GameHorizon-Bench 详情→
GameHorizon Suite introduces GameHorizon-Bench, a benchmark for evaluating gameplay capabilities across horizons with offline and online tracks, covering 21 games and 47 models.
最近发表的 Benchmark
按论文发表日整理公开资源证据，供研究者快速浏览、核对和继续阅读。
2,851公开目录
2,557有已核验代码
521有已核验数据集
资源归属核验不等于已安装、可运行或端到端复现。
浏览完整目录本期收录
论文发表日 2026-09-22 · 2 条BackTrend 详情→
BackTrend is a retrospective benchmark for evaluating scientific weak-signal prediction, containing 25 mature AI/ML topics and 66 human-validated weak signals, with tasks for LLMs, RAG systems, and agentic research systems.
显示该日全部 2 条；完整目录支持搜索和按代码/数据集核验状态筛选。
展开前一天的数据（2026-09-21）
2026-09-21
EvoPathBench 详情→
EvoPathBench is a benchmark for process-level evaluation of self-evolving agents, tracking individual capabilities across artifact checkpoints using public trading data and calibrated trajectories.
DUMA-Bench 详情→
DUMA-Bench is a benchmark and evaluation protocol for measuring LLM agent security under dual-control interaction, where both agent and user influence the environment. It extends tau^2-bench with adversarial environments covering eight vulnerability classes and evaluates 14 model
UK-PRBench 详情→
UK-PRBench is a new benchmark for paragraph-level precedent retrieval in UK case law, constructed from UK National Archives judgments. It evaluates state-of-the-art retrieval models and shows that this task remains challenging.
MemCalib 详情→
MemCalib is a benchmark for evaluating how LLM agents use memory in realistic memory-system scenarios, revealing over- and under-use issues. It also introduces an optimization algorithm, MemCalib-RL, but the primary contribution is the benchmark.
MCP-GRANITE 详情→
MCP-GRANITE is an open-source benchmark framework for evaluating MCP-based LLM agents under edge and IoT scenarios, with 81 multi-step scenarios across 9 domains at 4 granularity levels, measuring task completion, tool selection, argument accuracy, latency, and resource usage.
MCPGen 详情→
MCPGen is an executable benchmark for Model Context Protocol workflow development, containing 100 MCP projects across 16 domains, evaluating LLMs on three diagnostic tasks with static analysis, unit/integration tests, and end-to-end execution.