Andrew Parry
@MrParryParry
Followers
256
Following
4K
Media
18
Statuses
233
Terminally Ranking @TechAtBloomberg PhD Fellow Research @ir_glasgow / @TerrierTeam
Joined September 2022
🚨 New Pre-Print! You've just added your 600th model to your negative mining pool and filtered all false negatives. Does any of this even matter when we can apply distillation? . In this work with @debforit and @macavaney, we explore data selection in modern ranking. 🧵 Below
Disentangling Locality and Entropy in Ranking Distillation. @MrParryParry et al. separate example selection effects from teacher ranking entropy in neural ranking model optimization, showing complex hard-negative pipelines offer minimal gains. 📝
1
9
23
RT @ir_glasgow: Now at the Demo session @craig_macdonald and @MrParryParry are presenting the beautiful RAG capabilities of PyTerrier #SIG….
0
5
0
RT @DanielTian97: How does retrieval impact on Agentic RAG? Let's see what predicted intermediate retrieval quality can tell us! Our IR-RAG….
0
8
0
0
5
0
RT @_reachsumit: Am I on the Right Track? What Can Predicted Query Performance Tell Us about the Search Behaviour of Agentic RAG. @DanielTi….
0
4
0
Really like this work, if you haven't read it yet, have a look: PSA from Ian Soboroff! move to Krippendorff's Alpha if you want to check the annotator agreement!.
Now it’s @MrParryParry presenting the reproducibility efforts of a large team of researchers in relation to the shelf life of test collections #sigir2025
1
1
22
RT @McguireNiall: Come pop by our presentations if you are interested in how Brain Machine Interfaces🧠are impacting Information Retrieval!.
0
3
0
Presenting a few works at #SIGIR2025 in lovely Padova, looking forward to great discussions and food 🍝.
📢 We're off to #SIGIR2025 in Padova!. A large contingent of our students & staff will be at the main conference + tutorials, workshops & #ICTIR2025. Let’s connect if you’re around! 🤝. Also: we're hiring in #FinTech — DM us if you're interested!.#IR #Hiring #Research #ACMSIGIR
0
0
8
We have a demo at #SIGIR2025 this year. In the meantime, check out the pre-print and the codebase! . Pre-Print: Code:
0
1
2
We have many common datasets and a full evaluation stack, which allows familiar pt.Experiments to continue in a RAG setting.
1
0
1
Additionally, we support more advanced approaches such as FiD, KG-FiD and Search-o1/R1!
1
0
1
Pyterrier RAG can leverage the entire PyTerrier ecosystem as a retriever before the new Reader class integrates many common LLM packages with the PyTerrier data model.
1
0
1
❓Which retrieval library should you use for RAG? Which agent library? Where do you get datasets? We made PyTerrier RAG to stop your headaches! Led by @craig_macdonald /w (@JinyuanF , me and @mengzaiqiao )! 🧵 ⬇️
3
15
34
RT @mengzaiqiao: Pleased to share that Libra is now ranked at Top 3 on ReXrank (, a public leaderboard for chest X….
0
6
0
RT @_reachsumit: I published Vol. 106 of "Top Information Retrieval Papers of the Week" on Substack. 🔗 https://t.co….
0
3
0
Thanks to @_reachsumit for sharing! Full details, power law scaling and further discussions in our paper, have a read ➡️
0
0
3
🔍 We probe theoretical aspects of data selection and propose that focusing on rankings and their entropy under teacher targets can further improve effectiveness, where heuristic hardness cannot. However, one must be cautious of overfitting (see my talk at #SIGIR2025 this year!).
1
0
3
▶️ We derive a novel bound on model risk in terms of example locality and teacher entropy. Crucially, we show that localised examples from BM25 are generally sufficient, and ensembling approaches fail to tighten locality but can reduce ranking entropy.
1
0
3
Full details in our paper. I'd love to discuss broader applications of this work in Padua at #SIGIR2025 in July, so reach out! In the meantime, have a read! ➡️
0
2
9