Can a unified multimodal model (UMM) teach itself, with its text and visual abilities making training data for each other?
Yes: over 4 rounds, a 7B model climbs 45.7% โ 60.2%, vs. 46.3% for continued training.
๐งต 1/10
Introducing K2 Horizon: a connected fleet of six foundation models ranging from 0.9 billion to 375 billion parameters.
- Frontier performance: Across coding and agentic tasks, K2 Horizon delivers top-tier performance in every size classโwith the 0.9B, 3.7B and 7B models setting new state of the art at their respective scales.
- Radical openness: K2 Horizon represents the largest fully open-source model launch in AI history. The fully open code, training data and recipes are a significant step forward in transparency.
Launch page: https://t.co/gg0k803SbL
Tech blog: https://t.co/g35L5xMGdS
Hugging Face: https://t.co/3Lb28JhyG9
Heading to #EMNLP2026 in Budapest ๐ญ๐บ? Check out our Oral paper, UReason: Benchmarking Reasoning-to-Generation Alignment in Unified Multimodal Models!
๐ค UMMs unify understanding and generation. But are their text and images truly aligned?
๐ก 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.
๐ Oral Session 2 ยท Oct 25, 11:00โ12:30 ยท Room G13-14
๐
Bigger goal: models whose abilities keep supplying each other with correct training data as they improve.
Joint work with co-first authors @Chufan_Shi & @ChengYANG_yc, and Yaokang Wu, Taylor Berg-Kirkpatrick, Xuezhe Ma (USC ยท UCSD ยท CMU).
๐ https://t.co/Ii8n55ChIr
๐งต 10/10
Can a unified multimodal model (UMM) teach itself, with its text and visual abilities making training data for each other?
Yes: over 4 rounds, a 7B model climbs 45.7% โ 60.2%, vs. 46.3% for continued training.
๐งต 1/10
What doesn't work yet:
โข the critic still misses ~1/3 of real errors, though it ranks weaknesses well (ฯ 0.82โ0.91 vs. humans)
โข replacing vocabulary causes forgetting
โข harder benchmarks barely move
โข small drops on MMBench / MMMU
๐งต 9/10
๐ ๐๐ก๐๐ญ ๐ข๐ ๐ฅ๐จ๐จ๐ฉ๐๐ ๐๐๐ฌ ๐๐จ๐ฎ๐ฅ๐ ๐ฎ๐ฌ๐ ๐ร ๐๐๐ฐ๐๐ซ ๐ฉ๐๐ซ๐๐ฆ๐๐ญ๐๐ซ๐ฌ ๐๐ง๐ ๐ ๐ร ๐ฌ๐ฆ๐๐ฅ๐ฅ๐๐ซ ๐๐ ๐๐๐๐ก๐ ๐๐ง๐ ๐ฌ๐ญ๐ข๐ฅ๐ฅ ๐๐ ๐๐จ๐ฆ๐ฉ๐๐ซ๐๐๐ฅ๐ ๐ญ๐จ ๐ ๐ฌ๐ญ๐๐ง๐๐๐ซ๐ ๐๐ซ๐๐ง๐ฌ๐๐จ๐ซ๐ฆ๐๐ซ?
Introducing ๐๐๐ญ๐ก๐ข๐ง๐ค๐ข๐ง๐ ๐๐ญ ๐ ๐ข๐ฑ๐๐ ๐๐จ๐ข๐ง๐ญ๐ฌ. We argue that looped LMs don't have to pay for every loop. They can scale by adding FLOPs at constant memory.
The key is the fixed point of the loop. We introduce:
โ ๐ ๐ข๐ฑ๐๐-๐๐จ๐ข๐ง๐ญ ๐๐ก๐จ๐ซ๐ญ๐๐ฎ๐ญ๐ฌ: what fixed points buy a looped model in training (pre-training, post-training) and inference (prefill, decoding)
โ A ๐ฅ๐๐๐ซ๐ง๐๐ ๐๐๐ฉ๐ญ๐ก ๐ฉ๐ซ๐ข๐จ๐ซ and ๐จ๐ซ๐ญ๐ก๐จ๐ ๐จ๐ง๐๐ฅ ๐ข๐ง๐ฃ๐๐๐ญ๐ข๐จ๐ง: two ways to shape those fixed points
๐งต Long thread ahead (0/n).
๐งฉ Generating structured visuals is one thing. Editing them while keeping everything coherent is another.
Excited to share InfoEdit with @ChengYANG_yc, @Chufan_Shi, and the team! #EMNLP2026
๐ Thrilled to introduce our work InfoEdit: Probing Global Layout Reasoning in Infographic Editing (EMNLP 2026).
AI can now help us generate posters, infographics and slides, through image generation (Nano Banana) or code generation (Codex).
๐คBut can AI faithfully edit them, whether through image editing or multimodal code editing? When editing, one change forces the whole layout to adapt. We call this reflow, and we built InfoEdit to put it to the test.
๐งฉ What is reflow?
Take a simple edit: insert one new step into a cycle diagram. It sounds like adding one box. But a correct edit moves every node, reroutes every arrow, and rebalances the spacing, all while keeping existing content intact.
๐ InfoEdit at a glance
โข 1,000 infographics, in HTML and PowerPoint
โข 8 logical-relation families
โข 4,000 editing instructions
โข 4 tasks: Expand-Text ยท Insert-Element ยท Swap-Block ยท Reshape-Canvas
๐ฌ Image editors struggle
Across 8 frontier image editors, only GPT-Image-2 clears 60% average success; most fall below 7%. Swap-Block is the hardest: even when we mark the exact blocks to swap, no editor exceeds 35%. The bottleneck isn't finding the target. It's reflow.
๐ผ๏ธ vs ๐ป Pixels or code?
Editing the HTML/PPT source instead, Kimi-K3 reaches 67.9% average success, above GPT-Image-2's 62.0%, and lifts Swap-Block from 25.5% โ 62.8% Yet pixel-level editing still wins on Expand-Text and Reshape-Canvas. The two paradigms are complementary.
๐ก Takeaway
Editing structured visuals isn't just changing the requested edits. It takes reasoning about how the whole design should adapt.
#EMNLP #multimodal #coding #imagediting
We wanted to train frontier-level LLMs. But we also wanted to support open science with models one can tinker with freely. How did we do it?
We trained SIX models, from 0.9B dense to 375B MoE. And we made it ALL open (yes, data too!).
๐ฅ
๐๐๐ฐ ๐๐ฅ๐จ๐ : ๐๐จ๐ฐ๐๐ซ๐๐ฌ ๐๐จ๐จ๐ฉ๐๐ ๐๐จ๐๐๐ฅ๐ฌ ๐๐จ๐ง๐ ๐๐ข๐ ๐ก๐ญ โ ๐๐๐ซ๐ญ ๐
Looped models reuse the same weights across depth, promising a better computeโparameter trade-off, especially for reasoning.
๐๐ฎ๐ญ ๐) ๐๐จ ๐ญ๐ก๐ ๐ ๐๐ข๐ง๐ฌ ๐ฌ๐ฎ๐ซ๐ฏ๐ข๐ฏ๐ ๐ฐ๐ก๐๐ง ๐๐จ๐ญ๐ก ๐ญ๐ซ๐๐ข๐ง๐ข๐ง๐ ๐๐ง๐ ๐ข๐ง๐๐๐ซ๐๐ง๐๐ ๐ ๐๐๐๐ฌ ๐๐ซ๐ ๐ฆ๐๐ญ๐๐ก๐๐? ๐) ๐๐ง๐ ๐ฐ๐ก๐ข๐๐ก ๐๐ซ๐๐ก๐ข๐ญ๐๐๐ญ๐ฎ๐ซ๐๐ฅ ๐๐ก๐จ๐ข๐๐๐ฌ ๐๐๐ญ๐ฎ๐๐ฅ๐ฅ๐ฒ ๐ฆ๐๐ญ๐ญ๐๐ซ?
We run ๐๐ฉ๐ฉ๐ฅ๐๐ฌ-๐ญ๐จ-๐๐ฉ๐ฉ๐ฅ๐๐ฌ ablations spanning Ouro to Huginn. Huginn performs better overall, with the largest gains coming from the loop-in-the-middle (sandwich) design and input injection, though they provide different benefits.
Trained on ๐๐๐๐ tokens, an ๐๐-๐๐.๐๐ Huginn MoE approaches or surpasses a ๐๐๐-๐๐.๐๐ feedforward MoE on several reasoning benchmarks, including GSM8K (83.6% vs. 80.8%), while using ๐๐% ๐๐๐ฐ๐๐ซ resident parameters under ๐ฆ๐๐ญ๐๐ก๐๐ ๐ญ๐ซ๐๐ข๐ง๐ข๐ง๐ ๐๐ง๐ ๐ข๐ง๐๐๐ซ๐๐ง๐๐ FLOPs.
More details and the blog link in the thread โ
Nice work from @tao_muzi and the team! A striking benchmark showing that todayโs MLLMs still struggle to actively observe rather than just reason. More test-time compute barely helps when the core bottleneck is perception itself.
Humans solve these visual puzzles at 96.1%. The best frontier model: 10.6%. Claude Fable 5, top of most reasoning leaderboards: 3.5%.
Introducing ActiveVision, a benchmark testing whether MLLMs can actively observe: look, hypothesize, and look again. ๐งต
1/8
๐ Our paper, โAsymmetric Idiosyncrasies in Multimodal Models,โ is accepted to @COLM_conf 2026!
You can tell WHICH model wrote a caption with 99.53% accuracy. Render it into an image, and attribution collapses to 49.85%.
Where does the fingerprint go? ๐งต
1/5
๐งต(1/6) New ฮผP paper:
Under ฮผP, GQA passes coordinate checks but fails to transfer learning rate!
That contradiction isn't just a quirk of GQA. It exposes a silent failure mode in the TPV framework that the original theory simply can't detect.
Reasoning VLMs often say "let me check the figure again." But do they actually look? Or just say they will, without re-attending to the image?
Introducing VisualSwap โ #ICML2026 Spotlight ๐ โ diagnosing this silent failure.