senku14x/Validating-NLAs

1

stars

123

commits

Python

primary language

Aug 22, 2026

updated

README

Validating-NLAs

The NLA paper claims NLAs surface safety-relevant cognition that models represent internally but do not verbalize, with evaluation awareness as the headline case. They could not validate this against ground truth because a real model's internal beliefs are unobservable. This project replaces unobservable ground truth with constructed ground truth, using model organisms where we control exactly what cognition is present and whether it is verbalized, and then measures whether the NLA actually detects it.

Work In Progress :D

Contributors

70 commits

51 commits

2 commits

Similar repos

AnMaJ/LASR-Analysing-Faithfulness-of-Reasoning-to-Internal-Concepts

Repo for LASR project with Noah on Analysing Faithfulness of Reasoning to Internal Concepts

0

kmorhun/model-faithfulness

0

SolshineCode/NLAttack

Capability-floor evaluation harness for Natural Language Autoencoders (NLAs): concept-survival…

1

invi-bhagyesh/eval-evasion

training eval-aware model organisms

0

briteroses/uncertainty

experiments for measuring uncertainty in LLM latent beliefs -- @ NeurIPS 2023 ATTRIB

0

xu1998hz/llm_self_bias

This is the project to quantify the issues with LLM's self evaluation

9

PotatoChoudhary/crosswise

Reliability check for Natural Language Autoencoder explanations - catches fabricated claims against…

0

Lee-KocaTKa/1_ambig_evaluation

Codes for the step one of our research; vlm evaluation of ambiguity processing

0

Languages

Python

97.8%

Shell

2.2%

senku14x/Validating-NLAs

1

stars

123

commits

Python

primary language

Aug 22, 2026

updated

README

Validating-NLAs

The NLA paper claims NLAs surface safety-relevant cognition that models represent internally but do not verbalize, with evaluation awareness as the headline case. They could not validate this against ground truth because a real model's internal beliefs are unobservable. This project replaces unobservable ground truth with constructed ground truth, using model organisms where we control exactly what cognition is present and whether it is verbalized, and then measures whether the NLA actually detects it.

Work In Progress :D

Similar repos

AnMaJ/LASR-Analysing-Faithfulness-of-Reasoning-to-Internal-Concepts

Repo for LASR project with Noah on Analysing Faithfulness of Reasoning to Internal Concepts

0

kmorhun/model-faithfulness

0

SolshineCode/NLAttack

Capability-floor evaluation harness for Natural Language Autoencoders (NLAs): concept-survival…

1

invi-bhagyesh/eval-evasion

training eval-aware model organisms

0

briteroses/uncertainty

experiments for measuring uncertainty in LLM latent beliefs -- @ NeurIPS 2023 ATTRIB

0

xu1998hz/llm_self_bias

This is the project to quantify the issues with LLM's self evaluation

9

PotatoChoudhary/crosswise

Reliability check for Natural Language Autoencoder explanations - catches fabricated claims against…

0

Lee-KocaTKa/1_ambig_evaluation

Codes for the step one of our research; vlm evaluation of ambiguity processing

0

Contributors

70 commits

51 commits

2 commits

Languages

Python

97.8%

Shell

2.2%