yydxlv

浮生若梦 为欢几何

[Junlin Yang @junlin45300]

🚀 OpenRSI is a new open research series from @FrontisAI for concrete, testable progress toward recursive self-improvement (RSI).

As its first project—and also my first work as first author—I’m proud to present OpenMLE: an open full-stack AI4AI system for autoresearch, where evolutionary agents improve ML solutions through executable feedback.

OpenMLE has three components:
- OpenMLE-Gym: 5,758 executable tasks + evaluators
- OpenMLE-ERL: execution-grounded SFT + RL
- OpenMLE-Evo: experience-guided long-horizon search

🏆 The full system—our trained Frontis-MA1-35B model paired with OpenMLE-Evo-Max—reaches 71.21% Medal Average on MLE-Bench Lite: surpassing GPT-5.5 + Codex (68.18%) and just 1.52% from GPT-5.6 Sol + Codex and the 2.8T Kimi K3 + Claude Code (72.73%). Budget: 12 hours/task on one RTX 4090 capped at 12 GB VRAM.

🌍 On 10 held-out NatureBench Lite tasks, both components transfer:

• same framework, model swap: Match-SOTA 50% → 70%
• same base model, framework swap: Match-SOTA 20% → 50%

🔓 Paper, code, models, data, and analysis below. 🧵

[Kimbo @kimbochen]

Songlin Yang's video explanations
https://t.co/w2GR2tYb9i
https://t.co/uYhvjJz9Fa

Songlin Yang's blog post
Design intuition
https://t.co/8DnjDb2fpc
Kernel algorithm (I skipped to "A Chunkwise Algorithm for DeltaNet" section)
https://t.co/wLuwP4F6bN

FlashKDA
https://t.co/Qr2VPti1Kc
I prompted Codex to explain it to me, only after that the deep dive made sense
https://t.co/7tv9wA2A6H

vLLM serving explanations
https://t.co/3I6YimpUDV
https://t.co/uLddfZXp4G

The great Zhihu post for explaining AttnRes
Training
https://t.co/9ZuaHgJ8bp
Inference
https://t.co/njOLG8OJix

Jianlin Su's MoE 環遊記 series
https://t.co/7vJqhsJXo2
Explains the concept of MoE from math first principles

Papers
Kimi Linear, LatentMoE, Kimi K3 tech report

N/

[EverMind @evermind]

Our new paper.

Self-evolving agents have a measurement problem.
The hard part isn’t generating changes to an agent harness. It’s knowing which changes genuinely helped rather than got lucky, overfit, or never activated at all.

LLMs diagnose failures and propose changes across prompts, knowledge, runtime, tools, and configuration. But deterministic code owns the credit: validity checks, activation checks, paired significance testing, and a sealed test.

Across 7 domains, the evolved harnesses gained +9.0 to +15.5pp on the 6 statistically credited sealed tests, retaining 86–147% of the training gain.

What transfers is not one universal harness. It’s the diagnose-and-credit loop.

https://t.co/w56EndGp3n

[Kimi.ai @Kimi_Moonshot]

We've open-sourced FlashKDA, our high-performance CUTLASS-based implementation of Kimi Delta Attention kernels.

It delivers 1.72×–2.22× prefill speedup over the flash-linear-attention baseline on H20, and works as a drop-in backend for flash-linear-attention.

Explore on GitHub:
https://t.co/QScjWsJqSy

[Tom Huang @tuturetom]

感谢 @aiwarts 把 Humanize PPT skill贡献到 Open Design 社区，这个真的太强了💪

[GitTrend @GitTrend0x]

AI 网关天花板！

一个端点接 268+ 提供商、500+ 模型，Cursor/Claude Code 直接用

用 AI coding 时最烦的几个问题：
• 模型太多，切换麻烦
• 免费额度分散，经常超限
• Token 吃得太凶，成本高
• Agent 想自主调用模型却很难

OmniRoute 把这些问题一次性干掉了。它是一个免费开源的 AI 网关，通过一个本地端点（http://localhost:20128/v1），就能访问 268+ 提供商、500+ 模型（包括 Claude、GPT、Gemini、Kimi K3、GLM、DeepSeek 等）。支持 Claude Code、Cursor、Codex、Cline、Copilot 等主流 AI coding 工具。

//

核心黑科技：

• 智能路由 + 自动 fallback：18 种路由策略（优先级、加权、成本最优、融合等），配额感知，模型挂了自动切下一个
• RTK + Caveman 压缩：可叠加 11 种压缩引擎，节省 15-95% tokens，代码/JSON 几乎无损
• MCP/A2A 支持：AI Agent 可以直接控制网关，实现真正的自主调用
• 免费额度聚合：把散落在各家的免费额度集中管理，据称每月可省下巨量成本
• 本地优先：完全自托管，支持 Desktop/PWA/Termux，隐私安全
• 开箱即用：npm 一键安装，Dashboard 可视化配置，支持 43 种语言（含中文）

//

简单来说：把“模型碎片化”变成“统一智能入口”，让开发者真正实现 “Never stop coding”。目前已获 22k stars，由 500+ 贡献者共同打造，社区非常活跃。

GitHub：https://t.co/z4qcJNrGWm

去 star 一下，装上后把你的 Cursor 或 Claude Code 指向这个端点，model 填 auto，立刻感受丝滑体验。

[Jian Hu @hijkzzz]

1/ Still looking for a minimalist, high-performance framework for agentic RL research?

Meet Molt — an agentic-first, PyTorch-native reinforcement learning framework with roughly 9K lines of RL code for 700B models.

⭐ https://t.co/B4Kde1ihxl

[Mr Panda @PandaTalk8]

不要再去读 X 上又臭又长的 harness 工程的文章了，
跟这篇文章相比， X 上的那些文章垃圾都不如。

Lilian-weng 的这篇新的博客文章，
是我目前读过的最写的清晰、最容易理解的 harness 工程， 也是递归自我改进的 harness 工程写的最好的文章。
文章是一篇框架式的详细描述了 self-Improment 的 harness engineering 全部知识框架。

链接是我做的中文版本

https://t.co/o3NpmBaQo5

[青稞社区 @qingke_ai]

https://t.co/105kXRQY9B

[Sumit @_reachsumit]

WebSwarm: Recursive Multi-Agent Orchestration for Deep-and-Wide Web Search

Proposes a recursive multi-agent framework that dynamically delegates search nodes with modes suited to deep, wide, or combined web search tasks.

📝 https://t.co/K6oY8gudGC
👨🏽‍💻 https://t.co/uambmwP2AH

[kaize @0x_kaize]

Claude Code just dropped "Getting Started with Loops"

This is their first official document about Loop Engineering.

Spoiler: prompt engineering didn't survive.

Here's the full guide in one post:

1. Turn-based

Every prompt you send already runs as a loop: Claude gathers context, takes action, checks its own work, and repeats until it decides the task is done or realizes it needs your input.

2. Goal-based

You define what "done" looks like with /goal, and Claude keeps iterating toward it. Every time it tries to stop, a separate evaluator model checks your condition - if it's not met, Claude gets sent back to work until the goal is reached or the turn limit hits.

3. Time-based

This is the /loop command: нou set an interval and a prompt fires on schedule. For example, every 5 minutes Claude checks your PR, addresses review comments, and fixes failing CI.

4. Proactive

Here you set up an event once, and the loop triggers itself whenever that event happens - with zero human input.

Anthropic also shared optimization advice - and this part matters most:

- Loop quality depends on system quality.
- Tokens are the real cost

None of this is magic. It's the same familiar parts - /goal, /loop, Skills, Hooks - composed into systems that run without you.

I built exactly that kind of system in my article:

Loop that wakes up at 6 AM, finds work on its own, reviews itself, and leaves ready PRs for you. Step by step, from SKILL(.)md to cron trigger:

[Lilian Weng @lilianweng]

new post on harness engineering for AI self-improvement: https://t.co/ZYvGfVs61k

It is hard to forecast how much the future of RSI will rely on harnesses. Likely harness engineering will evolve in the direction of self-improvement and enable auto-research, and, in turn, smarter models keeps harnesses simple.

Even when many harness improvement get eventually internalized into core model, the need to specify goals and context will not disappear.

[Sumit @_reachsumit]

Autonomous Information Seeking: A Roadmap for Agentic Recommender Systems

@xylinhaha et al survey agentic recommenders, presenting a taxonomy based on autonomy & 3 paradigms: agent-assisted recommendation, agent-as-recommender, agent-as-user-simulator

📝 https://t.co/5B9szHNBYY

[Gill @gurtej__gill_]

This new paper from Stanford, Berkeley and NVIDIA feels like a massive piece of the puzzle when it comes to test time compute.

They’re proposing LLM as a Verifier, arguing that verification is the next major scaling axis we need to look at.

Instead of treating LLM judges like a black box that spits out a rigid discrete score, they extract the token logits to calculate a continuous probabilistic expectation.

It sounds nuanced, but it completely fixes the "tie rate" problem when you're trying to compare two complex long horizon agent trajectories.

By scaling up the score granularity, running repeated evaluations and decomposing the criteria, they get incredibly calibrated results.

The performance gains are wild: SOTA across the board, including hitting 86.5% on Terminal-Bench 2.0 (beating GPT-5.5) and crushing SWE-Bench Verified at 78.2%.

And this isn't just theoretical.

Because the feedback is so fine grained, they’re using it as a dense reward signal to massively speed up RL efficiency (like GRPO and SAC) and they’ve already built a practical extension for Claude Code to track agent progress in real time.

Definitely a work worth reading by Jacky Kwok, Chelsea Finn, Ion Stoica and the rest of the team.

Read the full paper here:
https://t.co/t8Pps8qynx

[How To Prompt @HowToPrompt__]

China has killed the entire vector database industry.

They open-sourced TencentDB Agent Memory. It gives any AI agent long-term memory that runs 100% locally.

No Pinecone. No cloud APIs. No repeating yourself every session.

- 61% fewer tokens
- PersonaMem accuracy: 48% → 76%
- Zero external API dependencies
- Runs on plain SQLite

Most memory systems compress your history into an opaque vector pile. when recall goes wrong, you're guessing. this one doesn't compress, it builds a semantic pyramid.

L0 Conversation → L1 Atom → L2 Scenario → L3 Persona.

Short-term state gets encoded as a Mermaid graph in your agent's context. verbose tool logs get offloaded to disk. when the agent needs proof, it drills back via node_id to the exact raw log.

no lossy compression. every layer is readable markdown you can just open and inspect.

5.1k stars. 100% Open Source.

[Xudong Han @Xudong07452910]

CMU 这门 NLP 课最硬核的地方：第一份作业就让你从零手搓一个 LLaMA。

卡内基梅隆大学公开了 2026 春季《Advanced NLP》课程，教授是 Sean Welleck。整套资源包括完整课程主页、课件、视频和配套代码，内容从 Tokenizer、Transformer、语言模型基础，一路讲到 RAG、多模态、RLHF、MoE、长上下文和 test-time scaling。

这门课最有价值的地方，不仅是教你怎么调用大模型 API，而且把大模型能力背后的工程路径也拆开给你看。

从「Build Your Own LLaMA」这种手搓模型作业，到后面的推理、效率、评测和 Agent，它训练的其实是一种更底层的能力：你能不能理解模型为什么能工作，为什么会变强，为什么有时会失败。

这也很符合现在 AI 的技术拐点。

过去大家更关心参数规模，谁的模型更大；现在越来越多关键进展，开始发生在推理侧、效率侧和系统侧：怎么让模型更会思考，怎么调度专家，怎么检索外部知识，怎么在有限算力下获得更好的输出。

所以这门课不只是 NLP 进阶课，更像一份 2026 年大模型研究者的训练路线图。

AI 时代真正的分水岭，可能不只是会不会使用模型。

更关键的是：

你停在应用层，还是愿意继续往下走到架构层。

Home: https://t.co/U96QszHIja
Youtubu: https://t.co/05dlefgqbe
Code: https://t.co/nqAsjPwatQ

[Jiaru "Rubin" Zou @Jiaru_Zou]

Excited to share that #LatentMAS has been accepted to ICML 2026 as a spotlight!

💻Code: https://t.co/jBN26NG1PY
📄Paper: https://t.co/vuv8nYBTic

We push multi-agent collaboration into the latent space — beyond human language.

Most multi-agent systems rely on text: agents reason in words, exchange messages, and repeatedly decode/re-encode information. But language can be slow, lossy, and unnecessarily constrained.

💡LatentMAS takes a different path: LLM agents reason and communicate directly through hidden embeddings.

No text decoding.
No extra training.
No token-level message passing.

Instead, agents collaborate through:
🧠 Autoregressive Latent Thoughts — hidden-state-level reasoning steps
🔁 Latent Communication — information sharing via KV-cache transfer
📌 Input-output Alignment — keeping latent representations in-distribution
🚀 Training-free Collaboration — plug-and-play with existing LLMs

Why it matters:
✅ Up to +14.6% better accuracy on complex reasoning tasks
⚡ 4-4.6x faster end-to-end inference
✂️ 70.8%–83.7% reduction in output token usage

A step toward multi-agent systems that collaborate not by speaking more, but by thinking together in latent space.

#MultiAgentSystems #ModelCollaboration #LatentReasoning #LLM #AgenticAI #ICML

[Huaxiu Yao @HuaxiuYaoML]

🦞 Coding agents live on your screen. Omni-modal agents live in your physical world. VisualClaw rides on your glasses, getting cheaper AND smarter every session, without ever retraining the VLM.

📉 −98.1% API cost vs full-frame upload.
📈 +15.80% peak accuracy on EgoSchema.
🤖 +3.2 macro on VisualClawArena with Claude Code.
🔒 VLM weights frozen throughout.

🎯 See — proactive frame filtering. An edge cascade decides what's worth showing the VLM, on-device. A 1-hour 1fps stream is 3,600 frames; we send 5–20.

📦 Streamline — adaptive skill memory. Hot/cold skills keep prompts lean even as the agent learns new behaviors at deployment.

🔁 Meta-Evolve — continuous self-evolution. Correct rollouts enter memory; failures trigger a memory-grounded skill evolver. The scaffold improves while it runs, not just while it trains.

🏟️ Also releasing VisualClawArena: a rigorous 5-stage multimodal agent benchmark with video clips, documents, user files, dynamic updates, and executable checks. Avg 24.4 rounds per scenario, 18.1 of them requiring vision.

Always on. Always learning. Always cheaper. ⚡

📄 https://t.co/nCt1QcYR6n
💻 https://t.co/9HKqBYRpBJ
🌐 https://t.co/UZx91yC7LY

@HaoqinT @cihangxie @yuyinzhou_cs @richardxp888 @ZhengBerkeley @itsJiaqiLiu @JimChenjw @jasoneshraghian

[Sumit @_reachsumit]

Querit-Reranker: Training Compact Multilingual Rerankers via Efficient Label-Free Distribution Adaptation

Presents a multilingual cross-encoder reranker family trained via a data-centric pipeline for label-efficient adaptation

📝https://t.co/8eRH7BfAgX
🤗https://t.co/14X9xR2duG

[Sumit @_reachsumit]

On the Memorization Behavior of LLMs in Generative Recommendation: Observations, Implications, and Training Strategies

Snap shows that LLM recommenders rely heavily on one-hop memorization, and teaches them richer relations.

📝https://t.co/D5JM50htbg
👨🏽‍💻https://t.co/lvc4oBQcov