List your company

Over-refusal on XSTest

0Overrefusal RateInitial
-48276600924Feb 28, 2025May 14, 2025Jul 28, 2025Oct 11, 2025Dec 25, 2025Mar 10, 2026May 25, 2026

Evaluation results

MethodLinks
InitialModel Size=1.5B2026.010——————
SafePathModel Size=7B2026.010——————
UnalignedModel Family=LLaMA-3-8B2026.050——————
Standard DPOModel Family=LLaMA-3-8B2026.050——————
SequentialModel Family=LLaMA-3-8B2026.050——————
Sqrt-CompetenceModel Family=LLaMA-3-8B2026.050——————
Curri-DPOModel Family=LLaMA-3-8B2026.050——————
UnalignedModel Family=Qwen3-8B2026.050——————
Standard DPOModel Family=Qwen3-8B2026.050——————
SequentialModel Family=Qwen3-8B2026.050——————
Sqrt-CompetenceModel Family=Qwen3-8B2026.050——————
Curri-DPOModel Family=Qwen3-8B2026.050——————
Staged-CompetenceModel Family=Qwen3-8B2026.050——————
UnalignedModel Family=Yi-1.5-9B2026.050——————
GPT-4oNBF steering=false2025.020.4——————
Staged-CompetenceModel Family=LLaMA-3-8B2026.050.4——————
Standard DPOModel Family=Yi-1.5-9B2026.051.2——————
SequentialModel Family=Yi-1.5-9B2026.051.2——————
Curri-DPOModel Family=Yi-1.5-9B2026.051.2——————
Sqrt-CompetenceModel Family=Yi-1.5-9B2026.052.4——————
GPT-4oNBF steering=true2025.022.6——————
No DefenseModel Family=Zephyr-7B2025.103——————
Staged-CompetenceModel Family=Yi-1.5-9B2026.053.6——————
o1NBF steering=false2025.023.9——————
sonnet-4-thinking2025.104——————
VanillaBackbone=LLaVA-1.5-7B2025.104——————
ECSOBackbone=LLaVA-1.5-7B2025.104.8——————
WildJailbreakModel Family=Zephyr-7B2025.105——————
ASCoT (open)Model Family=Zephyr-7B2025.105——————
No DefenseModel Family=Mistral-7B2025.105——————
WildJailbreakModel Family=Mistral-7B2025.105——————
Claude 3.5 SonnetNBF steering=false2025.025.2——————
ASTRABackbone=LLaVA-1.5-7B2025.105.4——————
o1NBF steering=true2025.025.7——————
o4-mini2025.106——————
ASCoTBase Model=LLaMA-3.1-8B, Access=open2025.106——————
No DefenseModel Family=Llama3.1-8B2025.106——————
WildJailbreakModel Family=Llama3.1-8B2025.106——————
ASCoT (open)Model Family=Llama3.1-8B2025.106——————
PAP TrainedModel Family=Zephyr-7B2025.106——————
Improved CoTModel Series=7B Models, Safety Training Data=10002025.056.7——————
CATModel Family=Mistral-7B2025.107——————
ASCoT (closed)Model Family=Mistral-7B2025.107——————
ASCoT (open)Model Family=Mistral-7B2025.107——————
MoRASBackbone=LLaVA-1.5-7B2025.107.2——————
Improved CoTModel Series=32B Models, Safety Training Data=10002025.057.3——————
Claude 3.5 SonnetNBF steering=true2025.027.4——————
GPT-3.5-turboNBF steering=false2025.027.8——————
GPT-3.5-turboNBF steering=true2025.027.8——————
Refusal TrainingModel Family=Zephyr-7B2025.109——————
Refusal TrainingModel Family=Mistral-7B2025.109——————
Refusal TrainingModel Family=Llama3.1-8B2025.1010——————
ASCoT (closed)Model Family=Zephyr-7B2025.1010——————
PAP TrainedModel Family=Mistral-7B2025.1010——————
PAP TrainedModel Family=Llama3.1-8B2025.1012——————
ASCoT (closed)Model Family=Llama3.1-8B2025.1012——————
STAR1-R1-Distill-32BModel Series=32B Models, Safety Training Data=10002025.0512.4——————
ETABackbone=LLaVA-1.5-7B2025.1015.6——————
RealSafe-R1-32BModel Series=32B Models, Safety Training Data=150002025.0516——————
CoCABackbone=LLaVA-1.5-7B2025.1024.8——————
LATModel Family=Mistral-7B2025.1030——————
LAT*Model Family=Mistral-7B2025.1030——————
CATModel Family=Llama3.1-8B2025.1032——————
LAT*Model Family=Zephyr-7B2025.1034——————
STAR1-R1-Distill-7BModel Series=7B Models, Safety Training Data=10002025.0534——————
RealSafe-R1-7BModel Series=7B Models, Safety Training Data=150002025.0536——————
SafeChainModel Size=1.5B2026.0140——————
SafePathModel Size=1.5B2026.0140——————
SafeChainModel Size=7B2026.0140——————
THINKSAFEModel Size=7B2026.0140——————
InitialModel Size=8B2026.0140——————
SafePathModel Size=8B2026.0140——————
CAT*Model Family=Mistral-7B2025.1040——————
LATModel Family=Zephyr-7B2025.1044——————
FigStepBackbone=LLaVA-1.5-7B2025.1046——————
CAT*Model Family=Zephyr-7B2025.1047——————
CAT*Model Family=Llama3.1-8B2025.1049——————
AutoSteerBackbone=LLaVA-1.5-7B2025.1062.4——————
DPO_WhispererAlignment=DPO, Data Recipe=ear-whisperer, Base Model=SFT_DB2025.0580.67——————
LATModel Family=Llama3.1-8B2025.1086——————
Mixtral (SFT_OG)LLM=Mixtral, SFT Status=SFT on original response data without CoTs2025.0587.6——————
Mixtral (SFT_DB)LLM=Mixtral, SFT Status=SFT on AIDSAFE-generated CoTs and responses2025.0591.84——————
SFT_DBTraining=Supervised Fine-Tuning, Data Generation=AIDSAFE2025.0591.84——————
Qwen (SFT_DB)LLM=Qwen, SFT Status=SFT on AIDSAFE-generated CoTs and responses2025.0593.6——————
Qwen (SFT_OG)LLM=Qwen, SFT Status=SFT on original response data without CoTs2025.0598——————
LAT*Model Family=Llama3.1-8B2025.1098——————
CATModel Family=Zephyr-7B2025.1098——————
Mixtral (Base)LLM=Mixtral, SFT Status=None2025.0598.8——————
Qwen (Base)LLM=Qwen, SFT Status=None2025.0599.2——————
STAR-1Model Size=1.5B2026.01120——————
THINKSAFEModel Size=1.5B2026.01120——————
InitialModel Size=7B2026.01120——————
SafeChainModel Size=8B2026.01160——————
THINKSAFEModel Size=8B2026.01160——————
STAR-1Model Size=7B2026.01240——————
SafeKeyModel Size=1.5B2026.01360——————
SafeKeyModel Size=7B2026.01720——————
SafeKeyModel Size=8B2026.01800——————
DirectRefusalModel Size=1.5B2026.01840——————
STAR-1Model Size=8B2026.011,200——————