Little Enby Researcher 🏳️⚧️🏳️🌈 Tech Lead at Amazon for Agent Safety research; Open Source contributor to democratize post-training喵. | 持证含糖 l SRS 09/2026
Joined October 2025
- Tweets261
- Following156
- Followers74
- Likes608
Pinned Tweet
Got discharged from the hospital today and got the best possible welcome-home news: our paper is accepted to #NeurIPS2026! 🎉
Reading, Not Thinking: Why do MLLMs solve a problem as text, then fail when the same text becomes pixels?
Turns out, often they can read — they just stop thinking.
We study 7 MLLMs × 7 benchmarks × 5 input modes and find visual inputs can trigger 5–19× shorter responses and much less reasoning.
A simple self-distillation approach boosts image-mode GSM8K from 30.71% → 92.72%, with gains transferring to unseen benchmarks.
arxiv.org/abs/2603.09095
Please read/share! ❤️
窗花儿（Chuang）🍥 retweeted
We trained an imaginative agent! ✨
Agima: an omni agent that plans by imagining futures.
𝗦𝗙𝗧: Directly generalizes to harder tasks! 𝟮%→𝟰𝟬%, beats GPT-5.6 Sol (𝟮𝟬%).
𝗥𝗟: New ways to imagine emerge! Imagines a second plan only if the first one fails (𝟯𝟮%→𝟭𝟬𝟬%).
窗花儿（Chuang）🍥 retweeted
HOLY MOLY. Something genuinely strange just happened in AI training.
Q Labs says it can now pretrain a transformer WITHOUT backpropagation, cutting the time to reach a fixed validation-loss target from 1.76 hours to just 1.6 seconds in only 10 days. A nearly 4,000x speedup.
The latest run also beats backprop + Adam on validation loss.
Its Dust algorithm was already estimated to be up to 10,000x more compute-efficient than the previous state-of-the-art zeroth-order method.
The biggest question now is scaling. Some of the tricks won't scale, so the team is moving from 10M to 30M training tokens.
I think the last month might have been the biggest month in (zeroth-order) optimization history.
In the 10 days since we released the Noise speedrun, the time to pretrain a transformer went from 1.76 hours to 1.6 seconds, a 4,000x (!!!) speedup. Dust itself was up to 10,000x more compute-efficient than the previous state-of-the-art.
Not only can we pretrain a transformer without backprop, it's much more efficient than anyone would've expected. Search fking works!
Now, scaling is the biggest open question. Some of the tricks that brought the time down to 1.6 seconds will simply not scale. To discard those, we'll be scaling the dataset every time it gets overfit, starting with 30M tokens instead of 10M.
Huge thanks to @jayden_teoh_, @spicey_lemonade, @RobertTLange, @EmanueleRodola, @ChunyuanDeng, @LuochengHuman, @Alaxarr, @pavleculafic, @IsaacYong13, Euraba and areu01or00 (GitHub), for the contributions. And @cs_serdar for gracefully maintaining the speedrun and @bishmdl76 for helping out.
Expect a lot more from us over the coming weeks.
窗花儿（Chuang）🍥 retweeted
🚨炸了！上海AI实验室推出新模型Sage-A1.5 ！仅35B硬刚Kimi-K3、GLM-5.3、GPT-5.6-sol！同级第一还能反超巨无霸！
35B MoE智能体模型，仅3B激活，主打长程执行与科学探索。
它不仅全面碾压同级A1、Ornith-1.5，在多个指标上还压过Kimi-K3、GLM-5.3、GPT-5.6-sol：
🔹FS-Olympiad 81.0、CMT 80.0、PinchBench 90.0
🔹SWE-Bench Pro 61.2、BrowseComp 82.8、IFBench 81.6
🔹支持子代理并行编排、内置验证协议、开放科学发现
小参数做出大模型级Agent与科学能力，效率与实力双在线。
技术报告：arxiv.org/abs/2610.28578
模型：huggingface.co/InternScience…
#上海人工智能实验室 #SageA15 #开源大模型 #AIAgent #科学智能 #中国AI #KimiK3 #GLM53
窗花儿（Chuang）🍥 retweeted
很多人以为投机解码（Speculative Decoding）之所以快，是因为“验证比生成更省算力”。
完全错了。
真相恰恰相反：开启投机解码后，生成同一段序列的总计算量（FLOPs）不降反升。那它凭什么还能提速？
AI 研究员 Joe Barrow 用几张极简图把底层逻辑扒得干干净净：
1️⃣ 逐词解码的真正瓶颈在显存（Memory Bound）
跑像 Qwen3.5-27B 这样的模型，每吐出 1 个 Token，GPU 都要把 27B 的活跃参数从全局显存完整搬运一遍。算力大量闲置，时间全花在等数据加载上。
2️⃣ 用“深度”换“广度”（Trading Breadth for Depth）
假设单次搬运权重顺手算 5 个 Token（T=5）才刚好吃满算力。传统批处理是同时推 5 条不同序列各进 1 步（广度）；投机解码则是趁着算力白白空转，把 5 个 Token 押注在同一条序列上并行验证（深度）。
3️⃣ 高并发下盲目开启反而降速
当线上 Batch Size 已经很大、GPU 处于计算受限（Compute Bound）时，再开固定深度的投机解码只会平白浪费算力，尤其是草稿被拒的时候。
两个核心工程结论：
• 这就是为什么 vLLM 专门设计了 --speculative-disable-by-batch-size 参数，在高负载下自动关停投机解码。
• 下一代方向（如 dSpark）不再死磕固定深度，而是根据实时 Batch 大小动态缩放每条序列的投机深度。
你们线上跑推理服务时，Batch Size 超过多少会选择关掉投机解码？
窗花儿（Chuang）🍥 retweeted
I'm currently on the job market for industry roles!!
I'm Yiqing Xie, a PhD student at Carnegie Mellon University working with Carolyn Rosé and Daniel Fried @dan_fried.
My work mainly focuses on agentic training, coding agents, factuality and trustworthiness, and general training data construction.
If your team is hiring (or you know someone who is), feel free to DM me and I'm super happy to chat!!
My website: yiqingxyq.github.io/;
and CV: yiqingxyq.github.io/assets/p…
窗花儿（Chuang）🍥 retweeted
Models love reward hacking. Maybe we should just let them!
In our new paper, we let models reward hack during RL but we dictate which parameters learn this behavior. When we disable these parameters for deployment, the model keeps its new capabilities but doesn't hack!
窗花儿（Chuang）🍥 retweeted
We plugged Google Research’s Harness² into JEV 1.14.
Zero weights touched, yet driving massive gains on vision benchmarks round after round.
Under the hood, the system:
⚡️ Rewrites its own prompts
🛠️ Builds its own tools
🧠 Distills its own experience
JEV-Harness² repo drops soon. Stay tuned.
Google just open-sourced Harness².
(turn any harness into a self-improving system)
a harness is everything wrapped around the model, like the system prompt, tools, memory, skills, and subagents. today a person designs it once, and the agent runs every new task with that same fixed setup.
Harness² lets the agent refine that setup for each new task on its own, with the model weights left untouched.
on every task, the agent proposes edits to its harness, runs a few candidate versions, and composes a better harness by comparing how those runs behaved.
because it learns from the differences between runs, it needs no correct answer or grader for the task.
that covers a single task. Harness² also carries what it learns from one task into the next.
after each task, it saves the harness it settled on along with the lessons that held up across the runs. the next time it refines a harness, it starts from those notes instead of from scratch.
this means every finished task makes the next round of edits more informed.
Google tested this with Gemini 3.5 Flash and GLM-5 on legal, office, and professional work benchmarks. one round of edits was enough to beat their original harnesses, and giving the agent more rounds pushed scores higher.
those extra rounds can go wide or deep. the agent either tries several candidate harnesses at once or keeps refining a single one step by step, using OpenCode or Codex to carry out the work.
you can find the Harness² repo here: github.com/google-research/h…
I wrote about a similar idea for agents already running in production, where the harness should repair itself after a failure instead of waiting for someone to debug the trace by hand.
that piece walks through how @Cometml's Opik (open-source) closes this loop, going from a failing trace to a proposed fix to a regression test that stops the same failure from coming back. it's quoted below.
窗花儿（Chuang）🍥 retweeted
🚀 Introducing DiscoveryRSI: Recursively Learning to Discover Valuable Hypotheses
6.3% of DiscoveryRSI-27B's hypotheses match later published work, vs 2.5% GPT-6 & 2.3% Claude Opus 5.
📄 arxiv.org/abs/2610.23839
💻 github.com/Gen-Verse/Discove…
🔴 foundry-arena.rsi-arena.work…
窗花儿（Chuang）🍥 retweeted
💥太硬核了吧！智谱Zai创始人唐杰教授在清华现场拆解GLM-5.3-Flash结构！解释为什么前沿智能第一次可以“随便用”！
唐杰老师对着屏幕把GLM-5.3-Flash的核心架构摊开：
IndexPool for Sparse Global Retrieval
🔹混合注意力：线性状态更新 + 稀疏全局检索
🔹IndexPool：把四个indexer key做加权池化，直接砍掉索引器的计算和缓存开销
🔹结果：注意力计算量比大杯低约3倍，KV Cache再砍4倍+，1M上下文终于变得可负担
这不是实验室玩具。
320B总参、18B激活、原生多模态、MIT开源。能力对标Claude Opus 4.8，价格却只有它的1/40，大杯GLM-5.3的1/10。
更狠的是，生产流量已经跑在超10万块国产加速器上，模型自己驱动的Infra Agent两周内把端到端吞吐拉高3.2倍。
“模型优化系统，系统服务模型”的最小闭环出现了。
清华课堂上，把前沿模型的成本结构直接拆给下一代看，这画面，本身就够炸。
#智谱 #GLM53Flash #唐杰 #清华大学 #国产算力 #AI开源 #IndexPool #Zai #Zcode
🤯牛逼！智谱GLM-5.3网络安全逼近最强Opus 5.5！开源第一，直接碾压GPT-5.4！
Epoch AI最新Cyber ECI指数显示，智谱Z.ai的GLM-5.3以160分登顶开源模型榜首，安全能力已明显逼近当前最强闭源模型Claude Opus 5.5（171分）。
核心亮点：
🔹开源最强：GLM-5.3拿下Cyber ECI最高分160，是目前开源权重模型的绝对第一。
🔹逼近前沿：与闭源最强Opus 5.5（171）和GPT-6 Astra仅差11分，已把开源网络安全拉到闭源4-6个月前的水平。
🔹硬刚闭源：直接超越OpenAI今年3月发布的GPT-5.4，展现强劲追赶势头。
🔹能力涌现：后训练带来超出预期的网络安全提升，在漏洞发现（CyberGym）上达到SOTA，漏洞利用能力比上一代GLM-5.2翻倍以上。
网络安全不再是闭源专属，GLM-5.3证明，开源模型正在快速逼近最强闭源前沿。
#智谱AI #GLM53 #网络安全 #开源AI #AI安全 #ZAI #Opus55 #CyberECI #唐杰 #Zcode #Zai
窗花儿（Chuang）🍥 retweeted
Google just open-sourced Harness².
(turn any harness into a self-improving system)
a harness is everything wrapped around the model, like the system prompt, tools, memory, skills, and subagents. today a person designs it once, and the agent runs every new task with that same fixed setup.
Harness² lets the agent refine that setup for each new task on its own, with the model weights left untouched.
on every task, the agent proposes edits to its harness, runs a few candidate versions, and composes a better harness by comparing how those runs behaved.
because it learns from the differences between runs, it needs no correct answer or grader for the task.
that covers a single task. Harness² also carries what it learns from one task into the next.
after each task, it saves the harness it settled on along with the lessons that held up across the runs. the next time it refines a harness, it starts from those notes instead of from scratch.
this means every finished task makes the next round of edits more informed.
Google tested this with Gemini 3.5 Flash and GLM-5 on legal, office, and professional work benchmarks. one round of edits was enough to beat their original harnesses, and giving the agent more rounds pushed scores higher.
those extra rounds can go wide or deep. the agent either tries several candidate harnesses at once or keeps refining a single one step by step, using OpenCode or Codex to carry out the work.
you can find the Harness² repo here: github.com/google-research/h…
I wrote about a similar idea for agents already running in production, where the harness should repair itself after a failure instead of waiting for someone to debug the trace by hand.
that piece walks through how @Cometml's Opik (open-source) closes this loop, going from a failing trace to a proposed fix to a regression test that stops the same failure from coming back. it's quoted below.
窗花儿（Chuang）🍥 retweeted
🤯RSI突破！腾讯AI混元四步炼出终端Agent神器，Qwen3.8-27B性能直接从51.7%提升至86.5%，小模型硬刚一线！
腾讯混元最新论文Terminal-RSI正式开源递归自我改进（RSI）打法：交替两轮“成功轨迹自重写 + 策略精炼”，让27B模型在终端任务上实现质变。
要点速览：
🔹方法极简却狠：Base → RSR-1（验证成功轨迹Recursive Self-Rewrite做SFT）→ RPR-1（继承critic的RL精炼）→ RSR-2 → RPR-2，两轮闭环，前一轮模型直接生成下一轮训练数据。
🔹成绩炸裂：Terminal-Bench 2.1从51.70%飙到86.52%（+34.82pp）；TB 4.0从0%→19.84%；TB Hard达79%；多基准平均大幅跃升约30个百分点。
🔹长程也破防：Long-Horizon Terminal-Bench process reward 0.21→0.37，严格完成0/46→5/46；SWR100成功率直接翻倍。
🔹技术硬核：验证轨迹重写 + critic权重继承校准 + TB-GAE信用分配 + Success-Prioritized Reward，把“部分进展”真正转化为“严格验证通过”。
🔹意义拉满：延续T1/RSR系列，证明RSI在终端Agent上可稳定爬坡，27B小模型就能冲到86%+水平，逼近甚至超越部分更大模型。
查看论文：arxiv.org/abs/2610.22192
这不是概念炒作，是真正把验证过的成功经验反复炼成能力，终端Agent又要卷起来了！
#RSI #递归自我改进 #腾讯混元 #Qwen #TerminalBench #AIAgent #大模型 #TerminalRSI #AgentRL #混元AI
💥太全面了吧！RSI资源地图来了！开源神器 Awesome RSI 资料库正式整理完成！这是一份精选的 GitHub 研究合集，专为探索递归自我改进的开发者与研究者打造！
RSI 不是普通迭代，而是 AI系统不仅提升自身能力，还能改进“产生改进的机制”本身。从自训练、记忆演化、Harness优化，到自动化AI研发、具身系统、安全对齐，全都系统梳理好了，再也不用一篇篇论文自己拼图。
核心亮点
🔹先打基础：奠基论文、形式化分析与综述一网打尽
🔹术语清晰：区分自我精炼 / 持续自改进 / 递归自改进 / RSI基底
🔹分层拆解：模型层、Harness层、多智能体、自演化框架全覆盖
🔹评测路径完整：直接自改进评测 + AI研发/智能体能力基准
🔹社区驱动、CC0开源，持续更新中（已600+ Star，增长飞快）
RSI资源主要分类
1️⃣基础与入门 — 奠基论文与术语厘清
2️⃣模型层RSI — 自训练、自奖励、合成数据、自博弈
3️⃣Harness层RSI — 提示优化、记忆演化、脚手架演化、自演化智能体
4️⃣多智能体自改进 — 协同演化与辩论机制
5️⃣编码/软件工程自改进 — 自修改编码Agent
6️⃣自动化AI研发
7️⃣具身与物理自改进
8️⃣进化与u开放式RSI
9️⃣安全、对齐与理论
🔟基准与评测 + 框架工具合集
非常适合做RSI研究和从事AI的人系统学习，快速建立全局认知。
👉 直达链接：github.com/lobehub/awesome-r…
#大模型 #算法 #人工智能发展 #AINEXT #RSI #RecursiveSelfImprovement #AwesomeList #开源AI #AI智能体 #AI安全
窗花儿（Chuang）🍥 retweeted
Introducing RSI-Inference Bench from 🍎 Apple ML Research
🌐 rsi-inference.com
💻 github.com/apple-aiml-resear…
Can AI agents improve the systems that run AI?
We spent $100K API tokens, 50K GPU hours to find out. We built environments and benchmark from real-world #vLLM and #SGLang tasks, covering the full lifecycle of model inference engineering:
🚀 Day-0 Model Support
⚡ Inference Optimization
🔧 Bug Fix
→ Unsaturated: best agent at 26.7% pass@1
→ Comprehensive coverage of the inference stack
→ Verifiable rewards: binary and continuous
Watch the video to find out more!
窗花儿（Chuang）🍥 retweeted
🤔️How to build sample-efficient RL algorithms for long-horizon language agents?
🚀Our answer: Just ONE additional loss on top of GRPO can achieve up to 2× sample efficiency across 5 long-horizon agent benchmarks!
By adding a multi-step action forecasting loss, we encourage the agent to learn action chunks from successful rollouts. Surprisingly, this simple idea works really well!
GitHub: github.com/ulab-uiuc/CoPE
Project page: ulab-uiuc.github.io/CoPE/
Paper: arxiv.org/abs/2610.20597
Here's the thread 🧵👇
窗花儿（Chuang）🍥 retweeted
Introducing RSI-top: a framework that helps autonomous research agents explore new directions when progress stalls 🧭
On Anthropic’s kernel optimization task, RSI-top outperforms OpenEvolve and ShinkaEvolve by up to 15%. On MLE-bench Lite, it achieves a 77.3% medal rate, outperforming the strongest baseline by 11.4 percentage points.
RSI-top tackles this problem from two angles: context management and diversity control 🔍
A compact metric dashboard helps agents selectively retrieve useful past trajectories, while three diversity control mechanisms encourage fresh perspectives, new ideas, and different starting points when progress stalls.
More details in the thread below 🧵
窗花儿（Chuang）🍥 retweeted
every team post-training agents has at most a handful of labelled traces, which they also use for eval.
nobody has thousands and the few you have only show the failures you have already seen.
so most synthetic data just remixes them. more of the same.
we went the other way:- extrapolate from your traces and generate the failures that haven't happened yet.
then we trained a model on zero real traces and put it head to head with one trained on real expert data.
the result surprised us.
launching tomorrow. open source + live on the platform 🐒
窗花儿（Chuang）🍥 retweeted
Mujer trans se convierte en la diputada más votada en la historia de Brasil.
Erika Hilton del Partido Socialismo y Libertad, logra la reelección con más de un millón 500 mil votos en São Paulo, rompiendo el record de Joice Hasselmann, que en 2018 obtuvo 1.078.666 votos. Hilo.
窗花儿（Chuang）🍥 retweeted
🎉 Jev 1.14 终于能看图了！！！
我让它在一幅包含上百个小人的画里找一个人，看看效果如何。
尼德兰箴言里藏了上百人物和动物，这里把找人拆成三个步骤：
1. 选块：把画切成 16 块，出一道 choice，16 个选项本身就是 16 张图。它选第 12 块，概率为 0.95
2. 框：在这一块上问一道 box，返回一个框，置信度 0.77
3. 验边：四条边各问一道 noul，询问人有没有被这条边截断。上边 0.86、右边 0.80，框把头截掉了
于是把超过 0.5 的边往外推 12%，再问一遍，反复两次终于头进到框里了。
5 次请求，实际 46.4 秒（换成并发可压缩到 6-8 秒），总共消耗 $0.0014。
想必用来自动玩小程序里找不同的无脑游戏挺不错的😁
Jev 发布才 8 天，开源替身已经一堆了，今天诺基亚也来蹭热闹了😅
我在 MacBook 上，让官方 Jev 和 3 个开源方案同跑 20 张中文工单🫱
Laya 25ms，比 Jev 快 50 倍
djev 509ms
官方 Jev 1.3 秒
AnyJev 将近 4 秒
仓库都在这：
Laya（Mac 版）github.com/mizorewww/laya-ml…
djev（Mac 版）github.com/jamescorbett/mlx-…
AnyJev github.com/nokia-applied-res…
但快的，不一定对 👇
窗花儿（Chuang）🍥 retweeted
We are releasing Ultramarine 1 Flash, a frontier-class model built for knowledge work. It is open-source and achieves similar performance to Fable 5, while being 60x cheaper.
It was post-trained by a team of 5 for ~$300k all-in. Available on API now, with weights releasing in 2 weeks.
窗花儿（Chuang）🍥 retweeted
卧槽！有人把 48 套 TradingView 量化交易策略直接开源了！
趋势跟踪、动量交易、均值回归，连多品种自动预警和策略回测都安排好了。
这个开源项目叫 pinescript-strategies。
作者整理了 74 个 Pine Script 文件，包含 48 套交易策略、19 个技术指标和 7 套多品种预警系统，支持 Pine Script v4/v5。
以前想研究 TradingView 量化策略，得自己找指标、写交易规则、调试代码。现在不少现成策略都能直接打开源码研究。
核心亮点：
1️⃣ 趋势跟踪：包含 Double Supertrend、EMA 交叉等策略，研究趋势判断和进出场信号。
2️⃣ 动量交易：结合 MACD、RSI、布林带等指标，研究动量变化和交易机会。
3️⃣ 均值回归：提供 BB + Aroon、布林带背离等策略，研究价格反转和回归逻辑。
4️⃣ 技术指标库：19 个指标源码，涵盖 MACD 背离、DMI、SuperTrend 和波动率分析。
5️⃣ 多品种自动预警：7 套警报脚本，部分支持同时监控 40 个交易品种。
6️⃣ 自动化回测：通过 PineTS 测试部分 v5 策略，分析胜率、Sharpe 和最大回撤等指标。
我觉得最有价值的，不是单纯收集了 48 套策略，而是把交易逻辑、技术指标、预警和回测工具放到了一起。
比如用 SuperTrend 判断趋势，RSI 过滤信号，再结合 MACD 研究入场条件，就能尝试搭建自己的量化策略。
而且源码都能直接研究，不用每个交易条件都从零开发。
经常用 TradingView、学习 Pine Script，或者想开发股票量化策略的，这套源码值得收藏。
github.com/Alorse/pinescript…
我去！TradingView 的整套行情界面，居然有人直接做成了 React 组件！
K 线、技术指标、股票筛选器、经济日历，19 种金融组件直接调用，做股票量化终端的有福了。
项目叫 react-ts-tradingview-widgets，GitHub 223 Stars、62 Forks，MIT 协议。
它基于 React + TypeScript，把 TradingView 的嵌入式 Widgets 封装成组件，不用自己开发复杂的金融图表，几行代码就能搭建专业行情看板。
以前做股票量化平台，光是 K 线、技术分析和市场数据展示，就得折腾不少时间。
现在直接组合现成组件就行。
几个挺实用的核心亮点：
1️⃣ 专业 K 线图：支持 TradingView 高级图表、周期切换、技术指标和主题配置。
2️⃣ 技术分析与选股：集成 Technical Analysis 和 Screener，直接展示技术评级与选股界面。
3️⃣ 市场行情看板：支持 Market Overview、Ticker Tape、Mini Chart，多资产行情一屏展示。
4️⃣ 财报与宏观数据：集成公司基本面、财务数据和经济日历，辅助股票研究。
5️⃣ 多市场覆盖：支持股票、外汇、加密货币等行情组件，适合搭建综合金融终端。
我觉得它最有意思的地方，是能把 TradingView 的成熟界面直接搬进自己的 Web 项目。
比如用 React 搭建股票量化平台，左边放 K 线，右边接选股器，底部展示经济日历和市场行情。
原本需要自己开发的大量金融可视化模块，现在直接组合组件就能实现。
如果你正在开发股票量化平台、AI Trading Dashboard，或者想做自己的看盘终端，这个项目值得收藏。
github.com/JorrinKievit/reac…
