
Postdoc & Former PhD @IELabGroup UQ 🇦🇺. BSc. @UMNComputerSci 🇺🇸 alumni. Information Retrieval, LLM.
Brisbane, Queensland
Joined October 2019
- Tweets336
- Following897
- Followers349
- Likes1.5K
Hang Li retweeted
To make an LLM a strong dense retriever, you often need training…
But what if you only have a few-examples 🥩🥦 and no training budget?
Just add RICE 🍚, a simple way to turn your LLM into an effective dense retriever that requires zero training 👨🍳
📄 arxiv.org/abs/2609.38099
📢 Our survey on late-interaction neural retrieval is now available as a preprint!
This survey provides a systematic overview of the field, covering scoring, efficient retrieval, representation compression, sparse/hybrid retrieval, query enhancement, multimodal retrieval, tooling, and emerging research directions. Works with Xiao Wang, Chuting Yu, Minghan Li, Binci Yang, and Ben He.
📄 Paper: ssrn.com/abstract=7412000
💻 Resources & updated reading list: github.com/Xiao0728/late-int…
Feedback and suggestions are very welcome!
#InformationRetrieval #NeuralIR #LateInteraction #MultiVectorRetrieval #ColBERT
I still find it astonishing that some labs evolve into not respecting these and fall behind
Why can Kimi ship K3? Let me tell my story.
Earlier this year, I left academia for industry. I talked to a lot of companies along the way. Here's what I saw:
1⃣Arrogance. They believe the AI war is over, and they won. No hunger for the future, and no hunger for talent.
2⃣Restlessness. Young labs short on foundation, either rushing to catch the frontier or pivoting away from the competition.
3⃣Fear. Strong teams with real experience, but from the second tier, they can't quite bring themselves to aim for #1.
4⃣Misalignment. Everyone is optimizing for their own credit, but nobody really cares whether the company can reach AGI.
Kimi was different.
Over many conversations with the founders, the same thing came through every time: a raw, genuine hunger for AGI.
I joined. The hunger was real.
We shipped K3. This is only the beginning.
Hang Li retweeted
I'm looking for a #PhD student who is interested in scalable #algorithms and data structures. Please share on if you know anyone who may be interested! study.uq.edu.au/study-option…
#rust #systems #algorithms #search #bioinformatics #UQ #Brisbane #toomanyhashtags
Okay so, we just found that over 50 papers published at @Neurips 2025 have AI hallucinations
I don't think people realize how bad the slop is right now
It's not just that researchers from @GoogleDeepMind, @Meta, @MIT, @Cambridge_Uni are using AI - they allowed LLMs to generate hallucinations in their papers and didn't notice at all.
It's insane that these made it through peer review👇
Hang Li retweeted
🎄 PyTerrier Advent 10/25: Dense retrieval often improves with pseudo-relevance feedback (Rocchio-style).
In PyTerrier_DR it’s easy: retrieve docs, a transformer to mix doc vectors w/ the query vector, and re-retrieve.
pyterrier.readthedocs.io/en/…
Hang Li retweeted
RA job in @ir_glasgow on an EU Horizon project, with @iadh @richardm_ and myself, on the use of LLMs in developing AI agents for the risk and ESG (Environmental, Social and Governance) assessment of sustainability-focused green investment projects. 👇
Hang Li retweeted
🎉 Excited to share the list of our @IELabGroup papers accepted at #SIGIR2025 @ACMSIGIR! Topics include efficiency in IR, domain-specific IR (SysRev, podcast), screenshot retrievers, hallucination detection, and more. See you in Padua 🇮🇹!
Details coming soon, stay tuned!
Hang Li retweeted
I published Vol. 96 of "Top Information Retrieval Papers of the Week" on Substack.
🔗 recsys.substack.com/p/the-co…
Hang Li retweeted
Pseudo-Relevance Feedback Can Improve Zero-Shot LLM-Based Dense Retrieval
@hanglics et al. show how LLMs can extract features from top-ranked documents, enabling smaller rankers with PRF to match larger rankers without PRF.
📝arxiv.org/abs/2503.14887
Hang Li retweeted
Rank-R1: Enhancing Reasoning in LLM-based Document Rerankers via Reinforcement Learning
@ShengyaoZhuang et al. introduce a LLM-based reranker that performs reasoning over queries and candidate documents before ranking.
📝arxiv.org/abs/2503.06034
Hang Li retweeted
We introduce Rank-R1, an LLM-based document reranker featured with reasoning abilities. Joint work with @xueguang_ma @bevan_koopman @lintool @guidozuc
Check out the paper: arxiv.org/pdf/2503.06034
and we open-sourced model checkpoints and our code at:
github.com/ielab/llm-rankers…
Hang Li retweeted
Our understanding of knowledge updating: Do not modify model parameters. arxiv.org/abs/2503.05212
Hang Li retweeted
RIP ”Needle in the haystack” 🪦
For all my RAG/vectorstore fanboys out there who keep betting against the models.
Hang Li retweeted
📣 📣 The #sigir2025 LiveRAG Challenge is open at:
sigir2025.dei.unipd.it/live-…
Prizes:
- First Prize: $5000
- Second Prize: $3000
- Third Prize: $2000
Deadlines:
- Application submission deadline: February 24, 2025
More info? Ask [email protected].
Hang Li retweeted
Anyone who thinks DeepSeek just came out of nowhere should see this graph.
For each model on this graph, weights, code, and detailed papers were released.
This is a team with a strong track record and has been working hard for a while. They didn't come out of nowhere.
Hang Li retweeted
Sharing this gem with the world, unearthed while cleaning out personal files: History of open-source IR systems, dating to 1985 with the release of SMART (yes, 40 years ago), and even further back! github.com/lintool/history-o…
Hang Li retweeted
Nice to see Terrier now integrated with our Vector PRF method!
Check out the paper: arxiv.org/pdf/2108.11044
Pyserini also supports it: github.com/castorini/pyserin…
Replying to @craig_macdonald
5/7. @macavaney has been polishing pyterrier_dr (github.com/terrierteam/pyter…) – our single-vector Dense Retrieval framework, for instance including PRF:
Hang Li retweeted
One key insight is that VLMs can open up many new possibilities for retrieval and RAG!
What can VLM brings to RAG beyond input modality change?
For “R”, our DSE dropped the document processing and improved relevancy modeling by preserving the content integration.
Now for “G”, we propose VISA. Aiming to take a step towards more verifiable and intuitive V-RAG.
arxiv.org/abs/2412.14457































