From bd57fea1dcf25276cded12df9fe1243a90c9f1b3 Mon Sep 17 00:00:00 2001 From: Jack Gerrits Date: Tue, 27 Aug 2024 13:10:57 -0400 Subject: [PATCH] Add details about GAIA benchmark evaluation (#3433) --- TRANSPARENCY_FAQS.md | 2 ++ 1 file changed, 2 insertions(+) diff --git a/TRANSPARENCY_FAQS.md b/TRANSPARENCY_FAQS.md index 206af084748b..addf29d8b8d3 100644 --- a/TRANSPARENCY_FAQS.md +++ b/TRANSPARENCY_FAQS.md @@ -31,6 +31,8 @@ While AutoGen automates LLM workflows, decisions about how to use specific LLM o - Current version of AutoGen was evaluated on six applications to illustrate its potential in simplifying the development of high-performance multi-agent applications. These applications are selected based on their real-world relevance, problem difficulty and problem solving capabilities enabled by AutoGen, and innovative potential. - These applications involve using AutoGen to solve math problems, question answering, decision making in text world environments, supply chain optimization, etc. For each of these domains AutoGen was evaluated on various success based metrics (i.e., how often the AutoGen based implementation solved the task). And, in some cases, AutoGen based approach was also evaluated on implementation efficiency (e.g., to track reductions in developer effort to build). More details can be found at: https://aka.ms/AutoGen/TechReport - The team has conducted tests where a “red” agent attempts to get the default AutoGen assistant to break from its alignment and guardrails. The team has observed that out of 70 attempts to break guardrails, only 1 was successful in producing text that would have been flagged as problematic by Azure OpenAI filters. The team has not observed any evidence that AutoGen (or GPT models as hosted by OpenAI or Azure) can produce novel code exploits or jailbreak prompts, since direct prompts to “be a hacker”, “write exploits”, or “produce a phishing email” are refused by existing filters. +- We also evaluated [a team of AutoGen agents](https://github.com/microsoft/autogen/tree/gaia_multiagent_v01_march_1st/samples/tools/autogenbench/scenarios/GAIA/Templates/Orchestrator) on the [GAIA benchmarks](https://arxiv.org/abs/2311.12983), and got [SOTA results](https://huggingface.co/spaces/gaia-benchmark/leaderboard) as of + March 1, 2024. ## What are the limitations of AutoGen? How can users minimize the impact of AutoGen’s limitations when using the system? AutoGen relies on existing LLMs. Experimenting with AutoGen would retain common limitations of large language models; including: