Research map / Vision-language and multimodal / Multimodal LLMs

Latent Implicit Visual Reasoning

CVPR 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 LIVR, which appends learnable latent visual reasoning tokens to an LMM and trains them with a two-stage visual-bottleneck attention mask and answer-only loss, achieving gains over direct SFT without any intermediate supervision.

Problem

Large multimodal models are text-centric and struggle with vision-centric reasoning, while existing fixes rely on costly explicit visual intermediates (helper images, boxes, crops, depth maps) that embed human priors about useful visual abstractions.

Approach

Append K learnable latent tokens to the prompt and force answer and prompt tokens to access the image only through the latents (visual bottleneck masking in stage 1), then unmask in stage 2; training uses only the end-task answer NLL, with no intermediate labels.

Key results

• On Qwen2.5-VL-3B, LIVR averages 67.85 vs 61.61 for Direct SFT across nine tasks (+6.24), with +13.02 on functional correspondence.
• On LVR-style broader benchmarks with Qwen2.5-VL-7B, LIVR-7B reaches 75.3 on MMVP (vs LVR 71.7) but 54.28 on BLINK-5 (vs LVR 55.37).
• On SAT Val, LIVR-3B reaches 85.6 vs 58.3 for direct SFT and 62.9 for ViGoRL-3B.

Surprising findings

• Latents-only variant (no bottleneck) ignores latents: accuracy is unchanged when latents are removed at evaluation (79.51 to 79.51).
• On Qwen3-VL-4B, LIVR slightly underperforms direct SFT on counting in the single-task setting (66.67 vs 66.86).

Limitations

• Latent tokens are less directly interpretable than text or explicit visual intermediates.
• Results are reported on relatively small backbones (3B-7B) and only mention scaling as future work.
• Hyperparameters were tuned on three tasks and held fixed; task-specific tuning not explored.

Experimental setup

Datasets: PixMo-Count COCO ArtBench-10 SPair-71k HPatches FunKPoint Multi-Illumination Dataset NIGHTS BLINK SAT-32k Visual-CoT MMVP

Benchmarks: BLINK PixMo-Count MMVP V* SAT Val RoboSpatial Visual Spatial Planning

Metrics: accuracy

Baselines: Zero-shot Direct SFT Mirage vanilla GRPO ViGoRL LVR PixelReasoner Vision-R1 PAPO Prompt tuning

Builds on: LLaVA Coconut Mirage LVR prompt tuning

Closest papers by meaning

• Machine Mental Imagery: Empower Multimodal Reasoning with Latent Visual Tokens (CVPR 2026 · similarity 0.92)
• Perception Tokens Enhance Visual Reasoning in Multimodal Language Models (CVPR 2025 · similarity 0.92)
• See It, Say It, Sorted: An Iterative Training-Free Framework for Visually-Grounded Multimodal Reasoning in LVLMs (CVPR 2026 · similarity 0.92)
• Visual Enhanced Depth Scaling for Multimodal Latent Reasoning (NeurIPS 2026 · similarity 0.91)
• UniVLR: Unifying Text and Vision in Visual Latent Reasoning for Multimodal LLMs (NeurIPS 2026 · similarity 0.91)
• SLVR: Structured Latent Visual Reasoning via Human-like Reasoning Flows (NeurIPS 2026 · similarity 0.91)