UC Berkeley
The University of California, Berkeley, a research university with a distinguished reputation in computer science and artificial intelligence.
Core Facts
vLLM是由UC Berkeley开发的高吞吐量大模型推理引擎，核心创新是PagedAttention技术
The LLM-as-Judge methodology was first systematically proposed by UC Berkeley and other institutions in 2023.
SGLang is a high-performance framework designed for large language model inference, developed by researchers from UC Berkeley and other institutions.
Timeline (last 90 days)
Minando is currently collaborating with colleagues at UC Berkeley to explore deployment paths for privacy-enhancing technologies (PETs) in the mainstream world, a path that first involves working with governments and NGOs to understand data-sharing needs and barriers, then customizing or redesigning PETs, and ultimately producing a privacy deployment handbook for the mainstream world.
According to a UC Berkeley test, seven frontier models showed no significant differences in SWE-Bench Lite scores across different harnesses such as Claude Code and Codex, and the newer the model, the smaller the differences.
Both studies found significant differences in per-task cost, even when eval score differences were small.
UC Berkeley research tested 7 frontier models on a variety of harnesses and found that for newer, smarter models, the choice of harness made no significant difference in SWE-bench Lite scores.
A UC Berkeley study ran seven frontier models on SWE-bench Lite across multiple harnesses such as Pi, Codex and Cloud Code, and found that on newer, smarter models the choice of harness had no significant effect on scores.
Both the UC Berkeley and the Apple/EPFL studies found significant differences in per-task cost under different harness choices.
A UC Berkeley study tested multiple harnesses such as Pi, Codex and Claude Code with seven frontier models and found that on newer, smarter models the choice of harness had no significant effect on SWE-bench Lite scores.
vLLM originates from UC Berkeley research on PagedAttention (Kwon et al.), and its core goal is to solve the memory waste problem in traditional inference serving.
UC Berkeley's research on multi-agent system failure taxonomy (MAST) shows that production-grade multi-agent workflows have failure rates as high as 41% to 86.7%.
The Dominant Resource Fairness (DRF) algorithm was proposed by UC Berkeley in 2011.
33 more timeline events
All Facts (20)
vLLM是由UC Berkeley开发的高吞吐量大模型推理引擎，核心创新是PagedAttention技术
90%VerifiedThe LLM-as-Judge methodology was first systematically proposed by UC Berkeley and other institutions in 2023.
80%VerifiedSGLang is a high-performance framework designed for large language model inference, developed by researchers from UC Berkeley and other institutions.
80%VerifiedLMSys Chatbot Arena was created by the LMSYS organization at UC Berkeley and has collected more than one million human votes since its launch in 2023.
75%VerifiedLM Arena (formerly Chatbot Arena), maintained by UC Berkeley's LMSYS team, uses an Elo rating system to measure model capabilities through user blind-test comparative voting.
70%UnverifiedChatbot Arena was introduced by a research team at UC Berkeley in 2023 and is one of the earliest platforms to implement large-scale LLM battle evaluation
90%UnverifiedMMLU was introduced in 2021 by UC Berkeley and other institutions as a large-scale multitask language understanding benchmark
90%UnverifiedCognition的三位创始人Scott Wu、Neal Wu和Walden Yan曾代表UC Berkeley参加ACM编程竞赛，并拥有IOI金牌荣誉
85%UnverifiedSGLang由UC Berkeley团队主导开发，是目前与vLLM并列的两大主流开源LLM推理框架之一
85%UnverifiedThe vLLM project was initiated by the UC Berkeley team.
70%UnverifiedTcl (Tool Command Language) was created by John Ousterhout during his tenure at the University of California, Berkeley.
60%UnverifiedMinando is currently collaborating with colleagues at UC Berkeley to explore deployment paths for privacy-enhancing technologies (PETs) in the mainstream world, a path that first involves working with governments and NGOs to understand data-sharing needs and barriers, then customizing or redesigning PETs, and ultimately producing a privacy deployment handbook for the mainstream world.
50%UnverifiedAccording to a UC Berkeley test, seven frontier models showed no significant differences in SWE-Bench Lite scores across different harnesses such as Claude Code and Codex, and the newer the model, the smaller the differences.
50%UnverifiedUC Berkeley research tested 7 frontier models on a variety of harnesses and found that for newer, smarter models, the choice of harness made no significant difference in SWE-bench Lite scores.
50%UnverifiedBoth studies found significant differences in per-task cost, even when eval score differences were small.
50%UnverifiedA UC Berkeley study ran seven frontier models on SWE-bench Lite across multiple harnesses such as Pi, Codex and Cloud Code, and found that on newer, smarter models the choice of harness had no significant effect on scores.
50%UnverifiedBoth the UC Berkeley and the Apple/EPFL studies found significant differences in per-task cost under different harness choices.
50%UnverifiedA UC Berkeley study tested multiple harnesses such as Pi, Codex and Claude Code with seven frontier models and found that on newer, smarter models the choice of harness had no significant effect on SWE-bench Lite scores.
50%UnverifiedvLLM originates from UC Berkeley research on PagedAttention (Kwon et al.), and its core goal is to solve the memory waste problem in traditional inference serving.
50%UnverifiedUC Berkeley's research on multi-agent system failure taxonomy (MAST) shows that production-grade multi-agent workflows have failure rates as high as 41% to 86.7%.
50%