←
Debasis Ganguly615 posts
Debasis Ganguly avatar

Debasis Ganguly

@debforit

Lecturer/Asst. Professor at the School of Computing, University of Glasgow (@UofGlasgow/@GlasgowCS/@IDAglasgow/@ir_glasgow)

◷ Joined February 2010▧ 97 Photos and videos
Debasis Ganguly avatar
Debasis Ganguly@debforit·Aug 21

🧵 Excited to share that our paper "Modeling Ranking Properties with In-Context Learning" has been accepted at #EMNLP2026! 🎉 Joint work with @nilanjansb, @MrParryParry, and Pabitra Mitra. Preprint: arxiv.org/abs/2505.17736

Debasis Ganguly avatar
Debasis Ganguly@debforit·Aug 10

📄 Robustness of IR Models to Collection Growth Led by PGR Akis Lionis, co-supervised with @macavaney Collections grow continuously; adding unrelated documents shouldn't hurt effectiveness. But does it? Finding: IR models suffer some degradation as collections expand. 📚

Debasis Ganguly avatar
Debasis Ganguly@debforit·Aug 10

📄 Generate to Accelerate: Improved Reranking via LLM-Generated Pivot Documents Led by external PhD student @nilanjansb (@IITKgp ), primarily supervised by Pabitra Mitra. What if rerankers could use a document that doesn't even exist? 🤔

Debasis Ganguly avatar
Debasis Ganguly@debforit·Aug 10

📄 Predicting Partial Answer Quality and Utility in Agentic Retrieval-Augmented Generation Led by PhD student @DanielTian97 , co-supervised with @craig_macdonald . Can we predict whether the next retrieval step in an agentic RAG pipeline will actually help? 🤖

Debasis Ganguly avatar
Debasis Ganguly@debforit·Aug 10

🧵 Exciting news from @ir_glasgow: 3 papers from our group have been accepted at #CIKM2026! 🎉 Two long papers and one short paper covering agentic RAG, efficient reranking, and retrieval robustness. Huge congratulations to the students and collaborators involved! 👏 🧵👇

Debasis Ganguly avatar
Debasis Ganguly@debforit·Jun 16

To live up to my ISI CS postgraduate roots, I finally wrote a theoretical IR paper 🙂🧮 to estimate how risky is stochastic reranking before we actually apply it? 🎲 Paper: arxiv.org/abs/2606.16970 (to appear in #ictir2026) Code: github.com/gdebasis/srea

Debasis Ganguly avatar
Debasis Ganguly@debforit·May 31

A single-authored paper always feels special, perhaps even more so these days, when most of my research time is devoted to working with students. Delighted to share that my paper, “A Theoretical Framework for Risk Analysis of Stochastic Rankers,” has been accepted at #ictir2026.

Debasis Ganguly avatar
Debasis Ganguly@debforit·May 28

While IR4RAG is more common, the other way is cool as well. Check out this excellent work on RAG4IR by one of my grad students. #tois #rag #qpp

I am pleased to share that our journal article, "RAQG-QPP: Query Performance Prediction with Retrieved Query Variants and Retrieval Augmented Query Generation," has been accepted by ACM Transactions on Information Systems (TOIS), a joint work with @debforit and @craig_macdonald.

Debasis Ganguly avatar
Debasis Ganguly@debforit·Jan 29

🚀📈 Does QPP actually improve IR? The first comparative study showing that QPP can improve retrieval effectiveness by selecting variable proportions of information across rankers. 📄 Preprint: arxiv.org/abs/2601.17339 🧠💬 Stay tuned for the talks at #ECIR2026, feedback welcome!

Debasis Ganguly avatar
Debasis Ganguly@debforit·Jan 29

⚖️🤖 QPP beyond single rankers How well do existing QPP models perform on a harder and more meaningful task? ➡️ Comparing multiple rankers for the same query — not just predicting absolute performance. 📄 Preprint: arxiv.org/abs/2601.17359 @PayelSantra17

Debasis Ganguly avatar
Debasis Ganguly@debforit·Jan 29

🔍📊 QPP for RAG: Can we predict answer quality? We study answer quality prediction in Retrieval-Augmented Generation, linking retrieval signals to downstream generation quality. 📄 Preprint: arxiv.org/abs/2601.14546 @DanielTian97 @craig_macdonald

Debasis Ganguly avatar
Debasis Ganguly@debforit·Jan 29

🧵📢 Preprints out! Ahead of ECIR’26 #ecir2026, I’m excited to share three new papers on Query Performance Prediction (QPP) — continuing my long-standing obsession with understanding when retrieval works (and when it doesn’t). More info⬇️