Adaptive Helpfulness–Harmlessness Alignment with Preference Vectors

ACL ARR 2025 July Submission1357 Authors

Abstract: Ensuring that large language models (LLMs) are both helpful and harmless is a critical challenge, as overly strict constraints can lead to excessive refusals, while permissive models risk generating harmful content. Existing approaches, such as reinforcement learning from human feedback (RLHF) and direct preference optimization (DPO), attempt to balance these trade-offs but suffer from performance conflicts, limited controllability, and poor extendability. To address these issues, we propose Preference Vector, a novel framework inspired by task arithmetic. Instead of optimizing multiple preferences within a single objective, we train separate models on individual preferences, extract behavior shifts as preference vectors, and dynamically merge them at test time. This modular approach enables fine-grained, user-controllable preference adjustments and facilitates seamless integration of new preferences without retraining. Experiments show that our proposed Preference Vector framework improves helpfulness without excessive conservatism, allows smooth control over preference trade-offs, and supports scalable multi-preference alignment.

Paper Type: Long

Research Area: Ethics, Bias, and Fairness

Research Area Keywords: Large Language Models, RLHF, Safety, Alignment, Task arithmetic

Contribution Types: Model analysis & interpretability

Languages Studied: English

Reassignment Request Area Chair: This is not a resubmission

Reassignment Request Reviewers: This is not a resubmission

A1 Limitations Section: This paper has a limitations section.

A2 Potential Risks: Yes

A2 Elaboration: 8

B Use Or Create Scientific Artifacts: Yes

B1 Cite Creators Of Artifacts: Yes

B1 Elaboration: 5.1, B.1

B2 Discuss The License For Artifacts: Yes

B2 Elaboration: B.1

B3 Artifact Use Consistent With Intended Use: Yes

B3 Elaboration: B.1

B4 Data Contains Personally Identifying Info Or Offensive Content: Yes

B4 Elaboration: B.1

B5 Documentation Of Artifacts: Yes

B5 Elaboration: B.1

B6 Statistics For Data: Yes

B6 Elaboration: B.1

C Computational Experiments: Yes

C1 Model Size And Budget: Yes

C1 Elaboration: 5.1, B.1

C2 Experimental Setup And Hyperparameters: Yes

C2 Elaboration: 5.1, B.2, B.6, B.7

C3 Descriptive Statistics: Yes

C3 Elaboration: 5, B.1

C4 Parameters For Packages: Yes

C4 Elaboration: B.1

D Human Subjects Including Annotators: Yes

D1 Instructions Given To Participants: Yes

D1 Elaboration: B.5

D2 Recruitment And Payment: Yes

D2 Elaboration: B.5

D3 Data Consent: Yes

D3 Elaboration: B.5

D4 Ethics Review Board Approval: No

D4 Elaboration: No, we asked our departments ethics board, and they informed us that no application was required for our work.

D5 Characteristics Of Annotators: Yes

D5 Elaboration: B.5

E Ai Assistants In Research Or Writing: Yes

E1 Information About Use Of Ai Assistants: Yes

E1 Elaboration: Yes. We used ChatGPT for proofreading during the writing process.

Author Submission Checklist: yes

Submission Number: 1357