Mitigating Attention Hacking in Preference-Based Reward Modeling via Interaction Distillation

• URL: http://arxiv.org/abs/2508.02618v1
• Date: Mon, 04 Aug 2025 17:06:23 GMT
• Title: Mitigating Attention Hacking in Preference-Based Reward Modeling via Interaction Distillation
• Authors: Jianxiang Zang, Meiling Ning, Shihan Dou, Jiazheng Zhang, Tao Gui, Qi Zhang, Xuanjing Huang,
• Abstract summary: "Interaction Distillation" is a novel training framework for more adequate preference modeling through attention-level optimization.<n>It provides more stable and generalizable reward signals compared to state-of-the-art RM optimization methods.
• Score: 40.79564929465515
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: The reward model (RM), as the core component of reinforcement learning from human feedback (RLHF) for large language models (LLMs), responsible for providing reward signals to generated responses. However, mainstream preference modeling in RM is inadequate in terms of token-level interaction, making its judgment signals vulnerable to being hacked by misallocated attention to context. This stems from two fundamental limitations: (1) Current preference modeling employs decoder-only architectures, where the unidirectional causal attention mechanism leads to forward-decaying intra-sequence attention within the prompt-response sequence. (2) The independent Siamese-encoding paradigm induces the absence of token-level inter-sequence attention between chosen and rejected sequences. To address this "attention hacking", we propose "Interaction Distillation", a novel training framework for more adequate preference modeling through attention-level optimization. The method introduces an interaction-based natural language understanding model as the teacher to provide sophisticated token interaction patterns via comprehensive attention, and guides the preference modeling to simulate teacher model's interaction pattern through an attentional alignment objective. Through extensive experiments, interaction distillation has demonstrated its ability to provide more stable and generalizable reward signals compared to state-of-the-art RM optimization methods that target data noise, highlighting the attention hacking constitute a more fundamental limitation in RM.

Related papers

• CTR-Sink: Attention Sink for Language Models in Click-Through Rate Prediction [42.92011330807996]
  $textitCTR-Sink$ is a novel framework introducing behavior-level attention sinks tailored for recommendation scenarios.<n>Inspired by attention sink theory, it constructs attention focus sinks and dynamically regulates attention aggregation via external information.
  arXiv  Detail & Related papers  (2025-08-05T17:30:34Z)
• AURORA: Augmented Understanding via Structured Reasoning and Reinforcement Learning for Reference Audio-Visual Segmentation [113.75682363364004]
  AURORA is a framework designed to enhance genuine reasoning and language comprehension in reference audio-visual segmentation.<n>AURORA achieves state-of-the-art performance on Ref-AVS benchmarks and generalizes effectively to unreferenced segmentation.
  arXiv  Detail & Related papers  (2025-08-04T07:47:38Z)
• Zero-Shot EEG-to-Gait Decoding via Phase-Aware Representation Learning [9.49131859415923]
  We propose NeuroDyGait, a domain-generalizable EEG-to-motion decoding framework.<n>It uses structured contrastive representation learning and relational domain modeling to achieve semantic alignment between EEG and motion embeddings.<n>It achieves zero-shot motion prediction for unseen individuals without requiring adaptation and superior performance in cross-subject gait decoding on benchmark datasets.
  arXiv  Detail & Related papers  (2025-06-24T06:03:49Z)
• A Self-Supervised Reinforcement Learning Approach for Fine-Tuning Large Language Models Using Cross-Attention Signals [0.0]
  We propose a novel reinforcement learning framework for large language models that does not rely on human in the loop feedback.<n>Instead, our approach uses cross attention signals within the model itself to derive a self supervised reward.
  arXiv  Detail & Related papers  (2025-02-14T01:44:04Z)
• HAAP: Vision-context Hierarchical Attention Autoregressive with Adaptive Permutation for Scene Text Recognition [17.412985505938508]
  Internal Language Model (LM)-based methods use permutation language modeling (PLM) to solve the error correction caused by conditional independence in external LM-based methods. This paper proposes the Hierarchical Attention autoregressive Model with Adaptive Permutation (HAAP) to enhance the location-context-image interaction capability.
  arXiv  Detail & Related papers  (2024-05-15T06:41:43Z)
• Collaborative Filtering Based on Diffusion Models: Unveiling the Potential of High-Order Connectivity [10.683635786183894]
  CF-Diff is a new diffusion model-based collaborative filtering method. It is capable of making full use of collaborative signals along with multi-hop neighbors. It achieves remarkable gains up to 7.29% compared to the best competitor.
  arXiv  Detail & Related papers  (2024-04-22T14:49:46Z)
• DELTA: Dynamic Embedding Learning with Truncated Conscious Attention for CTR Prediction [61.68415731896613]
  Click-Through Rate (CTR) prediction is a pivotal task in product and content recommendation. We propose a model that enables Dynamic Embedding Learning with Truncated Conscious Attention for CTR prediction.
  arXiv  Detail & Related papers  (2023-05-03T12:34:45Z)
• Progressively Guide to Attend: An Iterative Alignment Framework for Temporal Sentence Grounding [53.377028000325424]
  We propose an Iterative Alignment Network (IA-Net) for temporal sentence grounding task. We pad multi-modal features with learnable parameters to alleviate the nowhere-to-attend problem of non-matched frame-word pairs. We also devise a calibration module following each attention module to refine the alignment knowledge.
  arXiv  Detail & Related papers  (2021-09-14T02:08:23Z)
• Modeling long-term interactions to enhance action recognition [81.09859029964323]
  We propose a new approach to under-stand actions in egocentric videos that exploits the semantics of object interactions at both frame and temporal levels. We use a region-based approach that takes as input a primary region roughly corresponding to the user hands and a set of secondary regions potentially corresponding to the interacting objects. The proposed approach outperforms the state-of-the-art in terms of action recognition on standard benchmarks.
  arXiv  Detail & Related papers  (2021-04-23T10:08:15Z)
• Cost-effective Interactive Attention Learning with Neural Attention Processes [79.8115563067513]
  We propose a novel interactive learning framework which we refer to as Interactive Attention Learning (IAL) IAL is prone to overfitting due to scarcity of human annotations, and requires costly retraining. We tackle these challenges by proposing a sample-efficient attention mechanism and a cost-effective reranking algorithm for instances and features.
  arXiv  Detail & Related papers  (2020-06-09T17:36:41Z)
• Cascaded Human-Object Interaction Recognition [175.60439054047043]
  We introduce a cascade architecture for a multi-stage, coarse-to-fine HOI understanding. At each stage, an instance localization network progressively refines HOI proposals and feeds them into an interaction recognition network. With our carefully-designed human-centric relation features, these two modules work collaboratively towards effective interaction understanding.
  arXiv  Detail & Related papers  (2020-03-09T17:05:04Z)

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.