liruizhe94 Profile Banner
Ruizhe Li Profile
Ruizhe Li

@liruizhe94

Followers
994
Following
3K
Media
38
Statuses
227

Lecturer (Assistant Professor) @ABDNCompSci | Postdoc research fellow @ucl_wi_group | PhD CS @SheffieldNLP | mechanistic interpretability, multimodal LLMs

Aberdeen, Scotland
Joined February 2014
Don't wanna be here? Send us removal request.
@liruizhe94
Ruizhe Li
2 months
Very honoured to receive 🏆 Best Paper Award from @COLM_conf XLLM-Reason-Plan Workshop! We sincerely appreciate this valuable recognition from the organising committee and reviewers. We also appreciate the computational grant support from @Google for this work!
6
3
70
@liruizhe94
Ruizhe Li
2 months
We are very honored to receive the Best Paper Award! Reviewers’s comments are very insightful! We really appreciate such valuable recognition from the workshop committee!
@XllmReasonPlan
XLLM-Reason-Plan
2 months
Best Paper Award goes to Li et al., "Attributing Response to Context: A Jensen-Shannon Divergence Driven Mechanistic Study of Context Attribution in Retrieval-Augmented Generation" @liruizhe94
0
1
16
@liruizhe94
Ruizhe Li
2 months
@COLM_conf @Google @YuchenHu98 @Serena_pancakes @wangxieric Check the latest version of our work:
0
1
0
@liruizhe94
Ruizhe Li
2 months
@COLM_conf @Google I also really appreciate the huge help from my collaborators Chen Chen, @YuchenHu98, @Serena_pancakes, @wangxieric and Prof. Emine Yilmaz for this work!
1
0
1
@liruizhe94
Ruizhe Li
2 months
Based on super helpful and inspiring comments from those workshops, we further improved our work by adding more mechanistic experiments and discussion. Please refer to our latest version: https://t.co/TihL1Q8Sv1. Code:
Tweet card summary image
github.com
A Jensen-Shannon Divergence Driven Mechanistic Study of Context Attribution in Retrieval-Augmented Generation - ruizheliUOA/ARC_JSD
0
0
1
@liruizhe94
Ruizhe Li
2 months
Our ARC-JSD work was accepted at @NeurIPSConf Mechanistic Interpretability Workshop 2025. Our work also has opportunities to present at @COLM_conf 2025 Interplay and XLLM-Reason-Plan workshops. We really appreciate the computational resources funding support from @Google.
@liruizhe94
Ruizhe Li
6 months
🤔Is it possible to accurately and effectively attribute RAG response to relevant context without finetuning or further training surrogate model? 💡We propose an inference-time method called ARC-JSD using JSD for RAG context attribution, which only needs O(sent_num + 1)🚀
1
0
7
@liruizhe94
Ruizhe Li
3 months
Important Dates: Submission deadline: October 7, 2025 Notification: October 20, 2025 Camera-ready: October 25, 2025 Our workshop is supported by several HEIs from UK, US and China. Please feel free to distribute this call among your networks and interested colleagues.
0
0
1
@liruizhe94
Ruizhe Li
3 months
Our workshop invites contributions exploring practical and innovative solutions for real-world AI deployment, monitoring, and continuous improvement, including AI Agents, Multimodal LLMs, RAGs, HCI, etc. Details of AI4RWC workshop:
sites.google.com
Overview
1
0
1
@liruizhe94
Ruizhe Li
3 months
We'll be hosting AI4RWC: The 1st International Workshop on Artificial Intelligence for Real‑world Challenges, co‑located with the 24th International Conference on Web Intelligence and Intelligent Agent Technology (WI‑IAT 2025) , to be held in London, UK, from 15–18 November 2025
1
0
4
@liruizhe94
Ruizhe Li
4 months
Check out our multilingual instruction following datasets which is covering 30 languages across 6 language families with rich cultural information. Our datasets and paper are online already. Welcome to use!
@Chenyang_Lyu
Chenyang Lyu 吕晨阳
4 months
🚀 Our Marco-Bench-MIF dataset is finally online! Marco-Bench-MIF is a deeply localized multilingual benchmark for evaluating instruction-following in LLMs—now covering 30 languages across 6 language families, including many low-resource cases. 🔥 Unlike previous multilingual
0
0
10
@liruizhe94
Ruizhe Li
4 months
In-context error learning?
1
0
5
@liruizhe94
Ruizhe Li
4 months
I’ll present our mechainterp work at Hall X5 No. 132 board. Welcome to discuss research about mechainterp!
0
0
16
@liruizhe94
Ruizhe Li
4 months
Time to go to Vienna again! I’ll present one mechinterp work on 28th 17:00-18:30 Hall X4 X5 We have another work for multilingual instruction-following benchmark on 28th 14:00 at 1.15-16. Very honored to be involved in this oral work! Feel free to reach out & chat for mechinterp
1
0
9
@liruizhe94
Ruizhe Li
6 months
I’m honored to be part of our NLP research group and to have contributed multiple works in mechinterp, multimodal LLMs, AI in education, and diagnostic prediction in 2024. Our work spans a wide range of interdisciplinary topics. Check out the blog for more details.
@EhudReiter
Ehud Reiter
6 months
New blog: The Aberdeen NLP Research Group Learn about our NLP research group! https://t.co/uBDesr2AKN
0
0
9
@liruizhe94
Ruizhe Li
6 months
This work was collaborated with Chen Chen, @YuchenHu98, @Serena_pancakes, @wangxieric and Prof. Emine Yilmaz. Our paper: https://t.co/ZYH6mRBCm7 Our code:
Tweet card summary image
huggingface.co
0
0
6
@liruizhe94
Ruizhe Li
6 months
This finding confirms the contribution of MLP located using ARC-JSD above, and it is reasonable because Chinese is one of main language resources used in Qwen2 pre- and post-training.
1
0
1
@liruizhe94
Ruizhe Li
6 months
In our case study for located MLP layers in Qwen2 models, we identify several correct decoded tokens are gradually transferred from their Chinese format to the English version, such as 一只(A), 拥有(has) and 翅膀(wings) in the figure.
1
0
3
@liruizhe94
Ruizhe Li
6 months
In addition, we move forward to locate relevant attention heads and MLP layers using JSD from mechinterp view. We found that JSD-based mechinterp can identify context attribution-related attention heads and MLPs, which are mainly distributed around intermedium or higher layers.
1
0
3
@liruizhe94
Ruizhe Li
6 months
We evaluate our ARC-JSD on TyDi QA, Hotpot QA and MuSiQue datasets using Qwen2-1.5B/7B-IT and Gemma2-2B/9B-IT, which can achieve higher attribution acc than baseline.
1
0
3
@liruizhe94
Ruizhe Li
6 months
🤔Is it possible to accurately and effectively attribute RAG response to relevant context without finetuning or further training surrogate model? 💡We propose an inference-time method called ARC-JSD using JSD for RAG context attribution, which only needs O(sent_num + 1)🚀
1
3
10