The Year in AI Papers
The year/Independent research

Paper 2512.21218

Latent Implicit Visual Reasoning

Kelvin Li, Chuyi Shang, Leonid Karlinsky, Rogerio Feris, Trevor Darrell, Roei Herzig

Published
Dec 2025
Research lab
Independent
Citations
8
GitHub
Not linked

01 In brief

Summary

The paper introduces Latent Implicit Visual Reasoning (LIVR), a method to enhance visual reasoning in Large Multimodal Models (LMMs) without explicit intermediate supervision.

LIVR adds latent tokens to the model's vocabulary and trains them via a two-stage visual bottlenecking approach: Stage 1 forces visual information to pass through latent tokens by masking attention from answer and prompt tokens to image tokens; Stage 2 uses standard attention to integrate enriched latents.

This allows the model to learn task-relevant visual abstractions implicitly.

In data-matched experiments across nine perception-heavy tasks (e.g., counting, jigsaw, localization), LIVR outperforms direct supervised fine-tuning (SFT) on three LMM backbones (Qwen2.5-VL-3B, Qwen3-VL-4B, LLaVA-OneVision-1.5-4B) by averages of 6.24%, 3.43%, and 5.60%, respectively, and improves multi-task training on Qwen3-VL by 2.77%.

Compared to prior methods, LIVR matches or exceeds performance on benchmarks like Visual Spatial Planning (66.00 vs.

Mirage's 46.00 on Qwen2.5-VL-3B), SAT, BLINK, and RoboSpatial, without using helper images, bounding boxes, or chain-of-thought annotations.

Ablations confirm that both latent tokens and bottlenecking are necessary, with K=16 latents and a 4-epoch Stage 1 / 6-epoch Stage 2 schedule optimal.

Attention visualizations show latents focus on task-relevant regions.

02 From the paper

Abstract

While Large Multimodal Models (LMMs) have made significant progress, they remain largely text-centric, relying on language as their core reasoning modality. As a result, they are limited in their ability to handle reasoning tasks that are predominantly visual. Recent approaches have sought to address this by supervising intermediate visual steps with helper images, depth maps, or image crops. However, these strategies impose restrictive priors on what "useful" visual abstractions look like, add heavy annotation costs, and struggle to generalize across tasks. To address this critical limitation, we propose Latent Implicit Visual Reasoning (LIVR), a task-agnostic mechanism that trains LMMs to discover and use latent visual reasoning tokens without explicit intermediate supervision. These tokens attend globally and re-encode the image in a task-adaptive way, enabling the model to extract relevant visual information without hand-crafted supervision. LIVR consistently outperforms direct supervised fine-tuning across diverse vision-centric tasks and multiple LMM backbones. In broader comparisons, LIVR remains competitive with or outperforms prior text-based and explicit-visual-intermediate reasoning methods, while requiring no additional intermediate supervision such as helper images, bounding boxes, image crops, depth maps, or chain-of-thought annotations. Our project page can be found here: https://www.chuyishang.com/livr/