airlsyn

LLM & RL

[airlsyn @airlsyn]

能否支持 claude science 的逻辑呢

[xiaoyu @zhongxingyuyes]

Kun的设计稿也可以是原型交互稿

[OpenBMB @OpenBMB]

1/5 MiniCPM-V 4.6 (1.3B) is now live 🚀🚀
High-res visual processing, optimized for consumer-grade and mobile hardware. We’ve leveraged the latest LLaVA-UHD v4 technique to cut vision encoding costs by 55%, enabling native edge deployment with extreme efficiency.
🔥 Beats Gemma4-E2B-it and Qwen3.5-0.8B across key multimodal and Artificial Analysis benchmarks — scoring higher than Qwen3.5-0.8B using just 2.5% of its token budget.
⚡ TTFT (75.7ms) 2.2x Faster than Qwen3.5-0.8B even with 3136² high-res images.
🏗️ ~1.5x Token Throughput compared with Qwen3.5-0.8B on a single RTX 4090.
Try the model here:
🤗 Hugging Face:
https://t.co/CEkwKMSBwc
💻 GitHub:
https://t.co/iYDxpa52tn
🔭 Modelscope:
https://t.co/CHflKPLbvK
🌐 Web Demo:
https://t.co/DYUrtD0YzM
📱 App Demo:
https://t.co/SL7IOhm6zv

[airlsyn @airlsyn]

@zhijianliu_ Will you opensource the training code?

[airlsyn @airlsyn]

@oran_ge ColaOS 给的自画像，太牛了

[Huiqiang Jiang @iofu728]

🌩️Introducing FlashQLA: high-performance linear attention kernels on TileLang.
⚡ 2-3× fwd, 2× bwd speedup.
💻 Purpose-built for agentic on your personal devices.

1. Gate-driven auto intra-card CP.
2. Hardware-friendly reformulation.
3. TileLang fused warp-specialized kernels.

[艾略特 @elliotchen100]

非常激动宣布，我们打磨了很久很久的更新，她终于来了，不仅有产品，还有 Methods，Benchmark，Usecases 和官网等。

具体更新：

1. EverOS - 新思路，一站式 Agent 创建，测试和集成的开源库，其中包含了多个 Methods，多个 Benchmarks 和很多种 Usecases。

主打一站式服务，用户可用现成的 Methods 魔改 Usecases 成自己的 Agents，然后再用 Benchmark 对其去测试，提升和迭代。

一站式全能王 ，很方便。

2. 新 Benchmark - EvoAgentBench 开源，可以用来测试你自己配置的 OpenClaw Agent，Hermes Agent 和任何 Agent。

当有了中立 benchmark 测试的时候，方才知道谁是好龙虾。

3. 官网更新 - 新版官网，愿景和感觉，配色和调性，品味和新意，拉爆全场，外加新场景，新产品，新交互。

一个字：夯。

4. Academy - 我们是一个学识很强的团队，所以这个板块以后会是我们发过的所有模型，论文，methods 和benchmark 相关。

学就完了。

5. EverMe - C 端的产品，在来的路上，现在已经放出小彩蛋，可以期待期待。

新面貌，新风向，新动力，依然开源。

https://t.co/AUbhZlNBD3

[艾略特 @elliotchen100]

新论文发布了，HyperMem，长对话的记忆架构。🤓📝

目标是让对话系统更好地保存，组织和检索长期记忆，避免传统 RAG 或普通图记忆把相关信息拆得过碎。

我们的论文结论是：通过用超边显式表示多元素之间的高阶关联，HyperMem 在 LoCoMo 基准上取得了 92.73% 的 LLM-as-a-judge 准确率，达到了 SOTA。

https://t.co/gDNNKofcEA

[Weichen FAN @W30259893]

🔥BIG Update🔥for "The Prism Hypothesis" (UAE)

✨Paper: https://t.co/BAIRUngMOo
✨Code: https://t.co/t2k5N1CDDq

🔠languages and visuals👓 can coexist in one representation.

Now support both latent and pixel-wise Generation
(🔥JIT is supported!🔥)

💫#UAE💫 gives a strong foundation for NEO-Unify:

Understanding and generation share one representation making raw-input end-to-end learning a natural path! 👏👏👏

[Charlie O'Neill @oneill_c]

If you could abstract away the messiness of prompting from OPSD, it's got everything you'd want in a general learning algorithm: on-policiness (as opposed to SFT), and dense reward signal (as opposed to RL). We're bullish on OPSD, but @part_harry_ does a great job outlining some of the contours of the aforementioned messiness here

[Alex Weers @a_weers]

Finally finished!

If you're interested in an overview of recent methods in reinforcement learning for reasoning LLMs, check out this blog post: https://t.co/SHUyFF4rvP

It summarizes ten methods, tries to highlight differences and trends, and has a collection of open problems

[yan5xu @yan5xu]

完全面向 cli 的 agent 设计，也是我觉得最适合 LLM 的 agent 设计！

[PsiACE @repsiace]

Use diagrams to explain agent context and the technology behind bub — all you need is one tape.

https://t.co/SWKUArtv0P

[PsiACE @repsiace]

在 bub 的设计里，有一个名为 tape 的无尽长纸带，它存储的就是全量的历史事实，并且自主进行上下文装配，没有依赖向量、memory 等机制

我们使用两个简单的机制让 agent 自主管理它：

- anchor ，用来标记当前的阶段和状态以及需要关注的信息。
- handoff，切换到某个 anchor 并屏蔽无关的历史上下文

任务开始，选取合适的历史和通过工具/技能获取的信息装配。任务完成，上下文的生命周期结束。

我们在群聊多人多 Agent 环境进行了技术验证并且运行了超过两周时间，由 agent 自主管理上下文和 handoff/anchor 时机。

[AK @_akhaliq]

MiniCPM-o-4.5 streaming app is out on Hugging Face

app: https://t.co/93Q98d5okK

[OpenBMB @OpenBMB]

🚀 MiniCPM-o 4.5 WebDemo is now more accessible than ever!

We’ve revamped our local deployment to make it easier for everyone to experience the power of our latest Omni model. Whether you are a researcher or a dev, we have a path for you:

🐳 Option 1: DockerZero-setup, cross-platform support (macOS/Linux/Windows). Just pull and run! 🔗 https://t.co/BM178FqqZJ

⚡ Option 2: One-click ScriptNative deployment for macOS/Linux. Memory-efficient and perfect for secondary development. 🔗 https://t.co/7YKHk9cTnZ

Give it a spin and let us know your thoughts! Your feedback helps us grow. 🙏 https://t.co/KzzgiGYhVr

#MiniCPMo #OpenSource #LLM #EdgeAI #AI

[Huaxiu Yao ✈️ ICML 2026 @HuaxiuYaoML]

Why do most LLM agents hit a wall?
They don’t accumulate skills.

Introducing SkillRL📚 — recursive skill-augmented reinforcement learning that lets agents learn skills from failure and evolve over time.

🔥A 7B model:
• +41% over GPT-4o
• ~20% fewer training tokens
• 33% faster convergence

SkillRL bridges raw experience → policy improvement by distilling trajectories into structured, co-evolving skills during RL.

Most agents forget.
SkillRL evolves. 🔄

📄 Paper: https://t.co/6VoxpGoPR6
💻 Code: https://t.co/qVDnIaci2K

Great work @richardxp888, Jianwen Chen, Hanyang Wang, @JiaqiLiu835914, @lillianwei423, @AiYiyangZ, and nice collab. w/ @__YuWang__, @XujiangZhao, Haifeng Chen, Zeyu Zheng, @cihangxie.

[Dayal Kalra @dayal_kalra]

Excited to share work from my internship at MSL @AIatMeta! 🚀

We analyze Critical Sharpness: a scalable curvature measure requiring only ~6 forward passes to analyze LLM training dynamics at scale.

We extend this measure to introduce Relative Critical Sharpness, which measures the relative curvature between two landscapes. We use this to answer a major practical question: How much pre-training data should we mix during fine-tuning to avoid catastrophic forgetting?
🧵
(1/n)

[OpenBMB @OpenBMB]

🥳 Introducing MiniCPM-o 4.5
The first full-duplex omni-modal LLM in open-source community 🎬🎙️

🔥 Key Highlights:
• Full-duplex Omni-modal Live Streaming: The model can see, listen, and speak simultaneously in a real-time conversation without mutual blocking
• Proactive Interaction: Moving beyond reactive QA to performing proactive interaction, such as initiating reminders
• Leading Performance: Scoring 77.6 on OpenCompass, it outperforms GPT-4o & Gemini 2.0 Pro in vision-language tasks with 9B params

The best part? You can experience all above on your PC!

#MiniCPM #OpenSource #MultimodalAI #LLM

[MiniMax (official) @MiniMax_AI]

Q: Why choose CISPO instead of GSPO or GRPO? How well does CISPO adapt to MoE, and does changing the RL algorithm require architectural refactoring?

GRPO predates both, but in our attempts to reproduce R1-Zero it proved unreliable: PPO-style clipping caused token-level gradients to vanish, leading to unstable learning. GSPO can be a reasonable alternative in some settings, and the Meta paper provides a useful comparison.

We chose CISPO primarily for its empirical stability and favorable bias–variance trade-off.

Regarding MoE compatibility, our observations so far indicate that CISPO behaves similarly on MoE and dense models. At the algorithmic level, we do not see major discrepancies introduced by MoE when using CISPO.

As for architecture changes, switching RL algorithms does not require refactoring the core model architecture. That said, MoE models do introduce additional considerations during RL training, mainly due to the router mechanism.

Recent approaches (such as R3 with fixed routing) aim to improve MoE stability under RL. These are largely lower-level implementation choices and are mostly orthogonal to higher-level RL algorithms like CISPO, which primarily operate at the level of optimization dynamics rather than architectural design.

[Jiatao Gu @thoma_gu]

🚨We introduce Multiplex Thinking:

Humans think in high-bandwidth “soft” internal representations, while LLMs must commit token-by-token. Prior work adds continuous thought tokens to widen the channel, but often sacrifices the stochasticity that RL-style training needs.

Multiplex Thinking keeps both:
- hard-sample multiple tokens per step (stochastic, RL-friendly)
- soft-merge them into a continuous state (higher bandwidth)

Bonus: sampling run in parallel ⇒ near-zero overhead + shorter reasoning.

We’re releasing everything: paper+model+code. This project is led by my talented first-year PhD student @tyao923 at Penn (@CIS_Penn @PennEngineers) with collaborators at MSR (@donglixp @yaruhao @qx_dong @furu_wei)

Paper: https://t.co/5M3OGoH99T
Blog: https://t.co/bjTxmh1RVa
Code: https://t.co/UIb5oPTmzN
Model: https://t.co/xQlOCRDpDN