List your company
ResearchFormerly WizWand

Multi-turn Conversation Evaluation on MT-bench

81.1MT-Bench Score+RL (Skywork-Reward-V2-Llama-3.1-8B)
-3.0474418.7985340.644562.49047May 30, 2025Jul 26, 2025Sep 22, 2025Nov 18, 2025Jan 15, 2026Mar 13, 2026May 10, 2026
Updated 5mo ago

Evaluation results

MethodLinks
+RL (Skywork-Reward-V2-Llama-3.1-8B)Model=Qwen2.5-7B2025.0781.1——
+RL (Skywork-Reward-V2-Qwen3-4B)Model=Qwen2.5-7B2025.0779——
Instruct (official)Model=Qwen2.5-7B2025.0778.8——
+RL (Skywork-Reward-Gemma-2-27B-v0.2)Model=Qwen2.5-7B2025.0778.2——
+RL (Skywork-Reward-Llama-3-8B-v0.2)Model=Qwen2.5-7B2025.0776.8——
+SFTModel=Qwen2.5-7B2025.0767.3——
+RL (Skywork-Reward-V2-Llama-3.1-8B)Model=Llama-3.1-8B2025.0766.5——
Instruct (official)Model=Llama-3.1-8B2025.0765.7——
BaseModel=Qwen2.5-7B2025.0763.5——
+RL (Skywork-Reward-V2-Qwen3-4B)Model=Llama-3.1-8B2025.0762.8——
+RL (Skywork-Reward-Gemma-2-27B-v0.2)Model=Llama-3.1-8B2025.0758.5——
+RL (Skywork-Reward-Llama-3-8B-v0.2)Model=Llama-3.1-8B2025.0757.1——
+SFTModel=Llama-3.1-8B2025.0756.8——
BaseModel=Llama-3.1-8B2025.0752.8——
DPO + MaPPOModel=Qwen2.5-7B-Instruct2026.058.79——
ITModel=Qwen2.5-7B-Instruct2026.058.61——
DPOModel=Qwen2.5-7B-Instruct2026.058.56——
I-DPOModel=Qwen2.5-7B-Instruct2026.058.55——
I-DPO + MaPPOModel=Qwen2.5-7B-Instruct2026.058.54——
DPO + MaPPOModel=Llama-3-8B-Instruct2026.058.18——
DPO + MaPPOModel=Qwen2.5-3B-Instruct2026.058.13——
I-DPOModel=Qwen2.5-3B-Instruct2026.058.1——
DPOModel=Llama-3-8B-Instruct2026.058.07——
I-DPO + MaPPOModel=Llama-3-8B-Instruct2026.058.04——
DPOModel=Qwen2.5-3B-Instruct2026.058.02——
I-DPO + MaPPOModel=Qwen2.5-3B-Instruct2026.058.01——
I-DPOModel=Llama-3-8B-Instruct2026.058.01——
ITModel=Qwen2.5-3B-Instruct2026.057.92——
I-DPO + MaPPOModel=Qwen2.5-1.5B-Instruct2026.057.63——
Qwen2.5-7B-Writing-RLBackbone=Qwen2.5-7B, Stage=Writing-RL2025.067.62——
EvoPrefOptimization Paradigm=Multi-Objective Evolutionary, Base Model=Mistral-7B-Instruct-v0.2, LoRA Rank=16, LoRA Alpha=322026.057.62——
I-DPO + MaPPOModel=Mistral-7B-Instruct2026.057.59——
ORPOOptimization Paradigm=Gradient-Based, Base Model=Mistral-7B-Instruct-v0.2, LoRA Rank=16, LoRA Alpha=322026.057.58——
DPO + MaPPOModel=Qwen2.5-1.5B-Instruct2026.057.57——
SMS-EMOAOptimization Paradigm=Multi-Objective Evolutionary, Base Model=Mistral-7B-Instruct-v0.2, LoRA Rank=16, LoRA Alpha=322026.057.54——
EvoPref-BestOptimization Paradigm=Multi-Objective Evolutionary, Base Model=Mistral-7B-Instruct-v0.2, LoRA Rank=16, LoRA Alpha=322026.057.53——
ITModel=Llama-3-8B-Instruct2026.057.52——
DPOOptimization Paradigm=Gradient-Based, Base Model=Mistral-7B-Instruct-v0.2, LoRA Rank=16, LoRA Alpha=322026.057.51——
DPO + MaPPOModel=Mistral-7B-Instruct2026.057.51——
MOEA/DOptimization Paradigm=Multi-Objective Evolutionary, Base Model=Mistral-7B-Instruct-v0.2, LoRA Rank=16, LoRA Alpha=322026.057.5——
IPOOptimization Paradigm=Gradient-Based, Base Model=Mistral-7B-Instruct-v0.2, LoRA Rank=16, LoRA Alpha=322026.057.47——
CMA-ESOptimization Paradigm=Single-Objective Evolutionary, Base Model=Mistral-7B-Instruct-v0.2, LoRA Rank=16, LoRA Alpha=322026.057.43——
KTOOptimization Paradigm=Gradient-Based, Base Model=Mistral-7B-Instruct-v0.2, LoRA Rank=16, LoRA Alpha=322026.057.4——
I-DPOModel=Qwen2.5-1.5B-Instruct2026.057.39——
Qwen2.5-7B-WritingBench-SFTBackbone=Qwen2.5-7B, Stage=WritingBench-SFT2025.067.34——
DPOModel=Qwen2.5-1.5B-Instruct2026.057.29——
Qwen2.5-7B-InstructBackbone=Qwen2.5-7B, Stage=Instruct2025.067.21——
Zephyr Base Modellambda value=12025.057.07——
DSA:LLD:TAlambda value=02025.057.07——
ITModel=Qwen2.5-1.5B-Instruct2026.057.06——
I-DPOModel=Mistral-7B-Instruct2026.056.92——
DSA:LST+:TAlambda value=02025.056.89——
DPOModel=Mistral-7B-Instruct2026.056.86——
Llama3.1-8B-WritingBench-SFTBackbone=Llama3.1-8B, Stage=WritingBench-SFT2025.066.42——
Zephyr LoRA-alignedlambda value=02025.056.37——
Llama3.1-8B-Writing-RLBackbone=Llama3.1-8B, Stage=Writing-RL2025.066.29——
Llama3.1-8B-InstructBackbone=Llama3.1-8B, Stage=Instruct2025.066.16——
ITModel=Mistral-7B-Instruct2026.055.4——
Base (8B Instruct)Backbone=Llama 32026.050.794——
L3-LATBackbone=Llama 32026.050.763——
LPA (ours)Backbone=Llama 32026.050.734——
L3-CATBackbone=Llama 32026.050.686——
Base (7B chat-hf)Backbone=Llama 22026.050.632——
LPA (ours)Backbone=Llama 22026.050.618——
LPA-overfit (ours)Backbone=Llama 22026.050.606——
L2-CATBackbone=Llama 22026.050.555——
LPA-overfit (ours)Backbone=Llama 32026.050.467——
L2-LATBackbone=Llama 22026.050.189——
FedITFoundation model=TinyLlama, Strategy=Homo, Fine-tuning algorithm=FedIT2024.09—2.922.55
FedITFoundation model=Llama, Strategy=Homo, Fine-tuning algorithm=FedIT2024.09—3.073.73
FLORAFoundation model=TinyLlama, Strategy=Homo, Fine-tuning algorithm=FLORA2024.09—3.132.77
FLORAFoundation model=TinyLlama, Strategy=Heter, Fine-tuning algorithm=FLORA2024.09—3.142.71
FLORAFoundation model=Llama, Strategy=Homo, Fine-tuning algorithm=FLORA2024.09—4.213.93
FLORAFoundation model=Llama, Strategy=Heter, Fine-tuning algorithm=FLORA2024.09—4.143.64
LoRAFoundation model=TinyLlama, Strategy=Centralized, Fine-tuning algorithm=LoRA2024.09—2.342.79
LoRAFoundation model=Llama, Strategy=Centralized, Fine-tuning algorithm=LoRA2024.09—4.383.99
Zero-paddingFoundation model=TinyLlama, Strategy=Heter, Fine-tuning algorithm=Zero-padding2024.09—1.561.29
Zero-paddingFoundation model=Llama, Strategy=Heter, Fine-tuning algorithm=Zero-padding2024.09—3.513.26