RepBelief

Model bahasa mewakili kepercayaan diri dan orang lain

Repositori ini menyediakan kode untuk makalah "Model bahasa mewakili kepercayaan diri dan orang lain". Ini menunjukkan bahwa LLMS secara internal mewakili keyakinan diri mereka sendiri dan agen lain, dan memanipulasi representasi ini dapat secara signifikan memengaruhi teori kemampuan penalaran pikiran mereka.

Instalasi

 conda create -n lm python=3.8 anaconda
conda activate lm
# Please install PyTorch (<2.4) according to your CUDA version.
conda install pytorch==2.3.1 torchvision==0.18.1 torchaudio==2.3.1 pytorch-cuda=12.1 -c pytorch -c nvidia
pip install -r requirements.txt

Kemudian unduh model bahasa (misalnya Mistral-7B-instruct-V0.2, Deepseek-LLM-7B-CHAT) ke models/ . Anda juga dapat menentukan jalur file di lm_paths.json .

Ekstrak representasi

sh scripts/save_reps.sh 0_forward belief
sh scripts/save_reps.sh 0_forward action
sh scripts/save_reps.sh 0_backward belief

Menyelidik

Biner:

python probe.py --belief=protagonist --dynamic=0_forward --variable belief 
python probe.py --belief=oracle --dynamic=0_forward --variable belief

python probe.py --belief=protagonist --dynamic=0_forward --variable action 
python probe.py --belief=oracle --dynamic=0_forward --variable action

python probe.py --belief=protagonist --dynamic=0_backward --variable belief 
python probe.py --belief=oracle --dynamic=0_backward --variable belief

Multinomial:

python probe_multinomial.py --dynamic=0_forward --variable belief
python probe_multinomial.py --dynamic=0_forward --variable action
python probe_multinomial.py --dynamic=0_backward --variable belief

Evaluasi BigTom

sh scripts/0_forward_belief.sh
sh scripts/0_forward_action.sh
sh scripts/0_backward_belief.sh

Intervensi

Intervensi untuk Tugas Keyakinan Maju :

sh scripts/0_forward_belief_interv_oracle.sh
sh scripts/0_forward_belief_interv_protagonist.sh
sh scripts/0_forward_belief_interv_o0p1.sh

Intervensi lintas tugas:

sh scripts/cross_0_forward_belief_to_forward_action_interv_o0p1.sh
sh scripts/cross_0_forward_belief_to_backward_belief_interv_o0p1.sh

Kutipan

 @inproceedings { zhu2024language ,
    title = { Language Models Represent Beliefs of Self and Others } ,
    author = { Zhu, Wentao and Zhang, Zhining and Wang, Yizhou } ,
    booktitle = { Forty-first International Conference on Machine Learning } ,
    year = { 2024 }
}