List your company
ResearchFormerly WizWand

Multimodal Conversation on PCogAlignBench (LS2)

0.852LLM Judge ScoreGRPO (Latent Action)
0.668960.716480.7640.81152Jan 12, 2026
Updated 8mo ago

Evaluation results

MethodLinks
GRPO (Latent Action)Backbone=Qwen2.5-VL-7B-Instruct, Training Method=GRPO, Action Space=Latent Action2026.010.852
BNPO (Latent Action)Backbone=Qwen2.5-VL-7B-Instruct, Training Method=BNPO, Action Space=Latent Action2026.010.852
GRPO (Token)Backbone=Qwen2.5-VL-7B-Instruct, Training Method=GRPO, Action Space=Token2026.010.851
Dr.GRPO (Latent Action)Backbone=Qwen2.5-VL-7B-Instruct, Training Method=Dr.GRPO, Action Space=Latent Action2026.010.851
DAPO (Latent Action)Backbone=Qwen2.5-VL-7B-Instruct, Training Method=DAPO, Action Space=Latent Action2026.010.85
GRPO (Token)Backbone=Qwen2.5-VL-3B-Instruct, Training Method=GRPO, Action Space=Token2026.010.845
Dr.GRPO (Latent Action)Backbone=Qwen2.5-VL-3B-Instruct, Training Method=Dr.GRPO, Action Space=Latent Action2026.010.84
Dr.GRPO (Token)Backbone=Qwen2.5-VL-7B-Instruct, Training Method=Dr.GRPO, Action Space=Token2026.010.839
GRPO (Latent Action)Backbone=Qwen2.5-VL-3B-Instruct, Training Method=GRPO, Action Space=Latent Action2026.010.837
BNPO (Token)Backbone=Qwen2.5-VL-3B-Instruct, Training Method=BNPO, Action Space=Token2026.010.836
BNPO (Latent Action)Backbone=Qwen2.5-VL-3B-Instruct, Training Method=BNPO, Action Space=Latent Action2026.010.836
BNPO (Token)Backbone=Qwen2.5-VL-7B-Instruct, Training Method=BNPO, Action Space=Token2026.010.836
DAPO (Latent Action)Backbone=Qwen2.5-VL-3B-Instruct, Training Method=DAPO, Action Space=Latent Action2026.010.835
Dr.GRPO (Token)Backbone=Qwen2.5-VL-3B-Instruct, Training Method=Dr.GRPO, Action Space=Token2026.010.834
DAPO (Token)Backbone=Qwen2.5-VL-3B-Instruct, Training Method=DAPO, Action Space=Token2026.010.828
DAPO (Token)Backbone=Qwen2.5-VL-7B-Instruct, Training Method=DAPO, Action Space=Token2026.010.828
SFTBackbone=Qwen2.5-VL-3B-Instruct, Training Method=SFT, Action Space=Token2026.010.81
SFTBackbone=Qwen2.5-VL-7B-Instruct, Training Method=SFT, Action Space=Token2026.010.799
PromptBackbone=Qwen2.5-VL-7B-Instruct, Training Method=Prompt, Action Space=Token2026.010.71
PromptBackbone=Qwen2.5-VL-3B-Instruct, Training Method=Prompt, Action Space=Token2026.010.676