Skip to content
conftrace_
2026 ICML ICML 2026

Semantic-Enriched Latent Visual Reasoning

Multimodal latent-space reasoning replaces explicit “thinking with images” with reasoning in a compact latent space. Current methods mainly reshape low-level image features under visual supervision. Their latents capture appearance but carry limited high-level abstract semantics. We introduce Semantic-Enriched Latent Visual Reasoning (SLVR), a two-stage framework with a semantic attribute extraction head. In the first stage, this head converts region evidence into attribute-level semantic latents. These latents complement low-level visual features and support richer reasoning. In the second stage, Multi-query Group Relative Policy Optimization (M-GRPO) jointly optimizes multiple queries grounded in the same region. It builds on the invariance of intrinsic region attributes across different questions. M-GRPO constrains the degree of attribute variation across queries while preserving answer correctness. We construct SLV-Set, with approximately 400K region-level attribute annotations and 800K multi-query question answering samples, and introduce SV-QA for evaluation under semantic variation. Experiments show that SLVR improves the robustness and semantic consistency of latent visual reasoning. Our code and datasets are available at https://github.com/tinnel123666888/slvr.

More from ICML 2026

All ICML 2026 papers →