窗花儿（Chuang）🍥

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

[窗花儿（Chuang）🍥 @EnbyKid]

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.

https://t.co/SHJL5KHjj9

Please read/share! ❤️

[雅辛忒斯Hyacinthia An @anan_de_gushi]

@aoya_toboso_01 流血不流泪的男孩子？每月流血的不是女生吗😂

[amber水母🍥 @amber_medusozoa]

一个高阶层跨女mtf应该具备什么
我这里给出四点证据
①首先是触手
拥有十条以上长达5米的吸盘触手，可以长在身体的任何地方，长有巨大的黑色蝙蝠翅膀和章鱼头
②其次是超能力
拥有停止时间、操控死亡的力量，能够随意创造出绵延上千里的浩瀚黑色云雾
③第三是氛围感
出现时可以用自己虚空、漩涡般的身体将所有火和热卷走，周围的气温瞬间降低好几度，并充斥着刺骨的寒冷和凝滞的空气
④第四是锟斤拷？3$：
烫¥̵̧̢̬͖͎̥ͦ̍̌̚ͅ*̵̴̟̣̲̤̪̱̄͐ͨͪͯͫ̏ͅ{̗̖͕̬̐ͤ*̞̦̺��̺̙̎̐ͦ̇̐̀ͥ?̆͑̒̿̆҉̩̳%ͭͥ̌ͤ͗ͣͫ̔͊͏͙̬̯̪͉̤̦͉͘]̢̛̳͈͔͓̝̘ͨ͢€̘͈̰̮̱̖͛̈́̇̌͋̎̀͞ͅ<̛̹͇ͬ̏̎ͪ+̸͈͖͈̱̺̘͚̞̣͆̓͝?̧̪̱̺̠̝̟͗̔’̶̐̅̆̿̒̊ͨͫ͏̩̥̼w̰͖͎̣̘̃͆ͯͯͤ̽̈ͬb̧̛͍̪̩̦̯ͫ̒͊̉͒c̹̗̯̭̠̏̄̾̍ͫï͖͇̼̦̅͞͡¥͔̙̀ͤ͠;̨͚̬̏ͥ̂ͯ̐ͫͩ¥̵̧̢̬͖͎̥ͦ̍̌̚ͅ*̵̴̟̣̲̤̪̱̄͐ͨͪͯͫ̏ͅ{̗̖͕̬̐ͤ*̞̦̺̠̺̙̎̐ͦ̇̐̀ͥ?̆͑̒̿̆҉̩̳%ͭͥ̌ͤ͗ͣͫ̔͊͏͙̬̯̪͉̤̦͉͘]̢̛̳͈͔͓̝̘ͨ͢€̘͈̰̮̱̖͛̈́̇̌͋̎̀͞ͅ<̛̹͇ͬ̏̎ͪ+̸͈͖͈̱

看你身后

[冷夜yuki🍥 @coldfrigidyuki]

明明已经挺瘦的了，但是还是好男性化，肋骨宽，手臂粗，好恶心想hrt
#hrt #mtf

[Adithya S K @adithya_s_k]

1/ In RL environments, reward design is everything.

Here's what we learned from the 7,780 environments Xiaomi open-sourced for MiMo 🧵

[James Zou @james_y_zou]

Looped transformer meets agent swarm.

Our new @Neurips2026 paper shows how to loop compute across multiple agents.

Different LLMs recursively share latent thoughts w/ each other. Better performance with up to 75% fewer tokens.

Great work led by @Jiaru_Zou w/ awesome collaborators.

[meng shao @shao__meng]

斯坦福大学 2026 秋季新课 CME 295: Transformers & Large Language Models 已开课（第一讲公开）

@stanfordnlp 这门课程由双胞胎兄弟 @afshinea @shervinea 主讲，9 讲走完 LLM 全部版图：从 Transformer 架构、注意力变体、RoPE，到 LoRA 微调、RLHF/GRPO 训练，再到 Flash Attention 推理优化、MCP 与 AI Agent、扩散语言模型。

第一讲 Transformers 课件和课程视频都已公开：
https://t.co/YCTABF2nQs
https://t.co/ByY1n7JUBQ

[Larry Dial @classiclarryd]

New historic NanoGPT record at 39.9s (-27.7s) from @DevenPzak , obliterating the prior record of 67.6s!

This record introduces a new paradigm of thinking to NanoGPT: instead of optimizing matmuls or adding more expressive operations, optimize at the individual flop level with incredibly clever engineering and ML judgement. If a flop is low value on a particular step, skip it.
Specifically:
-(~8s) Sampled softmax. If a token doesn’t appear in a batch, skip its lm_head fwd/bwd some fraction of the time.
-Sparse values. Only run an optimizer step for ngram embeddings that occurred in the batch. Set beta1 to zero to enable this. Beta2 is applied retroactively when the row is later used.
-Sparse updates. Only update ngram and value embeddings once every 4 steps instead of once every 2.
-Sparse communication. Shard the n-gram table across GPUs, and only pass the rows receiving updates on each step.
-Sparse optimizer states. For the n-gram table, reduce from 2 floats in Adam optimizer per param, to 1 float per 768 params.
-Hand-rolled flash attention for 64 dim heads.

There are several additions that add accuracy too:
-(~4s) EMA during last 300 steps, combined with lifting final_lr to 0.3 instead of 0.15.
-(~1s) A new optimizer, Anvil2, which expands muon via a second tracked momentum buffer, improves the ortho coefficients, and modifies the cautious weight decay application.
-A couple additional dynamic skip connections in the network.

The most striking consequence of the ‘flop aware paradigm’ is you can grow parameters arbitrarily large,
only limited by the available memory, since you can selectively choose how to expend flops on those parameters on each step. NanoGPT has kept active parameters below 124M, but total is unbounded, and has grown to 640M through embedding sparsity over the last year. This PR takes that to its logical conclusion on the 8xH100, scaling up to 65B sparse embedding parameters, which accounts for 25% of the PR’s gains. At frontier scale, where one is not bounded by an 8xH100, one could imagine where this paradigm could lead.
https://t.co/Ycrzy6JFC3

As this was a very notable PR, I spoke with Deven for an hour to learn how he did it. Here’s his story on the changes: https://t.co/YEfM1VpOTi

[Tianyi Cui @tianyi]

DeepSeek 弹性计算团队大量 HC 招人！尤其需要资深工程师。 来看看这篇技术分享：《DeepSeek 弹性计算 (DSec)：面向大规模 Agent 训练的沙盒基础设施》 https://t.co/7rBokhowIA

[Bingxiang He @HBX_hbx]

[1/n] Introducing Diffusion Reward Models (DRM)

Reward Models are at the core of LLM post-training. But almost all of them make the same simplification:
they collapse human preference into a single score.

The problem is that human feedback is often not a point — different people can hold distinct, even conflicting, judgments about the same response.

So we asked: Can a Reward Model learn the distribution of human preference?

Here’s what we found 👇

📄 Paper: https://t.co/e0EMNa3I64
💻 Code: https://t.co/KVosB5D5PI
🤗 Models: https://t.co/wwdDpOtgbL

[Qiuyang Mang @MangQiuyang]

200 high-quality, long-horizon, open-ended RL environments are here!

FrontierSmith is a NeurIPS 2026 Spotlight (top 3.7%)! We’re releasing 200 problems for studying RL on challenging open-ended tasks:
https://t.co/90FjTjAjnv

And stay tuned: FrontierCS team will soon share something very interesting about training in these environments 👀

[Shanghua Gao @GaoShanghua]

RSI-Jev v3.0 is out, and this is the first version where RL actually helped.

Our AutoScientists loop ran 59 RL experiments on the 2B decision model. Most failed.

The pattern was pretty clear: when every example already gives the model the right answer, plain supervised training is very hard to beat.

What finally worked was giving RL a signal the labels didn’t contain: whether the model ranked the right memory first.

That made it find the right memory first 60% more often on a long-memory benchmark, with no regression elsewhere.

The lesson we’re taking from this: RL helps when the reward teaches something the labels can’t.

But honestly, the part we care about just as much is why the other 58 experiments failed, so we wrote all of that up too.
Report: https://t.co/YjaMKguYgC
Code, weights, experiments: https://t.co/KHKS4XtWzv

[Phillip Isola @phillip_isola]

Students sometimes ask me if it still makes sense, in this accelerating age, to pursue a PhD in AI.

Perhaps counterintuitively, I think it's a great time to do so.

I wrote up some thoughts on this here: https://t.co/RRqGR0qi1R

[サイバー環状線 @CyberKanjousen]

豆包……终究还是输在有建模上了啊😭

[Garrett Flip @yinshiren2024]

The two biggest weirdos in the zoo, hahaha.

[窗花儿（Chuang）🍥 @EnbyKid]

content warning: just myself posing with clothing on, nothing too explicit

[窗花儿（Chuang）🍥 @EnbyKid]

SRS手术完今天来复查

[沐安ねこ @show5560]

记得以前在学校经常被后面男生撩内衣带，那时我不敢反抗，也不敢告诉老师和爸妈
毕竟全部男生就我穿这玩意🥺

[SuSu_酥酥👅 @NFT_Chen]

🔥真牛哇！深度融合Jev的大模型来了！Open-Jev-27B 以 17.3ms 一把梭，6倍推理还不量化！

给决策模型换上专用 CUDA 内核后，大模型不仅判断的快，跑的也更快，open-jev-fast 是 jev.server 的即插即用加速后端，融合内核 + 前缀树 + CUDA Graphs，bf16 原精度。

核心亮点：
🔹示例推理：103.7ms → 17.3ms（相对 FLA 6×，相对裸 PyTorch 14.8×）

🔹JevBench：258ms → 42ms，P50 17.4ms，单卡 B300 可打 50Hz+

🔹自研 Gated DeltaNet / tree-attention，共享前缀只算一次

🔹精度几乎没掉：197/231 vs 原版 198/231

上手 4 步：
1️⃣安装 Open-Jev（commit 3308a15）+ FLA 0.5.2，下载 Open-Jev-27B-v1.1

2️⃣设置 OPEN_JEV_DIR / OJ_CKPT / CUDA_HOME（CUDA 13）

3️⃣跑 ./scripts/launch_server.sh，请求格式不变：POST :18791/v1/systemone

4️⃣用 bench/e2e_bench.py 自己打出 17.3ms

GitHub：https://t.co/juzojfONlD

#OpenJev #Jev #CUDA #推理加速 #大模型 #SystemOne #开源 #B300 #Qwen

[alphaXiv @askalphaxiv]

"Contrastive World Models"

If world model has to reconstruct every pixel, it'll pretty much waste most of its capacity modeling irrelevant background noise.

So this paper removes Google DeepMind's Dreamer pixel decoder and instead trains the latent state to identify features of the correct future observation.

It matches Dreamer on clean environments, but performs much better with moving distractors and natural-video backgrounds.

https://t.co/S5b0QWmwuH

[Ren @Ryrenz]

👁️ Qwen 团队刚开源了 Qwen-MM-Plugins，让你正在用的那个 agent 直接能看图、能看视频。

GitHub 2983 star，Apache 2.0 开源，8 月 3 日放出第一版，一个多月里又陆续加了八组插件。

最痛的场景大概是这样：让写代码的 agent 帮你调界面，它看不见截图，你只能用文字把「按钮在���上角、间距太大」描述一遍，来回好几轮还对不上。Qwen-MM-Plugins 的 core 插件默认不用任何 API 密钥就能读本地图片和视频帧，文档、代码、数据表、3D 模型都能直接看，让 agent 自己去看，比你描述快得多。

需要更重的能力再挂 api 插件，接 DashScope 或者兼容的自托管服务，OCR、目标定位、语音转文字、区分说话人、生成字幕、视频事件分析就都有了。想联网查资料有 search 插件，Serper、Exa、Tavily 这些都能接。

最有意思的是 mhs 这个插件，agent 能通过一套硬件标准接口去操作摄像头、传感器、灯和机械臂，主机端带安全限制和紧急停止，而且不需要云端���钥。安装环境支持 Claude Code、Codex、Gemini CLI、Qwen Code 这些，装进哪个里都行。

能看、能听、还能动手，agent 的感官算是配齐了。

GitHub：https://t.co/1bUwcxMM3L