Fetching the paper…

Reading the bibliography…

2023

Attention Sorting Combats Recency Bias In Long Context Language Models

Peysakhovich, Alexander, Lerer, Adam

Original
PDF
Post on X

Understand

Current language models often fail to incorporate long contexts efficiently during generation.

We show that a major contributor to this issue are attention priors that are likely learned during pre-training: relevant information located earlier in context is attended to less on average.
Yet even when models fail to use the information from a relevant document in their response, they still pay preferential attention to that document compared to an irrelevant document at the same position.
We leverage this fact to introduce ``attention sorting'': perform one step of decoding, sort documents by the attention they receive (highest attention going last), repeat the process, generate the answer with the newly sorted context.

Attention Sorting Combats Recency Bias In Long Context Language Models · Around

Built on

Neural machine translation by jointly learning to align and translate
Original
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio · 2014
Earlier work this paper cites.
Show, attend and tell: Neural image caption generation with visual attention
Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich Zemel, and Yoshua Bengio · 2015
Earlier work this paper cites.
Analysis methods in neural language processing: A survey
Yonatan Belinkov and James Glass · 2019
Earlier work this paper cites.
Transformer-xl: Attentive language models beyond a fixed-length context
Original
Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov · 2019
Earlier work this paper cites.
The curious case of neural text degeneration
Original
Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi · 2019
Earlier work this paper cites.
Is attention interpretable?
Original
Sofia Serrano and Noah A Smith · 2019
Earlier work this paper cites.
Attention interpretability across nlp tasks
Original
Shikhar Vashishth, Shyam Upadhyay, Gaurav Singh Tomar, and Manaal Faruqui · 2019
Earlier work this paper cites.
The elephant in the interpretability room: Why use attention as explanation when we have saliency methods?
Original
Jasmijn Bastings and Katja Filippova · 2020
Earlier work this paper cites.
Longformer: The long-document transformer
Original
Iz Beltagy, Matthew E Peters, and Arman Cohan · 2020
Earlier work this paper cites.
End-to-end object detection with transformers
Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko · 2020
Earlier work this paper cites.
Retrieval augmented language model pre-training
Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Mingwei Chang · 2020
Earlier work this paper cites.
Distilling knowledge from reader to retriever for question answering
Original
Gautier Izacard and Edouard Grave · 2020
Earlier work this paper cites.

Similar

Dense passage retrieval for open-domain question answering
Original
Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih · 2020
Cited alongside, same era.
Retrieval-augmented generation for knowledge-intensive nlp tasks
Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al · 2020
Cited alongside, same era.
Transformer protein language models are unsupervised structure learners
Roshan Rao, Joshua Meier, Tom Sercu, Sergey Ovchinnikov, and Alexander Rives · 2020
Cited alongside, same era.
Long range arena: A benchmark for efficient transformers
Original
Yi Tay, Mostafa Dehghani, Samira Abnar, Yikang Shen, Dara Bahri, Philip Pham, Jinfeng Rao, Liu Yang, Sebastian Ruder, and Donald Metzler · 2020
Cited alongside, same era.
Show your work: Scratchpads for intermediate computation with language models

Then

Extending context window of large language models via positional interpolation
Original
Shouyuan Chen, Sherman Wong, Liangjian Chen, and Yuandong Tian · 2023
Closest in time.
Lost in the middle: How language models use long contexts
Original
Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang · 2023
Closest in time.
Wizardcoder: Empowering code large language models with evol-instruct
Original
Ziyang Luo, Can Xu, Pu Zhao, Qingfeng Sun, Xiubo Geng, Wenxiang Hu, Chongyang Tao, Jing Ma, Qingwei Lin, and Daxin Jiang · 2023
Closest in time.
the-little-retrieval-test, 2023
Dimitris Papailiopoulos, Lee Kangwook, and Jy-yong Sohn · 2023
Closest in time.
Yarn: Efficient context window extension of large language models
Original

Beyond the bibliography

alphaXiv searches the wider corpus for related work and actual follow-ups.

Open on alphaXiv

alphaXiv is searching for related work…

Maxwell Nye, Anders Johan Andreassen, Guy Gur-Ari, Henryk Michalewski, Jacob Austin, David Bieber, David Dohan, Aitor Lewkowycz, Maarten Bosma, David Luan, et al · 2021

Cited alongside, same era.

Train short, test long: Attention with linear biases enables input length extrapolation

Ofir Press, Noah A Smith, and Mike Lewis · 2021

Cited alongside, same era.

Roformer: Enhanced transformer with rotary position embedding

Original

Jianlin Su, Yu Lu, Shengfeng Pan, Ahmed Murtadha, Bo Wen, and Yunfeng Liu · 2021

Cited alongside, same era.

Do long-range language models actually use long-range context?

Simeng Sun, Kalpesh Krishna, Andrew Mattarella-Micke, and Mohit Iyyer · 2021

Cited alongside, same era.

Flashattention: Fast and memory-efficient exact attention with io-awareness

Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré · 2022

Cited alongside, same era.

Chain-of-thought prompting elicits reasoning in large language models

Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al · 2022

Cited alongside, same era.

Memorizing transformers

Yuhuai Wu, Markus N Rabe, DeLesley Hutchins, and Christian Szegedy · 2022

Cited alongside, same era.

Prompt engineering for claude’s long context window, 2023

Anthropic · 2023

Cited alongside, same era.

Bowen Peng, Jeffrey Quesnelle, Honglu Fan, and Enrico Shippole · 2023

Closest in time.

In-context retrieval-augmented language models

Original

Ori Ram, Yoav Levine, Itay Dalmedigos, Dor Muhlgay, Amnon Shashua, Kevin Leyton-Brown, and Yoav Shoham · 2023

Closest in time.

Long-range language modeling with self-retrieval

Ohad Rubin and Jonathan Berant · 2023

Closest in time.

Toolformer: Language models can teach themselves to use tools

Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom · 2023

Closest in time.

Replug: Retrieval-augmented black-box language models

Weijia Shi, Sewon Min, Michihiro Yasunaga, Minjoon Seo, Rich James, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih · 2023

Closest in time.

Llama 2: Open foundation and fine-tuned chat models

Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al · 2023

Closest in time.

Larger language models do in-context learning differently

Jerry Wei, Jason Wei, Yi Tay, Dustin Tran, Albert Webson, Yifeng Lu, Xinyun Chen, Hanxiao Liu, Da Huang, Denny Zhou, et al · 2023

Closest in time.