# Inverse Reinforcement Learning with Dynamic Reward Scaling for LLM Alignment

arXiv: https://arxiv.org/abs/2503.18991 · HTML: https://mlsanity.com/paper/2503.18991/

## Key facts

- Cited by 20 papers in the map (data through 14 Oct 2026).
- Placed in Reward modeling and RLHF, one of 10 directions in LLM alignment / RLHF / preference optimisation.
- Builds on 65 of its 82 references that are in the map.
- First posted on arXiv 23 Mar 2025.

## Placed in

- LLM alignment / RLHF / preference optimisation / [Reward modeling and RLHF](https://mlsanity.com/alignment/reward-modeling-and-rlhf/)
- Offline RL and imitation learning from data / [Language-conditioned policies and VLAs](https://mlsanity.com/offlinerl/language-conditioned-policies-and-vlas/)

## Builds on (most cited first, 30 of 65)

- [Learning Transferable Visual Models From Natural Language Supervision](https://mlsanity.com/paper/2103.00020/) (2021)
- [Training language models to follow instructions with human feedback](https://mlsanity.com/paper/2203.02155/) (2022)
- [Proximal Policy Optimization Algorithms](https://mlsanity.com/paper/1707.06347/) (2017)
- [The Llama 3 Herd of Models](https://mlsanity.com/paper/2407.21783/) (2024)
- [Chain-of-Thought Prompting Elicits Reasoning in Large Language Models](https://mlsanity.com/paper/2201.11903/) (2022)
- [DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models](https://mlsanity.com/paper/2402.03300/) (2024)
- [Direct Preference Optimization: Your Language Model is Secretly a Reward Model](https://mlsanity.com/paper/2305.18290/) (2023)
- [Measuring Mathematical Problem Solving With the MATH Dataset](https://mlsanity.com/paper/2103.03874/) (2021)
- [Qwen2.5 Technical Report](https://mlsanity.com/paper/2412.15115/) (2024)
- [Universal and Transferable Adversarial Attacks on Aligned Language Models](https://mlsanity.com/paper/2307.15043/) (2023)
- [Membership Inference Attacks against Machine Learning Models](https://mlsanity.com/paper/1610.05820/) (2016)
- [Jailbreaking Black Box Large Language Models in Twenty Queries](https://mlsanity.com/paper/2310.08419/) (2023)
- [RealToxicityPrompts: Evaluating Neural Toxic Degeneration in Language Models](https://mlsanity.com/paper/2009.11462/) (2020)
- [Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!](https://mlsanity.com/paper/2310.03693/) (2023)
- [BeaverTails: Towards Improved Safety Alignment of LLM via a Human-Preference Dataset](https://mlsanity.com/paper/2307.04657/) (2023)
- [AutoDAN: Generating Stealthy Jailbreak Prompts on Aligned Large Language Models](https://mlsanity.com/paper/2310.04451/) (2023)
- [Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback](https://mlsanity.com/paper/2307.15217/) (2023)
- [Aligning Large Multimodal Models with Factually Augmented RLHF](https://mlsanity.com/paper/2309.14525/) (2023)
- [BBQ: A Hand-Built Bias Benchmark for Question Answering](https://mlsanity.com/paper/2110.08193/) (2021)
- [Automatic Chain of Thought Prompting in Large Language Models](https://mlsanity.com/paper/2210.03493/) (2022)
- [WildChat: 1M ChatGPT Interaction Logs in the Wild](https://mlsanity.com/paper/2405.01470/) (2024)
- [How Johnny Can Persuade LLMs to Jailbreak Them: Rethinking Persuasion to Challenge AI Safety by Humanizing LLMs](https://mlsanity.com/paper/2401.06373/) (2024)
- [ORPO: Monolithic Preference Optimization without Reference Model](https://mlsanity.com/paper/2403.07691/) (2024)
- [XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Models](https://mlsanity.com/paper/2308.01263/) (2023)
- [RRHF: Rank Responses to Align Language Models with Human Feedback without tears](https://mlsanity.com/paper/2304.05302/) (2023)
- [Catastrophic Jailbreak of Open-source LLMs via Exploiting Generation](https://mlsanity.com/paper/2310.06987/) (2023)
- [Trustworthy LLMs: a Survey and Guideline for Evaluating Large Language Models' Alignment](https://mlsanity.com/paper/2308.05374/) (2023)
- [A StrongREJECT for Empty Jailbreaks](https://mlsanity.com/paper/2402.10260/) (2024)
- [Safety-Tuned LLaMAs: Lessons From Improving the Safety of Large Language Models that Follow Instructions](https://mlsanity.com/paper/2309.07875/) (2023)
- [Measuring short-form factuality in large language models](https://mlsanity.com/paper/2411.04368/) (2024)

## Built on by (most cited first, 20 of 20)

- [A Comprehensive Survey on Trustworthiness in Reasoning with Large Language Models](https://mlsanity.com/paper/2509.03871/) (2025)
- [KnowRL: Exploring Knowledgeable Reinforcement Learning for Factuality](https://mlsanity.com/paper/2506.19807/) (2025)
- [Oyster-I: Beyond Refusal -- Constructive Safety Alignment for Responsible Language Models](https://mlsanity.com/paper/2509.01909/) (2025)
- [A Survey on Progress in LLM Alignment from the Perspective of Reward Design](https://mlsanity.com/paper/2505.02666/) (2025)
- [Obscure but Effective: Classical Chinese Jailbreak Prompt Optimization via Bio-Inspired Search](https://mlsanity.com/paper/2602.22983/) (2026)
- [VRPO: Rethinking Value Modeling for Robust RL under Noisy Supervision in LLM Post-Training](https://mlsanity.com/paper/2508.03058/) (2025)
- [Quantized but Deceptive? A Multi-Dimensional Truthfulness Evaluation of Quantized LLMs](https://mlsanity.com/paper/2508.19432/) (2025)
- [EcoAlign: An Economically Rational Framework for Efficient LVLM Alignment](https://mlsanity.com/paper/2511.11301/) (2025)
- [Pruning Weights but Not Truth: Safeguarding Truthfulness While Pruning LLMs](https://mlsanity.com/paper/2509.00096/) (2025)
- [Quark Medical Alignment: A Holistic Multi-Dimensional Alignment and Collaborative Optimization Paradigm](https://mlsanity.com/paper/2602.11661/) (2026)
- [Adversarial Orthogonal Disentanglement for LVLM Hallucination Mitigation](https://mlsanity.com/paper/2605.25377/) (2026)
- [DMRL: Data- and Model-aware Reward Learning for Data Extraction](https://mlsanity.com/paper/2505.06284/) (2025)
- [VisuoAlign: Safety Alignment of LVLMs with Multimodal Tree Search](https://mlsanity.com/paper/2510.15948/) (2025)
- [Dynamic Reward Scaling for Multivariate Time Series Anomaly Detection: A VAE-Enhanced Reinforcement Learning Approach](https://mlsanity.com/paper/2511.12351/) (2025)
- [Information-Dense Reasoning for Efficient and Auditable Security Alert Triage](https://mlsanity.com/paper/2512.08169/) (2025)
- [ShellForge: Adversarial Co-Evolution of Webshell Generation and Multi-View Detection for Robust Webshell Defense](https://mlsanity.com/paper/2601.22182/) (2026)
- [Membership Inference for Contrastive Pre-training Models with Text-only PII Queries](https://mlsanity.com/paper/2603.14222/) (2026)
- [MESA: Improving MoE Safety Alignment via Decentralized Expertise](https://mlsanity.com/paper/2606.00651/) (2026)
- [Coherence Maximization Improves Pluralistic Alignment](https://mlsanity.com/paper/2606.03110/) (2026)
- [Search-G1: Grounded Search Agents via Representation-Based Intrinsic Rewards](https://mlsanity.com/paper/2608.07531/) (2026)

## Abstract (arXiv)

Alignment is vital for safely deploying large language models (LLMs). Existing techniques are either reward-based (training a reward model on preference pairs and optimizing with reinforcement learning) or reward-free (directly fine-tuning on ranked outputs). Recent research shows that well-tuned reward-based pipelines remain the most robust, and single-response demonstrations can outperform pairwise preference data. However, there still exist two key challenges: (1) imbalanced safety datasets that overrepresent common hazards while neglecting long-tail threats; and (2) static reward models that ignore task difficulty, limiting optimization efficiency and attainable gains. To address these limitations, we propose DR-IRL, which Dynamically adjusts Rewards through Inverse Reinforcement Learning. We first train category-specific reward models using a balanced safety dataset of seven harmful categories as demonstration via IRL. Then we enhance Group Relative Policy Optimization (GRPO) by introducing dynamic reward scaling: adjusting rewards by task difficulty, data-level hardness by text encoder cosine similarity, and model-level responsiveness by reward gaps. Extensive experiments across various benchmarks and LLMs demonstrate that DR-IRL outperforms all baseline methods in safety alignment while maintaining usefulness.