MTS @Kimi_Moonshot, PhD student - MMLab@NTU. Prev @Tsinghua_Uni.
Joined October 2023
- Tweets179
- Following363
- Followers257
- Likes350
Yuhao Dong retweeted
🎉 Excited to share that our paper Apple-π has been accepted to NeurIPS 2026!
Looking forward to presenting our work and meeting everyone in Sydney 🇦🇺.
If you’re also attending NeurIPS and interested in our work, feel free to reach out — would love to chat and connect!
See you in Sydney! 🚀
#NeurIPS2026 #PhysicalAI #EmbodiedAI
Yuhao Dong retweeted
🎯Visual Faithfulness via Rubric-Based RL🎯
There is a credit-assignment failure in multimodal post-training, where visual facts are not grounded. 🧭V-Rubrics🧭 is here to fix with *auto rubric as reward* #EMNLP
- Project: shulin16.github.io/v-rubrics…
- Code: github.com/shulin16/v-rubric…
Vision-language models can produce fluent, plausible answers while still getting a visual fact or intermediate inference wrong.
We view this as a credit-assignment problem: a single outcome reward may tell us whether the final answer is acceptable, but it cannot identify which visual claims are grounded, which reasoning steps are valid, or which instructions were missed.
🚀 Excited to introduce V-Rubrics: Visual Faithfulness via Rubric-Based Reinforcement Learning, now accepted to @emnlpmeeting 2026 (main track).
V-Rubrics changes the unit of supervision. Instead of assigning one score to an entire response, we decompose it into atomic criteria for:
👁️ 𝗩𝗶𝘀𝘂𝗮𝗹 𝗙𝗮𝗶𝘁𝗵𝗳𝘂𝗹𝗻𝗲𝘀𝘀
🧠 𝗥𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴 𝗖𝗼𝗻𝘀𝗶𝘀𝘁𝗲𝗻𝗰𝘆
📋 𝗜𝗻𝘀𝘁𝗿𝘂𝗰𝘁𝗶𝗼𝗻 𝗙𝗼𝗹𝗹𝗼𝘄𝗶𝗻𝗴
These rubrics provide structured partial credit. When supporting evidence spans can be aligned, we further localize each component’s credit to the relevant response prefix during GRPO training.
📄 Paper: arxiv.org/abs/2608.25580
🌐 Project: shulin16.github.io/v-rubrics…
💻 Code: github.com/shulin16/v-rubric…
🤗 Dataset: huggingface.co/datasets/v-ru…
#VLM #MultimodalAI #ReinforcementLearning
Yuhao Dong retweeted
Vision-language models can produce fluent, plausible answers while still getting a visual fact or intermediate inference wrong.
We view this as a credit-assignment problem: a single outcome reward may tell us whether the final answer is acceptable, but it cannot identify which visual claims are grounded, which reasoning steps are valid, or which instructions were missed.
🚀 Excited to introduce V-Rubrics: Visual Faithfulness via Rubric-Based Reinforcement Learning, now accepted to @emnlpmeeting 2026 (main track).
V-Rubrics changes the unit of supervision. Instead of assigning one score to an entire response, we decompose it into atomic criteria for:
👁️ 𝗩𝗶𝘀𝘂𝗮𝗹 𝗙𝗮𝗶𝘁𝗵𝗳𝘂𝗹𝗻𝗲𝘀𝘀
🧠 𝗥𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴 𝗖𝗼𝗻𝘀𝗶𝘀𝘁𝗲𝗻𝗰𝘆
📋 𝗜𝗻𝘀𝘁𝗿𝘂𝗰𝘁𝗶𝗼𝗻 𝗙𝗼𝗹𝗹𝗼𝘄𝗶𝗻𝗴
These rubrics provide structured partial credit. When supporting evidence spans can be aligned, we further localize each component’s credit to the relevant response prefix during GRPO training.
📄 Paper: arxiv.org/abs/2608.25580
🌐 Project: shulin16.github.io/v-rubrics…
💻 Code: github.com/shulin16/v-rubric…
🤗 Dataset: huggingface.co/datasets/v-ru…
#VLM #MultimodalAI #ReinforcementLearning
Yuhao Dong retweeted
Excited to see demo-driven in-context learning take off. Our EMNLP paper #DemoICL studies exactly this on the video understanding side: learning dynamic, novel procedural knowledge from in-context video demonstrations, beyond static internal knowledge.
📄 arxiv.org/pdf/2602.08439
💻 github.com/dongyh20/Demo-ICL
Congrats to the @SkildAI team!
Our newly @emnlpmeeting paper #DemoICL tackles Demo-driven Video In-Context Learning, which aims to learn dynamic, novel procedural knowledge from *in-context video demonstrations*
- Paper: arxiv.org/pdf/2602.08439
- Code: github.com/dongyh20/Demo-ICL
Yuhao Dong retweeted
Learning from video demonstrations 👍
Our paper Demo-ICL made some initial explorations for video understanding task earlier this year, and now has been officially accepted to @emnlpmeeting 2026. Congrats to all the authors 🎉
- Paper: arxiv.org/pdf/2602.08439
- Code: github.com/dongyh20/Demo-ICL
Congrats to the @SkildAI team!
Our newly @emnlpmeeting paper #DemoICL tackles Demo-driven Video In-Context Learning, which aims to learn dynamic, novel procedural knowledge from *in-context video demonstrations*
- Paper: arxiv.org/pdf/2602.08439
- Code: github.com/dongyh20/Demo-ICL
Yuhao Dong retweeted
Congrats to the @SkildAI team!
Our newly @emnlpmeeting paper #DemoICL tackles Demo-driven Video In-Context Learning, which aims to learn dynamic, novel procedural knowledge from *in-context video demonstrations*
- Paper: arxiv.org/pdf/2602.08439
- Code: github.com/dongyh20/Demo-ICL
Yuhao Dong retweeted
Physical AI will scale with egocentric experience, just as LLMs scaled with tokens.
HOMIE Gen2 is our next step toward making real-world multimodal experience scalable for robots, world models and physical agents, just like Project Index @Figure_robot
The Experience Scaling Law is just getting started!
Vision is all about perception, perception, perception.
We present this valuable benchmark to the community for OOD evaluation, and hope we will not see a model suddenly scoring 75% in a few weeks😅
We are releasing PerceptionBench, a benchmark that isolates visual perception and evaluates it as a set of atomic capabilities - discovered from how today's models fail, rather than defined in advance.
From frontier-model failures across 42 benchmarks, we derive 10 atomic perceptual capabilities and construct 3,000 verified questions, each isolating a single capability and answerable by looking, with no reasoning or external knowledge required.
Blog: kimi.com/blog/perception-ben…
GitHub: github.com/MoonshotAI/Percep…
Hugingface: huggingface.co/datasets/moon…
Yuhao Dong retweeted
a great summary of the past year's work.
time to move on and carry on 🫡
Releasing the model weights and technical report of Kimi K3.
Kimi K3 is our most capable model: a 2.8T MoE model with native visual understanding and a 1M-token context window.
New model architecture: 2.5x the intelligence per unit of compute, not just more params.
Alongside Kimi K3, we're opening up more of the stack behind it — high-performance attention kernels, MoE communication library, and infrastructure for running agent environments at scale.
Model weights: huggingface.co/moonshotai/Ki…
Tech report: github.com/MoonshotAI/Kimi-K…
Tech blog: kimi.com/blog/kimi-k3
Yuhao Dong retweeted
Releasing the model weights and technical report of Kimi K3.
Kimi K3 is our most capable model: a 2.8T MoE model with native visual understanding and a 1M-token context window.
New model architecture: 2.5x the intelligence per unit of compute, not just more params.
Alongside Kimi K3, we're opening up more of the stack behind it — high-performance attention kernels, MoE communication library, and infrastructure for running agent environments at scale.
Model weights: huggingface.co/moonshotai/Ki…
Tech report: github.com/MoonshotAI/Kimi-K…
Tech blog: kimi.com/blog/kimi-k3
Yuhao Dong retweeted
Are video models Aristotle—or Newton? 🍎
Do they actually reason with physical laws, or only generate motion that looks plausible?
Apple-π is the first law-grounded benchmark: 400 videos, 10 mechanics tasks, Perception→Formulation→Deduction.
🌐 21yrm.github.io/Apple-PI-hom…
Yuhao Dong retweeted
Exciting update: Kimi K3 has landed at #4 on the Agent Arena leaderboard, matching Claude Opus 4.8 and GPT-5.6 Sol. If Kimi K3's weights are released on schedule by July 27, it will become the #1 open-weight model.
This release marks a major leap in agentic performance over Kimi K2.7 Code (#23 to #4). Based on 8K+ live agentic sessions, Kimi K3 leads on confirmed task success rate (#1). It also posts a strong +20.6% on praise vs. complaint (#3). It currently lags the field in steerability (#14) and bash recovery (#17).
Agent Arena measures models on millions of real-world, long-horizon agentic tasks. Models get web search, filesystem, and terminal tools to complete complex workflows: writing code, creating slide decks, researching the web, building apps, and analyzing documents.
We use causal tracing methodology to measure a model's net improvement, which indicates how much it improves outcomes relative to the average model.
Here's a primer on the 5 signals:
User-satisfaction proxies
- Confirmed Success: an explicit "yes that worked" feedback from the user
- Praise vs. Complaint: implicit sentiment in users reactions
- Steerability: can the model course-correct when you push back?
Tool-use proxies
- Bash Recovery: how it recovers from CLI errors (primary signal for tool use)
- Tool Hallucination: does it call tools that don't exist
Below we break down how Kimi K3 scored across the 5 signals, drawn from tasks submitted by a global community of users.
Congrats @Kimi_Moonshot on another big milestone!
Big news: Kimi-K3 by @Kimi_Moonshot is now #1 in the Frontend Code Arena with 1679 pts, surpassing Claude Fable 5.
This is a 17-place jump from Kimi-k2.6 (#18 -> #1).
In Frontend, Kimi-K3 ranked #1 in 6 of 7 domains: Brand & Marketing, Reference-Based Design, Data & Analytics, Consumer Product, Simulations, and Content Creation Tools, landing #2 only in Gaming behind Fable 5.
The full model weights will be released by July 27.
Congrats to the @Kimi_Moonshot team on this major milestone!
Yuhao Dong retweeted
windows-xp.kimi.site/
K3 发布前两天内部 vibe 出来的 win xp 模拟器，功能相当完整，有“千千静听”可以听歌，有“QQ2005”可以聊天，朋友们会在你的QQ空间里聊超女，新浪首页在报道神州6号……
在内部传开的那个下午，站起身来，发现一半的同事都在用 XP，瞬间有一些恍惚，此情此景，仿佛回到了学校的微机课，又像是当年的网吧。我回想起 20 年前中考后的夏天，每天在家听着《宁夏》和《突然很安静》，在不舍得开空调的炎热的下午，在家里的老电脑上刷冒险岛……
20 年弹指一挥间，感谢 K3，能在这个剧变的时代，让我怀念一下年轻的自己。
Having pure faith in RL.
That is how we achieve this.
Kimi K3 scores 57 on the Artificial Analysis Intelligence Index. Its intelligence is comparable to Opus 4.8 and GPT-5.5 but remains behind Fable 5 and GPT-5.6 Sol. Moonshot AI has expressed plans to release the 2.8T parameter model's weights, which would make it the leading open weights model
Key results:
➤ Strong agentic task performance: @Kimi_Moonshot's Kimi K3 reaches an Elo rating of 1668 on GDPval v2. This is a marked improvement over K2.6’s 1190, surpassing GLM-5.2 (1514), GPT-5.5 (1494), and Claude Opus 4.8 (1600). However, it still lags behind Claude Fable 5 (1760). Kimi K3 also scores an impressive 53% and takes the #1 position on AutomationBench-AA, our implementation of Zapier’s Agentic SaaS workflow evaluation.
➤ Second-highest performance on AA-Briefcase (agentic knowledge work): On our private long-horizon knowledge work evaluation, Kimi K3 reaches an overall Elo of 1547, +732 points from Kimi K2.6 and behind only Claude Fable 5. It is well-rounded: its rubric scoring and analytical quality almost reach Claude Fable 5’s scores, while GPT-5.6 Sol continues to outperform other leading models on presentation quality.
➤ Set to lead open weights models once weights are released: Moonshot AI has not yet released the weights but expressed plans to do so. Once available, Kimi K3 would clearly lead other open weights models including GLM-5.2 (51) and DeepSeek v4 Pro (44). However, at 2.8T parameters, it is significantly larger than its open weights peers (eg. GLM-5.2 at 753B params and DeepSeek V4 Pro at 1.6T), as well as the Kimi K2 to K2.6 models (1T params).
➤ Cost per task ($0.94) is similar to GPT-5.6 Sol ($1.04), ~1/2 the price of Opus 4.8 ($1.80) and higher than open weights peers: Moonshot AI’s pricing for K3 is significantly higher than their K2 pricing (K3’s output token price is $15/1M tokens while K2.6 was $4). This positions the model as cheaper on a cost per task basis than Opus 4.8, similar to GPT-5.6 Sol ($1.04) and more expensive than open weights peers, GLM-5.2 ($0.32) and DeepSeek V4 Pro ($0.04)
➤ Improved token efficiency alongside higher intelligence: Kimi K3’s token usage on the Artificial Analysis Intelligence Index decreased significantly, using 21% fewer output tokens than K2.6. The new model used approximately 132M output tokens to complete all nine evaluations, compared to approximately 166M for K2.6, while achieving higher scores.
➤ Native multimodal capabilities: Kimi K3, like K2.6, is released with native image and text multimodal input. If weights are released, this will position Kimi K3 as one of the leading open weights models with multimodal input capabilities
Other model details:
Context window: 1M
Size: 2.8T total parameters
Pricing: The first-party API is priced at $3.00/$15.00 per 1M input/output tokens, with cached input discounted 90% to $0.30 per 1M tokens.
Modality: Native multimodal input supports text and images, and the model remains text-only for output.
Accessibility: Accessible at launch through Moonshot’s first party API. Model weights are not yet released but Moonshot AI has expressed plans to do so.
Keen + Knowledgeable + Key-focused = K3
Introducing Kimi K3: Open Frontier Intelligence
🔹 2.8 Trillion Parameters, 1 Million Context, Native Multimodal
🔹 Kimi Delta Attention enables up to 6.3x faster decoding in million-token contexts
🔹 Attention Residuals deliver ~25% higher training efficiency at <2% additional cost
🔹 Built for long-horizon agentic coding and self-evolving workflows
Kimi K3 is now live on on Kimi.com, Kimi Work, Kimi Code, and the Kimi API.
Open Weights by July 27, 2026.
🔗 API: platform.kimi.ai
🔗 Tech blog: kimi.com/blog/kimi-k3
