Linear probes trained on diverse deception data to detect dishonest completions across model families (OLMo, Qwen, Gemma).
AI & ML interests
Frontier alignment research to ensure the safe development and deployment of advanced AI systems.
Recent Activity
View all activity
Papers
View all PapersObfuscated Policy, Obfuscated Activations, Blatant Deception, and Honest models trained in the Obfuscation Atlas paper
-
taufeeque/mbpp-hardcode
Viewer • Updated • 974 • 1.18k -
AlignmentResearch/obfuscation-atlas-Meta-Llama-3-8B-Instruct-kl0.001-det10-seed1-mbpp_probe
Updated • 2 -
AlignmentResearch/obfuscation-atlas-Meta-Llama-3-8B-Instruct-kl0.0001-det10-seed1-mbpp_probe
Updated • 2 -
AlignmentResearch/obfuscation-atlas-Meta-Llama-3-8B-Instruct-kl0.01-det10-seed1-mbpp_probe
Updated
Obfuscated Policy, Obfuscated Activations, Blatant Deception, and Honest models trained in the Obfuscation Atlas paper.
-
The Obfuscation Atlas: Mapping Where Honesty Emerges in RLVR with Deception Probes
Paper • 2602.15515 • Published -
taufeeque/mbpp-hardcode
Viewer • Updated • 974 • 1.18k -
AlignmentResearch/obfuscation-atlas-Meta-Llama-3-8B-Instruct-kl0.001-det10-seed1-mbpp_probe
Updated • 2 -
AlignmentResearch/obfuscation-atlas-Meta-Llama-3-8B-Instruct-kl0.0001-det10-seed1-mbpp_probe
Updated • 2
Holding model organisms that demonstrate shortcomings of black-box supervision of AI models
Linear probes trained on diverse deception data to detect dishonest completions across model families (OLMo, Qwen, Gemma).
Obfuscated Policy, Obfuscated Activations, Blatant Deception, and Honest models trained in the Obfuscation Atlas paper.
-
The Obfuscation Atlas: Mapping Where Honesty Emerges in RLVR with Deception Probes
Paper • 2602.15515 • Published -
taufeeque/mbpp-hardcode
Viewer • Updated • 974 • 1.18k -
AlignmentResearch/obfuscation-atlas-Meta-Llama-3-8B-Instruct-kl0.001-det10-seed1-mbpp_probe
Updated • 2 -
AlignmentResearch/obfuscation-atlas-Meta-Llama-3-8B-Instruct-kl0.0001-det10-seed1-mbpp_probe
Updated • 2
Obfuscated Policy, Obfuscated Activations, Blatant Deception, and Honest models trained in the Obfuscation Atlas paper
-
taufeeque/mbpp-hardcode
Viewer • Updated • 974 • 1.18k -
AlignmentResearch/obfuscation-atlas-Meta-Llama-3-8B-Instruct-kl0.001-det10-seed1-mbpp_probe
Updated • 2 -
AlignmentResearch/obfuscation-atlas-Meta-Llama-3-8B-Instruct-kl0.0001-det10-seed1-mbpp_probe
Updated • 2 -
AlignmentResearch/obfuscation-atlas-Meta-Llama-3-8B-Instruct-kl0.01-det10-seed1-mbpp_probe
Updated
Holding model organisms that demonstrate shortcomings of black-box supervision of AI models