"Using Cursor with DeepSeek to Quickly Get Started with Unfamiliar Components" by @xuanhun1 #DEVCommunity @ByteDanceOSS @deepseek_ai @cursor_ai #Cursor #OpenAI #DeepSeek

.@ByteDanceOSS introduced a fascinating development, Agent-R. Agent-R helps AI agents "reflect" on their mistakes as they work, learning from them dynamically. Using Monte Carlo Tree Search (MCTS) method, Agent-R fixes errors as they occur, helping the agent learn to self-correct in real time. What's special about Agent-R's workflow? ▪️ It identifies 4 types of reasoning paths: initial, bad, good, and revision. ▪️ Revision trajectories are fixed versions of the bad paths, which are used together with good paths for training the agent to self-correct. Overall, Agent-R's benefits include: - Dynamic error fixing - Self-reflection - Continuous learning - It significantly improves the agent's ability to perform long, complex tasks more effectively: +5.59% improvement over other methods. - Prevents the model from repeating mistakes or getting stuck in loops. - It's open source P.S. Interestingly, multi-task training with Agent-R was more effective than single-task training, so this method can be used to improve performance across different types of tasks. Check out all the links below👇

❓Struggling with serving high-throughput long-context LLMs? 📢 Introducing ShadowKV! 🚀 Achieve high-throughput long-context LLM inference with memory-efficient & low-latency sparse attention! 🌟 ShadowKV minimizes memory footprint by storing low-rank keys on the GPU and offloading values, supporting larger batch sizes without accuracy loss. ⚡ Benchmarked on Llama-3.1-128K, Llama-3-1M, GLM-4-1M, & more, ShadowKV boosts throughput up to 3.04x on an A100 with a 6x batch size or 6x longer sequences! 🔍 Dive in to see how ShadowKV balances speed, memory, and precision across over 128K tokens! 📜 Paper: 🔗Blog: 💻 Code: