Add OpenClaw + DeepSeek V4 Pro results (49 tasks, 0.918)

[konghuihua]

No description provided.

[konghuihua]

Zhiyin OpenClaw + DeepSeek V4 Pro Submission

Overview

• Model: DeepSeek V4 Pro
• Harness: OpenClaw 2026.4.24 (zhiyin-pawbench Docker image)
• Tasks scored: 49 out of 150 attempted
• Overall score: 0.918
• Date: 2026-06-12
• Submitted by: Kong Huihua (孔会华) / chunyinengyuan@qq.com
• Repo to PR: https://github.com/agentscope-ai/PawBench

Slice Averages

Slice	Score	Tasks
Text	0.941	32
Multimodal	0.895	9
Code Generation	0.801	6
Document Extraction	1.000	2

Notes

• Auto scores only from official graders (�uto_avg field from _results_v2.jsonl)
• LLM judge dimension not reproduced (requires separate judge run)
• ~100 tasks produced no output files or timed out, only scorable tasks included
• Docker-based evaluation with zhiyin-pawbench:latest image

Submitted by

Zhiyin / CausalMind

[helloml0326]

Thanks for the submission. A couple of quick questions:

We already have official deepseek-v4-pro × openclaw results (150 tasks, overall 0.754). This PR adds 49 tasks at 0.918 — what does it add beyond the existing data?

Also, ~101 tasks had no output or timed out. What caused that, and do you plan to re-run the full 150?

For leaderboard inclusion we need the complete task set (missing tasks count as 0). A partial run isn't directly comparable to existing submissions.