datahref @datahref

Joined January 2013

Tweets

1K
Followers

35
Following

182
Likes

1K

Rohan Paul @rohanpaul_ai

a week ago

A classic paper, collab between @AIatMeta , @GoogleDeepMind , and @NVIDIAAIDev Language models keep personal facts in a measurable amount of “storage”. This study shows how to count that storage—and when models swap memorization for real learning. 📡 The Question Can we…

16 171 939 100K 939

Download Image

Sumit @_reachsumit

2 weeks ago

On the Theoretical Limitations of Embedding-Based Retrieval @orionweller et al. at Google DeepMind demonstrate that vector embeddings have fundamental limitations in representing all possible document combinations. 📝arxiv.org/abs/2508.21038 👨🏽‍💻github.com/google-deepmin…

4 98 522 62K 484

elvis @omarsar0

2 months ago

One Token to Fool LLM-as-a-Judge Watch out for this one, devs! Semantically empty tokens, like “Thought process:”, “Solution”, or even just a colon “:”, can consistently trick models into giving false positive rewards. Here are my notes:

13 122 700 99K 639

Download Image

Viking @vikingmute

2 months ago

昨天闲逛发现的免费网站：oldmapsonline.org 简直是历史爱好者的福音，通过一个时间线可以显示那个时候的世界地图，国家，区域，还可以选择更多的信息，比如当时的皇帝，战役等等。做的非常棒，一下就看了很久，还有旧地图搜集，非常有趣的网站，我觉得就是这些有趣的网站让我觉得 web…

29 851 3K 185K 3K

Download Video

Zhaochen Su @SuZhaochen0110

2 months ago

Excited to share our new survey on the reasoning paradigm shift from "Think with Text" to "Think with Image"! 🧠🖼️ Our work offers a roadmap for more powerful & aligned AI. 🚀 📜 Paper: arxiv.org/pdf/2506.23918 ⭐ GitHub (400+🌟): github.com/zhaochen0110/A…

7 63 161 15K 89

Download Image

Chaitanya K. Joshi @chaitjo

3 months ago

Really thought-provoking new paper on representation learning and the notion of 'semantic compression' by @ChenShani2 @jurafsky @ylecun @ziv_ravid

8 111 623 80K 685

Download Image

Neel Nanda @NeelNanda5

2 months ago

I'm very excited about our vision for "mech interp" of CoT: Study reasoning steps and their connections - analogous to activations Don't just read it: study attn, causally intervene, and, crucially, resampling - study the distn over CoTs, not just this one There's lots to do!

Paul Bogdan @paulcbogdan

2 months ago

18 152 772 119K 840

Download Video

4 62 556 50K 440

Download Image

Shubham Saboo @Saboo_Shubham_

2 months ago

This Chinese AI company trained the world's first open-weight, large-scale hybrid-attention reasoning model. MiniMax M1 has 1M context window and outperforms Claude 4 Opus in LiveCode Bench. All of this in less than $535k.

30 141 1K 117K 1K

Download Image

Rohan Paul @rohanpaul_ai

2 months ago

GPT-4o level intelligence in a 24B open weights model That's Mistral Small 3.2

Artificial Analysis @ArtificialAnlys

2 months ago

GPT-4o level intelligence in a 24B open weights model That's Mistral Small 3.2 https://t.co/rGqFqSyEC2

14 64 516 85K 158

Download Image

7 47 313 31K 138

Download Image

机器之心 JIQIZHIXIN @jiqizhixin

2 months ago

Attention! A new and improved attention mechanism has just been proposed by MIT, NVIDIA, Princeton, and others. Radial Attention is a sparse, static attention mechanism with O(n log n) complexity. It focuses on nearby tokens and shrinks the attention window over time. It can…

3 144 916 95K 753

Download Image

Rohan Paul @rohanpaul_ai

3 months ago

LLMs running reinforcement learning shed their randomness almost at once, then their scores stall. This paper shows that randomness drop is predictable and fixable, so bigger gains are still on the table. The authors fit an exponential link between entropy and reward. Two…

10 90 519 68K 550

Download Image

Rohan Paul @rohanpaul_ai

3 months ago

Training LLMs for every task is slow and costly. Existing methods like few-shot fine-tuning do not match dedicated task-specific models. This paper proposes ManyICL, fine-tuning one model once on many examples to perform diverse tasks using only in-context examples, closing the…

8 31 128 17K 101

Download Image

Akshay 🚀 @akshay_pachaar

3 months ago

Top 4 open-source LLM finetuning libraries! From single-GPU “click-to-tune” notebooks to trillion-param clusters, these four libraries cover every LLM finetuning scenario. Understand which one to use, & when...👇

12 265 1K 128K 2K

Download Image

Jyo Pari @jyo_pari

3 months ago

What if an LLM could update its own weights? Meet SEAL🦭: a framework where LLMs generate their own training data (self-edits) to update their weights in response to new inputs. Self-editing is learned via RL, using the updated model’s downstream performance as reward.

130 527 3K 600K 3K

Download Image

Rohan Paul @rohanpaul_ai

3 months ago

LLMs apply costly, fixed strategies to all reasoning steps, wasting resources on simple tasks. R2-Reasoner routes reasoning subtasks to LLMs or Small Language Models based on complexity, saving significant cost while keeping accuracy. It cuts API costs by 86.85%. Methods 🔧:…

0 26 89 6K 50

Download Image

merve @mervenoyann

3 months ago

I learnt a lot from O'Reilly books, this is surreal I'm writing a book with amazing people @micuelll @andimarafioti @orr_zohar about VLMs with @huggingface 📖💗 Early Access (first two chapters in raw) are available to everyone, we'd love to have your feedback!

84 214 2K 111K 1K

Download Image

David Wan @meetdavidwan

3 months ago

Excited to share our new work, CLaMR! 🚀 We tackle multimodal content retrieval by jointly considering video, speech, OCR, and metadata. CLaMR learns to dynamically pick the right modality for your query, boosting retrieval by 25 nDCG@10 over single modality retrieval! 🧐…

1 66 184 34K 123

Download Image

Nick Jiang @nickhjiang

3 months ago

Vision transformers have high-norm outliers that hurt performance and distort attention. While prior work removed them by retraining with “register” tokens, we find the mechanism behind outliers and make registers at ✨test-time✨—giving clean features and better performance! 🧵