💬 NLP

Mechanistic Insights into Functional Sparsity in Multimodal LLMs via CoRe Heads

This paper reveals that Multimodal Large Language Models rely on functional sparsity, where a specialized subset of attention heads called CoRe heads are critical for extracting query-relevant visual features, as demonstrated by their significant impact on performance when ablated and their potential to accelerate inference.

Original authors: Ruoxi Sun, Quantong Qiu, Juntao Li, Zecheng Tang, Yihang Lou, Min Zhang

Published 2026-06-05

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ruoxi Sun, Quantong Qiu, Juntao Li, Zecheng Tang, Yihang Lou, Min Zhang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: Finding the "Special Agents" in a Crowd

Imagine a Multimodal Large Language Model (MLLM) as a massive, bustling newsroom. This newsroom has thousands of reporters (called attention heads) all working at once. Their job is to read a complex story (the text prompt) and look at a chaotic scene (an image or a video) to answer a specific question.

Usually, we think all these reporters are working together equally, scanning the whole room to find the answer. But this paper discovered something surprising: Most of the reporters are actually just looking around aimlessly or staring at the background noise.

Instead, there is a tiny, elite group of about 5% of the reporters who act as "Special Agents." These agents are the only ones who actually know where to look to find the specific answer. The authors call these the CoRe Heads (Context-aware Retrieval heads).

The Detective Work: How They Found the Agents

To prove this, the researchers invented a new way to measure attention called RAM (Retrieval Attention Mass).

The Analogy: Imagine you are looking for a specific red car in a busy parking lot.
- The "Bad" Reporters (Bottom Heads): They look at the sky, the pavement, the people walking by, and the trees. They are distracted by everything except the car.
- The "CoRe" Reporters: They ignore the sky and the people. They lock their eyes directly onto the red car.

The researchers measured how much attention each "reporter" paid to the correct object (the red car) versus the wrong stuff. They found that the CoRe Heads consistently pointed directly at the answer, while the others were just "noise."

The "What If" Experiment: Removing the Agents

To see if these Special Agents were actually necessary, the researchers tried a dangerous experiment: they turned them off.

The Result: When they silenced just the top 5% of the best reporters (the CoRe Heads), the newsroom collapsed. The model's ability to answer questions dropped drastically. It was like taking the only person who knew the way out of a maze and asking the rest of the crowd to guide you.
The Contrast: When they turned off the "distracted" reporters (the bottom 95%), the model kept working almost perfectly fine. This proved that the "noise" reporters were redundant, but the "Special Agents" were essential.

The "Bottleneck" Discovery

The paper also noticed a pattern as the models got bigger (from small to massive).

Small Models: The Special Agents were scattered all over the newsroom, a bit messy.
Big Models: As the model grew, the Special Agents moved into a specific, tight cluster in the middle-to-late sections of the building. They formed a bottleneck. The model realized, "Hey, we don't need everyone to look; we just need this specific team in this specific room to do the heavy lifting."

The Superpower: Speeding Things Up

Because the researchers proved that 95% of the reporters are just "filling space" and not doing the critical work, they tried a new strategy to make the model faster.

The Strategy: They told the 95% of distracted reporters, "You don't need to look at the whole room anymore. Just look at the immediate area right in front of you." Meanwhile, they let the 5% of Special Agents keep looking at the whole room.
The Result: This made the model much faster (up to 2x speed) without losing accuracy. It's like telling a crowd to stop shouting across the room and just whisper to their neighbors, while only the few experts shout the important news. The message still gets through, but the chaos (and the time it takes) is gone.

Summary

This paper reveals that Multimodal AI models aren't actually using their full brain power to find answers. They rely on a tiny, specialized team (CoRe Heads) to do the real work of finding visual clues, while the rest of the network just helps keep the lights on. By identifying and protecting this small team while simplifying the work of the rest, we can make these AI models significantly faster and more efficient.

Technical Summary: Mechanistic Insights into Functional Sparsity in Multimodal LLMs via CoRe Heads

1. Problem Statement

Multimodal Large Language Models (MLLMs) have demonstrated remarkable proficiency in complex vision-language tasks, yet the internal mechanisms governing how they extract query-relevant visual features from noisy, cluttered contexts remain opaque. While MLLMs map high-dimensional visual signals into semantic spaces, real-world inputs often contain redundant information. Robust multimodal reasoning depends on the model's ability to selectively isolate sparse, task-relevant visual cues. However, existing interpretability frameworks rely on coarse statistical heuristics (e.g., spatial entropy, attention aggregation) evaluated primarily on static images. These approaches fail to capture fine-grained, query-conditioned token-level attention in dense or dynamic spatiotemporal scenes, leaving a significant gap in understanding MLLM failure modes and internal efficiency.

2. Methodology

The authors propose a systematic framework to identify and characterize a specialized subset of attention heads responsible for cross-modal retrieval, termed Context-aware Retrieval (CoRe) heads.

2.1 Retrieval Attention Mass (RAM)

To isolate these heads, the authors introduce a token-level metric called Retrieval Attention Mass (RAM). Unlike generic attention metrics, RAM quantifies the attention mass allocated by a specific head $h$ from instructional query tokens ( $q$ ) to ground-truth visual entities ( $V^*$ ).
The RAM score for a head is defined as:
$M(h)_{RAM}(q \to V^*) = \mathbb{E}_{x \in q} \left[ \sum_{y \in \Omega(V^*)} A(h)_{x \to y} \right]$
where $\Omega(V^*)$ represents the tokenized spatial or temporal span of the key visual entities, and $A(h)_{x \to y}$ is the attention weight. High RAM values indicate targeted retrieval rather than generic attention accumulation or "attention sinks."

2.2 Unified Evaluation Protocol

The study employs a unified protocol across four diverse datasets covering different modalities and complexities:

RefCOCOg: Static image visual grounding (bounding boxes).
VidSTG: Video spatio-temporal grounding (temporal tubes).
MMLongBench: Long-context multi-evidence reasoning (evidence regions).
MMDocIR: Document-grounded retrieval (page/figure regions).

The authors map continuous annotations (spatial boxes, temporal tubes) to discrete token indices $\Omega(V^*)$ using architecture-specific strategies (e.g., boundary tagging for adapter-based models, uniform slicing for fixed-expansion models) to ensure precise attention measurement.

2.3 Causal Interventions and Acceleration

To validate the functional necessity of CoRe heads, the authors perform causal interventions via attention masking during inference. They ablate the top- $k$ CoRe heads (ranked by RAM) and compare performance against ablating bottom-ranked or random heads.
Furthermore, they propose a CoRe-guided Hybrid Attention mechanism for system-level acceleration. This strategy retains full global attention for the top- $k$ CoRe heads while restricting the remaining non-essential heads to local sliding windows (Stream Sparse Attention) during the prefill stage, aiming to reduce quadratic computational complexity.

3. Key Contributions and Results

3.1 Discovery of Functional Sparsity

The analysis reveals a profound structural property: functional sparsity. A small, specialized subset of attention heads (CoRe heads) acts as dedicated information extractors, while the vast majority of heads exhibit diffuse, global attention patterns.

Localization: CoRe heads precisely localize context-relevant visual entities (e.g., specific objects in a video or text regions in a document), filtering out background noise.
Differentiation: Lower-ranked heads (Bottom Heads) scatter attention across irrelevant regions and fail to ground instructions.

3.2 Structural Evolution and Scaling

Scaling Patterns: As model capacity increases (e.g., from 4B to 32B parameters), CoRe heads transition from a scattered distribution to a highly concentrated "bottleneck" in the deep layers (e.g., layers 44–56 in 32B models).
Architectural Divergence: Different model families exhibit distinct topologies. Qwen3-VL shows sparse localization, LLaVA-OneVision shows moderate dispersion, and InternVL3.5 displays dense activations, suggesting that structural sparsity is a dynamic allocation strategy responsive to visual representation density rather than an intrinsic property of the base LLM.

3.3 Cross-Domain Stability

CoRe heads exhibit a "globally consistent, locally variant" paradigm. Approximately 30 specific heads remain universally activated across diverse tasks (RefCOCOg, VidSTG, MMLongBench, MMDocIR), with high Spearman rank correlations (0.79–0.92) in activation patterns. These heads are stably concentrated in the middle-to-late layers.

3.4 Causal Necessity

Causal interventions confirm the non-redundant necessity of CoRe heads:

Performance Degradation: Ablating only the top 5% of CoRe heads causes a catastrophic drop in multimodal reasoning performance (e.g., accuracy plummeting from ~45% to ~27% in Qwen3-VL-4B).
Redundancy: Ablating bottom-ranked heads has minimal effect, indicating significant functional redundancy in the non-specialized heads.

3.5 Acceleration and Efficiency

Leveraging this sparsity, the proposed hybrid attention mechanism achieves significant speedups:

Speedup: Up to 2.3× prefill speedup on long-context tasks (e.g., MLVU, Video-MME) while maintaining robust performance.
Performance Retention: In some configurations (e.g., retaining 19.1% of heads in LLaVA-OneVision), the hybrid model marginally outperforms the fully dense baseline, likely due to an implicit regularization effect that filters out cross-modal noise.
Scalability: Larger models (e.g., Qwen3-VL-32B) sustain higher levels of sparsification (down to 4.9% of heads) with negligible performance loss, suggesting increased functional redundancy at scale.

4. Significance and Claims

The paper claims to reveal a structural principle of functional sparsity within MLLMs, refining the current understanding of mechanistic interpretability. Key implications include:

Mechanistic Insight: Multimodal reasoning is governed by a sparse, consistent subset of heads that act as a critical bottleneck for cross-modal extraction, distinct from the global feature aggregation performed by other heads.
Non-Redundancy: These specialized heads are causally indispensable; their removal cannot be compensated for by the rest of the network.
Optimization Potential: The discovery provides a theoretical foundation for efficient model optimization. By operating on a small fraction of critical heads, it is possible to significantly accelerate inference (particularly in long-context scenarios) without compromising semantic selection or task performance.

The authors note limitations, such as the reliance on explicit ground-truth annotations for RAM calculation and the current use of static head configurations for acceleration, suggesting future work could explore dynamic, input-adaptive routing strategies.