# Equilibrate RLHF: Towards Balancing Helpfulness-Safety Trade-off in Large Language Models

arXiv: https://arxiv.org/abs/2502.11555 · HTML: https://mlsanity.com/paper/2502.11555/

## Key facts

- Cited by 22 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 40 of its 58 references that are in the map.
- First posted on arXiv 17 Feb 2025.

## Placed in

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

## Builds on (most cited first, 30 of 40)

- [GPT-4 Technical Report](https://mlsanity.com/paper/2303.08774/) (2023)
- [Proximal Policy Optimization Algorithms](https://mlsanity.com/paper/1707.06347/) (2017)
- [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)
- [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://mlsanity.com/paper/2005.11401/) (2020)
- [Evaluating Large Language Models Trained on Code](https://mlsanity.com/paper/2107.03374/) (2021)
- [Direct Preference Optimization: Your Language Model is Secretly a Reward Model](https://mlsanity.com/paper/2305.18290/) (2023)
- [Measuring Massive Multitask Language Understanding](https://mlsanity.com/paper/2009.03300/) (2020)
- [Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge](https://mlsanity.com/paper/1803.05457/) (2018)
- [HellaSwag: Can a Machine Really Finish Your Sentence?](https://mlsanity.com/paper/1905.07830/) (2019)
- [Let's Verify Step by Step](https://mlsanity.com/paper/2305.20050/) (2023)
- [Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback](https://mlsanity.com/paper/2204.05862/) (2022)
- [Universal and Transferable Adversarial Attacks on Aligned Language Models](https://mlsanity.com/paper/2307.15043/) (2023)
- [Learning to summarize from human feedback](https://mlsanity.com/paper/2009.01325/) (2020)
- [Making the V in VQA Matter: Elevating the Role of Image Understanding in Visual Question Answering](https://mlsanity.com/paper/1612.00837/) (2016)
- [VQA: Visual Question Answering](https://mlsanity.com/paper/1505.00468/) (2015)
- [Large Language Models Encode Clinical Knowledge](https://mlsanity.com/paper/2212.13138/) (2022)
- [CommonsenseQA: A Question Answering Challenge Targeting Commonsense Knowledge](https://mlsanity.com/paper/1811.00937/) (2018)
- [Challenging BIG-Bench Tasks and Whether Chain-of-Thought Can Solve Them](https://mlsanity.com/paper/2210.09261/) (2022)
- [DeepSeek-Coder: When the Large Language Model Meets Programming -- The Rise of Code Intelligence](https://mlsanity.com/paper/2401.14196/) (2024)
- [Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!](https://mlsanity.com/paper/2310.03693/) (2023)
- [A General Language Assistant as a Laboratory for Alignment](https://mlsanity.com/paper/2112.00861/) (2021)
- [Ethical and social risks of harm from Language Models](https://mlsanity.com/paper/2112.04359/) (2021)
- [A General Theoretical Paradigm to Understand Learning from Human Preferences](https://mlsanity.com/paper/2310.12036/) (2023)
- [RACE: Large-scale ReAding Comprehension Dataset From Examinations](https://mlsanity.com/paper/1704.04683/) (2017)
- [BeaverTails: Towards Improved Safety Alignment of LLM via a Human-Preference Dataset](https://mlsanity.com/paper/2307.04657/) (2023)
- [C-Eval: A Multi-Level Multi-Discipline Chinese Evaluation Suite for Foundation Models](https://mlsanity.com/paper/2305.08322/) (2023)
- [Safe RLHF: Safe Reinforcement Learning from Human Feedback](https://mlsanity.com/paper/2310.12773/) (2023)
- [WildChat: 1M ChatGPT Interaction Logs in the Wild](https://mlsanity.com/paper/2405.01470/) (2024)
- [Catastrophic Jailbreak of Open-source LLMs via Exploiting Generation](https://mlsanity.com/paper/2310.06987/) (2023)

## Built on by (most cited first, 22 of 22)

- [Get Experience from Practice: LLM Agents with Record & Replay](https://mlsanity.com/paper/2505.17716/) (2025)
- [Pushing the Limits of Safety: A Technical Report on the ATLAS Challenge 2025](https://mlsanity.com/paper/2506.12430/) (2025)
- [USB: A Comprehensive and Unified Safety Evaluation Benchmark for Multimodal Large Language Models](https://mlsanity.com/paper/2505.23793/) (2025)
- [LLLMs: A Data-Driven Survey of Evolving Research on Limitations of Large Language Models](https://mlsanity.com/paper/2505.19240/) (2025)
- [Uni-DPO: A Unified Paradigm for Dynamic Preference Optimization of LLMs](https://mlsanity.com/paper/2506.10054/) (2025)
- [SafeSearch: Do Not Trade Safety for Utility in LLM Search Agents](https://mlsanity.com/paper/2510.17017/) (2025)
- [Beyond Safe Answers: A Benchmark for Evaluating True Risk Awareness in Large Reasoning Models](https://mlsanity.com/paper/2505.19690/) (2025)
- [MSR-Align: Policy-Grounded Multimodal Alignment for Safety-Aware Reasoning in Vision-Language Models](https://mlsanity.com/paper/2506.19257/) (2025)
- [UnsafeChain: Enhancing Reasoning Model Safety via Hard Cases](https://mlsanity.com/paper/2507.21652/) (2025)
- [Multiplayer Nash Preference Optimization](https://mlsanity.com/paper/2509.23102/) (2025)
- [Reinforcement Learning from Human Feedback with High-Confidence Safety Constraints](https://mlsanity.com/paper/2506.08266/) (2025)
- [QuadSentinel: Sequent Safety for Machine-Checkable Control in Multi-agent Systems](https://mlsanity.com/paper/2512.16279/) (2025)
- [Beyond Prediction: Reinforcement Learning as the Defining Leap in Healthcare AI](https://mlsanity.com/paper/2508.21101/) (2025)
- [Robust Policy Optimization to Prevent Catastrophic Forgetting](https://mlsanity.com/paper/2602.08813/) (2026)
- [Benchmarking Knowledge-Extraction Attack and Defense on Retrieval-Augmented Generation](https://mlsanity.com/paper/2602.09319/) (2026)
- [Don't Make Models Guess Security and Safety: Symbolic Guardrails for Domain-Specific AI Agents](https://mlsanity.com/paper/2604.15579/) (2026)
- [Ethics and Persuasion in Reinforcement Learning from Human Feedback: A Procedural Rhetorical Approach](https://mlsanity.com/paper/2505.09576/) (2025)
- [Alignment Tuning for Large Language Models: A Data-Centric Lens on Alignment Data Pipelines](https://mlsanity.com/paper/2605.26442/) (2026)
- [Step-Level Preference Learning for Generative Agents in Social Simulations](https://mlsanity.com/paper/2607.14485/) (2026)
- [SaFeR-ToolKit: Structured Reasoning via Virtual Tool Calling for Multimodal Safety](https://mlsanity.com/paper/2603.02635/) (2026)
- [COMPASS: Cognitive MCTS-Guided Process Alignment for Safe Search Agents](https://mlsanity.com/paper/2605.30838/) (2026)
- [Learning to Ideate for Scientific Impact](https://mlsanity.com/paper/2609.29802/) (2026)

## Abstract (arXiv)

Fine-tuning large language models (LLMs) based on human preferences, commonly achieved through reinforcement learning from human feedback (RLHF), has been effective in improving their performance. However, maintaining LLM safety throughout the fine-tuning process remains a significant challenge, as resolving conflicts between safety and helpfulness can be non-trivial. Typically, the safety alignment of LLM is trained on data with safety-related categories. However, our experiments find that naively increasing the scale of safety training data usually leads the LLMs to an ``overly safe'' state rather than a ``truly safe'' state, boosting the refusal rate through extensive safety-aligned data without genuinely understanding the requirements for safe responses. Such an approach can inadvertently diminish the models' helpfulness. To understand the phenomenon, we first investigate the role of safety data by categorizing them into three different groups, and observe that each group behaves differently as training data scales up. To boost the balance between safety and helpfulness, we propose an Equilibrate RLHF framework including a Fine-grained Data-centric (FDC) approach that achieves better safety alignment even with fewer training data, and an Adaptive Message-wise Alignment (AMA) approach, which selectively highlight the key segments through a gradient masking strategy. Extensive experimental results demonstrate that our approach significantly enhances the safety alignment of LLMs while balancing safety and helpfulness.