These are complete saved evaluations of Qwen3.5-9B-MR-v6-sft-final-iter2981, with four examples each from TCSBench 2026, IMO ProofBench, and ArxivMath 0626.
The model keeps up to 12 retained execution summaries alongside recursively compacted memory. Older summaries and failure notes are absorbed into memory. The compactor’s full chain of thought is available here; only its resulting memory is passed into later exploration.
Scores are the original saved judgments: normalized rubric points for TCS, official points out of 7 for IMO, and answer correctness for ArxivMath. No answers were regenerated or regraded. The displayed final answer follows the original raw-final grading convention; the parsed export is shown separately.
The exploration map shows execution calls, including those without retained summaries. A compaction marked “0 slots” incorporates discarded-attempt notes. Layer labels start at 1; saved compaction journal indices start at 0.
Downloads include the full original rollout, original judgment, and source checksum. Markdown, mathematics, and all trajectory data are served locally by this Space.