Dimensions overall score 8.0

GitHub Code Pulse

No public code linked for this paper yet.

• Autonomous Agents Bowen Ye1∗ Rang Li1∗ Qibin Yang1 Yuanxin Liu1 Linli Yao1 Hanglong Lv1 Zhihui Xie2 Chenxin An2 Lei Li2,† Lingpeng Kong2 Qi Liu2 Zhifang Sui1 Tong Yang1 1State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University 2The University of Hong Kong bwye@stu.pkuclaim 72%
• WebArena [50]❖ ✗ ❖ ✗ ✗ ✓ VisualWebArena [11]✓ ✗ ❖ ✗ ✗ ✓claim 72%
• ToolBench [42]✗ ✗ ❖ ✗ ✗ ✗ Terminal-Bench [18]✗ ✗ ✓ ✗ ✗ ✓ PinchBench [10]✗ ✗ ✗ ✗ ✗ ✓ Claw-Eval (Ours)✓ ✓ ✓ ✓ ✓ ✓ surfaced by Claw-Eval point to concrete and actionable directions for building agents that are not only technically capable but alse reliably deployable. 2 Related Work Agent benchmarksclaim 72%
• and TheAgentCompany [41] broaden coverage across heterogeneous environments. PinchBench [10] and WildClawBench [8] benchmark LLM agents on real-world tasks but couple evaluation to the full scaffold stack, precluding attribution of performance to the model itselfclaim 72%
• Think Act Obersve System Layer Tools Audit Logs Temporal Phases Spatial Layer Evidence 1 Evidence 2 Evidence 3 Agent Figure 1: Claw-Eval architecture. The framework operates across three temporal phases (Setup, Execution, Judge) and three spatial layers (Host Machine, Isolation Environment, Mock Services)claim 72%
• Social Science Law, Education, Public policy 13 Business Finance, Investment, Corporate strategy 15 Total 300 In theJudge phase, upon agent termination, grading artifacts (evaluation scripts, reference answersclaim 72%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card