Causal Probing for Internal Visual Representations in Multimodal Large Language Models

Researchers developed a causal probing framework to decode how Multimodal Large Language Models internally represent visual concepts, revealing that entities are encoded in localized regions while abstract concepts distribute globally across networks. The findings expose mechanistic drivers of scaling laws and uncover a disconnect between visual perception and reasoning capabilities in MLLMs.

This research addresses a fundamental gap in understanding how multimodal AI systems process and represent visual information at the mechanistic level. By using activation steering to systematically intervene on internal representations across four visual concept categories, the authors mapped the architecture of concept encoding in ways that traditional behavioral testing cannot reveal. The divergence between localized entity encoding and distributed abstract concept representation has direct implications for how researchers should approach model scaling and optimization going forward.

The findings challenge conventional scaling wisdom by demonstrating that depth matters asymmetrically across concept types. While entity localization remains stable regardless of model size, abstract concept encoding fundamentally requires deeper architectures—a distinction with significant engineering implications for deploying efficient multimodal systems. This mechanistic insight explains why larger models don't uniformly improve across all visual reasoning tasks.

The compensatory mechanism between perception and generation, revealed through reverse steering, suggests MLLMs allocate computational resources dynamically between understanding visual inputs and generating outputs. This finding has practical consequences for developers optimizing inference costs and latency. Perhaps most concerning for downstream applications, the perceived disconnect between visual perception and abstract reasoning exposes a critical limitation: MLLMs can recognize geometric relations but fail to execute the procedural reasoning required for problem-solving. This gap between pattern recognition and logical inference defines the current boundary of multimodal AI capability and suggests future model improvements must address reasoning architecture, not merely scale.

• →Entities are encoded in localized network regions while abstract concepts distribute globally, revealing distinct encoding strategies within MLLMs
• →Model scaling drives abstract concept improvement but leaves entity localization invariant, suggesting depth requirements vary by concept type
• →MLLMs exhibit a compensatory mechanism where blocking outputs increases latent activations, exposing dynamic resource allocation between perception and generation
• →Visual reasoning reveals a critical failure mode: MLLMs recognize geometric relations as static features but cannot execute procedural reasoning needed for problem-solving
• →Activation steering enables causal intervention on internal representations, providing new methodology for mechanistic interpretability research