# Do multimodal models imagine electric sheep?

arXiv: https://arxiv.org/abs/2605.09693 · HTML: https://mlsanity.com/paper/2605.09693/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- Builds on 36 of its 67 references that are in the map.
- First posted on arXiv 10 May 2026.

## Builds on (most cited first, 30 of 36)

- [Decoupled Weight Decay Regularization](https://mlsanity.com/paper/1711.05101/) (2017)
- [Mastering Diverse Domains through World Models](https://mlsanity.com/paper/2301.04104/) (2023)
- [Emerging Properties in Unified Multimodal Pretraining](https://mlsanity.com/paper/2505.14683/) (2025)
- [Autoregressive Model Beats Diffusion: Llama for Scalable Image Generation](https://mlsanity.com/paper/2406.06525/) (2024)
- [CoT-VLA: Visual Chain-of-Thought Reasoning for Vision-Language-Action Models](https://mlsanity.com/paper/2503.22020/) (2025)
- [Finite Scalar Quantization: VQ-VAE Made Simple](https://mlsanity.com/paper/2309.15505/) (2023)
- [Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task](https://mlsanity.com/paper/2210.13382/) (2022)
- [LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels](https://mlsanity.com/paper/2603.19312/) (2026)
- [Video models are zero-shot learners and reasoners](https://mlsanity.com/paper/2509.20328/) (2025)
- [Perception Tokens Enhance Visual Reasoning in Multimodal Language Models](https://mlsanity.com/paper/2412.03548/) (2024)
- [MIRAGE: The Illusion of Visual Understanding](https://mlsanity.com/paper/2603.21687/) (2026)
- [BabyVision: Visual Reasoning Beyond Language](https://mlsanity.com/paper/2601.06521/) (2026)
- [Are Video Models Ready as Zero-Shot Reasoners? An Empirical Study with the MME-CoF Benchmark](https://mlsanity.com/paper/2510.26802/) (2025)
- [A Very Big Video Reasoning Suite](https://mlsanity.com/paper/2602.20159/) (2026)
- [Emergent Representations of Program Semantics in Language Models Trained on Programs](https://mlsanity.com/paper/2305.11169/) (2023)
- [Mind the Gap: Benchmarking Spatial Reasoning in Vision-Language Models](https://mlsanity.com/paper/2503.19707/) (2025)
- [NitroGen: An Open Foundation Model for Generalist Gaming Agents](https://mlsanity.com/paper/2601.02427/) (2026)
- [Large Language Models Lack Understanding of Character Composition of Words](https://mlsanity.com/paper/2405.11357/) (2024)
- [When Visualizing is the First Step to Reasoning: MIRA, a Benchmark for Visual Chain-of-Thought](https://mlsanity.com/paper/2511.02779/) (2025)
- [Unified Latents (UL): How to train your latents](https://mlsanity.com/paper/2602.17270/) (2026)
- [Mull-Tokens: Modality-Agnostic Latent Thinking](https://mlsanity.com/paper/2512.10941/) (2025)
- [Visual Generation Unlocks Human-Like Reasoning through Multimodal World Models](https://mlsanity.com/paper/2601.19834/) (2026)
- [Visual Representations inside the Language Model](https://mlsanity.com/paper/2510.04819/) (2025)
- [Demystifying Video Reasoning](https://mlsanity.com/paper/2603.16870/) (2026)
- [Evaluating Multiview Object Consistency in Humans and Image Models](https://mlsanity.com/paper/2409.05862/) (2024)
- [Hidden in plain sight: VLMs overlook their visual representations](https://mlsanity.com/paper/2506.08008/) (2025)
- [MMGR: Multi-Modal Generative Reasoning Benchmark and Evaluation](https://mlsanity.com/paper/2512.14691/) (2025)
- [11Plus-Bench: Demystifying Multimodal LLM Spatial Reasoning with Cognitive-Inspired Analysis](https://mlsanity.com/paper/2508.20068/) (2025)
- [Solving and Generating NPR Sunday Puzzles with Large Language Models](https://mlsanity.com/paper/2306.12255/) (2023)
- [MentisOculi: Revealing the Limits of Reasoning with Mental Imagery](https://mlsanity.com/paper/2602.02465/) (2026)

## Built on by (most cited first, 1 of 1)

- [The Dichotomy Between Pattern Recognition and Step-by-Step Reasoning](https://mlsanity.com/paper/2610.09186/) (2026)

## Abstract (arXiv)

Yes. We find that large multimodal models develop mental imagery when solving spatial puzzles, and they do imagine sheep when solving sheep puzzles. We fine-tune a Qwen3.5 VLM to solve twelve diverse visual reasoning tasks -- including tangram, jigsaw, sokoban, 3D mental rotation, and rush hour -- that require understanding geometry, spatial relationships, and the consequences of actions. By supervising the model to predict the open-loop sequence of actions to solve a puzzle from an initial state, we show that the model's activations after each action encode meaningful visual information about the intermediate state. This finding suggests that an imperfect visual world model begins to form as a byproduct of learning to select correct actions, in the absence of any explicit visual supervision. Building on this observation, we propose two ways to sharpen and use the mental images formed by the model. We find that integrating as few as sixteen visual tokens per step into the chain of thought improves the average solve rate from 83% to 89%, with particularly strong gains on reasoning-heavy tasks such as jigsaw and 3D mental rotation.