Heading to #EMNLP2026 in Budapest ππΊ? Check out our Oral paper, UReason: Benchmarking Reasoning-to-Generation Alignment in Unified Multimodal Models!
π€ UMMs unify understanding and generation. But are their text and images truly aligned?
π‘ Our finding: UMMs usually get the reasoning right, but the image often doesn't follow it. Even more surprising: if we drop the reasoning and give the model only the final prompt it wrote itself, every model we tested does better, by a large margin.
π Oral Session 2 Β· Oct 25, 11:00β12:30 Β· Room G13-14
π
π¬ To diagnose this, we use reasoning-guided image generation: the model reasons in text first, then generates the image.
π A simple example:
Prompt: "A table had 10 red apples. A baker took 5, two children each grabbed 1, and the farmer threw away 1 rotten apple. Show the table now."
β’ Reasoning-guided generation: the model reasons correctly (10 β 5 β 1 β 1 β 1 = 2 apples left), then drawsβ¦ a bowl full of apples. β
β’ De-contextualized generation: keep only the refined prompt the model itself wrote, drop the rest of the reasoning, and it draws exactly 2. β
Same model, same visual intent. Only the context differs.
π It's not just one example. Across 8 open-source UMMs on UReason:
β’ Direct generation (no reasoning): only 4β8% accuracy
β’ Reasoning-guided helps every model: +0.4 to +19.2 points
β’ But de-contextualized beats reasoning-guided for every model, by up to +44.8 points (Bagel: 17.8% β 62.6%)
π§ So what goes wrong?
Not the reasoning itself: Bagel's reasoning chains are 93.4% correct. Instead, attention analysis suggests contextual interference: during generation, the intermediate thoughts compete with the final visual specification.
π― Takeaway: A unified architecture doesn't guarantee aligned modalities. UReason serves as a litmus test, and a challenging benchmark, for building more tightly aligned UMMs.
π This work could not have been done without my co-first author @Chufan_Shi, with an amazing team including @brian_cyan1@yaokangwu@tao_muzi@lemonnoel52@ivn1e @ Yong Liu @MaxMa1987@BergKirkpatrick
π Paper: https://t.co/VPp9Z6N9Qg
π Project: https://t.co/g6Zk0uXbtH
π» Code: https://t.co/qXvdXYJknW
See you in Budapest! ππΊ
#multimodal #imagegeneration #reasoning #LLM #UMM #EMNLP2026
Heading to #EMNLP2026 in Budapest ππΊ? Check out our Oral paper, UReason: Benchmarking Reasoning-to-Generation Alignment in Unified Multimodal Models!
π€ UMMs unify understanding and generation. But are their text and images truly aligned?
π‘ Our finding: UMMs usually get the reasoning right, but the image often doesn't follow it. Even more surprising: if we drop the reasoning and give the model only the final prompt it wrote itself, every model we tested does better, by a large margin.
π Oral Session 2 Β· Oct 25, 11:00β12:30 Β· Room G13-14
π
Excited to share our work InfoEdit at EMNLP 2026 β exploring whether AI can truly reason about global layout reflow when editing structured visuals like infographics, slides, and posters. π