Research map / Vision-language and multimodal / Multimodal LLMs

Semantic-Enriched Latent Visual Reasoning

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SLVR, a two-stage latent visual reasoning framework that adds attribute-level semantic latent supervision and a multi-query GRPO variant with latent consistency, plus the SLV-Set data and SV-QA benchmark.

Problem

Latent visual reasoning relies on visual-only supervision, producing latents lacking semantic richness and failing to support diverse region-level queries consistently.

Approach

Enrich region latents with a semantic latent trained to match attribute-profile embeddings, then align latents across multiple questions on the same region via Multi-query GRPO with cross-query consistency and stability regularization.

Key results

• SLVR-7B reaches 61.8 on OKVQA and 57.8 on VizWiz vs. LVR's 50.6 and 33.1, but trails LVR on GQA (55.6 vs 57.4) and AI2D (76.0 vs 77.3).
• On SV-QA, SLVR-7B Both Correct improves over LVR by +3.7 on V*, +3.2 on HRBench-4K and +4.0 on HRBench-8K.
• On VisualPuzzles, SLVR achieves 34.2 overall vs. 29.4 for LVR.

Surprising findings

• Text-only SFT+GRPO baselines still lag latent variants on the Both Correct metric, despite higher Q1 accuracy on HRBench-4K/8K.
• Human verification found a 7.29% error rate (29,457 erroneous annotations) in the automatically generated SLV-Set.

Limitations

• Auto-generated attribute and QA annotations contain a 7.29% error rate despite post-hoc verification.
• Gains over LVR are mixed; SLVR underperforms LVR on GQA and AI2D.
• SV-QA is small (591 pairs) and built from the same benchmarks used for evaluation of semantic variation.

Experimental setup

Datasets: Visual-CoT (ViSCoT) V* HRBench-4K HRBench-8K OKVQA GQA VizWiz ChartQA TextVQA AI2D VisualPuzzles

Benchmarks: OKVQA GQA VizWiz ChartQA TextVQA AI2D VisualPuzzles SV-QA (V*, HRBench-4K, HRBench-8K)

Metrics: accuracy Both Correct rate

Baselines: Qwen2.5-VL-7B GRIT DeepEyes Visual-SR1 LVR Monet

Builds on: Latent Visual Reasoning (LVR) GRPO Visual-CoT

Closest papers by meaning

• SLVR: Structured Latent Visual Reasoning via Human-like Reasoning Flows (NeurIPS 2026 · similarity 0.94)
• Latent Visual Reasoning (ICLR 2026 · similarity 0.94)
• Monet: Reasoning in Latent Visual Space Beyond Image and Language (CVPR 2026 · similarity 0.92)
• Leveraging Latent Visual Reasoning in Silence (NeurIPS 2026 · similarity 0.91)
• Vision-aligned Latent Reasoning for Multi-modal Large Language Model (ICML 2026 · similarity 0.90)
• Fill the GAP: A Granular Alignment Paradigm for Visual Reasoning in Multimodal Large Language Models (NeurIPS 2026 · similarity 0.90)