Cooperative or Competitive? Understanding the Interaction between Attention Heads From A Game Theory Perspective (2025.acl-long)
Copied to clipboard
| Challenge: | a number of attention-based large language models (LLMs) focus on individual head contributions, but the precise interaction mechanisms between attention heads remain poorly understood. |
| Approach: | They propose a game-theoretic attention calibration method that uses the Harsanyi dividend . they selectively retain heads demonstrating significant cooperative gains and apply fine-grained adjustments to remaining heads . |
| Outcome: | The proposed framework is based on the Harsanyi dividend, a concept from cooperative game theory. |
Similar Papers
Head-wise Shareable Attention for Large Language Models (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) suffer from huge number of parameters, which restricts their deployment on edge devices. |
| Approach: | They propose two methods that share parameters across attention heads to reduce memory usage and reduce performance drop by using coarse-grained weight sharing rules. |
| Outcome: | The proposed methods reuse pre-trained weights without retraining and then share, denoted as PostShare. |
Cross-layer Attention Sharing for Pre-trained Large Language Models (2026.tacl-1)
Copied to clipboard
Yongyu Mu, Yuzhang Wu, Yuchun Fan, Chenglong Wang, Hengyu Li, Jiali Zeng, Qiaozhi He, Murun Yang, Fandong Meng, Jie Zhou, Tong Xiao, Jingbo Zhu
| Challenge: | Existing studies focus on compressing the Key-Value cache or grouping attention heads, while overlooking redundancy between layers. |
| Approach: | They propose a lightweight substitute for self-attention in well-trained LLMs that uses feed-forward networks to align attention heads between adjacent layers and low-rank matrices to approximate differences in layer-wise attention weights. |
| Outcome: | The proposed model reduces redundancy by sharing weights across layers while maintaining high response quality while reducing redundant calculations within 53% 84% of the total layers. |
Exploring Attention Attractors in Large Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Existing studies have suggested that attention attractors function as "summary tokens" while others speculate that tokens with weaker semantics attract high attention, they act as attention sinks that offload excessive attention. |
| Approach: | They examine attention attractors, tokens that draw significantly high attention, in large language models. |
| Outcome: | The proposed models are able to capture long-range dependencies within a given context. |
Have Attention Heads in BERT Learned Constituency Grammar? (2021.eacl-srw)
Copied to clipboard
| Challenge: | Recent pre-trained language models have gained great success in many tasks, but what they have learned, and when they perform well remain unknown. |
| Approach: | They employ the syntactic distance method to extract implicit constituency grammar from attention weights of attention heads of BERT and RoBERTa. |
| Outcome: | The proposed models induce some grammar types much better than baselines, suggesting some heads act as a proxy for constituency grammar. |
Attention Basin: Why Contextual Position Matters in Large Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are sensitive to the contextual position of information in input. |
| Approach: | They introduce Attention-Driven Reranking (AttnRank) which estimates a model’s intrinsic positional attention preferences using a small calibration set and reorders retrieved documents or few-shot examples to align the most salient content with these high-attention positions. |
| Outcome: | Experiments on multi-hop QA and few-shot in-context learning tasks show that AttnRank achieves substantial improvements across 10 large language models of varying architectures and scales, without modifying model parameters or training procedures. |
Probing and Boosting Large Language Models Capabilities via Attention Heads (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing approaches to identifying capabilities rely on external signals with limited structural grounding . emergence of specific capabilities remains poorly understood . |
| Approach: | They propose a lightweight approach that links LLM capabilities to internal components by identifying correspondences at the level of attention heads. |
| Outcome: | The proposed approach improves accuracy on MMLU and BBH by 1 to 1.5 points over gradient-based method and 5 to 6 points over other intermediate-state baselines. |
Multi-Level Attention Aggregation for Language-Agnostic Speaker Replication (2024.eacl-short)
Copied to clipboard
| Challenge: | Recent advances in speech synthesis research have enabled the generation of natural-sounding speech, which has prompted a notable shift in TTS research towards the synthesis of speech in the voices of both seen and unseen speakers. |
| Approach: | They propose a multi-level attention aggregation approach that probes and amplifies various speaker-specific attributes in a hierarchical manner. |
| Outcome: | The proposed model achieves substantial speaker similarity and generalizes to out-of-domain (OOD) cases. |
Attention Mechanism with Energy-Friendly Operations (2022.findings-acl)
Copied to clipboard
| Challenge: | Empirical results show that attention mechanism can be improved from the energy consumption aspects. |
| Approach: | They propose to replace multiplications with either selective operations or additions to reduce energy consumption. |
| Outcome: | The proposed model achieves competitable accuracy while saving 99% and 66% energy during alignment calculation and the whole attention procedure. |
Cascaded Head-colliding Attention (2021.acl-long)
Copied to clipboard
| Challenge: | Existing frameworks for natural language processing ignore interactions among different heads, which wastes the capacity of the model. |
| Approach: | They propose a model which explicitly models interactions between attention heads through a hierarchical variational distribution. |
| Outcome: | The proposed model outperforms the baseline model on Wikitext-103 and WMT14 EN-DE on language modeling and translation tasks. |
Mitigating Attention Localization in Small Scale: Self-Attention Refinement via One-step Belief Propagation (2025.findings-emnlp)
Copied to clipboard
| Challenge: | a new framework for self-attention models is proposed to address this problem . it injects *multi-hop* relationships into the attention graph, allowing for better performance . |
| Approach: | They propose a framework that injects *multi-hop* relationships through a belief propagation process. |
| Outcome: | The proposed framework helps prevent entropy collapse in deeper layers and maintains GTD at task-appropriate levels. |