Source-linked AI summary
Mitigating Attention Sinks and Massive Activations in Audio-Visual Speech Recognition with LLMs
Anand, Umberto Cappellazzo, Stavros Petridis, Maja Pantic
TL;DR
Audio-visual LLM internals under fine-tuning remain poorly understood. This study analyzes BOS-aligned sinks and activations, then uses decorrelation to mitigate them, with consistent WER improvements under high compression across ASR, VSR, and AVSR.
Problem
The internal mechanisms of audio-visual LLMs under fine-tuning remain poorly understood, and attention sinks and massive activations had not been studied in speech recognition.
Source: §1 · p. 1Method
Results
Across ASR, VSR, and AVSR, decorrelation yields consistent WER improvements under high compression.
Source: §6 · p. 4Takeaways & Limitations
Abstract
from arXiv · showhide
Large language models (LLMs) have recently advanced auditory speech recognition (ASR), visual speech recognition (VSR), and audio-visual speech recognition (AVSR). However, understanding of their internal dynamics under fine-tuning remains limited. In natural language processing, recent work has revealed attention sinks, tokens that attract disproportionately high attention, and associated massive activations in which some features of sink tokens exhibit huge activation in LLMs. In this work, we are the first to study these phenomena in multimodal speech recognition. Through a detailed analysis of audio-visual LLMs, we identify attention sinks and massive activations not only at the BOS token but also at intermediate low-semantic tokens across ASR, VSR, and AVSR. We show that massive activations originate in the MLP layers and correspond to fixed feature indices across all sink tokens. We further show that intermediate sink tokens exhibit high cosine similarity to the BOS token, thereby amplifying attention and activation. Building on these insights, we introduce a simple decorrelation loss that reduces cosine similarity between BOS and other tokens, effectively mitigating intermediate sinks and massive activations. Furthermore, our method improves word error rate (WER) under high audio-visual feature downsampling while remaining stable at lower downsampling rates.
1. INTRODUCTION
The introduction identifies an understudied internal-dynamics gap in multimodal speech recognition with LLMs. It presents attention sinks and massive activations across ASR, VSR, and AVSR, then proposes decorrelation to mitigate them and improve WER under high compression.
1. INTRODUCTION
- Prior work found that BOS and semantically uninformative intermediate tokens attract disproportionate attention, but their interaction with massive activations remains unclear.Some sink features can reach magnitudes up to four orders larger than the median.Source: §1 · p. 1
- Despite LLM advances in ASR, VSR, and AVSR, their internal mechanisms under fine-tuning remain poorly understood, and these sink phenomena were unexplored in speech recognition.The introduction notes that intermediate sinks may divert attention from phonetic or lip-movement cues, while massive activations can over-amplify irrelevant features.Source: §1 · p. 1
- Intermediate attention sinks emerge during fine-tuning across ASR, VSR, and AVSR, and their hidden states align closely with BOS.Massive activations arise from the MLP as early as layer 2, with shared feature indices across BOS and intermediate sinks.Source: §1 · p. 1
- The decorrelation method improves WER across ASR, VSR, and AVSR under high compression while remaining comparable to the baseline at lower compression rates.It is designed to integrate with LoRA fine-tuning without inference-time cost.Source: §1 · p. 1
- A decorrelation loss penalizes BOS alignment to mitigate intermediate sinks and massive activations, explaining their co-existence through cosine similarity.The introduction presents this as the paper’s third key contribution.Source: §1 · p. 1
2. PRELIMINARIES
The preliminaries describe how Llama-AVSR converts audio and video into compressed LLM inputs and generates transcriptions autoregressively. They also outline the decoder blocks and token-wise MLP used in the analysis.
2. PRELIMINARIES
- Llama-AVSR encodes audio and video, temporally downsamples embeddings by average pooling, projects them into the LLM space, and concatenates them with an instruction prompt.For AVSR, compression rates (a, v) denote audio and video downsampling factors; unimodal ASR and VSR use a single factor.Source: §2 · pp. 1–2
- The LLM autoregressively generates the target output from the audio and video tokens, instruction prompt, and previously generated output tokens.The output sequence is generated one token at a time.Source: §2 · p. 2
- Each autoregressive decoder block contains multi-head self-attention followed by an MLP, which processes tokens independently in a gated form.The gated MLP combines projected features elementwise before applying a down projection.Source: §2 · p. 2
3. ANALYSIS OF AUDIO-VISUAL LLMS
Across ASR, VSR, and AVSR, the analysis finds BOS and intermediate attention sinks, with intermediate sinks appearing during fine-tuning on low-semantic tokens. These sinks co-occur with MLP-origin massive activations and align directionally with BOS; controlled rotations alter both phenomena.
3.1. Attention Sinks
- Intermediate sinks emerge after layer 2 during fine-tuning across ASR, VSR, and AVSR, often on low-semantic tokens; the BOS sink is already present in the pretrained LLM.The analysis notes that such tokens include modality markers and prompt tokens, and hypothesizes that under high compression they may divert attention from phonetic and lip-movement cues.Source: Figure 1 · p. 2, §3.1 · p. 2
3.2. Massive Activations
- Massive activations occur only at sink tokens in layers 2 through L−1, share feature indices across sink tokens, and originate in the layer-2 MLP.The reported MLP mechanism involves large positive gate values on fixed features, amplified by the up projection and propagated to the hidden state.Source: Figure 2 · p. 3, §3.2 · p. 3
3.3. Cosine Similarity with BOS
- From layer 2 onward, intermediate sink hidden states align strongly with BOS; sink tokens also align with one another, unlike other tokens.The paper links this directional alignment to shared activated features and excessive attention.Source: Figure 3 · p. 3, §3.3 · p. 3
- Rotating intermediate sinks toward nearby non-sink tokens removes their attention sinks and massive activations, while rotating a non-sink token toward BOS induces both.The reported induction occurs at token index 10 in Llama-AVSR (16, 5).Source: §3.3 · p. 3
4. PROPOSED METHOD
The method treats BOS alignment as the shared basis of intermediate attention sinks and massive activations, and penalizes that alignment with a decorrelation loss. The loss is combined with speech-recognition cross-entropy without changing the model architecture.
4. PROPOSED METHOD
- The proposed decorrelation loss discourages alignment between BOS and other tokens to mitigate intermediate attention sinks and massive activations.The method is motivated by the finding that intermediate tokens’ directional alignment with BOS underlies both phenomena.Source: §4 · p. 3
- Disaligning sink tokens from BOS removes sinks and massive activations, while aligning a token with BOS makes both emerge.Figure 4 illustrates these effects.Source: Figure 4 · p. 4
4.1. Decorrelation Loss
- The loss penalizes squared cosine similarity between BOS and non-BOS hidden states, excluding the first and last layers.Squaring the similarity yields smoother gradients and stronger penalties for highly aligned tokens.Source: §4.1 · p. 4
4.2. Final Training Loss
- The final training objective combines decorrelation regularization with standard cross-entropy for autoregressive speech recognition.The regularization strength is controlled by λ; the method requires no architectural changes and adds negligible computational overhead.Source: §4.2 · p. 4
5. EXPERIMENTAL RESULTS
The experiments report consistent WER improvements at high compression rates while performance remains comparable to baseline at lower rates. The comparison with ACT reports marginal improvements for speech-recognition tasks and says decorrelation addresses both sinks and massive activations.
5. EXPERIMENTAL RESULTS
- Experiments evaluate sink mitigation using Llama 3.2-3B, with similar trends reported for Llama 3.2-1B and Llama 2-7B.The experiments follow the training details and code provided in.Source: §5 · p. 4
5.1. Decorrelation Loss for Intermediate Attention Sinks
- WER improves consistently at high compression rates, while performance remains comparable to baseline at lower rates with decorrelation loss.The experiments use LRS2 for AVSR and ASR, and LRS3 for VSR; Table 1 reports Llama-AVSR WER results with and without the loss.Source: §5.1 · p. 4, Table 1 · p. 4
5.2. Comparison with Prior Sink Mitigation Methods
- ACT yields only marginal improvements for speech-recognition tasks and does not mitigate massive activations; decorrelation addresses both effects and gives larger gains under high compression.Table 2 compares the proposed method with ACT.Source: §5.2 · p. 4, Table 2 · p. 4
6. CONCLUSION
The conclusion presents the first study of attention sinks and massive activations in multimodal speech-recognition LLMs. It attributes these intermediate-token phenomena to directional alignment with BOS and reports that decorrelation mitigates them while improving WER under high compression.
6. CONCLUSION
- The paper presents the first study of attention sinks and massive activations in multimodal speech-recognition LLMs.Source: §6 · p. 4
- The analysis attributes intermediate sinks and massive activations to directional alignment between intermediate tokens and BOS.Source: §6 · p. 4
- Across AVSR, ASR, and VSR, the decorrelation loss yields consistent WER improvements under high compression without architectural changes.Source: §6 · p. 4