Lightnews — Scholar-powered news

Ishika Agarwal @wonderingishika.bsky.social · Feb 17

6/6 For more details, see:

Paper: arxiv.org/pdf/2502.09969
Code: github.com/agarwalishik...

Thank you so much to @dilekh.bsky.social and @convai-uiuc.bsky.social for their guidance and support during this project 🎉🎉

arxiv.org

4

Ishika Agarwal @wonderingishika.bsky.social · Feb 17

5/6 Finally, using our influence values, we pick a small subset & fine-tune the model. In our evaluation, we use 4 SOTA influence functions -- NN-CIFT achieves the same performance while using a model 34,000x smaller!

1 2

Ishika Agarwal @wonderingishika.bsky.social · Feb 17

4/6 Second, we train the InfluenceNetwork using basic mini-batch gradient descent, then let it estimate the influence for the remaining data. It has a very low error of 0.067!

1 2

Ishika Agarwal @wonderingishika.bsky.social · Feb 17

3/6 First, the neural network (called the “InfluenceNetwork”) needs to be trained. We compute influence values using existing methods -- but only for a tiny fraction of data (just 0.25%-5%).

1

Ishika Agarwal @wonderingishika.bsky.social · Feb 17

2/6 Estimating the value of data is expensive.

Past works use LLMs to estimate the influence of data -- we use small neural networks to *learn to estimate* influence, instead. This reduces costs and adapts to new data without heavy recomputation.

Here’s how it works:

1

Ishika Agarwal @wonderingishika.bsky.social · Feb 17

🚀Very excited about my new paper!

NN-CIFT slashes data valuation costs by 99% using tiny neural nets (205k params, just 0.0027% of 8B LLMs) while maintaining top-tier performance!

1 4 11

Ishika Agarwal @wonderingishika.bsky.social · Jan 23

Elated to announce that DELIFT has been accepted to ICLR'25 🎉 Looking forward to discussing it in Singapore!

Ishika Agarwal @wonderingishika.bsky.social · Nov 17

I'm so excited to share my latest paper called DELIFT along with Krishnateja Killamsetty, Lucian Popa, and Marina Danilevksy at IBM Research 🎉

We tackle expensive fine-tuning by selecting a small subset of informative data that targets a model's weaknesses.

3

Reposted by Ishika Agarwal

ConvAI @ UIUC @convai-uiuc.bsky.social · Dec 11

Congratulations to @dilekh.bsky.social for her ACL Fellowship! 🎉🎉🎉 www.aclweb.org/portal/conte...

ACL Fellows 2024 | ACL Member Portal

www.aclweb.org

2 11

Reposted by Ishika Agarwal

Manoj Agarwal @agarwalmk.bsky.social · Nov 24

The last response from Gemini in this thread may shock you: gemini.google.com/share/6d141b...

‎Gemini - Challenges and Solutions for Aging Adults

Created with Gemini

gemini.google.com

1 6

Ishika Agarwal @wonderingishika.bsky.social · Nov 24

Thank you Guneet! Would love to hear more about these stress tests :)

2

Ishika Agarwal @wonderingishika.bsky.social · Nov 24

👋

2

Ishika Agarwal @wonderingishika.bsky.social · Nov 20

Hey! Would love to be added :)

1

Reposted by Ishika Agarwal

Priyanka Kargupta @ EMNLP'24 @pkargupta.bsky.social · Nov 19

Can LLMs make us critical thinkers?

TreeInstruct reorients assistant-like LLMs to be instructors that guide students towards understanding their mistakes, without providing direct/indirect answers.

Check out aclanthology.org/2024.finding... (w/ @wonderingishika.bsky.social) to learn more!

1 1 2

Ishika Agarwal @wonderingishika.bsky.social · Nov 17

All around the theme of data-efficient NLP:

(1) using influence functions to improve language model performance from less data
(2) enabling language models to generate queries for things it doesn't know

lastpositivist.bsky.social @lastpositivist.bsky.social · Nov 17

Bluesky academics, lets get to know each other! Quote this & tell me: 1) a project you are working on & 2) an odd idea/theory you aren’t working on but keep thinking about

1. I came to hate my work and thinking so don't do it anymore.
2.

Embrace the Void @etvpod.bsky.social · Nov 17

Bluesky academics, lets get to know each other! Quote this & tell me: 1) a project you are working on & 2) an odd idea/theory you aren’t working on but keep thinking about

1. Convincing everyone that everything is luck, all the way down.

2. LLM’s can reason and understand in the external sense.

3

Ishika Agarwal @wonderingishika.bsky.social · Nov 17

For more details, see:
Paper: arxiv.org/pdf/2411.04425
Code: github.com/agarwalishik...

Thank you so much to Krishnateja, Lucian, and Marina for their help, mentorship, and guidance during this project! 🎉🎉

arxiv.org

Ishika Agarwal @wonderingishika.bsky.social · Nov 17

3. Continual fine-tuning: given a fine-tuned model, enabling it to integrate new and complementary information while mitigating catastrophic forgetting. We find that reducing the dataset helps remove samples that hinder performance, surpassing the performance of the full dataset.

1 1

Ishika Agarwal @wonderingishika.bsky.social · Nov 17

2. Task-specific fine-tuning: given an instruction-tuned model, refining the LLM's expertise in specific domains. We find that pruning the dataset removes noise and keeps relevant examples, achieving better performance than fine-tuning on the full dataset.

1 1

Ishika Agarwal @wonderingishika.bsky.social · Nov 17

1. Instruction tuning: given a base model, fine-tuning a model to follow general instructions. We find that performance drops are minimal when reducing the dataset by 70%.

1

Ishika Agarwal @wonderingishika.bsky.social · Nov 17

DELIFT quantifies the information present in a sample wrt an LLM's capabilities. Using submodular functions, DELIFT can automatically adapt the chosen subset based on the objectives in the 3 stages of language model fine-tuning:

1

Ishika Agarwal @wonderingishika.bsky.social · Nov 17

I'm so excited to share my latest paper called DELIFT along with Krishnateja Killamsetty, Lucian Popa, and Marina Danilevksy at IBM Research 🎉

We tackle expensive fine-tuning by selecting a small subset of informative data that targets a model's weaknesses.

2 1 8

Ishika Agarwal @wonderingishika.bsky.social · Nov 17

TreeInstruct is preferred 78.43% of the time. It solves 14.09% more bugs across all settings, and our questions are 14.18% better at addressing bugs, maintaining relevance, and ensuring logical conversation flow. TreeInstruct also adapts to human students of varying backgrounds.

Ishika Agarwal @wonderingishika.bsky.social · Nov 17

TreeInstruct estimates the knowledge a student needs to debug their code and devises a conversation plan. It then dynamically constructs a question tree based on its interactions with the student, navigating the knowledge state space till the student comprehends & fixes all bugs.

1

Ishika Agarwal @wonderingishika.bsky.social · Nov 17

github.com/agarwalishik...
We apply TreeInstruct to code debugging. Prior works directly give away bugs/fixes, assume single-turn conversations, and only work for one bug. We create a realistic, multi-bug dataset, where the bugs are mutually dependent.

GitHub - agarwalishika/TreeInstruct: TreeInstruct is a novel method that uses state space estimation and dynamic tree-based questioning for multi-turn Socratic instruction, applied to code debugging.

TreeInstruct is a novel method that uses state space estimation and dynamic tree-based questioning for multi-turn Socratic instruction, applied to code debugging. - agarwalishika/TreeInstruct

github.com

1

Ishika Agarwal @wonderingishika.bsky.social · Nov 17

Can LLMs make us critical thinkers?

TreeInstruct reorients LLMs to be instructors that guide students socratically to solve problems, instead of assistants that provide direct answers.

Check out our EMNLP2024 paper at arxiv.org/abs/2406.11709 (w/ @pkargupta.bsky.social) to learn more!

Instruct, Not Assist: LLM-based Multi-Turn Planning and Hierarchical Questioning for Socratic Code Debugging

Socratic questioning is an effective teaching strategy, encouraging critical thinking and problem-solving. The conversational capabilities of large language models (LLMs) show great potential for prov...

arxiv.org

1 2

Ishika Agarwal @wonderingishika.bsky.social · Nov 17

I'd love to be added - thank you!!

1