“OREO (Offline REasoning Optimization)) teaches LLMs to reason better by learning which steps actually matter in solving problems. OREO (Offline REasoning Optimization)) helps LLMs solve complex problems better by learning from both successes and failures, using a smart reward”
https://x.com/rohanpaul_ai/status/1876193273607536800
“NTP (Next Token Prediction) transforms complex multimedia data into simple sequential tokens for AI processing This paper introduces Next Token Prediction (NTP) as a unified framework for processing multiple types of data like images, audio, and text, transforming them into
https://x.com/rohanpaul_ai/status/1876197122707738706
“Self-improvement framework helps AI models reason about images without human guidance, AI learns to critique its own thinking process The paper introduces a systematic framework for improving multimodal reasoning in LLMs through self-evolving training, enhancing models’ ability
https://x.com/rohanpaul_ai/status/1877289514441252986
“this 2025 AI engineer reading list is amazing especially if you are confused on where to start, not a random list of paper titles, it includes 50 papers across 10 fields like: > LLMs > Benchmarks > Prompting > RAG > Agents > Vision > Diffusion > Finetuning
https://x.com/Hesamation/status/1874935860979962173
“Code for the Textualize website, built with React, TypeScript, and Next.js, hosted on Vercel”
https://x.com/tom_doerr/status/1876348788484301309
“ILLA Builder: A low-code platform for building internal tools like dashboards, CRUD apps, and admin panels, supporting real-time collaboration, automation, and self-hosting with Docker and Kubernetes”
https://x.com/tom_doerr/status/1877054737046000025
“Jujutsu (jj) is a Git-compatible version control system that uses changesets instead of commits, eliminates the index, and integrates conflict resolution from patch-based systems, designed for simpler and more efficient version control”
https://x.com/tom_doerr/status/1875072709568106775
“Memory layers store facts in LLMs using 10x less compute than traditional dense layers. Thereby, this Memory augmented LLMs achieve same accuracy as larger models with 90% less computation Memory layers add trainable key-value lookups to LLMs, enabling more parameters without
https://x.com/rohanpaul_ai/status/1876197634446422320
TangoFlux
https://tangoflux.github.io/
“🚀 DNA Data Storage Breakthrough: Storing 215,000 Terabytes in One Gram of DNA! Researchers at Peking University have unveiled a new DNA data storage method called “epi-bits” that uses enzymatic methylation to encode data as epigenetic modifications. This approach allows for
https://x.com/rohanpaul_ai/status/1877346006347690343
LongMemEval
https://xiaowu0162.github.io/long-mem-eval/
CrossEarth
https://cuzyoung.github.io/CrossEarth-Homepage/
DMesh++
https://sonsang.github.io/dmesh2-project/
Training AI models might not need enormous data centres
https://archive.md/x8KmO
[Distributed w/ TorchTitan] Breaking Barriers: Training Long Context LLMs with 1M Sequence Length in PyTorch Using Context Parallel – distributed / torchtitan – PyTorch Forums
https://discuss.pytorch.org/t/distributed-w-torchtitan-breaking-barriers-training-long-context-llms-with-1m-sequence-length-in-pytorch-using-context-parallel/215082
Tetsuwan Scientific is making robotic AI scientists that can run experiments on their own | TechCrunch
Tetsuwan Scientific is making robotic AI scientists that can run experiments on their own
[2501.01275v1] HybridTrack: A Hybrid Approach for Robust Multi-Object Tracking
https://arxiv.org/abs/2501.01275v1
sail-sg/sailor2: 🔱 Sailor2: Sailing in South-East Asia with Inclusive Multilingual LLMs
https://github.com/sail-sg/sailor2
“REINFORCE++ an improvement to the classical REINFORCE that integrates PPO-inspired techniques to achieve simpler, more stable, and efficient RLHF, leading to 30% faster training with comparable performance. What’s new with REINFORCE++ 1️⃣ No Critic Network: Unlike PPO,
https://x.com/_philschmid/status/1877010768689815696
“FlashAttention on a Napkin 🧠 Mathematical visualization framework systematically derives optimal GPU implementations. This paper introduces a diagrammatic approach to optimize deep learning algorithms for IO-awareness, particularly focusing on FlashAttention. It provides a
https://x.com/rohanpaul_ai/status/1875840471869882712
“LlamaIndex Workflows are a powerful way to crystallize an LLM-powered process into a controllable, repeatable form, and this is a great example of how that works! In this deep dive, Lingzhen Chen builds a system that: ➡️ Searches and summarizes academic papers from ArXiv ➡️
https://x.com/llama_index/status/1877044767047168503
[2412.20980v1] Efficient Parallel Genetic Algorithm for Perturbed Substructure Optimization in Complex Network
https://arxiv.org/abs/2412.20980v1
“🔍 After initial warm-up SFT for basic reasoning patterns, integrating PRMs into online RL presents unique challenges. Our implicit PRM approach tackles these head-on with novel solutions: (1) toke-level rewards (2) no need for extra PRM training (3) easy online update!
https://x.com/lifan__yuan/status/1874867820745703687
Human study on AI spear phishing campaigns — LessWrong
https://www.lesswrong.com/posts/GCHyDKfPXa5qsG2cP/human-study-on-ai-spear-phishing-campaigns
[2410.20268] Centaur: a foundation model of human cognition
https://arxiv.org/abs/2410.20268
“Turn any GitHub repository into a LLM prompt-friendly text ingest for LLMs. You can also replace hub with ingest in any github url to access the coresponding digest.
https://x.com/rohanpaul_ai/status/1875791921995771928
Magic Mirror
https://julianjuaner.github.io/projects/MagicMirror/
Commits · lucidrains/PaLM-rlhf-pytorch
https://github.com/lucidrains/PaLM-rlhf-pytorch/commits/main/
NousResearch/DisTrO: Distributed Training Over-The-Internet
https://github.com/NousResearch/DisTrO?tab=readme-ov-file
“Clean data beats complex architecture for better multilingual embeddings KaLM-Embedding achieves state-of-the-art performance in multilingual text embedding by focusing on superior training data quality and innovative data processing techniques. —– 🤔 Original Problem: →
https://x.com/rohanpaul_ai/status/1876213898199933396
Chirpy3D
https://kamwoh.github.io/chirpy3d/
