Negar Arabzadeh

1/ Thrilled to introduce T³: a corpus for RAG over reasoning tasks, built from thinking traces. We show that surprisingly RAG can improve reasoning— with the right corpus. Rag with Transformed Thinking Traces T³ gain by up to 43.9% on AIME 2025-2026. 🔗 arxiv.org/abs/2605.03344 🧵

Leaving #ICML2026 with some exciting news! Our paper DeepScholar-Bench has been accepted to #COLM2026! 🎉 arxiv.org/abs/2508.20033 If you're building Deep Research systems and looking for a challenging benchmark to evaluate them, check out our work!

Very excited about our talk today at 4pm in Hall B2 on Measuring Agents in Productions @icmlconf See you all soon!🤩 #ICML2026

Come join us in the poster session this morning in hall A!

We’re presenting Measuring Agents in Production (MAP) at @icmlconf today at 10:30 am! Come visit us at Hall A, Poster #4311 We have a little surprise for you at the poster stand 😉 ! #ICML2026 Paper: arxiv.org/abs/2512.04123

@naghmehfarzi Congrats!!!🥳

There is still a long way to go, but I just hit 100 citations on google scholar!

@binwu_cs See you in ICML and SIGIR then! 😀

Excited to attend #ACL2026, #ICML2026 & #SIGIR2026 this July! 🎉 Presenting 3 papers at ACL & ICML and organizing the AgentSearch Workshop at SIGIR. If you’re attending too, let’s connect! Looking forward to meeting you there.

📢 If you’re attending #SIGIR2026 in Melbourne, consider joining the Women in IR (WIR) event on Wednesday, July 22 (lunch time) @SIGIRConf 📝 It includes a low-effort poster session where you can introduce yourself, your research, and your interests using a simple template. bit.ly/4vHIQdh 📩 Submit your poster here: bit.ly/4v2p2jp 🌟 Highly recommend for early PhD students! It’s one of the best ways to meet other students and connect with senior researchers in a relaxed setting. RT is highly appreciated!

Join our #WWW2026 tutorial, Conversational Search: From Fundamentals to Frontiers in the Age of Agents, happening tomorrow, 30 June, from 10:00–13:30 UAE time in Hall 2. Website: convsearch.github.io/www2026 Presented by Fengran Mo, @maliannejadi, @JeffD, Jian-Yun Nie, and me

❤️🤍💚 The Iranian football team put in the greatest performance in their World Cup history. First time without a loss. Three goals ruled out by VAR as offside, three draws. The last VAR offside was two minutes before the end of their last match. They did all of this while their country was at war and their players were fighting visa battles. Kudus to your resilience and brilliance .

3/ QueryGym is also a toolkit. You can generate reformulations with a single API, then run the full pipeline: reformulate → retrieve → evaluate It supports Pyserini and PyTerrier (@TerrierTeam), and makes it easy to swap the LLM, prompt, retriever, method, or dataset without rewriting the pipeline. The goal is to make query reformulation results easier to compare, reproduce, and build on. Resources: 🔗 Website: querygym.com 🐙 Code: github.com/ls3-lab/QueryG… 📄 Docs: querygym.readthedocs.io 📜 Demo paper: arxiv.org/abs/2511.15996 📜 Reproducibility paper: arxiv.org/pdf/2604.27421 Feedback, issues, and contributions are very welcome! Shoutout to amazing Querygym team @amin_bigdelii @radin_hrad , Hai son Le, Mert Incesu, @claclarke and @ebrahim_bagheri #SIGIR2026 #Webconf2026 #WWW2026

2/ LLM-based query reformulation has exploded. Almost every paper reports gains, but evaluations are often not comparable: different LLMs, retrievers, prompts, decoding settings, datasets, and metrics. So even a simple question has been hard to answer: Which reformulation method actually works best, and under what conditions❓ QueryGym addresses this with one unified, reproducible evaluation framework. The leaderboard compares methods head-to-head across: 🔹 10 reformulation methods 🔹 4 LLMs: GPT-4.1, GPT-4.1-nano, Qwen2.5-72B, Qwen2.5-7B 🔹 3 retrievers: BM25, SPLADE++, BGE 🔹 21 datasets: MS MARCO/TREC DL, BEIR, BRIGHT 🌟In total: 1,150 fully reproducible runs.

1/ 🏋️🔍Excited to share that the QueryGym Leaderboard is live! 🔗 leaderboard.querygym.com QueryGym is an open-source toolkit + public leaderboard for LLM-based query reformulation. If you work on retrieval, RAG, agentic search, or deep research systems, and you are experimenting with query reformulation, QueryGym is built for you. 🧵👇 📍 Reproducibility study accepted at @SIGIRConf arxiv.org/pdf/2604.27421 📍 QueryGym demo accepted at @TheWebConf arxiv.org/abs/2511.15996

In T3, we recently showed that thinking traces can serve as a datastore for RAG 🧠📚 This work explores a related question in a more agentic setting 🤖: What if, instead of human users or static documents, we have a population of agents? Can they share their previous successful experiences with one another? Multi-Agent Transactive Memory (MATM) is a shared searchable repository where 30+ agents contribute successful trajectories, allowing other agents to retrieve and reuse collective experience at inference time 🔍 arxiv.org/abs/2606.19911

🚀 Beyond excited to share we're releasing LOTUSPlan, a new API & optimizer for higher performance LLM-powered data processing, from our team at Berkeley & Stanford. LOTUS now lets you write your LLM-based queries and optimize them for up to 2.4× lower cost and 4.6× higher accuracy for tasks like, agent trace analysis, LLM-judge evals, RAG, document extraction and deep research. ✨Checkout our our new blog: liana313.github.io/blog/lotusplan… 🧵

The web was never meant to be flattened into text. Yet most web RAG systems start by parsing HTML --- a complex and lossy process. 🔥 Introducing PixelRAG: the first RAG system that retrieves and reads 30M+ web pages as pixels. Instead of extracting text, PixelRAG retrieves screenshots and lets a VLM read them directly. PixelRAG not only preserves visual information, but also outperforms text-based RAG on text-only QA benchmarks by +18.1%. Why? (1) HTML-to-text conversion often discards layout, structure, tables, and other useful signals. (2) We continued pretraining a VLM on web page screenshots and turned it into a surprisingly strong visual retriever. (3) Recent VLMs are remarkably good at understanding web pages, often with better accuracy and token efficiency than text-only pipelines. Takeaway: HTML parsing may be one of the biggest self-inflicted bottlenecks in web RAG. Demo below 👇 Code: github.com/StarTrail-org/… Paper: github.com/StarTrail-org/… Playground: pixelrag.ai

Search is becoming increasingly agentic: systems plan, search, synthesize, cite, and revise. But, how should we study and evaluate these systems? 🤔 In TREC RAG 2026, we want to build a reusable collection for this new reality We’ve aligned on 4 core directions 🧵👇

Excited to announce that our paper “From Noise to Order: Learning to Rank via Denoising Diffusion” has been accepted to #ICTIR2026! 🎉📚 📄Paper: arxiv.org/pdf/2602.11453 💻 Code: github.com/sadjadeb/Diffu…

@ChuanMg @UWaterloo @claclarke Thanks Dr. Meng 🫢

@NegarEmpr @UWaterloo @claclarke Congrats Dr Arabzadeh!

Grateful that my PhD thesis was recognized as one of the top dissertations in the 2026 Faculty of Mathematics Doctoral Prize at the @UWaterloo ! 🎉 And it is always especially nice to hear kind words from your PhD supervisor @claclarke . I guess that feeling never really goes away, even after you graduate. 😊 uwaterloo.ca/computer-scien…

@sewon__min @UWaterloo @claclarke Thank you Sewon!

@NegarEmpr @UWaterloo @claclarke Congratulations Negar!

@tetsuyasakai @UWaterloo @claclarke Thanks Tetsuya 😀

@NegarEmpr @UWaterloo @claclarke Congrats! Last night Charlie was talking about you!