IC-588Topologically ordered context improves relational reasoning over random ordering across nearly all LLMs

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 generate linear chain comparison and syllogism problems and present the statements in three orders: topological (adjacent sentences share an entity), reverse topological, and random. Across nearly all models, randomizing the order causes a noticeable degradation in accuracy compared to topological order. Even reverse topological order improves over random, suggesting that any sorted structure helps. This mirrors the ordering effects documented in human transitive inference research.

Evidence
correlational
Caveat
The effect is described qualitatively as 'noticeable degradation' without a single unified effect size across all models and tasks.
Model
Gemma 2B, Gemma-9B, Gemma-27B, Mixtral 7x22B, Gemini Flash, Pro, GPT-4o
Concepts
Positional bias
Related work
Hotta et al. 2015 [builds-on]
Related findings
IC-586, IC-587, IC-589
Extraction
automatic-extraction