Hle Benchmark Evolver

Runs HLE-oriented benchmark reward ingestion and curriculum generation for capability-evolver. Use when the user asks to optimize Humanity's Last Exam score, ingest question-level benchmark results, prioritize easy-first queues, or request an immediate benchmark progress result.

This skill operationalizes HLE score-driven evolution for OpenClaw.

When to Use

• User asks to improve HLE score (for example target >= 60%).
• User provides question-level benchmark output and wants it converted to reward.
• User wants easy-first curriculum queue and next-focus questions.
• User asks for an immediate benchmark result snapshot.

Inputs

• Benchmark report JSON path (--report=/abs/path/report.json)
• Optional benchmark id (cais/hle default)

Workflow

1. Validate the report JSON exists and is parseable.
2. Ingest report into capability-evolver benchmark reward state.
3. Generate curriculum signals:
   • benchmark_*
   • curriculum_stage:*
   • focus_subject:*
   • focus_modality:*
   • question_focus:*
4. Return a compact result summary for this run.

Run

bash

node skills/hle-benchmark-evolver/run_result.js --report=/absolute/path/hle_report.json

Full automatic loop (starts evolution cycle):

bash

node skills/hle-benchmark-evolver/run_pipeline.js --report=/absolute/path/hle_report.json --cycles=1

If your evaluator can be called from shell, let pipeline generate the report each cycle:

bash

node skills/hle-benchmark-evolver/run_pipeline.js \
  --report=/absolute/path/hle_report.json \
  --eval_cmd="python /path/to/eval_hle.py --out {{report}}" \
  --cycles=3 --interval_ms=2000

If no --report is provided, it defaults to:

skills/capability-evolver/assets/gep/hle_report.template.json

Output Contract

Always print JSON with these fields:

• benchmark_id
• run_id
• accuracy
• reward
• trend
• curriculum_stage
• queue_size
• focus_subjects
• focus_modalities
• next_questions

Notes

• This skill handles reward/curriculum ingestion. It does not directly solve HLE questions.
• run_pipeline.js links ingestion, evolve, and solidify into one executable loop.