Demystifying Hidden-State Recurrence: Switchable Latent Reasoning with On-Policy Reinforcement Learning
Open paper
Benchmark
Model
Rank
Results
arithmetic-reasoning-on-gsm8k
Switch (ours, after Switch-GRPO)
–
Accuracy: 89.2
mathematical-reasoning-on-math-500
Switch (ours, after Switch-GRPO)
–
Accuracy: 79.3
Rank counts only results with a code link.