Compressing Many-Shots in In-Context Learning
Fuente:
arXiv
Salvato in:
| Autori principali: | , , , , , , , , , , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2025
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866917023816089600 |
|---|---|
| author | Khatri, Devvrit Kulkarni, Pranamya Gupta, Nilesh Varun, Yerram Peng, Liqian Yagnik, Jay Netrapalli, Praneeth Hsieh, Cho-Jui Go, Alec Dhillon, Inderjit S Kusupati, Aditya Jain, Prateek |
| author_facet | Khatri, Devvrit Kulkarni, Pranamya Gupta, Nilesh Varun, Yerram Peng, Liqian Yagnik, Jay Netrapalli, Praneeth Hsieh, Cho-Jui Go, Alec Dhillon, Inderjit S Kusupati, Aditya Jain, Prateek |
| contents | Large Language Models (LLMs) have been shown to be able to learn different tasks without explicit finetuning when given many input-output examples / demonstrations through In-Context Learning (ICL). Increasing the number of examples, called ``shots'', improves downstream task performance but incurs higher memory and computational costs. In this work, we study an approach to improve the memory and computational efficiency of ICL inference by compressing the many-shot prompts. Given many shots comprising t tokens, our goal is to generate a m soft-token summary, where m < t. We first show that existing prompt compression methods are ineffective for many-shot compression, and simply using fewer shots as a baseline is surprisingly strong. To achieve effective compression, we find that: (a) a stronger compressor model with more trainable parameters is necessary, and (b) compressing many-shot representations at each transformer layer enables more fine-grained compression by providing each layer with its own compressed representation. Based on these insights, we propose MemCom, a layer-wise compression method. We systematically evaluate various compressor models and training approaches across different model sizes (2B and 7B), architectures (Gemma and Mistral), many-shot sequence lengths (3k-6k tokens), and compression ratios (3x to 8x). MemCom outperforms strong baselines across all compression ratios on multiple classification tasks with large label sets. Notably, while baseline performance degrades sharply at higher compression ratios, often by over 20-30%, MemCom maintains high accuracy with minimal degradation, typically dropping by less than 10%. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2510_16092 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Compressing Many-Shots in In-Context Learning Khatri, Devvrit Kulkarni, Pranamya Gupta, Nilesh Varun, Yerram Peng, Liqian Yagnik, Jay Netrapalli, Praneeth Hsieh, Cho-Jui Go, Alec Dhillon, Inderjit S Kusupati, Aditya Jain, Prateek Machine Learning Artificial Intelligence Large Language Models (LLMs) have been shown to be able to learn different tasks without explicit finetuning when given many input-output examples / demonstrations through In-Context Learning (ICL). Increasing the number of examples, called ``shots'', improves downstream task performance but incurs higher memory and computational costs. In this work, we study an approach to improve the memory and computational efficiency of ICL inference by compressing the many-shot prompts. Given many shots comprising t tokens, our goal is to generate a m soft-token summary, where m < t. We first show that existing prompt compression methods are ineffective for many-shot compression, and simply using fewer shots as a baseline is surprisingly strong. To achieve effective compression, we find that: (a) a stronger compressor model with more trainable parameters is necessary, and (b) compressing many-shot representations at each transformer layer enables more fine-grained compression by providing each layer with its own compressed representation. Based on these insights, we propose MemCom, a layer-wise compression method. We systematically evaluate various compressor models and training approaches across different model sizes (2B and 7B), architectures (Gemma and Mistral), many-shot sequence lengths (3k-6k tokens), and compression ratios (3x to 8x). MemCom outperforms strong baselines across all compression ratios on multiple classification tasks with large label sets. Notably, while baseline performance degrades sharply at higher compression ratios, often by over 20-30%, MemCom maintains high accuracy with minimal degradation, typically dropping by less than 10%. |
| title | Compressing Many-Shots in In-Context Learning |
| topic | Machine Learning Artificial Intelligence |
| url | https://arxiv.org/abs/2510.16092 |