OpenCompass retweeted
🔥China’s Open-source VLMs boom—Intern-S1, MiniCPM-V-4, GLM-4.5V, Step3, OVIS
🧐Join the AI Insight Talk with @huggingface, @OpenCompassX, @ModelScope2022 and @ZhihuFrontier
🚀Tech deep-dives & breakthroughs
🚀Roundtable debates
⏰Aug 21, 5 AM PDT
📺Live: youtube.com/live/kh0WSMoVZmA
OpenCompass retweeted
大语言模型具备稳定推理能力吗?
「来自上海 AI 实验室 @OpenCompassX 的研究,通过创新的评估方法揭示了一个关键问题:尽管大语言模型在单次测试中可能表现出色(如 OpenAI 最新模型单次准确率达 66.5%),但在需要持续稳定输出的场景中,几乎所有模型的表现都会大幅下降(降幅普遍超过 60%),这说明当前 AI 模型的"思考能力"仍不够稳定和可靠」
论文目前在 HF Papers 当日排名居首位,关注度很高 @_akhaliq 👍
一、研究背景和问题
· 虽然大语言模型在数学推理等复杂任务上表现优秀,但基准测试成绩与实际应用效果存在差距
· 现有评估方法无法真实反映模型在实际应用中的稳定性
二、主要创新
· G-Pass@k评估指标:评估模型在多次尝试中保持正确答案的能力
· LiveMathBench测试集:收集最新数学竞赛题目,确保评估的时效性和难度
三、关键评测发现
· 通用大模型表现
- Gemini-1.5-Pro表现最佳:单次测试49.2%,但在严格标准下降至26.8%
- GPT-4和Claude-3.5表现相近:单次成绩约40%,严格测试均降至20%以下
· 数学专用模型表现
- Qwen2.5-Math-72B潜力最高:单次可达50.4%,但稳定性测试降至26.8%
- 其他专用模型表现不佳:DeepSeek-Math和NuminaMath在严格测试中均低于6%
· o1系列模型亮点
- OpenAI o1-mini最为稳定:单次66.5%,严格测试仍保持42.0%
- QwQ-32B紧随其后:严格测试保持33.3%的水平
展现出更好的推理稳定性
四、核心启示
· 稳定性挑战
- 几乎所有模型在严格测试下表现显著下降
- 平均性能衰减超过60%
- 即使是专门优化的数学模型也难以保持稳定表现
· 规模效应局限
- 增加模型参数量并不能显著提升推理稳定性
- 如72B和32B版本的Qwen模型表现相近
· 未来方向
- 需要更注重模型的稳定性而非单次表现
- o1系列的架构创新可能提供了提升稳定性的新思路
- 评估方法需要更贴近实际应用场景
OpenCompass retweeted
OpenCompass has established a leaderboard to evaluate complex reasoning capability of LMMs, consisting of four advanced multi-modal math reasoning benchmarks. Currently, Gemini-2.0-Flash took the 1st place. DM me to suggest more benchmarks and models to this LB.
OpenCompass retweeted
MMBench has been selected as one of the most influential papers at ECCV 2024, ranking second.🎉🎉🎉
paperdigest.org/2024/09/most…
OpenCompass retweeted
OpenCompass has just completed the evaluation of InternVL2-Llama3-76B. With excellent performance, it achieves the 1st place on Open VLM Leaderboard, outperforms GPT-4o by 1PT Avg. Score. @opengvlab @OpenCompassX @wangwenhai362
huggingface.co/spaces/openco…
OpenCompass retweeted
Excited to see that Open VLM Leaderboard has reached 50k visits @OpenCompassX
huggingface.co/spaces/openco…
OpenCompass retweeted
VLMEvalKit has just supported GPT-4o and we are running the evaluation, the results would be available in 1 / 2 days!
#OpenAI #GPT4o
github.com/open-compass/VLME…
OpenCompass retweeted
Welcome to OpenCompass, an open-source evaluation platform for large language model and multi-modality model. #ChatGPT #LargeLanguageModels #ACL2023 #ArtificialIntelligence
🧐Is your model the strongest in '#LLM Battle'?
🥳Try our open-source evaluation toolkit and update your model's performance on OpenCompass leaderboard!
- Efficient distributed evaluation
- Fully support for LMM and Multi-modality model
- 50+ datasets with about 300,000 questions
OpenCompass retweeted
🥳We are delighted to announce that our evaluation toolkit, OpenCompass, has been recommended by Meta (ai.meta.com/llama/get-starte…). @AIatMeta
☺️Welcome to try OpenCompass for your research and production of LLM.
👉github.com/open-compass/open…
#LLM #LLaMA #Meta #OpenAI #OpenCompass













