Reinforcement learning with verifiable rewards

Technique · Training · 4 items since May 2026

RL where rewards come from automatic checks, such as passing tests or matching a known answer.

Trends

Jul 25
Meituan's CAST turns solver values into rewards
Converts a game solver's change in state value into a dense turn-level reward for RLVR, provably equivalent to logit-free on-policy distillation, beating trained baselines on Sokoban, Minesweeper, and Rush Hour and transferring zero-shot to ALFWorld and WebShop.
Jun 11
RA-RFT retrieves demonstrations by reasoning benefit
Retrieval-Augmented Reinforcement Fine-Tuning ranks demonstrations by expected reasoning benefit rather than semantic similarity, then RL-fine-tunes on the retrieved traces, improving AIME 2025 average@32 by 7.1 and 2.8 points over GRPO for Qwen3-1.7B and Qwen3-4B.
May 16
Error diversity predicts RLVR training success
EDAS amplifies advantage penalties for dominant repeated errors and softens them for rare exploratory failures, adding an average 6.29 points over DAPO on Qwen3-8B across seven math benchmarks as a drop-in for any RLVR algorithm.
May 14
SU-01 reaches olympiad gold at 30B-A3B
A 30B-A3B MoE trained with reverse-perplexity curriculum SFT, coarse RLVR then proof-level RL, and test-time verification reaches gold on IMO 2025, USAMO 2026, and IPhO 2024–25, scoring 35/42 and matching the top human among 340 USAMO competitors.