Embodied task planning on BEHAVIOR-1K B-1000 1.0 benchmark leaderboard · SOTA2 Research

Embodied task planning on BEHAVIOR-1K B-1000 1.0

97.3Success Rate (SP)D-SFT+DAPO+Short-RL+GroupAdapt (τ=0.5)

Evaluation Results

Method
D-SFT+DAPO+Short-RL+GroupAdapt (τ=0.5)2026.06	97.3	97.3	99.39	0.2
D-SFT+DAPO+Short-RL2026.06	96.9	96.9	99.23	0
D-SFT+DAPO2026.06	96.2	96.2	98.64	0.2
DeepSeek-V4-Flash2026.06	93.8	96.9	98.36	3.6
GPT-5.42026.06	92.8	93.8	95.34	1
Gemma-4-31B-IT2026.06	92.8	95.4	96.49	4.6
Kimi-K2.62026.06	92.3	95.9	96.88	4.6
D-SFT2026.06	92.3	95.9	98.32	4.1
Qwen3.6-27B2026.06	91.2	94.8	97.21	4.1
GLM-5.12026.06	90.7	94.3	96.88	6.2
DeepSeek-V4-Pro2026.06	89.7	94.8	96.92	6.2
Qwen3.6-35B-A3B2026.06	88.1	93.3	95.17	9.3
Gemini-3.1-Pro2026.06	87.1	87.6	89.39	0.5
Qwen3.5-122B-A10B2026.06	86.6	90.7	93.8	7.2
Qwen3.5-27B2026.06	83	89.7	93.71	12.4
Qwen3-8B2026.06	77.3	83	89.84	13.4
Qwen3.5-35B-A3B2026.06	76.3	90.7	93.14	19.1