A team from MIT has uncovered the inner mechanics behind a persistent flaw in large language models that is their tendency to overlook information buried in the middle of documents. The findings, grounded in a rigorous theoretical framework and supported by extensive experiments, explain why these systems often favor text that appears at the beginning or end.
The researchers traced the root of this issue, referred to as “position bias”, to architectural design choices and how these models are trained to process sequences. Central to the analysis is how the attention mechanism, a core component of models like GPT-4 or LLaMA, handles the flow of information across multiple layers.
Using graph theory, the team demonstrated that attention patterns are not evenly distributed. Instead, certain tokens become dominant simply due to their position. When the model reads from left to right, earlier tokens often accumulate more influence as the layers deepen, even when their content is less relevant. This effect intensifies as more layers are added, creating a cascade where initial tokens disproportionately shape the model's decisions.
The study shows that even without adding any formal position tracking, the structure of the model itself introduces a preference for the start of the sequence. In experiments with synthetic retrieval tasks, the performance of the models dipped when key information was placed in the middle of the input. The retrieval curve followed a U-shape, strong at the start, weaker in the center, then improving slightly at the end.
This behavior wasn’t incidental. Controlled tests confirmed that position bias emerged even when the training data had no such leanings. In setups where the data favored certain positions, the models amplified those biases. When models were trained on sequences biased toward the beginning and end, they mirrored that pattern, heavily underperforming in the center.
The paper also explored how positional encoding schemes, tools designed to help the model track where a word appears, can partially counteract this effect. Techniques like decay masks and rotary encodings introduce a fading influence based on distance, nudging the model to attend more evenly across the sequence. However, these methods alone don’t eliminate the bias, especially in deeper networks where earlier layers already tilt the attention forward.
In practical terms, this means that users relying on AI models for tasks like legal search, coding assistance, or medical records review may unknowingly encounter blind spots. If key content appears mid-document, the model might miss or misjudge it, even if everything else in the system functions as intended.
The implications go beyond diagnostics. By showing that position bias is both an architectural and data-driven phenomenon, the researchers offer pathways to mitigate it. Adjustments in attention masks, fewer layers, and smarter use of positional encodings can help rebalance the focus. The study also suggests that fine-tuning models on more uniformly distributed data could be essential in high-stakes domains where omission carries risk.
The research not only maps the bias but explains its evolution. As tokens move through the model, their contextual representations are repeatedly reshaped. Those that appear earlier begin to dominate, not because they contain better information, but because they become more deeply embedded in the model's reasoning. In this sense, the bias is baked into the system’s logic.
Rather than treating this as a bug, the team sees it as an opportunity for improvement. Their framework doesn’t just diagnose; it provides tools to reshape how models perceive position. By better understanding these internal biases, developers can build systems that reason more fairly and consistently across the full length of input, beginning, middle, and end.
Image: DIW-Aigen
Read next: Why a Wrench Might Outlast Code in the Age of AI
The researchers traced the root of this issue, referred to as “position bias”, to architectural design choices and how these models are trained to process sequences. Central to the analysis is how the attention mechanism, a core component of models like GPT-4 or LLaMA, handles the flow of information across multiple layers.
Using graph theory, the team demonstrated that attention patterns are not evenly distributed. Instead, certain tokens become dominant simply due to their position. When the model reads from left to right, earlier tokens often accumulate more influence as the layers deepen, even when their content is less relevant. This effect intensifies as more layers are added, creating a cascade where initial tokens disproportionately shape the model's decisions.
The study shows that even without adding any formal position tracking, the structure of the model itself introduces a preference for the start of the sequence. In experiments with synthetic retrieval tasks, the performance of the models dipped when key information was placed in the middle of the input. The retrieval curve followed a U-shape, strong at the start, weaker in the center, then improving slightly at the end.
This behavior wasn’t incidental. Controlled tests confirmed that position bias emerged even when the training data had no such leanings. In setups where the data favored certain positions, the models amplified those biases. When models were trained on sequences biased toward the beginning and end, they mirrored that pattern, heavily underperforming in the center.
The paper also explored how positional encoding schemes, tools designed to help the model track where a word appears, can partially counteract this effect. Techniques like decay masks and rotary encodings introduce a fading influence based on distance, nudging the model to attend more evenly across the sequence. However, these methods alone don’t eliminate the bias, especially in deeper networks where earlier layers already tilt the attention forward.
In practical terms, this means that users relying on AI models for tasks like legal search, coding assistance, or medical records review may unknowingly encounter blind spots. If key content appears mid-document, the model might miss or misjudge it, even if everything else in the system functions as intended.
The implications go beyond diagnostics. By showing that position bias is both an architectural and data-driven phenomenon, the researchers offer pathways to mitigate it. Adjustments in attention masks, fewer layers, and smarter use of positional encodings can help rebalance the focus. The study also suggests that fine-tuning models on more uniformly distributed data could be essential in high-stakes domains where omission carries risk.
The research not only maps the bias but explains its evolution. As tokens move through the model, their contextual representations are repeatedly reshaped. Those that appear earlier begin to dominate, not because they contain better information, but because they become more deeply embedded in the model's reasoning. In this sense, the bias is baked into the system’s logic.
Rather than treating this as a bug, the team sees it as an opportunity for improvement. Their framework doesn’t just diagnose; it provides tools to reshape how models perceive position. By better understanding these internal biases, developers can build systems that reason more fairly and consistently across the full length of input, beginning, middle, and end.
Read next: Why a Wrench Might Outlast Code in the Age of AI
