Abdelrahman Abdallah

PhD candidate @disc_uibk

[Hamed Zamani @HamedZamani]

We are releasing the most significant technology transfer effort by the CIIR in the last decade. Please share with your network.

In this thread you will learn what Nage enables you to do.

🧵👇

[DSPy @DSPyOSS]

DSPy 3.4.0 out, with:

(1) native support for Jev and System One models - in the timeless DSPy syntax.

(2) a brand new optimizer, ReAnchor, specifically for calibrating outputs with confidence.

(3) lightning-fast import speeds for the LLM abstraction, via sibling library LM15)

[Sumit @_reachsumit]

SmallReason-ColBERT: An Ultra-Small Late-Interaction Retriever for Reasoning Intensive Retrieval

@perdactor et al. present a 32M retriever for reasoning-heavy queries, adding a learned token-weighting head.

📝https://t.co/IX9ZrP1UiB
👨🏽‍💻https://t.co/roDgX8pkPG

[Sumit @_reachsumit]

OBLIQ-IR: Training a Dense Retriever for Oblique Queries

@perdactor et al. train a dense retriever for oblique queries, where relevance depends on latent traits like stance or style rather than topic overlap.

📝 https://t.co/LAyyASU6VO
👨🏽‍💻 https://t.co/oLyMwaPLa4

[Sumit @_reachsumit]

EXCISE: Query-Side Exclusion for Late-Interaction Retrieval

Presents a query-time fix for "exclusion inversion" in ColBERT retrievers, where excluded topics get ranked higher, not lower.

📝 https://t.co/lGDJHGqKCv

[Omar Khattab @lateinteraction]

imo a significant part of the problem is the insistence of the community on misnaming the paradigm as “multi-vector retrieval”, we called it late interaction for a reason

the problem isn’t in having one vector; the problem is in the scoring function!

https://t.co/QupzaHoiSd

[Abdelrahman Abdallah @perdactor]

@srchvrs Thanks! should be out this week.

[Abdelrahman Abdallah @perdactor]

https://t.co/gk0VZuiMtV

[Abdelrahman Abdallah @perdactor]

Completely agree; we already have these in our limitations.

What we do fits a similarity definition. It doesn't let the query reshape the document representation, and as you say, late interaction is finer but still the same fixed space.

We say a version of this in the paper's Limitations: we narrow the first-stage bottleneck rather than close it. That's where our follow-up work is going — making the latent attribute explicit instead of hoping one fixed vector space encodes it. Promptriever is the right reference too.

We're still working on a new architecture for exactly that.

Genuinely useful question, thanks. Happy to compare notes.

[rahul @0interestrates]

most interesting thing happening rn is that you are allowed to use ai for everything. everything except writing. if you use ai for writing, they will unalive you in public

[Abdelrahman Abdallah @perdactor]

🎉 Our lab has 10 papers accepted at #EMNLP2026 — 9 Main Conference + 1 Findings!

This is ~1.5 years of work, and almost none of it landed on the first try. Several of these papers went through 4–6 submission cycles before they were accepted. Rejection was part of the process, not the end of it.

Main Conference

1️⃣ OBLIQ-IR: Training a Dense Retriever for Oblique Queries

Mahmoud Abdalla, 𝗔𝗯𝗱𝗲𝗹𝗿𝗮𝗵𝗺𝗮𝗻 𝗔𝗯𝗱𝗮𝗹𝗹𝗮𝗵, Shaimaa Sedek, Adam Jatowt
📄 Coming soon · 💻 Coming soon

2️⃣ MEMORA: A Memory-Enhanced Multimodal Committee Reranking Agent for Reasoning-Intensive Retrieval

Mohamed Mahmoud, Mostafa Farouk Senussi, Mahmoud Abdalla, Mahmoud SalahEldin Kasem, 𝗔𝗯𝗱𝗲𝗹𝗿𝗮𝗵𝗺𝗮𝗻 𝗔𝗯𝗱𝗮𝗹𝗹𝗮𝗵, Hyun Soo Kang
📄 Coming soon · 💻 Coming soon

3️⃣ SustainableQA: A Comprehensive Question Answering Dataset for Corporate Sustainability and EU Taxonomy Reporting

Mohammed Ali, 𝗔𝗯𝗱𝗲𝗹𝗿𝗮𝗵𝗺𝗮𝗻 𝗔𝗯𝗱𝗮𝗹𝗹𝗮𝗵, Adam Jatowt
📄 https://t.co/HvJQC0tPLZ · 💻 https://t.co/16pUx221nE

4️⃣ TEMPO: Realistic Multi-Domain Benchmark for Temporal Reasoning-Intensive Retrieval

𝗔𝗯𝗱𝗲𝗹𝗿𝗮𝗵𝗺𝗮𝗻 𝗔𝗯𝗱𝗮𝗹𝗹𝗮𝗵, Mohammed Ali, Muhammad Abdul-Mageed, Adam Jatowt
📄 https://t.co/rOt5FCopkl · 🌐 https://t.co/kbE3zqsOyn

5️⃣ Argus-Retriever: Vision-LLM Late-Interaction Retrieval with Region-Aware Query-Conditioned MoE for Visual Document Retrieval

𝗔𝗯𝗱𝗲𝗹𝗿𝗮𝗵𝗺𝗮𝗻 𝗔𝗯𝗱𝗮𝗹𝗹𝗮𝗵, Mahmoud Abdalla, Mohammed Ali, Adam Jatowt
📄 https://t.co/HsaSXT0OVS · 💻 https://t.co/KsdsVIR3Ys

6️⃣ Large Language Models Systematically Favor Popular Options: Evidence and Mitigation Across MCQs

𝗔𝗯𝗱𝗲𝗹𝗿𝗮𝗵𝗺𝗮𝗻 𝗔𝗯𝗱𝗮𝗹𝗹𝗮𝗵, Mohammed Ali, Bhawna Piryani, Mahmoud Abdalla, Adam Jatowt
📄 Coming soon · 💻 Coming soon

7️⃣ SmallReason-ColBERT: An Ultra-Small Late-Interaction Retriever for Reasoning-Intensive Retrieval

𝗔𝗯𝗱𝗲𝗹𝗿𝗮𝗵𝗺𝗮𝗻 𝗔𝗯𝗱𝗮𝗹𝗹𝗮𝗵, Mohammed Ali, Adam Jatowt
📄 Coming soon · 💻 Coming soon

8️⃣ MoCA-Agent: A Market-of-Claims Code Agent for Financial and Numerical Reasoning

𝗔𝗯𝗱𝗲𝗹𝗿𝗮𝗵𝗺𝗮𝗻 𝗔𝗯𝗱𝗮𝗹𝗹𝗮𝗵, AbdelRahim A. Elmadany, Sameh Al Natour, Hasan Cavusoglu, Adam Jatowt, Muhammad Abdul-Mageed
📄 https://t.co/SOsbIiAN20 · 💻 https://t.co/RuQorzl2LR

9️⃣ The Magnitude Mirage: Rethinking Confidence for Reasoning-Intensive Retrieval
Jamie Holdcroft, 𝗔𝗯𝗱𝗲𝗹𝗿𝗮𝗵𝗺𝗮𝗻 𝗔𝗯𝗱𝗮𝗹𝗹𝗮𝗵, Adam Jatowt
📄 Coming soon · 💻 Coming soon

Findings

🔟 REGREACT: Self-Correcting Multi-Agent Pipelines for Structured Regulatory Information Extraction
Mohammed Ali, 𝗔𝗯𝗱𝗲𝗹𝗿𝗮𝗵𝗺𝗮𝗻 𝗔𝗯𝗱𝗮𝗹𝗹𝗮𝗵, Adam Jatowt
📄 https://t.co/6YAjLOzH1a · 💻 Coming soon

Huge congratulations to everyone in the lab, and thanks to our supervisor @adammo . To anyone sitting on a pile of rejections right now: keep resubmitting. It works. 🙏

[Abdelrahman Abdallah @perdactor]

@TongletJ @danish037 No, that's why they should consider only the committed papers to EMNLP

[Abdelrahman Abdallah @perdactor]

@danish037 I don't feel it is correct because ARR in May is 17k and many papers were committed to AACL.

[Abdelrahman Abdallah @perdactor]

2/ The idea: a region-aware Mixture-of-Experts inside the document encoder.
A router reads each region's content, its 2D position, and a pooled query context z_q, then mixes 4 latent experts (+1 shared).
The page is now encoded differently per query → D(q). Still MaxSim.

[Abdelrahman Abdallah @perdactor]

1/👁️ Meet Argus-Retriever: the first late-interaction visual doc retriever where the document representation adapts to the query: D(q).

- 86.0 NDCG@5 on ViDoRe V1+V2 (SOTA open model)
- 📦 1024-dim head, 4.5× smaller index

[Sumit @_reachsumit]

Argus-Retriever: Vision-LLM Late-Interaction Retrieval with Region-Aware Query-Conditioned MoE for Visual Document Retrieval

Introduces a query-conditioned late-interaction visual document retriever.

📝 https://t.co/oB5oa4bgKc
👨🏽‍💻 https://t.co/SKzRERzgRh

[Sumit @_reachsumit]

Argus-Retriever: Vision-LLM Late-Interaction Retrieval with Region-Aware Query-Conditioned MoE for Visual Document Retrieval

Introduces a query-conditioned late-interaction visual document retriever.

📝 https://t.co/oB5oa4bgKc
👨🏽‍💻 https://t.co/SKzRERzgRh

[Abdelrahman Abdallah @perdactor]

Thrilled to share two paper acceptances! 🎉

📄 MM-BRIGHT — Multimodal Benchmark for Reasoning-Intensive Retrieval → #KDD2026

🛠️ Rankify — Python toolkit for Retrieval, Re-Ranking & RAG (670+ ⭐) → #ACL2026 Demo

Thanks to @ supervisor & collaborators 🚀

#NLP #RAG #IR

[Abdelrahman Abdallah @perdactor]

@antoine_chaffin Sure i will try it now

[Aamir @aaxsh18]

another late interaction w