DyCo-RL: Dynamic Cross-Modal Coordination for Visual Reasoning

A new method called DyCo-RL aims to fix a core weakness in how multimodal AI models handle visual reasoning by forcing them to coordinate visual and textual information during the generation process, according to a paper submitted on 6 Jun 2026 [1]. Reinforcement Learning with Verifiable Rewards (RLVR) has become a standard approach for improving visual reasoning in Multimodal Large Language Models (MLLMs), but existing techniques optimize only for the final answer, ignoring the step-by-step coordination between seeing and reading [1]. The researchers behind DyCo-RL found that during Chain-of-Thought reasoning, models often fail to switch dynamically between extracting visual evidence and synthesizing textual context, a breakdown they link directly to incorrect answers [2]. DyCo-RL addresses this by operating in two stages. First, it computes the Fisher–Rao geodesic distance between consecutive attention distributions within a single modality, such as vision or text. A larger distance signals a significant restructuring of attention, indicating active information extraction from that modality [3]. Based on this measurement, each token is assigned a functional role as either visually-oriented or text-oriented [1]. Second, the method evaluates how well a token’s actual attention allocation matches its assigned role. This alignment score is then used for advantage reweighting during policy optimization, amplifying learning signals for well-coordinated tokens and dampening the influence of misaligned ones [3]. The technique is algorithm-agnostic and was tested as a plug-in with four representative RLVR algorithms [1]. Experiments applied DyCo-RL to the Qwen2.5-VL-3B and Qwen2.5-VL-7B models and reported consistent improvements across seven benchmarks covering both visual-centric and mathematical reasoning tasks [2]. The work adds to a growing body of research focused on disentangling perception and reasoning in multimodal systems. A separate framework, PRCO, introduced a dual-role setup where an Observer generates question-conditioned evidence captions and a Solver produces the final answer, with each role receiving its own reward signal [4]. Another approach, R-C2, enforces cross-modal cycle consistency by requiring a model to perform backward inference, switch modalities, and reconstruct the answer, yielding a dense, label-free reward that improved reasoning accuracy by up to 7.6 points [5]. A third line of work, CalibRL, reinterprets expert supervision as a calibration baseline rather than direct imitation, using a LeakyReLU-based asymmetric activation to preserve policy entropy while guiding exploration [6]. The DyCo-RL authors argue their contribution is distinct because it integrates coordination directly into the RLVR optimization loop rather than restructuring the model’s architecture or reward sources [1]. The code repository is publicly available [3].