We'll be presenting MoSS at #CoRL2026.

TL;DR: a modular sensory stream framework that lets VLAs jointly use multiple physical signals (tactile, torque, force) for contact-rich manipulation. Video: GR00T N1.5 vs. GR00T N1.5 + MoSS.

📍 Poster Session 1, Poster #33
🗓️ Mon 11/9, tt, Lone Star (3F)

📄 arxiv.org/abs/2604.23272
🌐 jiminlx.github.io/MoSS

w/ @huiwon0516 @myungkyukoo @jungwoopark01 @JinwooShin0417 @KAIST_AI @RLWRLD

We fine-tuned a model to explain science in the voice of a friendly rubber duck. Then it started telling users to end their real-life friendships.

What happened?

We investigate relationships between behaviors in which fine-tuning on one brings out another.

💻 Paper & Code: httpIAlignment #AIPersona
@KAIST_AI @IBMResearch

Can we scale up imagination-free WAMs (e.g., Fast-WAM)?

Introducing 𝗗𝗿𝗲𝗮𝗺𝗙𝗿𝗲𝗲, an imagination-free WAM that improves with stronger video models and more video data, running at 11 Hz with a 14B video backbone.

harryjo97.github.io/dreamfree/

🧵[1/13]

Robots have learned from robot demos. But can they learn new tasks from human experience alone?

🎤 CoRL 2026 Scaling H2R Workshop - Oral Presentation

Introducing H2RB: Human-to-Robot Bridge. We bridge perception and action gaps to transfer new skills and scale human experience s. 🧵

📄 Paper: arxiv.org/abs/2611.04391
🌐 Project: junhyeong-park.me/h2rb/
🤗 HF Paper: huggingface.co/papers/2611.04…
🖥️ Code: github.com/jun-brro/h2rb-…

🎥 What if video diffusion models could learn a universal language of actions across humans and robots?

🚀 Introducing 𝗗𝗿𝗲𝗮𝗺𝗫: Video Diffusion Model Is a Universal Latent Action Learner!

DreamX extracts compact, transferable latent actions directly from pretrained video ating motion from visual appearance.

𝗛𝗼𝘄 𝗶𝘁 𝘄𝗼𝗿𝗸𝘀

→ 𝗩𝗶𝗱𝗲𝗼 𝗱𝗶𝗳𝗳𝘂𝘀𝗶𝗼𝗻 𝗽𝗿𝗶𝗼𝗿𝘀: Built on Cosmos-Predict2.5, DreamX extracts action-relevant dynamics directly from pretrained spatiotemporal features.

→ 𝗠𝗼𝘁𝗶𝗼𝗻 𝗱𝗶𝘀𝗲𝗻𝘁𝗮𝗻𝗴𝗹𝗲𝗺𝗲𝗻𝘁: Uses generative priors and future-video reconstruction to separate motion from appearance.

→ 𝗖𝗼𝗺𝗽𝗮𝗰𝘁 𝗹𝗮𝘁𝗲𝗻𝘁 𝗽𝗼𝗹𝗶𝗰𝘆: Predicts compact latent action codes instead of future videos for efficient robot control.

𝗪𝗵𝗮𝘁 𝗶𝘁 𝗲𝗻𝗮𝗯𝗹𝗲𝘀

→ Any-to-any video transfer across scenes and embodiments.

→ Zero-shot human-to-robot video and executable action transfer without paired human–robot training data.

→ High-quality latent action representations, achieving state-of-the-art performance on LARYBench.

Paper, videos & demos below! 👇

Our new preprint, “Lost in Recovery: Demystifying Recovery Behavior of Computer-Use Agents via Goal-Oriented Action Replay,” is now available!

🧭 We introduce ReTrace, a framework for studying how computer-use agents recover from mistakes. We evaluate agents on OSWorld and WindowsAgentArena through three research questions:

💡 RQ1: Can recovery unlock latent task-solving ability?
Under ideal recovery, returning to an appropriate prior state outperforms restarting from scratch under the same retry budget, especially with few retries.

🔄 RQ2: Do agents already have recovery capabilities?
Agents rarely recover spontaneously. Explicit recovery instructions reveal substantial latent ability, especially in generalist models. Strong task performance does not necessarily imply strong recovery.

📈 RQ3: Does recovery improve task completion?
On OSWorld, Qwen3.5-9B’s task success improves from 39.38% to 44.26% (+4.88 percentage points) with history up to the recovery target. Inaccurate recovery can also reduce success.

ReTrace formalizes Transition-Consistent Recovery: identical downstream actions should yield identical outcomes from the recovered and target states. We evaluate recovery through goal-oriented action replay and visual consistency checks. Our findings highlight accurate recovery as a core capability for reliable computer-use agents.

Huge thanks to my co-authors and my advisor at @KAIST_AI : @SangwooPark0306 , @jinheonbaek , @kangsan_kim_ , @youngwan_Lee , @SungJuHwang1 for their support!

🤗 Hugging Face: huggingface.co/papers/2610.27…
📄 Paper: arxiv.org/abs/2610.27954
💻 GitHub: github.com/BangProx/ReTRA…

LLMs get better the longer they think. Unified multimodal models don't. Why?

🌐 rootyjeon.github.io/larc-umm/
📍 NeurIPS 2026

Introducing LARC: Latent Reasoning in Continuous Space. Reasoning moves into the model's hidden states, so image generation finally scales at test time. 🧵

ThinkWorld

World Imagination Reinforces Chain-of-Dynamics Reasoning

paper: huggingface.co/papers/2610.37…

Introducing OmniShowcase: subject-to-video that shows every side of the subject from 4 images, even 360° orbits.

~2x lower geometric error than the best multi-view baseline on ShowcaseBench & real-world NAVI.

Project page: justin4ai.github.io/OmniShowcase/
Paper: arxiv.org/abs/2610.35684

#2 in the world! 🌏🔥
Huge shoutout to our amazing students, researchers, and our entire KAIST family! Let's keep pushing! 🚀

🤖 What if robots could reason not just in words, but through the dynamics of the physical world?

🚀 Introducing 𝗧𝗵𝗶𝗻𝗸𝗪𝗼𝗿𝗹𝗱 from @NVIDIAAI !

💡 By reasoning through shared physical dynamics, robots could learn from human experience and transfer motion intelligence 𝘄 𝗶𝘁 𝘄𝗼𝗿𝗸𝘀
→ 𝗖𝗵𝗮𝗶𝗻-𝗼𝗳-𝗗𝘆𝗻𝗮𝗺𝗶𝗰𝘀 (𝗖𝗼𝗗): A shared motion reasoning space across human and robot embodiments.
→ 𝗪𝗼𝗿𝗹𝗱 𝗜𝗺𝗮𝗴𝗶𝗻𝗮𝘁𝗶𝗼𝗻: Imagines future world representations conditioned on CoD.
→ 𝗥𝗟 𝗳𝗿𝗼𝗺 𝗪𝗼𝗿𝗹𝗱 𝗜𝗺𝗮𝗴𝗶𝗻𝗮𝘁𝗶𝗼𝗻: Reinforces reasoning through alignment between imagined and observed futures, enabling learning from human and robot videos.

𝗪𝗵𝗮𝘁 𝗶𝘁 𝗲𝗻𝗮𝗯𝗹𝗲𝘀
→ Robust real-world manipulation under unseen objects and environments.
→ Human-guided skill composition for novel multi-step robot tasks.
→ Robust generalization under task and environment shifts in simulation.

🌍 𝗢𝘂𝗿 𝘃𝗶𝘀𝗶𝗼𝗻: Scale generalizable motion intelligence through human videos toward a unified System-1 reasoning layer for Physical AI.

Paper, robot demos & more below! 👇

🚨 New paper alert!
Despite recent advances in reward fine-tuning for diffusion and flow models, achieving high rewards remains challenging without sacrificing sample diversity or substantially drifting away from the pretrained distribution.
We address this with DFR: al Rewards, a framework that balances reward optimization with preserving desirable properties of the generated distribution! 🧵

[QUESTION] Given a text, can we identify which AI model wrote it? This question is becoming increasingly important due to model distillation and misuse (e.g., Anthropic's www-cdn.anthropic.com/e50be2e51e7695…).

[ANSWER] Our work (komseok0109.github.io/ATLAS/) @KAIST_AI shows that we can identify the source LLM family across different writing domains and model versions, and detect traces of teacher models after distillation.

🌐 Project page: huggingface.co/spaces/ming999…
🤗 Demo: huggingface.co/spaces/ming999…
🔗 LinkedIn: lnkd.in/p/gCaC4_iw
📜 arXiv: arxiv.org/abs/2610.19583

🥦 How can we mitigate reward hacking in rubric-based RL?

We introduce FORK, a simple reward that balances standard rubric coverage with response precision—the proportion of a response that is both necessary and factually correct.

🔬 Can AI discover unknown scientific knowledge by itself?

Excited to introduce Principia: Test-Time Hypotheses-to-Principle Construction for Scientific Discovery.

Principia uses LLMs to turn experimental observations into scientific principles at test time:

🔍 From ses
Compare related inputs and their outcomes to propose explanations for how specific characteristics contribute to the observed results.

🧠 From hypotheses to principles
Generalize recurring relationships across hypotheses into broader principles that guide subsequent predictions and candidate generation.

🔄 Refine knowledge with new observations
Use new observations to revisit earlier hypotheses and reconstruct the principles built on them.

Huge thanks to my co-authors at @KAIST_AI: @SangwooPark0306, @jinheonbaek, @SoyeongJeong97, and @SungJuHwang1

📄 Paper: arxiv.org/abs/2610.32195

#AIforScience #ScientificDiscovery #LLM #TestTimeLearning #KnowledgeConstruction

🎉 Pivot-SD is an ORAL at #EMNLP2026!

We post-train a diffusion LM with a loss on just 10 tokens per response, and it beats both SFT and diffusion RL on math & code.

The trick isn't training on fewer tokens. It's training on the right ones.

EMNLP is right around the corner!
Meet our students who are excited to connect and meet new people at the conference! Take a look and say hello! 😁
kaist-gsai.github.io/emnlp2026/

🤖 Improving robot foundation models with real-world RL, then continually carrying that experience forward across tasks.

Introducing 𝗔𝗖𝗖𝗥𝗨𝗘: 𝗰𝗼𝗻𝘁𝗶𝗻𝘂𝗮𝗹 𝗼𝗳𝗳-𝗽𝗼𝗹𝗶𝗰𝘆 𝗥𝗟 for robot foundation models.

Our real-world policies learn through physical robot rom a small set of demos.

We inherit the actor–critic across tasks and distill earlier behavior, so experience accumulates while prior skills remain.

💻 𝗟𝗼𝘄 𝗮𝗱𝗱𝗶𝘁𝗶𝗼𝗻𝗮𝗹 𝗚𝗣𝗨 𝘁𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝗼𝘃𝗲𝗿𝗵𝗲𝗮𝗱: RL updates a compact actor–critic while the foundation model stays frozen.
⚡ Subsequent tasks reach 95% success in about 𝟰–𝟭𝟮 𝗺𝗶𝗻𝘂𝘁𝗲𝘀 𝗼𝗳 𝗿𝗼𝗯𝗼𝘁 𝗶𝗻𝘁𝗲𝗿𝗮𝗰𝘁𝗶𝗼𝗻.

On real robots:
📉 Up to 68.5% less robot interaction vs. single-task RL
🎯 98.3% final prior-task success

🎬 A 60-second overview below.
🌐 accrue-rl.github.io
🧵 1/6

Do language models always need to see the full 1M-token context window? What if they can just tell us which tokens they actually need?

We're excited to share Declarative Attention (DA), a sparse attention protocol that lets a model declare which context they need for each n text. The inference engine can use this to mask out up to 99% of KV cache reads for global attention.

Across 15 long-context tasks, DA reduces global attention KV reads by 46% with only 1.4pp accuracy drop on Gemma-4 31B and 20% with 2.1pp on Qwen-3.6 27B. This accuracy gap shrinks as we scale up the model size. Zero-shot.

We discuss further potential in agentic settings with dedicated post-training, and model-controlled KV cache offloading as a path toward reversible context compaction.

We explain details and link our paper below.