Member of Technical Staff @KeenableAI | Previous Research Fellow at @Ielabgroup | @UWAEMS alumni RAG, IR, NLP, Model Efficiency BlueSky: dylanshuaiwang
Brisbane, Queensland
Joined February 2018
- Tweets165
- Following212
- Followers230
- Likes579
Wang shuai (Dylan) retweeted
🚀 Introducing SkimSearchAgent!
A research framework for deep-search agents:
⚖️ Fair, controlled benchmarking
🔁 Reproducible experiments
⚡ Easy setup & execution
🧩 Modular & extensible
Swap components. Run experiments. Compare fairly.
🔗 ielab.io/skim-search-agent/
🚀 Introducing SkimSearchAgent!
A research framework for deep-search agents:
⚖️ Fair, controlled benchmarking
🔁 Reproducible experiments
⚡ Easy setup & execution
🧩 Modular & extensible
Swap components. Run experiments. Compare fairly.
🔗 ielab.io/skim-search-agent/
Wang shuai (Dylan) retweeted
For IR folks: @xueguang_ma and I quickly tested Jev as a pointwise/pairwise/setwise/listwise reranker on the classic DL19 and DL20, reranking the top 100 BM25 results. It is pretty good and cheap.
Wang shuai (Dylan) retweeted
ITER: Interaction-Aware Retrieval for Agentic Search
Presents a retriever for deep-research agents that conditions search on prior sub-queries and visited docs, surfacing evidence instead of re-returning already-seen material.
📝arxiv.org/abs/2608.27912
👨🏽💻github.com/ielab/ITER
🔎 Deep-research agents read too much.
We introduce SIEVE: Search → Inspect → Fetch, using document structure to decide what to read.
Higher accuracy with 20.7–50.6% fewer tokens.
📄 arxiv.org/abs/2608.02751
💻 github.com/ielab/skim-search…
🌐 ielab.io/skim-search-agent/
Wang shuai (Dylan) retweeted
Search, Inspect, Fetch: Exploiting Boolean Retrieval for Deep-Research Agents
@dylan_wangs et al. filter docs via Boolean queries, then fetches just the relevant section, beating Search-Visit with 20-50% fewer tokens.
📝 arxiv.org/abs/2608.02751
👨🏽💻 github.com/ielab/skim-search…
DiffRetriever
Turn a diffusion LM into a retriever: both as dense/sparse/hybrid retriever.
→Works both zero-shot or tuned!
→Beats PromptReps, DiffEmbed, RepLLaMA when tuned
→~15× faster than AR multi-token (zero-shot) and 3x AR multi-token (tuned)
arxiv.org/abs/2605.07210
how it works: append K [MASK]s to a retrieval prompt; the diffusion LM fills them in one bidirectional forward pass → K dense (ColBERT MaxSim) + K sparse (max-pooled logits)
joint work with Yin Yu @ShengyaoZhuang @BevanKoopman @GuidoZuccon (UQ + CSIRO)
What if you can turn any paper into a live demo with one command?
for example: paper-demo-agent demo 1706.03762
It reads the paper and builds demo end-to-end(app, slides, project page, diagram...).
10 output types · 6 LLM
Picture show example paper
⭐ github.com/wshuai190/PaperDe…
🚀 AutoBool Demo is live!
An 4B-LLM that automatically generates Boolean queries for systematic reviews: outperforms GPT-4o & O3 on extensive evaluation
🤗 huggingface.co/spaces/wshuai…
📄 Accepted in EACL 2026 Main Oral: arxiv.org/abs/2602.00005
#SystematicReview #EBM #NLP #Pubmed
on top of that, if you also worry only about getting the correct Mesh Terms, you can also use our Mesh-Suggester Tool
🤗 huggingface.co/spaces/ielabg…
📷 In WSDM2023 demo paper: dl.acm.org/doi/10.1145/35395…
The old cloud service clashed so moved it to hugging-face.
Wang shuai (Dylan) retweeted
Beyond Chunk-Then-Embed: A Comprehensive Taxonomy and Evaluation of Document Chunking Strategies for Information Retrieval
Provides a systematic evaluation of doc chunking strategies for dense retrieval, finding that optimal methods are task-dependent.
📝arxiv.org/abs/2602.16974
I'm here in Sydney for WWW2025! see us this Thursday, at WWW conference at Posterboard-21, Short Paper & Demo Parkside Ballroom if you are interested!
Today in @SIGIRConf at M3.1, I will briefly introduce our newly published collection FeB4RAG! The collection is the first collection to evaluate Federated search in RAG. Our resource support evaluation of Resource Selection, Merging, and Human Evaluation on Answer Generation!
Wang shuai (Dylan) retweeted
Thank you so much @guidozuc ! It was great to collaborate with you and ieLab members @dylan_wangs @ShengyaoZhuang ! Hope to see you soon and work together again. 😁
Yesterday I presented our #ECIR2025 paper "An Investigation of Prompt Variations for Zero-shot LLM-based Rankers", with @shuoqi_sun @ShengyaoZhuang @dylan_wangs. Yes, it was in a church! Slides at ielab.io/files/guido-2025040… and paper at arxiv.org/abs/2406.14117
Wang shuai (Dylan) retweeted
Yesterday I presented our #ECIR2025 paper "An Investigation of Prompt Variations for Zero-shot LLM-based Rankers", with @shuoqi_sun @ShengyaoZhuang @dylan_wangs. Yes, it was in a church! Slides at ielab.io/files/guido-2025040… and paper at arxiv.org/abs/2406.14117
Wang shuai (Dylan) 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!
Wang shuai (Dylan) 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
Share our paper "ReSLLM: Large Language Models are Strong Resource Selectors for Federated Search" with @ShengyaoZhuang , @bevan_koopman and @guidozuc has been accepted to TheWebConference 2025 as a short.
Preprint: arxiv.org/abs/2401.17645
Github: github.com/ielab/SLAT-RsLLM
