirms the vulnerability is real and not an artifact of benchmark construction. If they don't, the effect may be specific to how current models were trained on historical Wikipedia versions.
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
Study questions temporal reasoning split in language models
A new study challenges the foundational assumption that temporal reasoning in language models splits cleanly into arithmetic and semantic tasks. Researchers tested whether models trained on one type of temporal task transfer skills to similar problems, finding minimal cross-dataset gains within each category. The work suggests current benchmarks and training approaches may be built on…
MentionsQwen3.5 · Wikipedia · LLMs
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 “Collection-Date Assignment Redirects Answers in Temporal QA”. 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.