Yujia Qin
@TsingYoga
Followers
2,986
Following
276
Media
41
Statuses
234
Explore trending content on Musk Viewer
Brazil
• 1338090 Tweets
#BUS1stFANCON_KnockKnockKnock
• 364381 Tweets
アニサマ
• 192122 Tweets
GULF NEW SONG
• 67175 Tweets
TEDDYTINT x POND
• 49575 Tweets
Djokovic
• 29719 Tweets
음악중심
• 28521 Tweets
札幌2歳
• 25415 Tweets
Briggitte
• 24639 Tweets
Ticketmaster
• 22238 Tweets
JAMESSU FM IN QINGDAO
• 21275 Tweets
Stanford
• 17564 Tweets
Popyrin
• 17358 Tweets
SUNSET X VIBES WATCH PARTY
• 15011 Tweets
레이브 1위
• 14330 Tweets
台風一過
• 13956 Tweets
メドゥーサ
• 10137 Tweets
Last Seen Profiles
是时候把数据scale down了
LLaMA3告诉大家一个悲观的现实：模型架构不用动，把数据量从2T加到15T就可以暴力出奇迹。这一方面告诉大家基座模型长期来看就是大厂的机会；另一方面，考虑到scaling
16
151
701
🥳🛠️Introducing ToolBench!🤖🎉
🌟Large-scale instruction tuning SFT data to equip LLMs with general tool-use capability
🔖 We release 98k data with 312k real API calls. We also release a capable model ToolLLaMA that matches ChatGPT in tool use
Github:
9
89
441
视觉-语言模型(VLM)领域在研究些什么？🧐
VLM是一个从去年末开始快速发展的领域，对研究者来说尚有大量“金矿”未被发掘，且当前探索仍然非常初步，对大模型的初学者上手难度较小🥰
以下是帮你快速掌握VLM领域目前发展的文章推荐📰：
1.
12
74
313
我们离理想的AutoGPT还有多远？[1/4]
AutoGPT[1]已经163k star了，AutoGPT的开发者雕花了一年多，但它仍然停留在demo阶段，算不上产品（即使面向开发者)。这和传统开源软件的发展轨迹相差甚远，核心原因是Agent的上限由底座模型决定
20
55
278
🚀 Introducing XAgent: The next evolution of AI agents designed for intricate task-solving. XAgent completely outperforms AutoGPT and GPT-4 on various tasks and benchmarks.
💡 XAgent's dual-loop mechanism bridges the gap between high-level planning and detailed task execution.
2
47
230
It's Time to Scale Down the Data
LLaMA3 revealed a grim reality: without changing the model architecture, increasing the data from 2T to 15T can forcefully produce miracles. This tells us that, in the long run, foundational models are only an opportunity for big companies. On
1
44
213
最近KV-Cache层间复用的工作有点火，YOCO（You Only Cache Once）逼出了好几篇在研的工作，感觉半年内差不多能再看到几十篇后续的优化工作。
去年从StreamingLLM开始，大家变着法子在Decoder-LM上雕花SparseAttention。
今年KV-cache复用从inter-sentence
5
41
208
量化不是LLM掉点的原罪，post-train才是？
量化对模型性能的损失大家有目共睹，尤其是对reasoning能力要求很高的任务，LLM量化完（即使再post-train）几乎就不能看了
但其实量化不是LLM掉点的原罪。去年的BitNet [1]就已经告诉大家，量化后的模型呈现和vanilla Transformer一致的scaling
0
29
180
📰 像看报纸一样轻松读懂Github项目代码
🔍 一键解析热门项目，了解最in技术和场景
💡 连接你的创意和实现,Coder灵感之源
🧑💻 每个项目都有专家级助手随时解惑
🌉 连接开发者,共创AI时代
GitRead带你冲浪最前沿,一起开启代码
#GitRead
立即体验 成为项目阅读小能手!
5
33
148
What is the Vision-Language Model (VLM) field researching?🧐
VLM is a rapidly developing field since the end of last year, with a large amount of "gold mines" yet to be discovered for researchers. The current exploration is still very preliminary, and it is relatively easy for
9
37
109
大模型对“数数”这件事做得一直很不好😔
比如，“让模型续写137个字”、“统计这段话里出现he”的次数这种任务，在不显示CoT的情况下，效果都非常差，而且看起来不是scaling能解决的
😍最近看到两篇很有意思的从position embedding (PE)角度帮助Transformer“数数”的工作分享一下
3
16
106
🥳Successfully defended my thesis! 🎓✨
It's been an incredible journey at Tsinghua Univ. Thank you to everyone who supported me along the way!
Here’s to new beginnings ahead!
#ThesisDefense
#PhDone
7
2
95
Is post-training the original sin of LLM performance drop, not quantization?
It is widely acknowledged that quantization leads to a significant loss in model performance, especially for tasks that require high reasoning abilities. Quantized LLMs (even after post-training) are
2
17
83
Thanks for sharing🥰!
Highlights of our work:
1 🛠️16000+ real APIs from RapidAPI
2 🤔Better reasoning strategy (DFSDT) for LLMs than CoT
3 🤖ToolLLaMA achieves comparable performance than turbo-16k
4 🔍An API retriever
...
Codes, models, demo are out:
ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs
paper page:
Despite the advancements of open-source large language models (LLMs) and their variants, e.g., LLaMA and Vicuna, they remain significantly limited in performing
8
161
616
1
21
65
🚀Plenty of adaptation methods have been proposed for tuning PLMs, and all of them can find high-performance minima.
🧐However, what is the connection among various minima reached under different adaptation configurations?
🥳Check our paper on EMNLP 2022!
#NLProc
3
5
50
Thanks for sharing🥰! Highlights of our work:
(1) 🛠️16000+ real APIs from RapidAPI
(2) 🤔Better reasoning strategy (DFSDT) for LLMs than ReACT
(3) 🤖ToolLLaMA achieves comparable performance than turbo-16k
(4) 🔍An API retriever
(5) 🏁ToolEval for auto-evaluation
...
ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs
ToolLLaMA exhibits comparable performance to ChatGPT
repo:
abs:
11
154
633
0
11
46
Just got accepted to main conf of NAACL 2022 🤗
Thank all the co-authors for the work!
#NAACL2022
@naaclmeeting
Knowledge Inheritance for Pre-trained Language Models
pdf:
abs:
github:
pre-training framework, knowledge inheritance, combines both self-learning and teacher-guided learning to efficiently train larger PLMs
1
7
44
4
6
42
📌📌📌
Humans possess an extraordinary ability to create and utilize tools 🪛🔨🔧
With the advent of foundation models, does AI🤖 have the potential to be equally adept and capable as its creators? 🧐
Check out our survey on
#ToolLearning
📜
Paper:
3
7
30
以前听别人说“相信N年内（e.g., N<3）实现AGI”的话会觉得可笑
最近自己反思了下，其实这大多不是理性判断，而是感性信念
举个例子，gpt4出来后大家才能“超越3.5，接近4”，gpt5不出来没人能（实际）超越4；同样的，sora出来了才会有可灵这么牛逼的东西跟进；
9
1
30
How Far Are We from the Ideal AutoGPT? [1/4]
AutoGPT has already garnered 163k stars, and its developers have been working on it for over a year. However, it remains in the demo stage and can't be considered a fully-fledged product (even for developers). This trajectory differs
1
7
29
Heading to Vienna tomorrow for
#ICLR2024
, will be presenting our ToolLLaMA on 5/8 4:30 pm. Feel free to grab a coffee for (1) LLM agent, (2) GUI automation (VLM), and (3) repo-level coding/debugging.
1
1
25
最近聊了几位大厂的算法一号位，了解到因为合规问题，国内几个大厂是**一直不能**用GPT/Claude API来蒸馏数据的（他们确实是一点没用），post-training的落后容易导致前期没有声音。
2
1
24
@tianle_cai
@Google
Very interesting and inspiring work! We've also done similar things (i.e., investigating LLM's tool creation capabilities) recently, check out:
1
5
22
Amazing experience using
@AdeptAILabs
Adept Experiments. Glad to see Adept first introduce the concept of AI workflow in the industry🥳
Actually we've also released a paper about the next generation of RPA -> "Agentic Process Automation" (APA). Personally I believe APA (autogpt
1
3
21
Why could different parameter-efficient tuning methods with distinct structures achieve comparable performance to fine-tuning? Are they intrinsically connected? Check our recent work on Findings of
#emnlp2022
🎇
#NLP
1
0
16
The corresponding codes:
We have recently integrated BabyAGI into BMTools.
1
7
14
😂
大模型对“数数”这件事做得一直很不好😔
比如，“让模型续写137个字”、“统计这段话里出��he”的次数这种任务，在不显示CoT的情况下，效果都非常差，而且看起来不是scaling能解决的
😍最近看到两篇很有意思的从position embedding (PE)角度帮助Transformer“数数”的工作分享一下
3
16
106
3
0
14
Thanks for sharing!! We've also created a must-read paper list on tool learning, and would be continually updating it in the future.
#ToolLearning
Link:
0
2
14
Just discovered ScrapeGraphAI - a Python library that makes web scraping a breeze! 🕷️
You tell it what info you want, and it builds the scraping pipeline for you.
Perfect for devs who hate writing complex scraping logic.
Check it out if you're into AI-powered data
0
2
14
Wanna tune few-shot NLP tasks with 5 free parameters😉?
Check our recent preprint: Exploring Low-dimensional Intrinsic Task Subspace via Prompt Tuning. We explore how can PLMs effectively adapt to broad NLP tasks differing a lot superficially.
1
3
14
Glad that Andrew also likes the idea of agentic workflow, which was initially proposed in our paper from last year🤪
I think AI agentic workflows will drive massive AI progress this year — perhaps even more than the next generation of foundation models. This is an important trend, and I urge everyone who works in AI to pay attention to it.
Today, we mostly use LLMs in zero-shot mode, prompting
213
1K
5K
1
2
13
Check out our continually updated paper list on Tool Learning🛠️🔧🔨🪛🔩🪚🪓⛏️:
Welcome contributing~
1
4
13
GitRead is on Product Hunt!
It helps read codes like a newspaper
0
1
12
Large models have always struggled with "counting" 😔
For example, tasks like "having the model continue writing for 137 characters" or "counting the occurrences of 'he' in this passage" perform very poorly without showing the CoT, and it doesn't seem like scaling can solve
0
0
9
News from our XAgent Team:
🥰XAgent on DockerHub: Official release of XAgent Container Images now available! 📷 Check it out.
🥳Localhost Models on HuggingFace: XAgentLlama-7B and 34B models are live! 📷 Explore here.
😊XAgentGen Released: Enhance your experience with our new
0
1
9
超级印钞机巨头们在早期都被质疑过“怎么赚钱”🤔 技术创新->商业模式成熟需要时间⌛
当年纽约时报嘲笑Google商业模式，现在看多搞笑的事但在当时被质疑很正常
Google和Meta的成长史告诉我们：对AI商业模式，多点耐心没坏处。伟大的事物需要时间酝酿，也给我们点儿时间！
0
1
9
The first self-evolving Agent based on XAgent~
BTW, XAgent-2.0 is ready soon~
📢Release our work on agent self-evolution!
We propose the ICE experience learning strategy that makes agent deployment significantly more effective and efficient.
ICE is also integrated into XAgent 2.0 about to be released soon!
Find out more here:
1
19
49
0
0
8
🤗Many **amazing** findings of delta tuning would be released in the next few months 💅
🎉Thrilled to introduce our latest work, Delta Tuning: A Comprehensive Study of Parameter Efficient Methods for Pre-trained Language Models. We perform a comprehensive theoretical analysis and experimental validation of the parametrically efficient paradigm for
#PLMs
.
1
23
80
1
0
8
New Work Released🥰!
We combine ELO rating system 🎮into autonomous decision making, which significantly improves the efficiency and effectiveness of machine tool use.
Introducing JuDec:
🤖 Autonomous decision making: Eliminating the need for task-specific expert knowledge.
🧠 Optimal Solution Searching: Assessing solutions quantitatively.
📊 Effectiveness and Efficiency: Over 10% improvements with lower cost.
Paper:
0
0
4
0
2
8
Hope to see mistral-pretrain in the near future~
We just released mistral-finetune, the official repo and guide on how to fine-tune Mistral open-source models using LoRA:
Also released Mistral-7B-Instruct-v0.3 with support for function calling with Apache 2.0 license:
7
139
761
1
1
7
Our XAgent receives 5k stars🔯 within a month🥳
The next version (70% codes added) will be released soon😉
🚀 Introducing XAgent: The next evolution of AI agents designed for intricate task-solving. XAgent completely outperforms AutoGPT and GPT-4 on various tasks and benchmarks.
💡 XAgent's dual-loop mechanism bridges the gap between high-level planning and detailed task execution.
2
47
230
0
0
7
孔哥做的很赞的真实评测，推荐大家以此为准聊各家模型长文本实际效果🤗
0
0
6
未来我们还能用到Meta的开源模型么？😅
传统软件开源，程序员在上面接着开发优化；大模型这种大多数程序员没法参与二次开发的，本身价值就收窄很多，出现底层枯竭上层应用开花
这个投入产出比，加上政治因素，很难想象未来还能一直免费用LLaMA。另一个角度上或许给国内持续预训练的厂商一波机会？
0
0
6
Many thanks for sharing our work🥳
Tell Me More! Towards Implicit User Intention Understanding of Language Model Driven Agents
Trains a powerful model that proactively assesses task vagueness, inquires user intentions, and refines them into actionable goals before starting downstream agent task execution
repo:
0
9
42
0
0
5
Best 8b-VLM ever, surpassing GPT4V on benchmark. Very exciting to see how LLaMA-3V would become in the coming months!🤩
🚀 Excited to introduce MiniCPM-Llama3-V 2.5! With 8B parameters, it’s our latest breakthrough, outperforming top models like GPT-4V. 📈
💪 Superior OCR capabilities
🔑 Supports 30+ languages
HuggingFace:
GitHub:
1
36
81
0
1
5
LLM领域有个所有人都在讲但从来没有被验证过的“数据飞轮”
最近看了一个做擦边聊天的产品收的数据，突然意识到这个时代用户数据其实并没有特别重要，用户也给不了高质量的标注，质量还不如合成数据。
0
0
5
Many long-context methods have been proposed recently. Despite their superficial differences in the design, most of them can be unified through the lens of memory augmentation under one framework (i.e., they are actually the same!!). See our new work UniMem!!🥳
Introducing UniMem: A unified framework redefining long-context processing in large language models (LLMs)!
🌟 By integrating a spectrum of long-text strategies within a memory enhancement framework, UniMem harmonizes these techniques across four key dimensions: memory
0
3
8
0
1
5
Our thinking on Next-Generation AI Agent: XAgent + Zapier😍
ProAgent: From Robotic Process Automation to Agentic Process Automation
paper page:
From ancient water wheels to robotic process automation (RPA), automation technology has evolved throughout history to liberate human beings from arduous tasks. Yet, RPA
0
37
149
1
0
5
Scanning AI papers on arxiv every day is very troublesome. We have made a website to facilitate everyone to match the most relevant/highest quality papers. If you are interested, please give us some feedback. We will quickly optimize.
最近关于PM要不要读论文的讨论越来越热烈了
不管怎么说算法都要读论文的
一位算法朋友为了满足自己需要搞了网站
- 用算法选出每天最值得看的10篇AI论文
- 也可以自己搜索感兴趣的论文话题
- 论文解决的问题都用中文列好了
- 未来研究思路启发灵感
如果你也想AI论文，可以试试
11
24
84
1
1
5
最近看到国内某大佬的观点，感觉很有道理：
大模型商业化之路崎岖,关键在于质量达标。这其实为中国创业者带来新的机遇:
1.打磨垂直领域数据
2.人机协作弥补短板
3.瞄准容错高的场景
AI赋能消费电子前景可期,依托华强北的模式需本土合伙人
0
0
4
Glad to see this work come out. Our ToolLLaMA has been upgraded lately (released soon), can't wait to test how well it performs against NexusRaven-13B
🔥NexusRaven-13B, a new SOTA OSS LLM for function calling.
📊Match GPT-3.5 in zero-shot for generic domain and beat it by 4% in security software.
🛠️Outperform GPT-4 by up to 30% with retrieval augmentation for software unseen during training.
Blog: (1/3)
3
26
105
0
0
4
Thanks for sharing this work!
Enabling LLMs with tool-use capabilities is where I am noticing the greatest potential for companies to go big with LLMs.
Gorilla is a good popular example but I have seen a ton of other examples, especially from people building with AI-powered agents. I also think this is one
6
143
538
0
0
4
All the data is automatically generated by OpenAI API and filtered by us, the whole data creation process is easy to scale up. We built ToolBench based upon BMTools:
0
0
4
不开放接口/用户即用的Agent产品都是一眼丁真
0
0
4
今天调GPT-4o的API做个QA测试，发现回答的内容里有前几天刚发的yolo-v10，所以GPT-4o的API已经默认把GPT4All-tools merge进来了嘛？
Today, while testing the GPT-4o API for QA, I found that it referenced the recently released YOLO-v10. Does this mean the GPT-4o API has already integrated
0
0
4
UniMem is accepted to
@COLM_conf
Check our findings in unifying memory (long-context) architecture for Transformer!
Many long-context methods have been proposed recently. Despite their superficial differences in the design, most of them can be unified through the lens of memory augmentation under one framework (i.e., they are actually the same!!). See our new work UniMem!!🥳
0
1
5
0
1
4
✨ Features of ToolBench
1. Multi-step tool invocations and diverse scenarios like weather inquiry and PPT automation
2. Support both single-tool (LangChain) and multi-tool (AutoGPT) settings
3. Include model's inner thought process
4. Diverse real-world tools and APIs
0
0
4
@imxiaohu
So cheap, why not free？Let's witness another cash-burning war, like in China's internet companies 10 years ago 😅
1
0
3
Whispers in the machine... 🤖🌌
ChatTTS emerges from the digital shadows. A voice synthesizer unlike any other. It laughs. It pauses. It speaks in tongues. 100,000+ hours of whispered secrets.
Dare you give AI its voice?
#ChatTTS
#AIAwakens
#DigitalWhispers
Link:
0
2
3
经朋友提醒，发现自己有一篇老论文从Google Scholar的收录里消失了（google scholar里原来是可以搜到的，有这篇论文的收录链接，个人profile里也能看到这篇论文），目前arxiv链接 / semantic scholar还是正常的，求助下情况该怎么解决？
4
0
3
之前训VLM模型的一个认知是，对于OCR能力，像大模型这种seq2seq的架构是很难学会的（即使堆了再多的数据）。
比如GPT-4o / llava-llama3对一些简单的中英文还是会识别错
相比之下传统OCR算法/工具简单粗暴效果还好。
所以路线上一直坚信应该站在已有的OCR工具肩膀上去训模型。不知道大家什么看法？
2
0
3
Very impressive step towards AutoAgent ㊗️
After 5 month dedicated work from >15 researchers & developers, we're thrilled to introduce 🚀OPEN-SOURCE language model Agents🚀!
Try demos: 🥑
Stay tuned for open-source code, model, framework, evaluation & more at !
6
50
177
0
0
3
???
Introducing Claude 3.5 Sonnet—our most intelligent model yet.
This is the first release in our 3.5 model family.
Sonnet now outperforms competitor models on key evaluations, at twice the speed of Claude 3 Opus and one-fifth the cost.
Try it for free:
450
2K
7K
0
0
3
Very powerful tool to implement your multi-agent environment!
Pokémon: In the game, agents can visit shops, train their Pokémon at the gym, and interact with one another. As a player, you take on the role of an agent and engage with others at any time. There are 6 characters in the Pokémon environment who appeared in Pokemon Emerald. 6/n
0
0
1
0
1
2
@_florianmai
@gg42554
Same here. My March ARR submission (4, 4.5, 2.5, meta 4) only gets into Findings of EMNLP, with 2.5 dude giving ridiculous reviews. It seems PCs are too busy scanning the submission. Not sure whether ARR is worth considering anymore. 😞
1
0
2
We strive to investigate three facets: (1) the effects that affect PLM's mode connectivity, (2) the change of mode connectivity during pre-training, and (3) the change of task knowledge along the path connecting two minima.
Paper:
1
0
2
In the found unified subspace, the minima of all DETs have excellent transferring performance, forming a low-loss / high-performance manifold in the parameter space. Such a phenomenon indicates that DETs are mode connected intrinsically.
#mode_connectivity
0
0
1
hh
0
0
1
From the perspective of parameter space, PLMs provide generic initialization, starting from which high-performance minima could be found. For the first time, we investigate the geometric connections of different minima through the lens of mode connectivity.
1
0
1
To fathom the connections among various delta tuning (aka parameter-efficient tuning) methods (DETs), we hypothesize that the adaptations of different DETs could all be reparameterized as low-dimensional optimizations in a unified optimization subspace.
1
0
1
😅
《Sora 的买家秀，当魔术背后的真相被揭开，走下神坛？》
SORA能够生成整个视频，一次性可长达一分钟，这在技术上是一个巨大的进步，尤其是它在保持视频中主体一致性方面的能力。
在网上所放出的精挑细选的影片中，Sora 让人印象深刻，但同时大家也都知道这是 cherry pick 的卖家秀。
12
33
122
0
0
1
@LChoshen
@jang_yoel
Merging adaptors actually works, but not that well compared with full-parameter fine-tuning. To attain good linear connectivity, it is always better to use the same initialization. Also, finding a merging point using a non-linear path is another choice.
0
0
1
@hhhaoran
感谢关注~ IPT / ELLE这两篇都很可惜，审稿结果两级分化很严重。
IPT我们已针对低分审稿人严重误解的点做了详细的解释/修改，后续会继续完善相关论文;
ELLE由于同期工作过多，我们打算接受findings
0
0
1
类似的，的受众是不是图的是免费能搜代码的GPT4？
下午和
@thinkanyai
的朋友聊天
他对
@perplexity_ai
能跑出来有个洞察还蛮有趣的。
在去年那个时间点有个奇妙的时间差
ChatGPT 只能用 bing
Google 的大模型当时又不太行
只有 PPLX，能用GPT4+Google提供最好的AI搜索体验。
4
1
24
0
0
1
Agent as Tool Retriever🫡
LLM Agent for Large-Scale API Calls
Cool research paper presenting AnyTool, an LLM-based agent that can utilize 16000+ APIs from Rapid API.
Proposes a simple framework consisting of
- a hierarchical API-retriever to identify relevant API candidates to a query
- a solver to
5
170
786
0
0
1
@omarsar0
Temporarily, folks could use our dataset, which supports real-world APIs, and multi-step, multi-tool processes with reasoning traces and real tool executions. 100k+ SFT data released at
0
0
1