Researchers have introduced CogAlign, a framework designed to align multimodal large language models (MLLMs) with human cognitive patterns. The approach utilizes 66 semantic dimensions derived from the THINGS project, which analyzed 1,854 object concepts based on 4.7 million human judgments.

The CogAlign process involves three stages: identifying key semantic dimensions using the SPoSE algorithm, generating reasoning chains via an LLM, and fine-tuning the target model on 309,329 examples. This method aims to teach models to reason using human-like criteria rather than relying on standard reinforcement learning techniques.

Testing on the Qwen3-VL model family showed that CogAlign outperformed simple prompting strategies, which often led to logical errors or repetitive loops. In comparative tests, fine-tuned Qwen3-VL models surpassed the 70-billion parameter Centaur model, which reached 51.96% accuracy in human-alignment tasks.

The framework also demonstrated improvements in physical reasoning benchmarks. For instance, the Qwen3-VL-4B model increased its score in the physical properties category of the MMBench benchmark from 67% to 75%, while the 8B version improved from 61% to 77%.