Progressive Reasoning with Primitive Correction for Compositional Zero-Shot Learning

Compositional Zero-Shot Learning (CZSL) requires models to recognize unseen combinations of attributes and objects (e.g., 'ancient building') based on learned primitives. Existing methods often fail because they either treat attributes and objects as independent or use unidirectional pipelines where an initial error in object detection cascades into an incorrect attribute prediction. This paper asks: can a bidirectional, multi-step reasoning process mitigate this error propagation?

The authors propose Progressive Reasoning with Primitive Correction (PRPC). Instead of a single-pass classification, PRPC treats CZSL as a structured, five-step Chain-of-Thought (CoT) reasoning process. The model first identifies the object, then the attribute, and then performs a 'correction loop' where the attribute is used to re-verify the object, and the updated object is used to re-verify the attribute. The framework is trained in two stages: first, supervised fine-tuning (SFT) to teach the model to follow the reasoning template, and second, reinforcement learning (RL) using the Group Relative Policy Optimization (GRPO) objective to provide step-level rewards for accurate intermediate reasoning.

PRPC achieves state-of-the-art performance across three major CZSL benchmarks (MIT-States, C-GQA, and VAW-CZSL). The results demonstrate that the bidirectional correction loop significantly reduces error accumulation compared to unidirectional baselines. Ablation studies confirm that the later verification steps (Steps 3-4) are critical for performance, and that step-wise reinforcement learning is more effective than standard supervised training alone for improving compositional consistency.

This work shifts the paradigm of CZSL from simple similarity-based classification to explicit visual reasoning. By demonstrating that MLLMs can be trained to perform iterative, self-correcting logic, the paper provides a robust template for handling compositional tasks where context and primitive dependencies are essential for accurate recognition.