Reasoning Palette: Modulating Reasoning via Latent Contextualization for Controllable Exploration for (V)LMs

Benchmark Model Rank Results
arithmetic-reasoning-on-gsm8kQwen3-8B-Base with Reasoning Palette (RLOO, Linear Decay)–Accuracy: 93.03
mathematical-reasoning-on-math-500Qwen3-8B-Base with Reasoning Palette (RLOO, Linear Decay)–Accuracy: 72.2
mathematical-reasoning-on-minerva-mathQwen3-8B-Base with Reasoning Palette (RLOO, Linear Decay)–Accuracy: 43.77
mathematical-reasoning-on-olympiadbenchQwen3-8B-Base with Reasoning Palette (RLOO, Linear Decay)–Accuracy: 46.61