version was submitted on February 3, 2026.
The authors introduce C3PO, a training-based framework combining chain-of-thought compression with contrastive preference optimization. Their stated premise is that reasoning mechanisms can increase reliance on language priors over visual inputs, producing traces with fewer visual cues and redundant text.
C3PO filters selected reasoning tokens, uses preference pairs built with AI feedback, and adds hallucination-inducing prompts to create negative training signals. The paper reports consistent reductions across multimodal reasoning models and evaluations including CHAIR, POPE, AMBER, and a GPT-4-based benchmark.
The source does not establish independent replication, production performance, the cost of training or inference, or whether the method improves reliability on real user workloads.