Mental Imagery That Matters: Curing Latent Collapse in Multimodal ReasoningMi Luo1, Rohan Dugad1, Zihui Xue1, Alex Dimakis2,3, Kristen Grauman1
1UT Austin
2UC Berkeley
3Bespoke Labs
NeurIPS 2026
TL;DR: Latent-reasoning MLLMs often ignore their own visual latents. We call this Mental Simulation Collapse. AnchorSim trains models to use them, and improves visual reasoning on 9 benchmarks.
Models can now think in continuous visual latents between text steps. But does the answer depend on them?
We perturb a model's own latents and re-score its answer. Corrupting the visual latents barely moves the answer, while the same perturbations on text are usually far more damaging.
Two training signals make generated latents consequential. The extra branches are dropped at inference.
| Model | BLINK Jigsaw | BLINK Avg | V* | MMVP | EMMA | VisuLogic |
|---|---|---|---|---|---|---|
| GPT-4o* | - | 60.0 | 66.0 | 70.7 | - | - |
| Claude-4-Sonnet* | - | 39.6 | 15.2 | - | - | - |
| Vision-R1* | 27.3 | 51.0 | 80.1 | 72.6 | 24.3 | 15.2 |
| PixelReasoner* | 52.7 | - | 80.6 | - | 23.5 | 23.4 |
| LVR* | 52.7 | 52.5 | 81.7 | 72.0 | - | - |
| CoVT | 62.7 | 56.0 | 78.0 | 77.3 | 22.7 | 24.6 |
| SkiLa | 45.3 | 53.0 | 84.3 | 75.3 | 21.5 | 24.1 |
| Qwen2.5-VL-7B | 61.3 | 55.5 | 77.5 | 74.0 | 24.3 | 25.6 |
| Qwen2.5-VL-7B (interleaved) | 64.7 (+3.4) | 58.0 (+2.5) | 79.1 (+1.6) | 75.7 (+1.7) | 25.5 (+1.2) | 25.8 (+0.2) |
| AnchorSim-7B (Ours) | 66.0 (+4.7) | 59.4 (+3.9) | 82.7 (+5.2) | 78.3 (+4.3) | 26.0 (+1.7) | 27.5 (+1.9) |
| Model | CV-Bench 2D | CV-Bench 3D | SAT-R | VSP | CoMT | Avg |
|---|---|---|---|---|---|---|
| Vision-R1* | 8.0 | 37.6 | - | - | - | - |
| CoVT | 76.0 | 84.7 | 64.7 | 6.0 | 27.8 | 48.6 |
| SkiLa | 76.7 | 74.4 | 56.7 | 4.3 | 32.3 | 47.5 |
| Qwen2.5-VL-7B | 76.6 | 73.3 | 59.7 | 9.3 | 39.5 | 48.9 |
| Qwen2.5-VL-7B (interleaved) | 76.7 (+0.1) | 77.3 (+4.0) | 65.3 (+5.6) | 74.8 (+65.5) | 56.3 (+16.8) | 59.7 (+10.8) |
| AnchorSim-7B (Ours) | 80.4 (+3.8) | 86.4 (+13.1) | 68.7 (+9.0) | 78.0 (+68.7) | 58.0 (+18.5) | 62.4 (+13.5) |
Table 1. Comparison across 9 visual reasoning benchmarks. Best in bold underline, second-best in bold. Numbers in parentheses are gains over Qwen2.5-VL-7B. Avg uses BLINK Avg and the CV-Bench mean. *Results are reported from prior work.
AnchorSim-7B has the best average and ranks first on 7 of the 9 benchmarks and second on BLINK and V*.
@inproceedings{luo2026anchorsim,
title={Mental Imagery That Matters: Curing Latent Collapse in Multimodal Reasoning},
author={Luo, Mi and Dugad, Rohan and Xue, Zihui and Dimakis, Alex and Grauman, Kristen},
booktitle={Advances in Neural Information Processing Systems},
year={2026}
}