DeepSeek V4 Flash 0731 推理基准成绩发布
DeepSeek 发布 V4 Flash 0731 模型，在 ARC-AGI 推理基准上取得新成绩。最高档位 ARC-AGI-1 得分 89.0%，ARC-AGI-2 得分 61.4%，单任务成本仅 0.02-0.04 美元。模型提供 Max/High/Low 三档推理强度，成本与性能呈线性关系。 
ARC Prize
ARC 奖基金会（ARC Prize Foundation）是一个非营利组织，通过基准测试和奖金推动开源通用人工智能（AGI）研究。
评论
点赞
收藏ARC-AGI 排行榜：从被动流体智力到实时交互适应
ARC Prize 发布 ARC-AGI 排行榜，涵盖 ARC-AGI 1/2/3 三个版本。ARC-AGI-3 转向评估 AI 在新型交互环境中的即时适应能力。榜单可视化了任务成本与性能的效率关系，区分了推理系统趋势线、基础 LLM 单次推理及 Kaggle 竞赛方案。仅展示运行成本低于 1 万美元的系统，部分结果为预览或估算。 
Inkling：ARC Prize 榜单上得分最高的开源权重模型
Thinking Machines 发布开源模型 Inkling，在 ARC-AGI-1 上得分 79.5%，ARC-AGI-2 上得分 36.5%，刷新了开放权重模型在 ARC Prize 榜单上的最高纪录。该模型单任务推理成本仅约 0.3 至 0.64 美元，且尚未在 ARC-AGI-3 上进行评估。 
GPT-5.6系列：ARC-AGI测试结果
OpenAI发布GPT-5.6系列在ARC-AGI基准测试的成绩。Sol模型在最大推理力度下，于ARC-AGI-3公共测试集平均得分13.33%，并在ft09任务中以87%正确率成为首个赢得该任务的模型。文章详细列出了Sol、Terra、Luna三款模型在不同推理层级（Max至Low）下的具体得分与通过率数据。 
ARC Prize 2026：ARC-AGI-3 第一阶段里程碑大奖揭晓
ARC Prize 2026 公布 ARC-AGI-3 基准测试第一阶段里程碑奖（$37.5K）及前三名开源方案。冠军 Tufa Labs 的 "The Duck" 采用小型开源 LLM，通过实时 REPL 编写运行 Python 代码解决交互式推理游戏，利用图像、ASCII 和分割工具灵活感知状态，并通过消息淘汰机制实现无限上下文下的持续游玩。该方案展示了让模型驱动轻量级 Harness 处理长视界规划与探索的新思路。 
Analyzing GPT-5.5 & Opus 4.7 with ARC-AGI-3
本文通过 ARC-AGI-3 基准测试，对比分析了 GPT-5.5 与 Opus 4.7 的推理过程，发现三大共性失败模式：模型能感知局部操作效果却无法构建正确世界模型；错误地将新环境类比为已知游戏（如 Tetris、Sokoban）；即使解决单个关卡也未能真正理解机制并迁移到下一关。研究还指出两者压缩策略差异——Opus 倾向快速形成错误理论，GPT-5.5 则难以有效压缩信息。该分析工具已开源，旨在推动更透明的 AI 能力评估。
Measuring Human Performance on ARC-AGI-3
Measuring Human Performance on ARC-AGI-3 AGI is here when a system can learn like a human. However there is still a gap between what humans can learn and what AI can learn. ARC Prize Foundation exists...
Announcing ARC-AGI-3
Announcing ARC-AGI-3 A New Challenge for Frontier Agentic Intelligence Today we're excited to announce the release of ARC-AGI-3, a series of hundreds of interactive environments and thousands of game-...
Arc Prize 2025 Results and Analysis: Year of the Refinement Loop
ARC Prize 2025 Results & Analysis Year of the Refinement Loop We've officially wrapped Year 2 of ARC Prize! While the Grand Prize remains unclaimed, we're excited to announce the ARC Prize 2025 Score ...
Arc Prize Verified
Announcing ARC Prize Verified Today we're announcing **ARC Prize Verified**, a program to increase the rigor of evaluating frontier systems on the ARC-AGI benchmark. In addition to certified score ver...
Hidden drivers of HRM's performance on ARC-AGI
The Hidden Drivers of HRM's Performance on ARC-AGI We scored on hidden tasks, ran ablations, and found that performance comes from an unexpected source On June 8, 2025, the Hierarchical Reasoning Mode...
Arc-AGI-3 Preview: 30-day learnings
ARC-AGI-3 Preview: 30-day learnings Highlighting the gap between humans and AI with Interactive Benchmarks ARC-AGI-3, the first Interactive Reasoning Benchmark by ARC Prize Foundation On July 17, we r...
ARC Prize Foundation Statement on the US AI Action Plan
ARC Prize Foundation Statement on the US AI Action Plan Last week, the White House released its AI Action Plan. We’re encouraged to see the White House take transparency and measurement seriously as c...
We tested every major AI reasoning system. There is no clear winner.
We tested every major AI reasoning system. There is no clear winner. Reflecting on Frontier AI Reasoning Systems It's been six months since OpenAI achieved a breakthrough high score on ARC-AGI-1 with....
Analyzing o3 and o4-mini with ARC-AGI
Analyzing o3 and o4-mini with ARC-AGI ARC Prize Foundation is a nonprofit committed to serving as the **North Star for AGI** by building open reasoning benchmarks that highlight the gap between what’s...