In this work the authors adopt the Relational Match‑to‑Sample (RMTS) paradigm from comparative psychology to systematically evaluate abstract reasoning in state‑of‑the‑art Vision‑Language Models (VLMs). Experiments are run on a parametrically controlled stimulus set and cover three leading commercial APIs (GPT, Claude, Gemini) as well as three open‑source families (Qwen3.5, Gemma‑4, InternVL3). Four levers are identified that shift models from object matching to relational matching: capability tier, model scale, number of objects per scene, and the presence of per‑object noise. Varying these factors produces a developmental‑like trajectory that mirrors the human relational shift.
To probe the internal mechanism, a per‑layer representational similarity analysis combined with causal mediation analysis reveals two competing circuits implementing abstract reasoning: an early circuit that organizes images by surface object features, and a later circuit that organizes them by abstract relations. Extending the study to the ARC‑AGI‑1 benchmark, ablating the relational heads discovered on RMTS degrades performance more than ablating random heads, indicating that the relational circuit is recruited beyond the controlled stimuli.
The mechanistic perspective offered here advances our understanding of how abstract reasoning is realized in VLMs.
Review