🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

It's coming along 🌈

#changelog And we're back. We had some downtime the last few hours due to some unexpected changes our hosting provider made that didn't play well with Emergent Mind's codebase. 👉 updates.emergentmind.com/and-we're-back…

Upvotes welcome🔺

New research introduces GemFilter, an algorithm that uses early layers of LLMs to filter and compress inputs, achieving a 2.4× speedup and 30% GPU memory reduction while maintaining performance, thus enhancing both efficiency and interpretability.: emergentmind.com/search?q=2409.…

New research introduces MaskLLM, a learnable pruning method that reduces computational overhead in LLMs by using Gumbel Softmax sampling to create efficient and transferable sparsity masks, resulting in significant improvements in model perplexity and adaptability across various

#changelog More detailed answers. 👉 updates.emergentmind.com/more-detailed-…

New research introduces KV-Compress, a novel method that significantly improves memory efficiency for LLMs by achieving up to 64x compression rates with minimal performance impact, thereby increasing throughput and supporting larger decoding batches.: emergentmind.com/search?q=2410.…

New research introduces the L-Mul algorithm, which replaces floating point multiplications with integer additions, achieving higher precision with significantly less computation and energy use, demonstrating near lossless application in transformer models.:

#changelog Weekly updates. ✅ Rolled out various updates to make the responses more detailed. ✅ Added support for LaTeX ✅ Re-added a Let's Chat link to the footer. 👉 updates.emergentmind.com/weekly-updates…

👋

New research presents an end-to-end reinforcement learning method that enables language models to improve code synthesis by effectively utilizing feedback, achieving state-of-the-art results in competitive programming tasks with fewer samples.: emergentmind.com/search?q=2410.…

#changelog Find the latest research on any topic. Many of you have been trying to find recent research on specific topics, such as "RAG papers in the last 2 weeks". 👉 updates.emergentmind.com/find-the-lates…

New research shows that the o1 system from OpenAI surpasses previous LLMs in reasoning tasks but still retains sensitivity to task probability, indicating that optimization can reduce but not eliminate this limitation.: emergentmind.com/search?q=2410.…

New research introduces Diff Transformer, a model that enhances relevant context attention by subtracting noise, significantly outperforming traditional Transformers in language modeling, long-context tasks, and robustness, while mitigating issues like hallucination and

New research finds that training LLMs on elementary cellular automata shows greater model intelligence with complex rules, suggesting a sweet spot of complexity fosters reasoning and prediction abilities.: emergentmind.com/search?q=2410.…

New research introduces MLE-bench, highlighting that OpenAI's o1-preview with AIDE scaffolding reaches Kaggle bronze medal standards in 16.9% of competitions, offering insights into AI agents' ML engineering capabilities.: emergentmind.com/search?q=2410.…

New research highlights that UGround, a visually-grounded model for GUI agents, significantly outperforms existing models by using a human-like visual perception approach, achieving up to 20% better results without relying on text-based inputs.: emergentmind.com/search?q=2410.…

New research presents a novel method to train LLMs for explicit thinking, enhancing their performance on reasoning and non-reasoning tasks like marketing and health, without additional human data.: emergentmind.com/search?q=2410.…

❤️

New research presents Astute RAG, a novel method that improves the reliability of Retrieval-Augmented Generation by resolving knowledge conflicts and outperforming previous methods, even under worst-case scenarios.: emergentmind.com/search?q=2410.…