HAIR: Hardness-Aware Inverse Reinforcement Learning with Introspective Reasoning for LLM Alignment

• URL: http://arxiv.org/abs/2503.18991v2
• Date: Tue, 06 May 2025 13:47:34 GMT
• Title: HAIR: Hardness-Aware Inverse Reinforcement Learning with Introspective Reasoning for LLM Alignment
• Authors: Ruoxi Cheng, Haoxuan Ma, Weixin Wang,
• Abstract summary: HAIR (Hardness-Aware Inverse Reinforcement Learning with Introspective Reasoning) is a novel alignment approach inspired by shadow models in membership inference attacks.<n>Our approach consists of two main components: (1) construction of a balanced safety Chain-of-Draft dataset for seven harmful categories using structured prompts that leverage introspective reasoning capabilities.<n> Comprehensive experiments across four harmlessness and four usefulness benchmarks demonstrate that HAIR achieves state-of-the-art performance.
• Score: 0.0351124620232225
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: The alignment of large language models (LLMs) with human values remains critical yet hindered by four key challenges: (1) scarcity of balanced safety datasets, (2) alignment tax, (3) vulnerability to jailbreak attacks due to shallow alignment, and (4) inability to dynamically adapt rewards according to task difficulty. To address these limitations, we introduce HAIR (Hardness-Aware Inverse Reinforcement Learning with Introspective Reasoning), a novel alignment approach inspired by shadow models in membership inference attacks. Our approach consists of two main components: (1) construction of a balanced safety Chain-of-Draft (CoD) dataset for seven harmful categories using structured prompts that leverage the introspective reasoning capabilities of LLMs; and (2) training of category-specific reward models with Group Relative Policy Optimization (GRPO), dynamically tuning optimization to task difficulty at both the data and model levels. Comprehensive experiments across four harmlessness and four usefulness benchmarks demonstrate that HAIR achieves state-of-the-art performance, outperforming all baseline methods in safety while maintaining high levels of usefulness.

Related papers

• DARLR: Dual-Agent Offline Reinforcement Learning for Recommender Systems with Dynamic Reward [14.323631574821123]
  Model-based offline reinforcement learning has emerged as a promising approach for recommender systems.<n>DarLR is proposed to dynamically update world models to enhance recommendation policies.<n>Experiments on four benchmark datasets demonstrate the superior performance of DARLR.
  arXiv  Detail & Related papers  (2025-05-12T06:18:31Z)
• More is Less: The Pitfalls of Multi-Model Synthetic Preference Data in DPO Safety Alignment [80.04449725137177]
  Direct Preference Optimization (DPO) has emerged as a simple, yet effective alternative to reinforcement learning from human feedback.<n>Our study reveals a striking, safety-specific phenomenon associated with DPO alignment.<n>Using solely self-generated responses for both chosen and rejected pairs significantly outperforms configurations that incorporate responses from stronger models.
  arXiv  Detail & Related papers  (2025-04-03T00:36:40Z)
• Improving LLM General Preference Alignment via Optimistic Online Mirror Descent [57.622821649679786]
  Reinforcement learning from human feedback (RLHF) has demonstrated remarkable effectiveness in aligning large language models (LLMs) with human preferences.<n>In this paper, we drop the Bradley-Terry (BT) model assumption and study LLM alignment under general preferences, formulated as a two-player game.<n>We show that our approach achieves an $O(T-1)$ bound on the duality gap, improving upon the previous $O(T-1/2)$ result.
  arXiv  Detail & Related papers  (2025-02-24T05:24:52Z)
• MM-RLHF: The Next Step Forward in Multimodal LLM Alignment [59.536850459059856]
  We introduce MM-RLHF, a dataset containing $mathbf120k$ fine-grained, human-annotated preference comparison pairs. We propose several key innovations to improve the quality of reward models and the efficiency of alignment algorithms. Our approach is rigorously evaluated across $mathbf10$ distinct dimensions and $mathbf27$ benchmarks.
  arXiv  Detail & Related papers  (2025-02-14T18:59:51Z)
• Few-shot Steerable Alignment: Adapting Rewards and LLM Policies with Neural Processes [50.544186914115045]
  Large language models (LLMs) are increasingly embedded in everyday applications.<n> Ensuring their alignment with the diverse preferences of individual users has become a critical challenge.<n>We present a novel framework for few-shot steerable alignment.
  arXiv  Detail & Related papers  (2024-12-18T16:14:59Z)
• A Systematic Examination of Preference Learning through the Lens of Instruction-Following [83.71180850955679]
  We use a novel synthetic data generation pipeline to generate 48,000 instruction unique-following prompts.<n>With our synthetic prompts, we use two preference dataset curation methods - rejection sampling (RS) and Monte Carlo Tree Search (MCTS)<n>Experiments reveal that shared prefixes in preference pairs, as generated by MCTS, provide marginal but consistent improvements.<n>High-contrast preference pairs generally outperform low-contrast pairs; however, combining both often yields the best performance.
  arXiv  Detail & Related papers  (2024-12-18T15:38:39Z)
• Solving the Inverse Alignment Problem for Efficient RLHF [0.0]
  We define the 'inverse alignment problem' in language model training.<n>We investigate whether repeatedly fine-tuning a reward model on subsets of the offline preference dataset aligned with a periodically frozen policy improves upon vanilla RLHF.
  arXiv  Detail & Related papers  (2024-12-13T19:47:38Z)
• In-Dataset Trajectory Return Regularization for Offline Preference-based Reinforcement Learning [15.369324784520538]
  We propose In-Dataset Trajectory Return Regularization (DTR) for offline preference-based reinforcement learning.<n>DTR mitigates the risk of learning inaccurate trajectory stitching under reward bias.<n>We also introduce an ensemble normalization technique that effectively integrates multiple reward models.
  arXiv  Detail & Related papers  (2024-12-12T09:35:47Z)
• SeRA: Self-Reviewing and Alignment of Large Language Models using Implicit Reward Margins [30.767203592231496]
  Self-Reviewing and Alignment (SeRA) is a cost-efficient and effective method that can be readily combined with existing DAAs. SeRA comprises of two components: (1) sample selection using implicit reward margins, which helps alleviate over-fitting to some undesired features, and (2) preference bootstrapping using implicit rewards to augment preference data with updated policy models.
  arXiv  Detail & Related papers  (2024-10-12T04:17:28Z)
• Self-supervised Preference Optimization: Enhance Your Language Model with Preference Degree Awareness [27.43137305486112]
  We propose a novel Self-supervised Preference Optimization (SPO) framework, which constructs a self-supervised preference degree loss combined with the alignment loss. The results demonstrate that SPO can be seamlessly integrated with existing preference optimization methods to achieve state-of-the-art performance.
  arXiv  Detail & Related papers  (2024-09-26T12:37:26Z)
• Contrastive Policy Gradient: Aligning LLMs on sequence-level scores in a supervised-friendly fashion [43.77763433288893]
  We introduce Contrastive Policy Gradient, or CoPG, a simple and mathematically principled new RL algorithm that can estimate the optimal policy even from off-policy data.<n>We show this approach to generalize the direct alignment method IPO (identity preference optimization) and classic policy gradient.<n>We experiment with the proposed CoPG on a toy bandit problem to illustrate its properties, as well as for finetuning LLMs on a summarization task.
  arXiv  Detail & Related papers  (2024-06-27T14:03:49Z)
• Provably Mitigating Overoptimization in RLHF: Your SFT Loss is Implicitly an Adversarial Regularizer [52.09480867526656]
  We identify the source of misalignment as a form of distributional shift and uncertainty in learning human preferences.<n>To mitigate overoptimization, we first propose a theoretical algorithm that chooses the best policy for an adversarially chosen reward model.<n>Using the equivalence between reward models and the corresponding optimal policy, the algorithm features a simple objective that combines a preference optimization loss and a supervised learning loss.
  arXiv  Detail & Related papers  (2024-05-26T05:38:50Z)
• Multi-Reference Preference Optimization for Large Language Models [56.84730239046117]
  We introduce a novel closed-form formulation for direct preference optimization using multiple reference models. The resulting algorithm, Multi-Reference Preference Optimization (MRPO), leverages broader prior knowledge from diverse reference models. Our experiments demonstrate that LLMs finetuned with MRPO generalize better in various preference data, regardless of data scarcity or abundance.
  arXiv  Detail & Related papers  (2024-05-26T00:29:04Z)
• LIRE: listwise reward enhancement for preference alignment [27.50204023448716]
  We propose a gradient-based reward optimization approach that incorporates the offline rewards of multiple responses into a streamlined listwise framework. LIRE is straightforward to implement, requiring minimal parameter tuning, and seamlessly aligns with the pairwise paradigm. Our experiments demonstrate that LIRE consistently outperforms existing methods across several benchmarks on dialogue and summarization tasks.
  arXiv  Detail & Related papers  (2024-05-22T10:21:50Z)
• SPO: Multi-Dimensional Preference Sequential Alignment With Implicit Reward Modeling [34.32744849352087]
  We propose a method that sequentially fine-tunes large language models to align with human preferences. We theoretically derive closed-form optimal SPO policy and loss function. Empirical results on LLMs of different size and multiple evaluation datasets demonstrate that SPO successfully aligns LLMs across multiple dimensions of human preferences.
  arXiv  Detail & Related papers  (2024-05-21T12:47:17Z)
• Fine-Tuning Language Models with Reward Learning on Policy [68.70065254564642]
  Reinforcement learning from human feedback (RLHF) has emerged as an effective approach to aligning large language models (LLMs) to human preferences. Despite its popularity, (fixed) reward models may suffer from inaccurate off-distribution. We propose reward learning on policy (RLP), an unsupervised framework that refines a reward model using policy samples to keep it on-distribution.
  arXiv  Detail & Related papers  (2024-03-28T10:02:10Z)
• WARM: On the Benefits of Weight Averaged Reward Models [63.08179139233774]
  We propose Weight Averaged Reward Models (WARM) to mitigate reward hacking. Experiments on summarization tasks, using best-of-N and RL methods, shows that WARM improves the overall quality and alignment of LLM predictions.
  arXiv  Detail & Related papers  (2024-01-22T18:27:08Z)
• Stabilizing RLHF through Advantage Model and Selective Rehearsal [57.504894664689]
  Large Language Models (LLMs) have revolutionized natural language processing, yet aligning these models with human values and preferences remains a significant challenge. This challenge is characterized by various instabilities, such as reward hacking and catastrophic forgetting. We propose two innovations to stabilize RLHF training: 1) Advantage Model, which directly models advantage score and regulates score distributions across tasks to prevent reward hacking; and 2) Selective Rehearsal, which mitigates catastrophic forgetting by strategically selecting data for PPO training and knowledge rehearsing.
  arXiv  Detail & Related papers  (2023-09-18T23:06:32Z)
• Contrastive Self-supervised Sequential Recommendation with Robust Augmentation [101.25762166231904]
  Sequential Recommendationdescribes a set of techniques to model dynamic user behavior in order to predict future interactions in sequential user data. Old and new issues remain, including data-sparsity and noisy data. We propose Contrastive Self-Supervised Learning for sequential Recommendation (CoSeRec)
  arXiv  Detail & Related papers  (2021-08-14T07:15:25Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.