Zhichao Xu Brutus
@zhichaoxu_ir
Followers
203
Following
2K
Statuses
305
CS PhD @UtahNLP. Prev Intern @Visa @GoogleAI @Dataminr. Opinions my own.
Joined April 2020
Are compressed LLMs less toxic and biased against different demographic groups❓In this new📜, we study 4 pruning methods and 3 quantization methods and evaluate on 7 bias/toxicity benchmarks. (Un)surprising answer is: they are not less toxic/biased
1
4
11
In this insightful work led by @cychang9 , we propose a method to reduce the noise from retrieval. Strong performance boost to naive RAG pipeline💙
MAIN-RAG: Multi-Agent Filtering Retrieval-Augmented Generation Proposes a training-free RAG framework using multiple LLM agents to collaboratively filter retrieved documents, improving retrieval precision while maintaining high recall. 📝
0
1
5
Excited by the recent development of state space models and linear attention models, we conducted a more rigorous evaluation with passage reranking and long document reranking.
State Space Models are Strong Text Rerankers Shows Mamba-based models achieve comparable reranking performance to transformers while being more memory efficient, with Mamba-2 outperforming Mamba-1. 📝
0
1
10
Are state space models (e.g., Mamba) competitive in text ranking❓ Performance-wise, yes. Efficiency-wise, they are better than transformers w/ multi-head attention, but lag behind transformers w/ IO-aware optimization (e.g., flash attention).
0
1
14
Hongyi is one of the smartest MLSys researchers I know,please consider applying!
I have three Ph.D. student openings in my research group at @RutgersCS starting in Fall 2025. If you are interested in working with me on efficient algorithms and systems for LLMs, foundation models, and AI4Science, please apply at: The deadline is January 1, 2025. Please be sure to indicate in your application that you are interested in working with me (this is very important! And no need to email me directly). I look forward to your applications!
0
0
1
@spacemanidol @SnowflakeDB hi! great work! can you elaborate more about the rationales of choosing GTE and BGE as base models?
0
0
0
@srchvrs @neuralmagic matches our finding that 4 bit quantization does not experience much performance degradation and bias/toxicity score
1
0
1
1
0
1
RT @gowthami_s: A piece of unsolicited advice. I see a lot of internship calls these days from many companies on X. Usually one chooses an…
0
18
0
so it concludes my trip to EMNLP where I presented two posters and gave a short talk invited by @JinaAI_ about using Mamba model for ranking until next time
0
1
12
paper link 🗓️happening this afternoon poster session 16:00-17:30 at Riverfront look forward to talking about LLM-as-judge and dialogue analysis.
Multi-dimensional Evaluation of Empathetic Dialogue Responses from my summer @GoogleAI , we propose a finegrained taxonomy of empathetic dialogue responses, and followup with exploring automatic evaluation metrics.
0
1
1
paper link 🗓️Poster session 03 (Tue 14:00 - 15:30), look forward to talking about bias, fairness and model compression😃
Beyond Perplexity: Multi-dimensional Safety Evaluation of LLM Compression We pinpoint a critical problem in model compression literature - a lack of multi-faceted safety evaluation. Some of the findings are surprising and exciting.
0
1
3
paper link I'll be at poster session 09 (Wed 16:00 - 17:30), look forward to talking about dialogues, LLM-as-judge and other fun topics
Multi-dimensional Evaluation of Empathetic Dialogue Responses from my summer @GoogleAI , we propose a finegrained taxonomy of empathetic dialogue responses, and followup with exploring automatic evaluation metrics.
0
2
4