Share your thoughts, 1 month free Claude Pro on usSee more

Tool Use on RobustBench-TC Perturbed 1.0 (test)

52.9Accuracy (Perturbed)

Qwen3-14B

Updated 21d ago

Evaluation Results

Method	Links
Qwen3-14B 2026.05		52.9	0.039	0.143	0.464	0.25
MUA-RL-32B 2026.05		52.7	0.028	-0.01	0.397	0.218
Qwen3-32B 2026.05		52.4	0.024	0.104	0.366	0.377
LoopTool-32B 2026.05		50.8	0.034	0.12	0.509	0.396
o4-mini (OpenAI) 2026.05		50.2	0.049	0.157	0.326	0.276
MUA-RL-8B 2026.05		47.9	0.023	0.156	0.401	0.229
LoopTool-8B 2026.05		47.6	0.049	0.173	0.429	0.296
MUA-RL-14B 2026.05		47	0.011	0.09	0.334	0.202
ToolRL-DR-Full (ours) 2026.05		46.3	0.009	0.147	0.331	0.231
ToolRL-DR-Mixed (ours) 2026.05		46.1	0.049	0.152	0.364	0.232
Qwen3-8B 2026.05		44.8	0.019	0.171	0.409	0.293
DeepSeek-R1-Distill-14B 2026.05		44.2	0.013	0.16	0.318	0.213
Qwen3.5-9B 2026.05		43.7	0.04	0.168	0.411	0.417
ToolRL-Qwen2.5-1.5B 2026.05		43.2	0.033	0.174	0.481	0.379
TL-CodeLLaMA-2 2026.05		40.5	0.024	0.166	0.456	0.333
ToolRL-Qwen2.5-3B 2026.05		40.4	0.02	0.143	0.45	0.315
ToolRL-Llama3.2-3B 2026.05		39	0.025	0.171	0.463	0.296
Qwen2.5-3B-Instruct 2026.05		39	0.014	0.195	0.446	0.237
Qwen2.5-1.5B-Instruct 2026.05		36.7	0.04	0.2	0.405	0.205
Llama-3.2-3B-Instruct 2026.05		34.1	0.009	0.177	0.372	0.174
SFT-Clean-4k-3B 2026.05		26.7	0.019	0.184	0.443	0.577