Mila's EMNLP 2026 paper: thinking-images boost cross-view spatial reasoning
aagrawalAA · x · 2026-10-24
A Mila / Université de Montréal / McGill / ServiceNow AI Research paper accepted to EMNLP 2026 trains unified multimodal models to generate a "thinking-image" mid-reasoning for cross-view spatial reasoning (answering spatial questions from two views of one 3D room).
Key findings:
- Generating a thinking-image beats no-thinking and text-CoT baselines, in-domain and on five real-world out-of-domain benchmarks.
- Standard training lets models ignore their own image; the proposed View Dropout makes the image causally used, not decorative.
- Of three image types, only the panorama is both informative and learnable, and generalizes best out-of-domain.
The team also released an 8K cross-view QA dataset, code, and an arXiv paper. Poster: Oct 26, 16:45–18:15, Hall H.
More from Research
- Dessa's dialog-only voice training isn't new: dGSLM did it in 2022 with just 2k hours — rdesh26 · 2026-10-24
- NYU talk argues 30+ years of consciousness research may be partly wrong — birchlse · 2026-10-24
- Psycholinguistic analysis finds Claude Opus 5 the most out-of-distribution model tested — gleech · 2026-10-24
- Google Research tests 180 agent setups: multi-agent hurts sequential tasks by 39-70% but boosts parallel ones 80.9% — ombre-ai · 2026-10-24
- Loopholing: pushing self-conditioning all the way in discrete diffusion language models — NandoDF · 2026-10-24
- Review-it tool suggests 10 best-fit journals and conferences for your paper — Faheem_uh · 2026-10-24