Interactive Program Induction - Query Strategy - Diversity: leaderboard

Appendix Table6, full anonymized CodeARC split, one seed invocation, 30-I/O budget including seed. Prior majority uses no oracle queries and three seed runs; other strategies use five runs. Native instruction-tuned models, shared input distribution and Mokav judge. LM-only and EIG repeat primary protocols and are not duplicated. No information-gain or query-count proxy. Metric: Program accuracy (0–1). Source: https://arxiv.org/html/2611.18485v1

Metric: Program accuracy (0–1). Source: arxiv.org. Saturation forecast: Around August 2027. 6 models tracked. Updated 17 Nov 2026.

Top models

#ModelScoreOverall rank
1Qwen 2.5 32B Instruct0.316#1257
2OLMo 3 7B Instruct0.235#6916
3Qwen 2.5 7B Instruct0.213#2706
4Llama 3.1 8B Instruct0.204#4576
5Llama 3.2 3B Instruct0.166#7230
6Ministral-8B-Instruct-24100.162#4569

No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.

Interactive version: theaggregate.ai/benchmark?slug=interactive-program-induction-query-strategy-diversity · How It Works · Data refreshed daily, snapshot 2026-11-17.