Attention Head Specialization

Recent large audio language models (LALMs) demonstrate remarkable capabilities in processing extended multi-modal sequences, yet incur high inference costs. Token compression is an effective method that directly reduces redundant tokens in the sequence. Existing compression methods usually assume that all attention heads in LALMs contribute equally to various audio tasks and calculate token importance by averaging scores across all heads. However, our analysis demonstrates that attention heads exhibit distinct behaviors across diverse audio domains. We further reveal that only a sparse subset of attention heads actively responds to audio, with completely different performance when handling semantic and acoustic tasks. In light of this observation, we propose HeadRouter, a head-importance-aware token pruning method that perceives the varying importance of attention heads in different audio tasks to maximize the retention of crucial tokens. HeadRouter is training-free and can be applied to various LALMs. Extensive experiments on the AudioMarathon and MMAU-Pro benchmarks demonstrate that HeadRouter achieves state-of-the-art compression performance, exceeding the baseline model even when retaining 70% of the audio tokens and achieving 101.8% and 103.0% of the vanilla average on Qwen2.5-Omni-3B and Qwen2.5-Omni-7B, respectively.

When do transformers commit to a decision, and what prevents them from correcting it? We introduce prolepsis: a transformer commits early, task-specific attention heads sustain the commitment, and no layer corrects it. Replicating \citeauthor{lindsey2025biology}'s (\citeyear{lindsey2025biology}) planning-site finding on open models (Gemma2 2B, Llama3.2 1B), we ask five questions. (Q1)~Planning is invisible to six residual-stream methods; CLTs are necessary. (Q2)~The planning-site spike replicates with identical geometry. (Q3)~Specific attention heads route the decision to the output, filling a gap flagged as invisible to attribution graphs. (Q4)~Search requires ${\leq}16$ layers; commitment requires more. (Q5)~Factual recall shows the same motif at a different network depth, with zero overlap between recurring planning heads and the factual top-10. Prolepsis is architectural: the template is shared, the routing substrates differ. All experiments run on a single consumer GPU (16,GB VRAM).

Multimodal news often loses its image or text as it spreads across Web platforms. Existing methods treat this as information loss, assuming that the original veracity label remains valid. This assumption fails for fake news reusing authentic material, whose remainder may be genuine. Detectors should therefore decouple modality cues and adapt their contributions to modality availability. However, coarse news labels provide little guidance for modality decoupling, while training on complete inputs can induce modality bias. To address these challenges, we propose Head-wise Modality Decoupling and Regulation (HMDR) for multimodal large language models. First, Modality Head Decoupling decouples modality representations through soft head grouping by attention preference. Second, Modality Head Regulation regulates head contributions through compensation for imbalanced modality attention. Finally, Modality Invariance Learning promotes invariance by aligning representations of each modality under replacement of the other. Experiments on two benchmarks show that HMDR improves detection under missing modalities while maintaining performance on complete inputs.

From customer feedback to social media, understanding human sentiment in text is central to how machines can interact meaningfully with people. However, despite notable progress, accurately capturing sentiment remains a challenging task, which continues to motivate further research in this area. To this end, we introduce Non-Differential Transformer (NDT). It is inspired by (but in contrast to) the state-of-the-art Differential Transformer (DT) model. While standard Transformers can struggle with irrelevant context, the sota DT model uses attention map subtraction, potentially for noise cancellation. We explore an alternative motivation, hypothesizing that benefits may arise from enabling different attention components to specialize on distinct concepts within the text, similar to multiplexing information channels or mixture models, rather than primarily canceling noise via subtraction. Guided by this concept-multiplexing (ConPlex) view, the specific architecture presented in this paper employs a purely additive strategy. It uses only positive weights, learned during training, to ensure constructive combination of these specialized attention perspectives. This design choice explores positive only integration, though our broader framework also shows promise with less constrained linear combinations involving both positive and negative weights. Our model computes attention via this positively weighted sum of multiple distinct attention maps. This allows the model to constructively integrate diverse signals and potentially capture more complex contextual relationships. Competitive performance is achieved by the proposed model for Sentiment Analysis while tested on multiple datasets. We conclude by presenting our results, challenges and future research agenda in this important area of research.

Multi-head attention enables transformer models to represent multiple attention patterns simultaneously. Empirically, head specialization emerges in distinct stages during training, while many heads remain redundant and learn similar representations. We propose a theoretical model capturing this phenomenon, based on the multi-index and single-location regression frameworks. In the first part, we analyze the training dynamics of multi-head softmax attention under SGD, revealing an initial unspecialized phase followed by a multi-stage specialization phase in which different heads sequentially align with latent signal directions. In the second part, we study the impact of attention activation functions on performance. We introduce the Bayes-softmax attention, which achieves optimal prediction performance in this setting.

Prompt learning is a dominant paradigm for adapting pre-trained Vision-Language Models (VLMs) to downstream tasks. However, existing methods often rely on a simplistic, layer-centric view, assuming shallow layers capture general features while deep layers handle task-specific knowledge. This assumption results in uncontrolled interactions between learnable tokens and original tokens. Task-specific knowledge could degrades the model's core generalization and creates a trade-off between task adaptation and the preservation of zero-shot generalization. To address this, we challenge the layer-centric view and propose DeAR, a framework that achieves fine-grained VLM adaptation by Decomposing Attention head Roles. We posit that the functional specialization within VLMs occurs not between layers, but at the finer-grained level of individual attention heads in the deeper layers. Based on this insight, we introduce a novel metric, Concept Entropy, to systematically classify attention heads into distinct functional roles: Attribute, Generalization, and Mixed. Guided by these roles, we introduce specialized attribute tokens and a Role-Based Attention Mask mechanism to precisely control information flow, ensuring generalization heads remain isolated from task-specific knowledge. We further incorporate a Task-Adaptive Fusion Strategy for inference. Extensive experiments on fifteen datasets show that DeAR achieves a strong balance between task adaptation and generalization, outperforming previous methods across various tasks.

Retrieval heads, a subset of attention heads in Transformers, were studied in English, showing its crucial role in retrieving information from the context. We expand the study of retrieval heads to multilingual context and find that while nearly half of all retrieval heads are often shared across multiple languages, language-specific retrieval heads also emerge. We further design a cross-lingual needle-in-the-haystack task, to identify $\textit{Retrieval Transition Heads (RTH)}$ that retrieve key information and further map it to target-language output. Our experiments reveal that RTHs do not always overlap with RH, and can be more vital for Chain-of-Thought reasoning in multilingual LLMs. Across four multilingual benchmarks (MMLU-ProX, MGSM, MLQA, and XQuAD) and two model families (Qwen2.5-7B Instruct and Llama3.1-8B Instruct), we demonstrate that masking RTH induces bigger performance drop than masking Retrieval Heads (RH). Specifically, for Llama3.1-8B Instruct, masking the top-25 RTHs results in an average 37.3-point drop (against 19.5-point drop from RH masking) in reasoning accuracy (MMLU-ProX, MGSM) and 9.0 F1-score reduction (against 5.1 drop for RH) in extractive QA (MLQA, XQuAD). Our work advances the understanding of multilingual LMs by isolating the attention heads responsible for mapping to target languages.

This paper studies simple transformers trained on a high-order Markov chain, where the model must incorporate information from multiple past positions, each with different statistical importance. We show that transformers learn the task incrementally, with each stage corresponding to learning how to copy information from a subset of positions via a sparse attention pattern. Notably, the learning dynamics transition from a competitive phase, where all heads focus on the statistically most important positions, to a cooperative phase, where different heads specialize in different patterns. We model these dynamics with simplified differential equations and prove stage-wise convergence of the resulting system. Functionally, these stages correspond to a sequence of increasingly expressive misspecified models, with the full model class reached only at the end. Overall, we give a theoretical account of how structured attention patterns and head specialization emerge in stages without an explicit curriculum, with implications for generalization in sequential tasks.

Some transformer attention heads appear to function as membership testers, dedicating themselves to answering the question "has this token appeared before in the context?" We identify these heads across four language models (GPT-2 small, medium, and large; Pythia-160M) and show that they form a spectrum of membership-testing strategies. Two heads (L0H1 and L0H5 in GPT-2 small) function as high-precision membership filters with false positive rates of 0-4% even at 180 unique context tokens -- well above the $d_\text{head} = 64$ bit capacity of a classical Bloom filter. A third head (L1H11) shows the classic Bloom filter capacity curve: its false positive rate follows the theoretical formula $p \approx (1 - e^{-kn/m})^k$ with $R^2 = 1.0$ and fitted capacity $m \approx 5$ bits, saturating by $n \approx 20$ unique tokens. A fourth head initially identified as a Bloom filter (L3H0) was reclassified as a general prefix-attention head after confound controls revealed its apparent capacity curve was a sequence-length artifact. Together, the three genuine membership-testing heads form a multi-resolution system concentrated in early layers (0-1), taxonomically distinct from induction and previous-token heads, with false positive rates that decay monotonically with embedding distance -- consistent with distance-sensitive Bloom filters. These heads generalize broadly: they respond to any repeated token type, not just repeated names, with 43% higher generalization than duplicate-token-only heads. Ablation reveals these heads contribute to both repeated and novel token processing, indicating that membership testing coexists with broader computational roles. The reclassification of L3H0 through confound controls strengthens rather than weakens the case: the surviving heads withstand the scrutiny that eliminated a false positive in our own analysis.

Large-scale pretraining of transformers has been central to the success of foundation models. However, the scale of those models limits our understanding of the mechanisms at play during optimization. In this work, we study the training dynamics of transformers in a controlled and interpretable setting. On the sparse modular addition task, we demonstrate that specialized attention circuits, called clustering heads, can be implemented during gradient descent to solve the problem. Our experiments show that such pathways naturally emerge during training. By monitoring the evolution of tokens via a visual sandbox, we uncover a two-stage learning and the occurrences of loss spikes due to the high curvature of normalization layers. Our findings provide several insights into patterns observed in more practical settings, such as the pretraining of large language models.