PRCO splits perception and reasoning roles to sharpen multimodal model training
Researchers introduce PRCO, a reinforcement learning framework that gives visual evidence extraction its own reward signal, aiming to fix a credit-assignment problem in multimodal reasoning models.
In plain words
Multimodal AI models answer questions about images, but training them with a single reward for the final answer can leave weak image understanding uncorrected. PRCO trains one policy in two roles: an Observer writes a question-relevant caption of the visual evidence, and a Solver answers from that caption. Each role gets its own reward, so the captioning step improves alongside the answering step.
What was announced
The authors of the arXiv paper, writing in its abstract, say that reinforcement learning with verifiable rewards has boosted reasoning in multimodal large language models, but common approaches optimize perception and reasoning together with one reward based only on the final answer. In their view, this shared reward blurs credit assignment: models can pick up better reasoning patterns while upstream extraction of visual evidence stays unreliable. They present PRCO (Perception-Reasoning Coevolution) as a dual-role framework built on a single shared policy.
PRCO assigns the two roles distinct signals. The Solver is optimized with verifiable rewards on the final answer, while the Observer earns a utility reward tied to how well the Solver succeeds downstream using its caption. The authors report experiments across eight multimodal reasoning benchmarks in which PRCO improves average accuracy by more than 7 points over the base model across model scales, ahead of prior open-source RL-tuned baselines. These results come from the paper's own abstract; the excerpt does not include the full experiments, and no independent evaluation is cited.
What it does not show
- The evidence is limited to the paper's abstract, so benchmark details, model versions and comparison settings cannot be verified from the supplied passages.
- The reported accuracy gains are the authors' own claims; no independent confirmation is provided in the sources.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.