Table layout format flips LLM lookup accuracy by 50+ points
Source published ·Modelwire updated
Original coverage: arXiv cs.CL ↗·How Modelwire adds context
The development
Researchers have identified a critical dependency in how instruction-tuned language models process tabular data: the serialization format fundamentally shapes which relational queries succeed. Testing across three models with synthetic 256-cell tables reveals that row-major layouts dominate on tall tables by up to 54 points, while column-major formats win decisively on wide tables by up to 59 points. This finding exposes a structural brittleness in LLM table reasoning that scales with aspect ratio, suggesting that model performance on structured lookup tasks is not robust but format-contingent. The work matters for practitioners building retrieval systems and for researchers understanding why language models struggle with consistent tabular reasoning despite strong general capabilities.
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 paper doesn't just show that serialization matters; it quantifies a scaling relationship: performance gaps widen predictably with table aspect ratio, suggesting this isn't a minor quirk but a fundamental constraint baked into how transformers tokenize and attend to structured data.
This joins a recent cluster of work exposing brittleness in LLM structured reasoning. The PremiseSQL benchmark from October 15th revealed that frontier models fail to refuse malformed queries, and the Mamba retrieval paper from October 14th showed that architectural choices (not just capacity) determine whether models can route information correctly. Together, these papers suggest that LLMs lack robust mechanisms for handling data outside natural language, whether the problem is recognizing invalid premises, selecting relevant context, or preserving relational structure across format changes. The serialization finding is narrower but cuts deeper: it implies that even when a model has the capacity to solve a task, the surface form of the input can silently degrade performance without triggering any failure signal.
If the same three models tested here (identities withheld in the summary but likely public in the paper) show consistent aspect-ratio sensitivity on real-world enterprise tables from Kaggle or similar sources, the finding generalizes beyond synthetic data. If they don't, this remains a lab artifact. Watch whether any major LLM provider (OpenAI, Anthropic, Google) ships a table-aware tokenizer or preprocessing layer in the next six months as a direct response.
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
Benchmark reveals LLMs generate SQL for impossible queries instead of refusing
A new benchmark exposes a critical blind spot in production LLM systems: the inability to recognize and refuse flawed user queries. PremiseSQL introduces 1,800 premise-failure test cases across four failure modes, revealing that frontier models often generate plausible but misleading SQL rather than abstaining. This matters because deployed NL2SQL systems currently lack guardrails against logically…
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 “Table Serialization Determines Which Relations Language Models Preserve”. 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.