Positional Bias in Nested Defeasible Reasoning: Order Sensitivity and Unfaithful Intermediate Steps in Small Language Models

Status: manuscript in preparation.

In formal reasoning, a valid conclusion does not depend on the order of its premises. This work asks whether language models hold to that in nested defeasible reasoning, where a general rule is overridden by an exception, which is in turn overridden by a more specific one. It also asks whether models that reach the correct final answer get there through correct intermediate steps.

Results will be posted here once the paper is out.

Recommended citation: M. A. Khan. (2026). "Positional Bias in Nested Defeasible Reasoning: Order Sensitivity and Unfaithful Intermediate Steps in Small Language Models." Manuscript in preparation.