Reading
This article introduces OpenSeeker-v2, a cutting-edge search agent fully developed by an academic team and trained solely via supervised fine-tuning (SFT). It outperforms industrial-grade models using complex CPT+SFT+RL processes in four authoritative benchmark tests.
Section 01
This article introduces the OpenSeeker-v2 search agent developed by an academic team. Trained solely through supervised fine-tuning (SFT) using 10.6k data points, it outperforms industrial-grade models adopting complex CPT+SFT+RL processes in four authoritative benchmark tests.
Section 02
Deep search capability is the core competitiveness of cutting-edge LLM agents but has long been monopolized by tech giants. The mainstream industrial approach uses a resource-intensive four-stage process: pre-training → continuous pre-training (CPT) → SFT → reinforcement learning (RL), requiring thousands of GPU hours and millions of dollars in investment—something academic teams can hardly afford.
Section 03
The academic team broke through via high-quality training trajectories + SFT method, with three key data synthesis strategies: 1. Knowledge graph scale expansion to increase information depth and breadth, approaching real complex scenarios; 2. Toolset scale expansion to enhance tool usage flexibility and collaboration capabilities; 3. Strict low-step filtering to retain training trajectories with high information density and strong reasoning efficiency.
Section 04
Using a 30B model and the ReAct paradigm, OpenSeeker-v2 performed excellently in four benchmarks: BrowseComp English 46.0% (Tongyi DeepResearch 43.4%), Chinese 58.1% (Tongyi 46.7%); Humanity's Last Exam 34.6% (Tongyi 32.9%); xbench 78.0% (Tongyi 75.0%).
Section 05
Section 06
There is room for improvement: 1. Scale limitation—explore larger models; 2. Expand multimodal capabilities (images, videos); 3. Improve real-time information acquisition efficiency; 4. Enhance safety and controllability.
Section 07
OpenSeeker-v2 proves that academic teams can compensate for resource shortages with ingenuity, marking an important milestone in AI research democratization. Open-source model weights will推动 the development of search agent technology toward an open and inclusive direction.
Keep going with more reads from the same topic.
Splinter is a minimalist, high-performance key-value (KV) and vector storage system enabling zero-latency inter-process communication via shared memory and atomic operations. With only 766 lines of core code, it supports millions of operations per second and 768-dimensional vector storage, offering a new architectural approach for local LLM inference and data-intensive applications.
Folkering OS is the world's first AI-native bare-metal operating system, entirely written in Rust no_std without relying on Linux, POSIX, or libc. It can generate commands from scratch, compile them into WASM, and run them in 10 seconds, achieving true self-evolution.
This project explores how to use Large Language Models (LLMs) to detect logical vulnerabilities in Ethereum smart contracts, especially business logic flaws that are difficult for traditional static analysis tools to capture, such as economic manipulation and execution order errors.
This article introduces an open-source project for building modern large language models from scratch. The project evolves from a GPT-2-style basic architecture to a Llama 2/3-style production-level implementation, covering complete tutorial implementations of key technologies such as RMSNorm, RoPE, SwiGLU, GQA, and MoE.