# Few Tokens, Big Leverage: Preserving Safety Alignment by Constraining Safety Tokens during Fine-tuning

arXiv: https://arxiv.org/abs/2603.07445 · HTML: https://mlsanity.com/paper/2603.07445/

## Key facts

- Cited by 4 papers in the map (data through 14 Oct 2026).
- Placed in Safety alignment and jailbreak attacks, one of 10 directions in LLM alignment / RLHF / preference optimisation.
- Builds on 24 of its 45 references that are in the map.
- First posted on arXiv 8 Mar 2026.

## Placed in

- LLM alignment / RLHF / preference optimisation / [Safety alignment and jailbreak attacks](https://mlsanity.com/alignment/safety-alignment-and-jailbreak-attacks/)
- LLM safety and security / [Safety alignment, refusal and guardrails](https://mlsanity.com/safety/safety-alignment-refusal-and-guardrails/)

## Builds on (most cited first, 24 of 24)

- [The Llama 3 Herd of Models](https://mlsanity.com/paper/2407.21783/) (2024)
- [Llama 2: Open Foundation and Fine-Tuned Chat Models](https://mlsanity.com/paper/2307.09288/) (2023)
- [Training Verifiers to Solve Math Word Problems](https://mlsanity.com/paper/2110.14168/) (2021)
- [Direct Preference Optimization: Your Language Model is Secretly a Reward Model](https://mlsanity.com/paper/2305.18290/) (2023)
- [Finetuned Language Models Are Zero-Shot Learners](https://mlsanity.com/paper/2109.01652/) (2021)
- [Universal and Transferable Adversarial Attacks on Aligned Language Models](https://mlsanity.com/paper/2307.15043/) (2023)
- [Character-level Convolutional Networks for Text Classification](https://mlsanity.com/paper/1509.01626/) (2015)
- [Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!](https://mlsanity.com/paper/2310.03693/) (2023)
- [LIMA: Less Is More for Alignment](https://mlsanity.com/paper/2305.11206/) (2023)
- [BeaverTails: Towards Improved Safety Alignment of LLM via a Human-Preference Dataset](https://mlsanity.com/paper/2307.04657/) (2023)
- [Safety Alignment Should Be Made More Than Just a Few Tokens Deep](https://mlsanity.com/paper/2406.05946/) (2024)
- [A StrongREJECT for Empty Jailbreaks](https://mlsanity.com/paper/2402.10260/) (2024)
- [Shadow Alignment: The Ease of Subverting Safely-Aligned Language Models](https://mlsanity.com/paper/2310.02949/) (2023)
- [PKU-SafeRLHF: Towards Multi-Level Safety Alignment for LLMs with Human Preference](https://mlsanity.com/paper/2406.15513/) (2024)
- [LoRA Fine-tuning Efficiently Undoes Safety Training in Llama 2-Chat 70B](https://mlsanity.com/paper/2310.20624/) (2023)
- [Latent Adversarial Training Improves Robustness to Persistent Harmful Behaviors in LLMs](https://mlsanity.com/paper/2407.15549/) (2024)
- [Removing RLHF Protections in GPT-4 via Fine-Tuning](https://mlsanity.com/paper/2311.05553/) (2023)
- [Lisa: Lazy Safety Alignment for Large Language Models against Harmful Fine-tuning Attack](https://mlsanity.com/paper/2405.18641/) (2024)
- [Antidote: Post-fine-tuning Safety Alignment for Large Language Models against Harmful Fine-tuning](https://mlsanity.com/paper/2408.09600/) (2024)
- [Bypassing the Safety Training of Open-Source LLMs with Priming Attacks](https://mlsanity.com/paper/2312.12321/) (2023)
- [NLSR: Neuron-Level Safety Realignment of Large Language Models Against Harmful Fine-Tuning](https://mlsanity.com/paper/2412.12497/) (2024)
- [Mitigating Fine-tuning based Jailbreak Attack with Backdoor Enhanced Safety Alignment](https://mlsanity.com/paper/2402.14968/) (2024)
- [Nudging: Inference-time Alignment of LLMs via Guided Decoding](https://mlsanity.com/paper/2410.09300/) (2024)
- [AsFT: Anchoring Safety During LLM Fine-Tuning Within Narrow Safety Basin](https://mlsanity.com/paper/2506.08473/) (2025)

## Built on by (most cited first, 4 of 4)

- [Multilingual Safety Alignment via Self-Distillation](https://mlsanity.com/paper/2605.02971/) (2026)
- [Ablating Safety: Mechanisms for Removing Alignment in Language Models for Security Applications](https://mlsanity.com/paper/2605.17413/) (2026)
- [Beyond Token Positions: Safety Alignment Across Denoising Steps in Diffusion Language Models](https://mlsanity.com/paper/2609.00495/) (2026)
- [SafeRI: Recognition and Intervention for Token-Level Safety Intervention in Large Vision Language Models](https://mlsanity.com/paper/2609.03544/) (2026)

## Abstract (arXiv)

Large language models (LLMs) often require fine-tuning (FT) to perform well on downstream tasks, but FT can induce safety-alignment drift even when the training dataset contains only benign data. Prior work shows that introducing a small fraction of harmful data can substantially compromise LLM refusal behavior, causing LLMs to comply with harmful requests. Existing defense methods often rely on model-wide interventions, such as restricting which parameters are updated or injecting additional safety data, which can limit generality and degrade downstream task performance. To address these limitations, we propose a fine-tuning framework called Preserving Safety Alignment via Constrained Tokens (PACT), which stabilizes the model's confidence on safety tokens. Our approach is motivated by the empirical observation that safety-aligned behavior is reflected in the model's token-level output confidence and is often concentrated on a small subset of safety-related tokens. During downstream fine-tuning, we regularize the fine-tuned model to match the aligned reference model's confidence on safety-related tokens at each response step, while leaving non-safety tokens largely unconstrained to allow effective task adaptation. This targeted constraint prevents alignment drift without imposing global restrictions that typically trade off with model utility. Our code is available at {https://github.com/Glresearch1/PACT}.