HSPR Recovers Full Agent Benchmark Scores From Small Subsets via Hierarchical Posterior Inference
SUDA · hf · 2026-10-20
Researchers at Soochow University and Singapore Management University propose HSPR (Hierarchical Score Posterior Recovery), which combines source-fitted response factors from historical responses with uncertain group success rates, using adaptive task selection and posterior updating to accurately estimate full benchmark scores for software agents from small task subsets, cutting evaluation cost.
More from coding & agent
- MemPilot: RL-orchestrated on-demand multimodal memory for LLM agents — NanyangTechnologicalUniversity · 2026-10-20
- Scale AI unveils Sauron, an evidence-first verifier to cut RL reward errors — ScaleAI · 2026-10-20
- Agent hits level 49 in World of Warcraft with a live map dashboard — max_paperclips · 2026-10-20
- Capability ≠ Permission: a five-principle framework for AI agent autonomy — AiVentureRO · 2026-10-20
- Anthropic's Felix Rieseberg: general models keep absorbing custom AI agents' capabilities — Olivier__OG · 2026-10-20
- Lawyers are Claude's #1 power users as legal usage grows 108x — how to build your own legal team with it — rubenhassid · 2026-10-20