WORTH Benchmark: Evidence Discovery under Not-Fully-Cooperative Information -- A Two-Episode Pilot

A paper proposes the WORTH Benchmark, a two-episode pilot designed to evaluate how models autonomously discover evidence and how information presentation within narratives guides or obscures judgment, rather than supplying task-relevant information as available premises. The authors argue existing reasoning evaluations rarely examine these two dimensions concurrently.

Existing reasoning evaluations typically provide task-relevant information to models as available premises, rarely examining concurrently how models discover evidence autonomously and how information presentation within narratives guides or obscures judgment. This paper proposes the WORTH Benchmark: