LLM reviewers favor more confident wording even when scientific content is unchanged

What is changing?

A new preprint from researchers in Germany shows that LLMs used for scientific review can be influenced by writing tone. The team rewrote 504 ICLR 2026 papers in more confident or more cautious language while preserving their claims, evidence and structure, then asked six models to review them under conference-style criteria.

The researchers report that every model systematically scored the confident versions higher. Claude Sonnet 4.6 favored confident wording seven times more often than cautious wording; it accepted a confident version while rejecting the cautious version of the same content in 21% of cases, compared with 2.8% in the reverse direction. The results show that wording alone can move a paper across an acceptance threshold.

The study is an arXiv preprint and has not yet undergone peer review. The authors recommend stress-testing LLM reviewers with content-preserving rewrites and examining not only overall scores but also which aspects written reviews praise or criticize.