💬 NLP

On the Existence and Behavior of Secondary Attention Sinks

This paper identifies and characterizes "secondary attention sinks" as distinct, transient phenomena emerging in middle layers of transformer models due to specific MLP-induced vector alignments, revealing their deterministic patterns across model scales and their dynamic interplay with weakening primary sinks.

Original authors: Jeffrey T. H. Wong, Cheng Zhang, Louis Mahon, Wayne Luk, Anton Isopoussu, Yiren Zhao

Published 2026-03-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jeffrey T. H. Wong, Cheng Zhang, Louis Mahon, Wayne Luk, Anton Isopoussu, Yiren Zhao

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

Imagine a large language model (like the AI behind this chat) as a massive, multi-story library where information travels from the ground floor to the top floor. As a story or a math problem moves through this library, it passes through many "rooms" (layers), and in each room, a team of librarians (the attention mechanism) decides which parts of the story are most important to remember.

For a while, researchers knew about one specific librarian who always got a lot of attention: the BOS Token (the "Start of Sequence" token). Think of this as the Librarian at the Front Desk. No matter what story you tell, the Front Desk Librarian always gets a huge spotlight, even if the story is about something completely unrelated to the front desk. This is called an "Attention Sink." It's like a black hole that sucks up attention, and it helps the library run smoothly.

But this new paper, accepted to a top AI conference, discovered something surprising: There are other librarians who also get sucked into the spotlight, but they are very different.

Here is the breakdown of their discovery using simple analogies:

1. The "Secondary Sinks": The Pop-Up Stars

The authors found a new type of attention sink they call Secondary Sinks.

The Old Way (Primary Sinks): The Front Desk Librarian (BOS) is always there from the very first room to the very last. They are the permanent star.
The New Way (Secondary Sinks): Imagine that halfway up the building, in the middle floors, a random, boring-looking book (a token with no real meaning, like a comma or a space) suddenly gets a massive spotlight. It wasn't important at the start, but suddenly, in the middle of the story, the librarians start staring at it intensely.
The Difference: These "Secondary Sinks" don't stay forever. They appear in the middle, shine brightly for a few floors, and then fade away. They are like pop-up stars that appear only during specific parts of the story.

2. How They Are Made: The "Transformer" Machine

How does a boring token suddenly become a star? The paper found the "factory" that makes them.

The Factory: It happens in a specific machine in the middle of the building called an MLP (a type of processing unit).
The Process: Imagine a conveyor belt carrying many different items (tokens). Most items stay normal. But there is a specific machine in the middle that grabs a few random items and paints them with a special "glow" that makes them look exactly like the Front Desk Librarian.
The Result: Once these items pass through this machine, they align perfectly with the Front Desk Librarian's direction. They become "fake Front Desks" for a little while.

3. The "Sink Levels": A Hierarchy of Fame

The researchers noticed these pop-up stars aren't random; they follow a pattern.

Level 1: The permanent Front Desk Librarian (BOS).
Level 2, 3, etc.: Different groups of "pop-up stars" that appear at different heights in the building.
The Rule: In bigger, smarter models (like the 32B or 14B parameter models mentioned), these levels are very organized. It's like a well-rehearsed play where specific actors know exactly when to step into the spotlight and when to leave. In smaller models, this doesn't happen as much.

4. The "Compensation" Effect: Why Do They Exist?

Why would the library create these temporary stars?

The Theory: The permanent Front Desk Librarian (BOS) gets tired. As the story gets longer and more complex, the Front Desk Librarian's attention starts to fade or get "diluted" in the middle of the building.
The Fix: The Secondary Sinks appear right when the Front Desk Librarian gets weak. They act as substitutes. They take over the job of holding the attention steady so the story doesn't fall apart. It's like a backup singer stepping in when the lead singer's voice gets a little raspy, ensuring the song keeps going smoothly.

5. The "Strength" of the Star

The paper also found that the "brightness" of these pop-up stars depends on how hard the factory machine works.

If the machine pushes the token really hard (high mathematical "norm"), the star shines brighter and stays in the spotlight for more floors.
If the machine is gentle, the star is dimmer and disappears quickly.

Why Does This Matter?

Understanding these "Secondary Sinks" is like finding the hidden wiring in a house.

Better AI: If we know how these temporary stars work, we can build better AI that doesn't get confused by long stories or complex math problems.
Efficiency: We might be able to save computer memory (by knowing which "stars" we can ignore) or make the AI faster.
Reasoning: The paper noticed that models trained specifically for math and reasoning (like the Qwen and DeepSeek models they tested) have these secondary sinks much more often. This suggests that these "pop-up stars" might be a secret weapon the AI uses to solve hard problems.

In short: The AI isn't just staring at the start of the sentence. In the middle of a complex thought, it creates temporary "focus points" to help it keep its balance, acting like a team of substitute librarians stepping in to keep the story on track.

1. Problem Statement

Attention Sinks are tokens (typically the Beginning-of-Sequence or BOS token) that receive disproportionately high attention weights despite having limited semantic relevance. While prior research established that BOS tokens act as "primary sinks" emerging early and persisting throughout the network, recent studies suggested other tokens could also become sinks.

However, previous works treated these non-BOS sinks as analogous to the BOS sink (emerging at the same layer and persisting indefinitely). This paper identifies a gap in understanding: Secondary Attention Sinks. These are a distinct class of sinks that:

Emerge primarily in middle layers of the network rather than the first layer.
Persist for a variable number of layers (shorter lifetimes) rather than the entire depth.
Draw a smaller, yet significant, amount of attention mass compared to primary sinks.
Often appear on semantically uninformative tokens (e.g., spaces, punctuation, numbers) at arbitrary positions in the generation sequence.

The authors aim to characterize these secondary sinks, understand their causal formation, and analyze their impact on the attention mechanism.

2. Methodology

The authors conducted extensive empirical experiments across 11 model families (including Qwen2/2.5/3, QwQ, DeepSeek, LLaMA-3.1, Phi-4, etc.) with varying scales and training objectives (base, instruct, math, reasoning).

Key Methodological Steps:

Detection Protocol:
- Generated reasoning traces using DeepSeek-14B on datasets like AIME24 and Math.
- Fed these traces into 11 different model families.
- Identified sinks by computing the pairwise cosine similarity between hidden states ( $h_t$ ) and the BOS token ( $h_0$ ) at each layer. Tokens with similarity $> 0.95$ were classified as sinks.
- Verified findings using $\ell_2$ -norm analysis of hidden states, keys, and values.
Causal Formation Analysis:
- Layer-wise Tracking: Tracked tokens that eventually became secondary sinks ("future sinks") through specific layers (e.g., Layer 22 in DeepSeek-14B) to pinpoint the exact transformation point.
- Component Isolation: Analyzed the contributions of Multi-Head Self-Attention (MHSA) and Multi-Layer Perceptron (MLP) modules by swapping outputs (hidden states, attention outputs, MLP outputs) with average uninformative tokens.
- PCA & Linear Hypothesis: Performed Principal Component Analysis (PCA) on MLP inputs to test if future sinks share a common key that retrieves the sink direction.
- Clustering: Used t-SNE to visualize the separation between "future sinks" and "normal uninformative tokens" across network depth.
Quantification: Defined Sink Levels based on the creation layer ( $l_{start}$ ) and the lifetime (number of layers the sink persists).

3. Key Contributions & Findings

A. Discovery of Secondary Sinks and Sink Levels

The paper establishes that attention sinks are not monolithic. They can be categorized into:

Primary Sinks: Correspond to the BOS token; emerge early and persist throughout the network.
Secondary Sinks: Emerge in middle layers (e.g., Layer 22 in DeepSeek-14B) and have variable lifetimes.

Sink Levels: The authors define "levels" by the tuple $(l_{start}, \text{lifetime})$ . In larger models (e.g., QwQ-32B, Qwen3-14B), these levels become more deterministic and frequent, with specific layers acting as "creation points" for distinct sink groups.

B. Causal Mechanism: The Role of Middle-Layer MLPs

The formation of secondary sinks is driven by specific MLP modules in the middle layers:

Transformation Point: Tokens behave normally until they reach a specific layer ( $l_{start}$ ). At this layer, the MLP transforms the token's representation.
Direction Alignment: The MLP maps diverse, orthogonal token representations into a vector that aligns with the primary sink direction (the BOS direction).
Amplification: The MLP acts as an amplifier for components aligned with the sink direction and suppresses misaligned components. This results in a large $\ell_2$ -norm for the hidden state, which correlates with high attention scores.
Early Preparation: While the sink effect manifests at $l_{start}$ , the decision to create the sink begins in earlier layers (e.g., Layer 19), where "future sinks" start to cluster separately from normal tokens in the representation space.

C. Correlation with Norms and Lifetime

$\ell_2$ -Norm as a Predictor: The $\ell_2$ $ℓ_{2}$ -norm of the MLP output at $l_{start}$ $l_{s t a r t}$ determines two critical properties:
1. Sink Score: The magnitude of the attention weight.
2. Lifetime: How many subsequent layers the token remains a sink.
- Higher norms lead to stronger sinks and longer persistence.
Compensatory Effect: The primary (BOS) sink strength decays as the network depth increases, reaching a minimum in the middle layers. This decay coincides exactly with the emergence of secondary sinks, suggesting a compensatory mechanism where secondary sinks maintain the attention stability required for the model to function.

D. Model Scale and Training Dependence

Scale: Secondary sinks are rare or non-existent in small base models. They become prominent in larger models and those with extensive post-training on reasoning/math data (e.g., Qwen-Math, QwQ).
Determinism: As model scale increases, the location ( $l_{start}$ ) and lifetime of sinks become more deterministic, forming distinct "levels" rather than random occurrences.

4. Results Summary

Detection: Secondary sinks were found in Qwen2, Qwen2.5, Qwen3, QwQ, and DeepSeek families, but notably absent in CodeLlama and LLaMA-3.1 (despite using similar architectures), suggesting specific training data or positional embedding nuances (RoPE base frequencies) play a role.
Token Types: Secondary sinks frequently occur on semantically uninformative tokens (spaces, newlines, digits) at arbitrary positions, not just fixed positions.
Quantitative Impact: In models like Qwen3-14B, up to 6 distinct sink levels were identified. The ratio of secondary sink scores to BOS sink scores grows log-linearly with the MLP output norm.

5. Significance and Implications

Theoretical Insight: This work challenges the view that attention sinks are solely a BOS phenomenon. It reveals a dynamic, multi-layered mechanism where the model actively constructs "reference points" (sinks) at various depths to stabilize attention.
Mechanism of Reasoning: The strong correlation between secondary sinks and reasoning/math-trained models suggests these sinks may be a byproduct or a necessary component of complex reasoning capabilities, acting as internal coordinate systems for high-dimensional representations.
Practical Applications:
- KV-Cache Optimization: Understanding that sinks have variable lifetimes could lead to more efficient KV-cache compression strategies that do not blindly preserve all tokens up to the end of the sequence.
- Model Robustness: Identifying the specific MLP modules responsible for sink formation could help in debugging model instability or improving quantization (by preserving these critical "sink" vectors).
- Architecture Design: Future models might be designed to control or suppress secondary sinks if they are found to interfere with specific tasks, or enhance them to improve reasoning stability.

In conclusion, the paper provides a rigorous characterization of Secondary Attention Sinks, demonstrating they are a distinct, causally formed phenomenon driven by middle-layer MLPs, which serve to compensate for the decay of primary sinks and are strongly linked to model scale and reasoning capabilities.