Tasks every model fails teach RL nothing: training signal lives at 1-6/8 solve rate
geoffwolfe · x · 2026-11-10
ReasonCoreAI argues that training benchmarks shouldn't aim for "hard": tasks where every model fails (or passes) give RL no gradient signal. Real learning signal comes from mid-difficulty tasks a model solves 1 to 6 times out of 8 attempts.
But building and certifying such tasks is expensive: one terminal-bench-science task can take a frontier agent up to 3 hours, certification requires 8 runs—redone in full whenever a new model ships or a single file changes. That's up to 24 agent-hours per task before it's worth handing to a lab.
More from Research
- New cryo-EM structure: pSTAT1 bound to importin alpha 5 (PDB 11GJ) — chaitjo · 2026-11-10
- ARLoop: fully-open 1.2B looped transformer beats Huginn-3.5B and Qwen3.5-2B on all 8 benchmarks — hhsun1 · 2026-11-10
- RAGCap-Bench: A Fine-Grained Benchmark for Intermediate Capabilities in Agentic RAG — kalyan_kpl · 2026-11-10
- RBA gives LLaVA-OneVision-2 a 3x lossless training speed-up, merged upstream — liuziwei7 · 2026-11-10
- Networks aren't scale invariant — so why has scale invariance motivated so many useful methods? — jasondeanlee · 2026-11-10
- SPACE-Tag preprint: high-throughput spatial chromatin profiling via CUT&Tag — anshulkundaje · 2026-11-10