Progressive Reasoning with Primitive Correction for Compositional Zero-Shot Learning

2026-07-08 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, long

Summary

PRPC, a Progressive Reasoning framework with Primitive Correction, addresses challenges in Compositional Zero-Shot Learning (CZSL) by explicitly modeling bidirectional dependencies between attributes and objects. This framework mitigates error propagation common in prior unidirectional methods. PRPC formulates CZSL as a structured, Q&A-style Chain-of-Thought (CoT) reasoning process, guiding a Multimodal Large Language Model (MLLM) like Qwen-VL through five predefined semantic steps. These steps include initial object and attribute prediction, followed by mutual correction where attributes refine object recognition and vice versa. To enhance reasoning reliability, PRPC incorporates reinforcement learning post-training using a GRPO-based objective, which provides step-level rewards. Extensive experiments on three CZSL benchmarks, including MIT-States and C-GQA, demonstrate that PRPC achieves leading performance, validating its progressive reasoning and bidirectional correction strategy.

Key takeaway

For Machine Learning Engineers developing compositional zero-shot learning systems, consider adopting a progressive, bidirectional reasoning approach. Your models can significantly improve accuracy by implementing mutual correction steps between attribute and object predictions, rather than relying on unidirectional inference. Leverage Chain-of-Thought prompting with MLLMs and integrate reinforcement learning with step-level rewards to refine intermediate reasoning, leading to more robust and consistent compositional understanding.

Key insights

Bidirectional primitive correction via step-wise Chain-of-Thought reasoning enhances compositional zero-shot learning performance.

Principles

• Contextual dependency is crucial for attribute-object pairs.
• Bidirectional correction reduces error propagation.
• Structured reasoning improves compositional generalization.

Method

PRPC uses a 5-step Q&A-style CoT process with an MLLM (e.g., Qwen-VL) for object/attribute prediction and mutual correction. RL post-training with GRPO loss provides step-level rewards.

In practice

• Implement 5-step reasoning for attribute-object classification.
• Use RL with step-level rewards for MLLM fine-tuning.
• Generate CoT supervision with advanced LLMs like GPT-4o.

Topics

• Compositional Zero-Shot Learning
• Multimodal Large Language Models
• Chain-of-Thought Reasoning
• Reinforcement Learning
• Attribute-Object Recognition
• Error Propagation Mitigation

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• OpenAI CLIP: The Model That Learnt Zero-Shot Image Recognition Using Text — CLIP connects vision and language through contrastive learning on internet-scale image-text pairs, enabling zero-shot image classification.
• LoRi: Low-Rank Distillation for Implicit Reasoning — Hidden-state reasoning trajectories exhibit low-rank structure, enabling efficient distillation for implicit reasoning.
• Plug-and-Adapt: Multimodal Coreference Resolution at First Sight with a Pretrained Alignment Model — Plug-and-Adapt repurposes a pre-trained vision-language alignment model for zero-shot multimodal coreference resolution, outperforming existing methods.
• Self-Questioning Vision-Language Models: Reinforcement Learning for Compositional Visual Reasoning — Rewarding VLMs for generating intermediate sub-questions enables them to autonomously discover compositional reasoning strategies for complex visual tasks.
• Grounding vs. Compositionality: On the Non-Complementarity of Reasoning in Neuro-Symbolic Systems — Symbol grounding alone is insufficient for compositional generalization; explicit reasoning objectives are required.
• One and Only at SemEval-2026 Task 2: Evaluating Zero-Shot Autonomous LLM Agents and Heuristic Proxies in Ecological Affect Forecasting — Zero-shot LLM reasoning combined with statistical priors can be applied to ecological affect forecasting, with task-dependent performance.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.