IC-587Real-world knowledge acts as a shortcut in LLM relational reasoning, causing worse-than-chance performance on logically valid but factually incongruent statements

Andrew Liu, Henry Prior, Gargi Balasubramaniam, Rivka Moroshko, Amir Zait, Ilia Labzovsky, Danny Karmon, Ishita Dasgupta, Kim Stachenfeld, Kenneth Marino

SourceReCogLab: a framework testing relational reasoning & cognitive hypotheses on LLMs

The authors construct comparison problems using 540 real-world objects with known size/weight, creating congruent (consistent with reality), incongruent (contradicting reality), and random-string (no semantic prior) variants. Across all models, congruent statements outperform incongruent ones. Some models exhibit worse-than-chance (0.5) correctness on incongruent comparison problems, meaning they are actively misled by their real-world priors. Random-string performance lands between the two extremes, confirming that the effect is driven by factual coherence rather than surface form.

Evidence
correlational
Key metric
worse-than-chance (0.5) correctness on incongruent comparison problems; 540 objects used for congruency construction
Caveat
Object sizes and weights were estimated by Gemini Pro, introducing potential noise in the congruency labels.
Model
Gemma 2B, Gemma-9B, Gemma-27B, Mixtral 7x22B, Gemini Flash, Pro, GPT-4o
Concepts
Shortcut
Related work
Lampinen et al. 2024 [builds-on]
Related findings
IC-586, IC-588, IC-589
Extraction
automatic-extraction