I am a PhD at UEC in Japan. My research interests include deep reinforcement learning and its control applications.
Joined November 2022
- Tweets1.9K
- Following283
- Followers17
- Likes6
Colin retweeted
🤖🧠NEW PAPER🧠🤖
(The result of an 8-year project!)
LLMs seem very different from symbolic systems. Yet LLMs excel in symbolic domains (e.g., language/code/math). How do they do it?
Our finding: LLM representations have implicit symbolic structure!
Link in thread ⬇️
1/n
![Overview of the paper.
Title: The Emergent Symbolic Structure of Artificial Neural Networks
Authors: Tom McCoy, Paul Soulos, Tal Linzen, Paul Smolensky
Left: Neural networks encode information in vectors (there is then an image of a vector), yet they excel at tasks long thought to require symbolic structure (there is then an image of a symbolic representation, specifically a syntax tree). How do LLMs do it?
Right: We find that LLM representations can be closely approximated with symbolic structures. This approximation lets us edit the structure of an LLM’s output by editing the structure of its internal representations, as shown. There is then an image of two edits to LLMs. In the first one, the original input is 3 + 6 * 8, with an answer of 51. But if we swap the positions of the 3 and the 6, the output becomes 30. In the second one, the original input is a Python command repeating the list [Z, U] three times, producing [Z, U, Z, U, Z, U]. But if we edit the input in a way that adds](https://pbs.twimg.com/media/HRJ5Jr5W0AAQ-sN.jpg?name=orig)
ALT Overview of the paper. Title: The Emergent Symbolic Structure of Artificial Neural Networks Authors: Tom McCoy, Paul Soulos, Tal Linzen, Paul Smolensky Left: Neural networks encode information in vectors (there is then an image of a vector), yet they excel at tasks long thought to require symbolic structure (there is then an image of a symbolic representation, specifically a syntax tree). How do LLMs do it? Right: We find that LLM representations can be closely approximated with symbolic structures. This approximation lets us edit the structure of an LLM’s output by editing the structure of its internal representations, as shown. There is then an image of two edits to LLMs. In the first one, the original input is 3 + 6 * 8, with an answer of 51. But if we swap the positions of the 3 and the 6, the output becomes 30. In the second one, the original input is a Python command repeating the list [Z, U] three times, producing [Z, U, Z, U, Z, U]. But if we edit the input in a way that adds
Colin retweeted
A while back I made a blog about rectified flows.
It is quite relevant to this post.
alechelbling.com/blog/rectif…
Colin retweeted
Most latent generative models still follow the same two-stage recipe: train a VAE for reconstruction, freeze it, then force the generative model to learn whatever latent space it gets. But reconstruction-optimal latents are not necessarily generation-friendly — so why not let the generative model shape the latent space directly?
ByteDance Seed presents GenFirst to make stable end-to-end latent learning work.
This work revisit the VAE objective and identify a prior–entropy imbalance as the key cause of latent collapse in naive end-to-end training. After fixing collapse with explicit entropy preservation, they find that generation and reconstruction have strongly asymmetric learning dynamics.
This leads to Generation Before Reconstruction (GenFirst): first let the generative objective shape a generation-friendly latent space, then strengthen reconstruction to recover visual details.
The results are strong. With end-to-end SiT + VAE training, GenFirst achieves 0.97 gFID with CFG / 1.45 without CFG on ImageNet-256, and reaches the final SiT result with at least 70× fewer training steps. This is the first diffusion model to reach FID < 1.0 without an FD loss. On text-to-image generation, it reaches 0.90 GenEval.
More importantly, this end-to-end framework goes beyond image generation: it learn a shared visual latent space that supports generation, representation learning, and reconstruction, and further extend end-to-end latent learning to continuous text latents, enabling unified text-to-image and image-to-text diffusion modeling.
GenFirst: Generation Before Reconstruction for Stable End-to-End Latent Generative Modeling
Paper: arxiv.org/abs/2608.29335
Project: tom-zgt.github.io/GenFirst/
HF Daily Paper: huggingface.co/papers/2608.2…
Colin retweeted
杜克大学2026年暑期学期讲义《扩散模型的数学导论》
杜克大学 Jianfeng Lu 他为 SLMath 数学暑期学校准备的讲义(2026年7月),面向具备概率论基础但没有随机微分方程或扩散模型背景的研究生。
没事的无聊的 想补数学的可以保存下来看看
arxiv.org/pdf/2607.01693
Colin retweeted
The only tutorial you should attend at #ECCV26 (kidding 😂).
But I am incredibly psyched to be doing this with the best bunch out there. I strongly feel this is a timely tutorial!
We will present general approaches alongside our learnings from (post)-training impactful models like Flux2/3.
Tutorial website: pdm-tut.github.io/
Save your calendars!
Colin retweeted
本地大模型的显存壁垒，今天被正式宣告物理终结了，
单张 24G 的 RTX 4090，竟然真把 125B 的 Qwen 3.8 Flash Next 给跑起来了，
而且上下文直接拉到了 25 万，解码速度稳在每秒 21 个词
以前跑这种百亿千亿级的巨无霸，动辄要租几张 A100 数据中心集群，
但这次海外极客在 Ubuntu 上用一张普通 4090 加 110G 白菜价 DDR4 内存，
硬生生跑出了生产级的速度:
预填充每秒 364 词，解码每秒 21 词，
最狠的是没开 MTP、没开 dflash，连 KV 缓存都没做任何量化
很多人好奇 24G 显存怎么塞得下 125B 还带 25 万上下文，
核心全靠 llama.cpp 最新分支里一个叫 -cmoe 的物理外挂:
它把 512 个专家层全部扔给廉价的系统内存，
把最吃计算的注意力机制留在 4090 显存里，
显存占用瞬间从 24G 暴跌到 11.6G，而且解码速度几乎毫无损耗
空出来的整整 12G 显存，全被拿去放上下文，
直接把窗口一路顶到了 250,000 的绝对上限，跑完甚至还剩 5 个多 G 的显存余量
再加上把 batch 设到 4096，提示词处理速度直接翻倍狂飙，
以前大家觉得消费级显卡只能跑跑 7B 到 27B 的小模型，
现在 MoE 架构加上内存卸载，直接把数据中心级别的智能焊死在了家用电脑里
本地折腾私有大模型和超长 Agent 的时代，
真的被这套组合拳给彻底砸开大门了啊
Colin retweeted
Our paper, “The Geometry of Low-Resource Language Representations” (w/ @janbuys), has been accepted to #EMNLP Main!
We study the geometry of representation spaces across LLM layers and analyse differences between low- and high-resource languages.
Paper: arxiv.org/abs/2608.23358
Colin retweeted
去年想搞清楚LLM到底怎么工作的，翻了不少教程，要么只讲概念不写代码，要么直接上transformers库把底层全封装了，看完还是懵。
这个项目（Train LLM From Scratch）用纯PyTorch从零实现Transformer，没有trl、peft、transformers这些依赖，
每一步都拆开给你看：数据怎么变成token、模型怎么搭（MLP→注意力→多头→Transformer块→完整模型）、预训练怎么跑、SFT怎么教模型回答问题、Reward Model怎么打分、DPO/PPO/GRPO怎么对齐，整个链路清清楚楚。
代码里每个模块都有解释和预期输出，单GPU就能跑百万到十亿参数的模型，硬件参考也标得挺细。
🔗 地址：github.com/FareedKhan-dev/tr…
Colin retweeted
🏛️ We're excited to hold The UMI Arena workshop at CoRL 2026 — with a real-robot policy competition! 🏆
📦 20,000 hours of UMI data, shared robots, one public leaderboard.
Three tracks: 🤖 policies, 🔧 devices, 📄 papers.
Nov 12 — come play! 🎉 umi-arena.airoa.io 🧵
Colin retweeted
Introducing LpWM: A Case for Sparse Representations in World Models
Dense Gaussian representations are a choice, not a requirement. We find that sparse representations can make latent dynamics easier to model for planning.
📄arxiv.org/abs/2608.22764
💻github.com/YilunKuang/lpworl…
Colin retweeted
ICML has released the videos! Here's my talk: icml.cc/virtual/2026/invited…
Relatedly, @sayashk and I have collected the essays that we think are particularly helpful for understanding the AI as Normal Technology framework, organized into Foundational essays, Applications of the framework, and Technical research and explainers normaltech.ai/p/at-a-glance
I had the honor of giving a keynote at the International Conference on Machine Learning in Seoul last week titled “What will be left for us to work on?” I addressed the widespread anxiety about how we should adapt as AI capabilities increase. I was thrilled by the talk’s reception, so I have made my slides available, annotated with a lightly edited transcript: cs.princeton.edu/~arvindn/ta…
I made three arguments. First, the "AI as Normal Technology" framework is a correct and useful as a way to think about AI’s impacts, unless and until there is some future discontinuity such as through recursive self-improvement. Second, even though we should take recursive self-improvement seriously, there is no milestone that companies might achieve in the lab that will suddenly put us all out of work. Third and finally, jobs of the future will be radically different, and a lot of adaptation will be needed. I shared my thinking about what this might look like and ended with a vision of human/AI “co-superintelligence”.
Colin retweeted
“LeFlow: Generative Latent Flow Planning for World Models”
World models can predict future states, but planning still usually requires expensive online optimization from scratch for every new goal.
This paper learns the planning process itself, generating trajectories directly in the world model’s latent space, decoding them into actions, then verifying them with the frozen world model.
LeFlow gets better success than CEM across four control benchmarks while cutting planning time by 4.5-14.4x
alphaxiv.org/abs/2608.24855v…
Colin retweeted
打破 AI 信息差
每天刷 X、Reddit、YouTube、GitHub、Hacker News，最浪费时间的就是信息太碎：同一个热点重复刷，真正有价值的新东西却藏在不同平台里。
很多内容你看到的时候，可能已经是几周前的旧讨论。
分享一个最近在用的开源神器：Last 30 Days
它的核心也很直接：只研究最近 30 天，全网真正正在讨论什么。
· 自动搜索 Reddit、X、YouTube、TikTok、Hacker News、GitHub 等平台最近一个月的内容
· 结合点赞、评论、互动这些真实信号，优先筛出真正有人关注的话题
· 自动归纳最近什么开始升温、哪些工具突然变火、哪些观点反复出现
· 做产品研究时，可以直接看最近一个月真实用户都在夸什么、吐槽什么
· 做自媒体时，可以直接拿它找最近正在起量的选题和新玩法
· 最后是整理成一份可以直接继续交给 AI 使用的研究结果
用法也很简单，直接给它一个主题，比如：/last30days AI Video
它就会把最近 30 天相关的新工具、新玩法、社区讨论和趋势变化一起扒出来。
GitHub：github.com/mvanhorn/last30da…
信息差时代，真正值钱的不是知道得多，而是更早知道什么正在发生。
感兴趣的朋友可以自己研究一下～ DYOR
Colin retweeted
🎁 Hugging Face 把一整套 AI Agent 课程免费开源了，从原理一路带到自己做出一个能跑的智能体
GitHub 揽获 3.1 万 stars，课程署名作者是 Ben Burtenshaw、Joffrey Thomas、Thomas Simonini 和 Sergio Paniego。
想入门 agent 的人最常卡在同一个地方：教程东一篇西一篇，看完一堆概念，还是不知道自己该动手做什么。这套课把顺序替你排好了，从智能体和大模型基础起步，中间过函数调用微调、可观测性与评估、智能体式 RAG，最后是游戏里的智能体。
框架不是只讲一个，smolagents、LlamaIndex 和 LangGraph 三套都带你走一遍。收尾是一个真项目：自己建智能体、测试、跑自动评估，通过了发认证，还能上学员排行榜。前置要求只有 Python 基础和对大模型的基本了解。
学 agent 最难的从来不是看不懂，是不知道下一步该学什么。
GitHub：github.com/huggingface/agent…
Colin retweeted
💡Videos and images should not be treated merely as inputs to understand or outputs to render. They can be first-class substrates for problem solving beyond language.
🎉Introducing VBVR-Pro: an infrastructure suite for visual intelligence: trainable, verifiable, RL-optimizable, and experimentally controllable.
🔗Homepage: video-reason.com/
Colin retweeted
We've just released code+model for Surflo!🚀
Turn any number of photos into a 3D surface with Flow Matching.
Our latent representation encodes the full scene and can be queried for downstream applications like 3D segmentation.
🔗anttwo.github.io/surflo/
💻github.com/Anttwo/Surflo
What if you could turn any number of photos (3, 8, 15, or even 60) into one clean 3D surface (pts & mesh) with Flow Matching?
Check out our new work, Surflo: Consistent 3D Surface Flow Model with Global State. 🧵
1/n
🔗anttwo.github.io/surflo/
Colin retweeted
Platonic Representation Hypothesis on World Models
TL;DR: World models initialized with different visual encoders converge toward shared internal representations when trained to predict the same environment dynamics. ViT-based models show increasing alignment and functional compatibility via model stitching, while ResNet remains distinct.
arxiv.org/abs/2608.23720
Colin retweeted
arxiv.org/abs/2608.24814
Effective learning rate, the ratio of learning rate and weight norm, governs training dynamics. This could allow transferring the settings across norm control methods (by matching effective learning rate).
Colin retweeted
"On-Policy Self-Distillation in Diffusion Models"
Diffusion RL gets a reward only after the final image, even though the model actually acts through many intermediate denoising steps.
DiffusionOPSD turns that endpoint reward into explicit reward-improving targets at the model’s own intermediate states, then distills those targets back into the model and continually refreshes them on-policy.
Across two diffusion backbones and ten evaluators, it gets the best held-out score in 19 of 20 settings while cutting training GPU-hours by up to 63%.
The key idea is treating diffusion alignment less like noisy trajectory-level RL and more like continuously refreshed self-distillation with direct intermediate supervision.
alphaxiv.org/abs/2608.24646v…
Colin retweeted
Fantastic initiative my dream came true (see my article here paoloai.substack.com/p/stop-…)
Now I just have to move to China!