Leveraging Self-Attention for Input-Dependent Soft Prompting in LLMs
Paper
•
2506.05629
•
Published
•
36
Abhilash Nandy
abhi1nandy2
AI & ML interests
Natural Language Processing, Computer Vision
Recent Activity
upvoted
a
paper
2 months ago
Leveraging Self-Attention for Input-Dependent Soft Prompting in LLMs
posted
an
update
2 months ago
A New Update!
Paper "Leveraging Self-Attention for Input-Dependent Soft Prompting in LLMs" accepted at ACL 2025 (Main) Conference
HuggingFace Paper Link - https://huggingface.co/papers/2506.05629
Authors - Ananth Muppidi, Abhilash Nandy, Sambaran Bandyopadhyay
Abstract -
The performance of large language models in domain-specific tasks necessitates fine-tuning, which is computationally expensive and technically challenging. This paper focuses on parameter-efficient fine-tuning using soft prompting, a promising approach that adapts pre-trained models to downstream tasks by learning a small set of parameters. We propose a novel Input Dependent Soft Prompting technique with a self-Attention Mechanism (ID-SPAM) that generates soft prompts based on the input tokens and attends different tokens with varying importance. Our method is simple and efficient, keeping the number of trainable parameters small. We show the merits of the proposed approach compared to state-of-the-art techniques on various tasks and show the improved zero shot domain transfer capability.
commented on
a paper
2 months ago
Leveraging Self-Attention for Input-Dependent Soft Prompting in LLMs
Organizations
None yet
upvoted
a
paper
2 months ago
posted
an
update
2 months ago
Post
290
A New Update!
Paper "Leveraging Self-Attention for Input-Dependent Soft Prompting in LLMs" accepted at ACL 2025 (Main) Conference
HuggingFace Paper Link - Leveraging Self-Attention for Input-Dependent Soft Prompting in LLMs (2506.05629)
Authors - Ananth Muppidi, Abhilash Nandy, Sambaran Bandyopadhyay
Abstract -
The performance of large language models in domain-specific tasks necessitates fine-tuning, which is computationally expensive and technically challenging. This paper focuses on parameter-efficient fine-tuning using soft prompting, a promising approach that adapts pre-trained models to downstream tasks by learning a small set of parameters. We propose a novel Input Dependent Soft Prompting technique with a self-Attention Mechanism (ID-SPAM) that generates soft prompts based on the input tokens and attends different tokens with varying importance. Our method is simple and efficient, keeping the number of trainable parameters small. We show the merits of the proposed approach compared to state-of-the-art techniques on various tasks and show the improved zero shot domain transfer capability.
Paper "Leveraging Self-Attention for Input-Dependent Soft Prompting in LLMs" accepted at ACL 2025 (Main) Conference
HuggingFace Paper Link - Leveraging Self-Attention for Input-Dependent Soft Prompting in LLMs (2506.05629)
Authors - Ananth Muppidi, Abhilash Nandy, Sambaran Bandyopadhyay
Abstract -
The performance of large language models in domain-specific tasks necessitates fine-tuning, which is computationally expensive and technically challenging. This paper focuses on parameter-efficient fine-tuning using soft prompting, a promising approach that adapts pre-trained models to downstream tasks by learning a small set of parameters. We propose a novel Input Dependent Soft Prompting technique with a self-Attention Mechanism (ID-SPAM) that generates soft prompts based on the input tokens and attends different tokens with varying importance. Our method is simple and efficient, keeping the number of trainable parameters small. We show the merits of the proposed approach compared to state-of-the-art techniques on various tasks and show the improved zero shot domain transfer capability.
commented
a
paper
2 months ago
commented
2
papers
3 months ago
REFINE-AF: A Task-Agnostic Framework to Align Language Models via Self-Generated Instructions using Reinforcement Learning from Automated Feedback
Paper
•
2505.06548
•
Published
•
30
•
6
REFINE-AF: A Task-Agnostic Framework to Align Language Models via Self-Generated Instructions using Reinforcement Learning from Automated Feedback
Paper
•
2505.06548
•
Published
•
30
•
6
authored
a
paper
3 months ago
posted
an
update
3 months ago
Post
543
PAPER - REFINE-AF: A Task-Agnostic Framework to Align Language Models via Self-Generated Instructions using Reinforcement Learning from Automated Feedback
Huggingface Paper Link - REFINE-AF: A Task-Agnostic Framework to Align Language Models via Self-Generated Instructions using Reinforcement Learning from Automated Feedback (2505.06548)
AUTHORS - Aniruddha Roy, Pretam Ray, Abhilash Nandy, Somak Aditya, Pawan Goyal
ABSTRACT -
Instruction-based Large Language Models (LLMs) have proven effective in numerous few-shot or zero-shot Natural Language Processing (NLP) tasks. However, creating human-annotated instruction data is time-consuming, expensive, and often limited in quantity and task diversity. Previous research endeavors have attempted to address this challenge by proposing frameworks capable of generating instructions in a semi-automated and task-agnostic manner directly from the model itself. Many of these efforts have relied on large API-only parameter-based models such as GPT-3.5 (175B), which are expensive, and subject to limits on a number of queries. This paper explores the performance of three open-source small LLMs such as LLaMA 2-7B, LLama 2-13B, and Mistral 7B, using a semi-automated framework, thereby reducing human intervention, effort, and cost required to generate an instruction dataset for fine-tuning LLMs. Furthermore, we demonstrate that incorporating a Reinforcement Learning (RL) based training algorithm into this LLMs-based framework leads to further enhancements. Our evaluation of the dataset reveals that these RL-based frameworks achieve a substantial improvements in 63-66% of the tasks compared to previous approaches.
Huggingface Paper Link - REFINE-AF: A Task-Agnostic Framework to Align Language Models via Self-Generated Instructions using Reinforcement Learning from Automated Feedback (2505.06548)
AUTHORS - Aniruddha Roy, Pretam Ray, Abhilash Nandy, Somak Aditya, Pawan Goyal
ABSTRACT -
Instruction-based Large Language Models (LLMs) have proven effective in numerous few-shot or zero-shot Natural Language Processing (NLP) tasks. However, creating human-annotated instruction data is time-consuming, expensive, and often limited in quantity and task diversity. Previous research endeavors have attempted to address this challenge by proposing frameworks capable of generating instructions in a semi-automated and task-agnostic manner directly from the model itself. Many of these efforts have relied on large API-only parameter-based models such as GPT-3.5 (175B), which are expensive, and subject to limits on a number of queries. This paper explores the performance of three open-source small LLMs such as LLaMA 2-7B, LLama 2-13B, and Mistral 7B, using a semi-automated framework, thereby reducing human intervention, effort, and cost required to generate an instruction dataset for fine-tuning LLMs. Furthermore, we demonstrate that incorporating a Reinforcement Learning (RL) based training algorithm into this LLMs-based framework leads to further enhancements. Our evaluation of the dataset reveals that these RL-based frameworks achieve a substantial improvements in 63-66% of the tasks compared to previous approaches.
commented
a
paper
3 months ago
posted
an
update
4 months ago
Post
1600
Little late to the party 🥳, but I’m thrilled to finally share our AAAI 2025–accepted work! Check out the project homepage here: https://midas-pro-mds.github.io/
🚀 What’s MiDAS-PRo all about?
We tackle the challenge of coherent, non-redundant multi-document summarization with source attribution via a three-stage LLM-based pipeline:
1. Plan a hierarchical document organization
2. Reason by generating entities/topics
3. Summarize the collection into a cohesive narrative
All “planning” and “reasoning” steps are framed as code-completion tasks, guided by graph attention network-based in-context example selection—boosting both automated and human evaluation scores!
đź”— Resources
- Paper: https://ojs.aaai.org/index.php/AAAI/article/view/34676
- Slides: https://drive.google.com/file/d/1lWqQtHRnpn-g2IQ3guloj_2j8sDm4z0V/view?usp=drivesdk
- Poster: https://drive.google.com/file/d/1EQqgwbcS7xkVx38y0qPvdRh0gQyeOH5u/view?usp=drivesdk
- Video: https://youtube.com/shorts/6ecxLLUpWJE?si=BAluAeP4-_eCmfu7
Big thanks to my mentor and co-author Sambaran Bandyopadhyay at Adobe Research for the guidance (Summer 2024 internship days FTW!) 🙏.
#AAAI2025 #MultiDocumentSummarization #LLM #Research #NLP
🚀 What’s MiDAS-PRo all about?
We tackle the challenge of coherent, non-redundant multi-document summarization with source attribution via a three-stage LLM-based pipeline:
1. Plan a hierarchical document organization
2. Reason by generating entities/topics
3. Summarize the collection into a cohesive narrative
All “planning” and “reasoning” steps are framed as code-completion tasks, guided by graph attention network-based in-context example selection—boosting both automated and human evaluation scores!
đź”— Resources
- Paper: https://ojs.aaai.org/index.php/AAAI/article/view/34676
- Slides: https://drive.google.com/file/d/1lWqQtHRnpn-g2IQ3guloj_2j8sDm4z0V/view?usp=drivesdk
- Poster: https://drive.google.com/file/d/1EQqgwbcS7xkVx38y0qPvdRh0gQyeOH5u/view?usp=drivesdk
- Video: https://youtube.com/shorts/6ecxLLUpWJE?si=BAluAeP4-_eCmfu7
Big thanks to my mentor and co-author Sambaran Bandyopadhyay at Adobe Research for the guidance (Summer 2024 internship days FTW!) 🙏.
#AAAI2025 #MultiDocumentSummarization #LLM #Research #NLP
- 1 reply
updated
a
Space
6 months ago
updated
a
dataset
9 months ago
authored
a
paper
10 months ago
commented
2
papers
10 months ago
A Pointer Network-based Approach for Joint Extraction and Detection of Multi-Label Multi-Class Intents
Paper
•
2410.22476
•
Published
•
29
•
3
YesBut: A High-Quality Annotated Multimodal Dataset for evaluating Satire Comprehension capability of Vision-Language Models
Paper
•
2409.13592
•
Published
•
52
•
9
updated
a
Space
11 months ago
commented
3
papers
11 months ago
YesBut: A High-Quality Annotated Multimodal Dataset for evaluating Satire Comprehension capability of Vision-Language Models
Paper
•
2409.13592
•
Published
•
52
•
9
YesBut: A High-Quality Annotated Multimodal Dataset for evaluating Satire Comprehension capability of Vision-Language Models
Paper
•
2409.13592
•
Published
•
52
•
9
YesBut: A High-Quality Annotated Multimodal Dataset for evaluating Satire Comprehension capability of Vision-Language Models
Paper
•
2409.13592
•
Published
•
52
•
9
authored
a
paper
11 months ago
commented
a
paper
11 months ago