List your company
ResearchFormerly WizWand

General Reasoning on MMLU-R

84.4Accuracy (MMLU-R General Reasoning)ROSA2
-3.30039219.46797942.2363565.004721Sep 27, 2025Oct 23, 2025Nov 18, 2025Dec 14, 2025Jan 9, 2026Feb 4, 2026Mar 2, 2026
Updated 5mo ago

Evaluation results

MethodLinks
ROSA2Model=Qwen3-8B2026.0384.4—
ROSAModel=Qwen3-8B2026.0375.8—
TextGradModel=Qwen3-8B2026.0370.6—
ROSA2Model=Qwen2.5-7B-Base2026.0363—
TextGradModel=Qwen2.5-7B-Base2026.0360.2—
ROSAModel=Qwen2.5-7B-Base2026.0360.2—
ROSA2Model=DeepSeek-R1-Distill-Llama-8B2026.0359.4—
BaselineModel=Qwen3-8B2026.0357—
ROSA2Model=Qwen3-0.6B-Instruct2026.0350—
ROSAModel=Qwen3-0.6B-Instruct2026.0348.8—
TextGradModel=Qwen3-0.6B-Instruct2026.0346.4—
TextGradModel=DeepSeek-R1-Distill-Llama-8B2026.0343.4—
ROSAModel=DeepSeek-R1-Distill-Llama-8B2026.0342.8—
BaselineModel=Qwen2.5-7B-Base2026.0339.8—
BaselineModel=Qwen3-0.6B-Instruct2026.0324—
BaselineModel=DeepSeek-R1-Distill-Llama-8B2026.0323.6—
ROSA2Model=Qwen2.5-0.5B-Instruct2026.0318.4—
TextGradModel=Qwen2.5-0.5B-Instruct2026.0312.4—
ROSAModel=Qwen2.5-0.5B-Instruct2026.0311.4—
BaselineModel=Qwen2.5-0.5B-Instruct2026.039.4—
ROSAModel=Qwen3-8B, Update Location=LM Head, Reward Model=Model-based2025.090.7036—
ROSAModel=Qwen2.5-7B-Instruct, Update Location=LM Head, Reward Model=Model-based2025.090.6847—
ROSAModel=Qwen3-8B, Update Location=Hidden States, Reward Model=Rule-based2025.090.6837—
ROSAModel=Qwen2.5-7B-Instruct, Update Location=Hidden States, Reward Model=Rule-based2025.090.6731—
ROSAModel=Qwen3-8B, Update Location=LM Head, Reward Model=Rule-based2025.090.6727—
ROSAModel=Qwen2.5-7B-Instruct, Update Location=LM Head, Reward Model=Rule-based2025.090.6217—
BaselineModel=Qwen3-8B2025.090.5135—
ROSAModel=DeepSeek-R1-Distill-Llama-8B, Update Location=LM Head, Reward Model=Model-based2025.090.4579—
BaselineModel=Qwen2.5-7B-Instruct2025.090.4536—
ROSAModel=DeepSeek-R1-Distill-Llama-8B, Update Location=Hidden States, Reward Model=Rule-based2025.090.4218—
ROSAModel=DeepSeek-R1-Distill-Llama-8B, Update Location=LM Head, Reward Model=Rule-based2025.090.4114—
ROSAModel=Qwen3-0.6B, Update Location=LM Head, Reward Model=Model-based2025.090.4068—
ROSAModel=Qwen3-0.6B, Update Location=Hidden States, Reward Model=Rule-based2025.090.36—
ROSAModel=Qwen3-0.6B, Update Location=LM Head, Reward Model=Rule-based2025.090.334—
BaselineModel=DeepSeek-R1-Distill-Llama-8B2025.090.3046—
BaselineModel=Qwen3-0.6B2025.090.186—
ROSAModel=Qwen2.5-0.5B-Instruct, Update Location=LM Head, Reward Model=Model-based2025.090.1372—
ROSAModel=Qwen2.5-0.5B-Instruct, Update Location=Hidden States, Reward Model=Rule-based2025.090.11—
ROSAModel=Qwen2.5-0.5B-Instruct, Update Location=LM Head, Reward Model=Rule-based2025.090.0907—
BaselineModel=Qwen2.5-0.5B-Instruct2025.090.0727—
BaseBackbone=Qwen2.5-3B-Instruct, Protocol=Single-turn baselines2026.05—76.8
BaseBackbone=Llama3.1-8B-Instruct, Protocol=Single-turn baselines2026.05—84.5
DRIFT-5turnBackbone=Qwen2.5-3B-Instruct, Protocol=Ours, Max turns=52026.05—84.6
DRIFT-5turnBackbone=Llama3.1-8B-Instruct, Protocol=Ours, Max turns=52026.05—87.6
PPOBackbone=Qwen2.5-3B-Instruct, Protocol=Single-turn baselines2026.05—78.8
PPOBackbone=Llama3.1-8B-Instruct, Protocol=Single-turn baselines2026.05—87
SCoRe-2turnBackbone=Qwen2.5-3B-Instruct, Protocol=Multi-turn baselines (Online), Max turns=22026.05—85.4
SCoRe-2turnBackbone=Llama3.1-8B-Instruct, Protocol=Multi-turn baselines (Online), Max turns=22026.05—86.6
SFTBackbone=Qwen2.5-3B-Instruct, Protocol=Single-turn baselines2026.05—77
SFTBackbone=Llama3.1-8B-Instruct, Protocol=Single-turn baselines2026.05—85.3
SFT-5turnBackbone=Qwen2.5-3B-Instruct, Protocol=Multi-turn baselines (Offline), Max turns=52026.05—78.2
SFT-5turnBackbone=Llama3.1-8B-Instruct, Protocol=Multi-turn baselines (Offline), Max turns=52026.05—85.4
STaR-2turnBackbone=Qwen2.5-3B-Instruct, Protocol=Multi-turn baselines (Offline), Max turns=22026.05—84.4
STaR-2turnBackbone=Llama3.1-8B-Instruct, Protocol=Multi-turn baselines (Offline), Max turns=22026.05—86.1
UFO-5turnBackbone=Qwen2.5-3B-Instruct, Protocol=Multi-turn baselines (Online), Max turns=52026.05—87
UFO-5turnBackbone=Llama3.1-8B-Instruct, Protocol=Multi-turn baselines (Online), Max turns=52026.05—90.9