Mitigate Position Bias in Large Language Models via Scaling a Single Dimension
- URL: http://arxiv.org/abs/2406.02536v1
- Date: Tue, 4 Jun 2024 17:55:38 GMT
- Title: Mitigate Position Bias in Large Language Models via Scaling a Single Dimension
- Authors: Yijiong Yu, Huiqiang Jiang, Xufang Luo, Qianhui Wu, Chin-Yew Lin, Dongsheng Li, Yuqing Yang, Yongfeng Huang, Lili Qiu,
- Abstract summary: This paper first explores the micro-level manifestations of position bias, concluding that attention weights are a micro-level expression of position bias.
It further identifies that, in addition to position embeddings, causal attention mask also contributes to position bias by creating position-specific hidden states.
Based on these insights, we propose a method to mitigate position bias by scaling this positional hidden states.
- Score: 47.792435921037274
- License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
- Abstract: Large Language Models (LLMs) are increasingly applied in various real-world scenarios due to their excellent generalization capabilities and robust generative abilities. However, they exhibit position bias, also known as "lost in the middle", a phenomenon that is especially pronounced in long-context scenarios, which indicates the placement of the key information in different positions of a prompt can significantly affect accuracy. This paper first explores the micro-level manifestations of position bias, concluding that attention weights are a micro-level expression of position bias. It further identifies that, in addition to position embeddings, causal attention mask also contributes to position bias by creating position-specific hidden states. Based on these insights, we propose a method to mitigate position bias by scaling this positional hidden states. Experiments on the NaturalQuestions Multi-document QA, KV retrieval, LongBench and timeline reorder tasks, using various models including RoPE models, context windowextended models, and Alibi models, demonstrate the effectiveness and generalizability of our approach. Our method can improve performance by up to 15.2% by modifying just one dimension of hidden states. Our code is available at https://aka.ms/PositionalHidden.
Related papers
- Eliminating Position Bias of Language Models: A Mechanistic Approach [119.34143323054143]
Position bias has proven to be a prevalent issue of modern language models (LMs)
We find that causal attention generally causes models to favor distant content, while relative positional encodings like RoPE prefer nearby ones.
We propose to ELIMINATE position bias caused by different input segment orders (e.g., options in LM-as-a-judge, retrieved documents in QA) in a TRAINING-FREE ZERO-SHOT manner.
arXiv Detail & Related papers (2024-07-01T09:06:57Z) - Discovering Bias in Latent Space: An Unsupervised Debiasing Approach [6.67121343477106]
Question-answering capabilities of foundation models are highly sensitive to prompt variations.
We propose to rectify this bias directly in the model's internal representation.
We empirically show that SteerFair significantly reduces instruction-tuned model performance variance across prompt modifications.
arXiv Detail & Related papers (2024-06-05T21:29:09Z) - Position-Aware Parameter Efficient Fine-Tuning Approach for Reducing Positional Bias in LLMs [18.832135309689736]
Recent advances in large language models (LLMs) have enhanced their ability to process long input contexts.
Recent studies show a positional bias in LLMs, demonstrating varying performance depending on the location of useful information.
We develop a Position-Aware PAPEFT approach which is composed of a data augmentation technique and an efficient parameter adapter.
arXiv Detail & Related papers (2024-04-01T19:04:17Z) - Position bias in features [0.0]
Document-specific historical click-through rates can be important features in a dynamic ranking system.
This paper describes the properties of several such features, and tests them in controlled experiments.
arXiv Detail & Related papers (2024-02-04T22:15:30Z) - A Frustratingly Easy Improvement for Position Embeddings via Random
Padding [68.75670223005716]
In this paper, we propose a simple but effective strategy, Random Padding, without any modifications to existing pre-trained language models.
Experiments show that Random Padding can significantly improve model performance on the instances whose answers are located at rear positions.
arXiv Detail & Related papers (2023-05-08T17:08:14Z) - The Curious Case of Absolute Position Embeddings [65.13827063579728]
Transformer language models encode the notion of word order using positional information.
In natural language, it is not absolute position that matters, but relative position, and the extent to which APEs can capture this type of information has not been investigated.
We observe that models trained with APE over-rely on positional information to the point that they break-down when subjected to sentences with shifted position information.
arXiv Detail & Related papers (2022-10-23T00:00:04Z) - Mitigating Representation Bias in Action Recognition: Algorithms and
Benchmarks [76.35271072704384]
Deep learning models perform poorly when applied to videos with rare scenes or objects.
We tackle this problem from two different angles: algorithm and dataset.
We show that the debiased representation can generalize better when transferred to other datasets and tasks.
arXiv Detail & Related papers (2022-09-20T00:30:35Z) - Unveiling The Mask of Position-Information Pattern Through the Mist of
Image Features [75.62755703738696]
Recent studies show that paddings in convolutional neural networks encode absolute position information.
Existing metrics for quantifying the strength of positional information remain unreliable.
We propose novel metrics for measuring (and visualizing) the encoded positional information.
arXiv Detail & Related papers (2022-06-02T17:59:57Z)
This list is automatically generated from the titles and abstracts of the papers in this site.
This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.