galwert/AnchorRep-Llama-3-8B-Instruct

AnchorRep — Llama-3-8B-Instruct

LoRA defense adapter for meta-llama/Meta-Llama-3-8B-Instruct, trained against frozen anchor microsoft/Phi-3-medium-4k-instruct to reduce cross-model jailbreak transfer. Companion artifact for the AnchorRep paper (NeurIPS 2026).

Intended use

Defensive research on cross-model jailbreak robustness for open-weight LLMs. Apply the adapter on top of the unmodified base model to obtain a hardened version. Not for production deployment without independent safety validation.

Quick start

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "meta-llama/Meta-Llama-3-8B-Instruct"
adapter = "galwert/AnchorRep-Llama-3-8B-Instruct"

tokenizer = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype="float32", device_map="auto")
model = PeftModel.from_pretrained(model, adapter)

Merge the adapter into the base weights for a single deployable model:

model = model.merge_and_unload()

Hyperparameters

Parameter	Value	Role
Refusal-direction weight (α)	0.15	refusal projection
Coherency weight (β)	1.0	benign output preservation
CKA repulsion weight (γ)	2.0	anchor repulsion
LM weight (δ)	0.0	next-token preservation
KL weight (ε)	0.4	benign KL preservation
CKA scope	harmful_only	prompts contributing to repulsion
Training steps	200
LoRA rank / alpha	32 / 64
Target modules	q,k,v,o,up,down,gate_proj
Layer	mid (50% depth)
Precision	fp32
Seed	42

Full training config in training_config.json. Per-loss ablations and hyperparameter ranges are in the paper appendix.

Reported metrics

Metric	Value
Cross-model GCG transfer ASR (self / anchor / other)	1% / 2% / 0%
Benign Garble Rate (OR-Bench)	0%
Δ XSTest refusal	-2.4
Δ OR-Bench refusal	+1.9
Δ MT-Bench	+0.18

ASR is reported after manual verification per the paper protocol.

Limitations

• Does not block same-model adaptive embedding-space attacks (e.g., Embedding PGD).
• TAP (semantic reframing) partially bypasses on this model.
• Single-anchor design; an adversary jointly optimizing across multiple surrogates could partially circumvent the defense.
• Refusal-direction stability (across seeds, template counts, and alternative phrasings) and multilingual robustness (English, Spanish, Chinese) were evaluated in the paper; languages beyond these three remain untested.

See the paper Limitations section for full discussion.

License

The LoRA delta is intended for use with Meta-Llama-3-8B-Instruct and is subject to the Meta Llama 3 Community License of the base model. Model card text and metadata are released under CC BY 4.0.

Citation

@inproceedings{anchorrep2026,
  title={AnchorRep: Defending LLMs Against Cross-Model Adversarial Transfer via Representation Repulsion},
  author={Wertheizer, Gal and Himelstein, Rom and Peretz, Tomer and Mendelson, Avi},
  booktitle={NeurIPS},
  year={2026}
}

Companion GitHub repository (training, evaluation, audit logs, GCG suffixes): https://github.com/galwert/AnchorRep