Verify evals on Papers with Code

[NielsRogge]

Hi,

Niels here from the open-source team at Hugging Face. Congratulations on your work!

I've made the paper and 10 verified paper-native evaluations available on Papers with Code.

The paper has results on Agents, Coding Agents, and World Knowledge task pages.

The GLM-5 result currently ranks second on τ²-Bench.

Would it be possible to verify these results and let me know if any score, model name, benchmark protocol, or openness metadata should be corrected? The imported rows are tied to the paper or its official release artifacts; comparison-table baselines were not added.

You can also edit the task, methods, project page, and GitHub URL directly from the paper page using your Hugging Face account.

If you'd like to showcase the results in your repository README, you can copy these live leaderboard badges (or use the “Copy PwC badge” button in the Results section):

Kind regards,

Niels

[zRzRzRzRzRzRzR]

I verified the 10 imported evaluations against Table 7 of the tech report (arXiv:2602.15763v2) — the scores all match the paper-native results, so the import looks correct.

One thing worth adding to the τ²-Bench protocol metadata: the 89.7 was obtained with minor prompt adjustments for Retail/Telecom and the Airline domain fixes from the Claude Opus 4.5 system card (see Section 6.1.3 of the report), not the fully vanilla protocol. Flagging it so side-by-side comparisons on the leaderboard stay apples-to-apples. Otherwise looks good — thanks!

[NielsRogge]

Thanks a lot for reviewing! I've made similar ones available for GLM-5.2: #123.

I'll update the τ²-Bench results to indicate those differences, thanks.