claudini/claudini/methods/original/reg_relax

name, full_name, reference, paper_url, code

name	full_name	reference	paper_url	code
RR	Regularized Relaxation	chacko2024adversarial	https://arxiv.org/abs/2410.19160	https://github.com/sj21j/Regularized_Relaxation

RR — Regularized Relaxation

Paper: Chacko et al., "Adversarial Attacks on Large Language Models Using Regularized Relaxation" (2024) Links: arXiv | Code \cite{chacko2024adversarial}

Algorithm

Continuous embedding-space optimization using AdamW with decoupled weight decay as the regularization mechanism. Unlike PGD (which operates on simplex-projected logits) or PEZ (which uses straight-through estimation), RR directly optimizes soft embeddings with weight decay pulling them toward zero to prevent drift into invalid regions of embedding space.

1. Initialize suffix embeddings from random tokens, add Gaussian noise (σ=0.1)
2. Each step:
   • Forward pass with soft embeddings concatenated into the input sequence
   • Compute CE loss on target tokens
   • Backprop through soft embeddings only
   • Clip gradients (max_norm=1.0)
   • AdamW step (weight_decay=0.05 provides L2 regularization)
   • Exponential LR decay: lr = lr_init × 0.99^step
3. Discretize via normalized L2 nearest-neighbour projection (both soft embeddings and token embedding matrix are L2-normalized before distance computation)

Hyperparameters

Parameter	Default	Paper
lr	0.1	0.1 (Llama-2/Vicuna); model-dependent
weight_decay	0.05	0.05
lr_decay	0.99	0.99
max_norm	1.0	1.0
init_noise_std	0.1	0.1
optim_length	20	20

Notes

• The paper reports model-dependent learning rates (Falcon: 0.7, MPT: 0.8, Mistral: 0.6); we use 0.1 (Llama-2/Vicuna default) as the general default.
• Weight decay in AdamW is decoupled: param -= lr * weight_decay * param, equivalent to L2 regularization toward zero.
• The normalized L2 projection is equivalent to cosine similarity on the unit sphere (minimizing ||x̂ - ŵ||₂ = minimizing √(2 - 2·cos(x,w))).