Question on Table 2 skill retrieval reproducibility

[LvAoAo]

Thanks for your great job.

I checked the current public codebase to understand how the skill retrieval ablation in Table 2 is implemented.

What I could confirm from the released code:

• I could not find a public implementation of UCB or KM / k-means based skill retrieval in the current repository.
• The released skill-based training examples use a fixed SkillProvider with skill_all=false by default. This appears to be rule-based task matching, not Full, not UCB, and not KM.
• skill_all=true would correspond to concatenating all available skills into the prompt / teacher context, so that seems closest to the Full row.

Because of that, I am not sure how to reproduce the UCB and KM rows reported in the paper, and I am also unsure whether the main reported results in the paper rely on KM retrieval or on the currently released rule-based matching.

Could you clarify:

1. Is there a plan to open-source the KM / UCB skill retrieval code or the exact evaluation scripts / configs used for Table 2?
2. Which retrieval setting should be considered the default one for reproducing the paper's main results? (I noticed that 'KM' is the one for main table)

Thanks.