Auto-regressive dynamical system as opposed to vision classifiers
python3 scripts/sgcg.py
–dataset datasets/100_squad_train_v2.0.jsonl
–model meta-llama/Meta-Llama-3-8B-Instruct
–k 20
–max_parallel 30
–grad_batch_size 50
–num_iters 30
SUM:
Aman Bhargava from Caltech and Cameron Witkowski from the University of Toronto have mapped out the reachable space of a language model using control theory, revealing that prompt engineering can significantly influence model outputs. Their research suggests that a deeper exploration of control theory concepts could lead to more reliable and capable language models.