Some questions regarding your research

Although you control the final fine-tuning set size and use a comparable rollout budget, the current ablations only demonstrate the effectiveness of the overall FDE/EIR data-synthesis pipeline over root-level parallel sampling. You seem to not isolate the contribution of the proposed heuristic node-selection mechanisms. In particular, comparisons against random or uniform intermediate-node expansion, random failed-prefix recovery, and greedy priority-based selection without the UCB term are missing. Therefore, the observed gains may stem from prefix reuse, increased state/error diversity, or a higher-quality candidate data pool, rather than specifically from the proposed fragility and recovery-node selection criteria.