PhD candidate @disc_uibk
Innsbruck, Austria
Joined November 2013
- Tweets1.7K
- Following565
- Followers227
- Likes1.4K
Abdelrahman Abdallah retweeted
how did natural language processing researchers deal with the rise of natural language processing models and their application to almost every other domain?
โฆ where do these people think the grads leading almost every frontier lab training team come from
New open retrieval model: UIBK-Embedding-v1 ๐
179M parameters. A dense retriever and a ColBERT retriever in one model, one index, one search.
๐ฅ Best model under 1B parameters on BEIR
๐ฅ #1 on NanoBEIR among models up to 6B
Blog: huggingface.co/blog/abdoelsaโฆ
Top 10 of the 156 models with complete official MTEB results. The nine models above it have 1.5B to 7.6B parameters, at least 8 times more.
It beats LateOn (57.22) and DenseOn (56.20), the strongest late-interaction and dense models of its size.
How it works: the dense embedding is written into the list of token vectors as 3 extra vectors, and a one-dimensional tag keeps the two kinds apart. One standard MaxSim then gives both scores.
Blog: huggingface.co/blog/abdoelsaโฆ
Model: huggingface.co/DataScience-Uโฆ
Abdelrahman Abdallah retweeted
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.
๐งต๐
Abdelrahman Abdallah retweeted
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)
Abdelrahman Abdallah retweeted
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.
๐arxiv.org/abs/2609.29652
๐จ๐ฝโ๐ปgithub.com/DataScienceUIBK/Sโฆ
Abdelrahman Abdallah retweeted
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.
๐ arxiv.org/abs/2609.29649
๐จ๐ฝโ๐ป github.com/DataScienceUIBK/oโฆ
Abdelrahman Abdallah retweeted
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.
๐ arxiv.org/abs/2608.05497
Abdelrahman Abdallah retweeted
imo a significant part of the problem is the insistence of the community on misnaming the paradigm as โmulti-vector retrievalโ, we called it late interaction for a reason
the problem isnโt in having one vector; the problem is in the scoring function!
youtu.be/Z2TmdcylyEc?si=hrtqโฆ
๐ Our lab has 10 papers accepted at #EMNLP2026 โ 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 4โ6 submission cycles before they were accepted. Rejection was part of the process, not the end of it.
Main Conference
1๏ธโฃ OBLIQ-IR: Training a Dense Retriever for Oblique Queries
Mahmoud Abdalla, ๐๐ฏ๐ฑ๐ฒ๐น๐ฟ๐ฎ๐ต๐บ๐ฎ๐ป ๐๐ฏ๐ฑ๐ฎ๐น๐น๐ฎ๐ต, Shaimaa Sedek, Adam Jatowt
๐ Coming soon ยท ๐ป Coming soon
2๏ธโฃ MEMORA: A Memory-Enhanced Multimodal Committee Reranking Agent for Reasoning-Intensive Retrieval
Mohamed Mahmoud, Mostafa Farouk Senussi, Mahmoud Abdalla, Mahmoud SalahEldin Kasem, ๐๐ฏ๐ฑ๐ฒ๐น๐ฟ๐ฎ๐ต๐บ๐ฎ๐ป ๐๐ฏ๐ฑ๐ฎ๐น๐น๐ฎ๐ต, Hyun Soo Kang
๐ Coming soon ยท ๐ป Coming soon
3๏ธโฃ SustainableQA: A Comprehensive Question Answering Dataset for Corporate Sustainability and EU Taxonomy Reporting
Mohammed Ali, ๐๐ฏ๐ฑ๐ฒ๐น๐ฟ๐ฎ๐ต๐บ๐ฎ๐ป ๐๐ฏ๐ฑ๐ฎ๐น๐น๐ฎ๐ต, Adam Jatowt
๐ arxiv.org/abs/2508.03000 ยท ๐ป github.com/DataScienceUIBK/Sโฆ
4๏ธโฃ TEMPO: Realistic Multi-Domain Benchmark for Temporal Reasoning-Intensive Retrieval
๐๐ฏ๐ฑ๐ฒ๐น๐ฟ๐ฎ๐ต๐บ๐ฎ๐ป ๐๐ฏ๐ฑ๐ฎ๐น๐น๐ฎ๐ต, Mohammed Ali, Muhammad Abdul-Mageed, Adam Jatowt
๐ arxiv.org/abs/2601.09523 ยท ๐ tempo-bench.github.io
5๏ธโฃ Argus-Retriever: Vision-LLM Late-Interaction Retrieval with Region-Aware Query-Conditioned MoE for Visual Document Retrieval
๐๐ฏ๐ฑ๐ฒ๐น๐ฟ๐ฎ๐ต๐บ๐ฎ๐ป ๐๐ฏ๐ฑ๐ฎ๐น๐น๐ฎ๐ต, Mahmoud Abdalla, Mohammed Ali, Adam Jatowt
๐ arxiv.org/abs/2606.04300 ยท ๐ป github.com/DataScienceUIBK/Aโฆ
6๏ธโฃ Large Language Models Systematically Favor Popular Options: Evidence and Mitigation Across MCQs
๐๐ฏ๐ฑ๐ฒ๐น๐ฟ๐ฎ๐ต๐บ๐ฎ๐ป ๐๐ฏ๐ฑ๐ฎ๐น๐น๐ฎ๐ต, Mohammed Ali, Bhawna Piryani, Mahmoud Abdalla, Adam Jatowt
๐ Coming soon ยท ๐ป Coming soon
7๏ธโฃ SmallReason-ColBERT: An Ultra-Small Late-Interaction Retriever for Reasoning-Intensive Retrieval
๐๐ฏ๐ฑ๐ฒ๐น๐ฟ๐ฎ๐ต๐บ๐ฎ๐ป ๐๐ฏ๐ฑ๐ฎ๐น๐น๐ฎ๐ต, Mohammed Ali, Adam Jatowt
๐ Coming soon ยท ๐ป Coming soon
8๏ธโฃ MoCA-Agent: A Market-of-Claims Code Agent for Financial and Numerical Reasoning
๐๐ฏ๐ฑ๐ฒ๐น๐ฟ๐ฎ๐ต๐บ๐ฎ๐ป ๐๐ฏ๐ฑ๐ฎ๐น๐น๐ฎ๐ต, AbdelRahim A. Elmadany, Sameh Al Natour, Hasan Cavusoglu, Adam Jatowt, Muhammad Abdul-Mageed
๐ arxiv.org/abs/2606.11537 ยท ๐ป github.com/UBC-NLP/MoCA-Agenโฆ
9๏ธโฃ The Magnitude Mirage: Rethinking Confidence for Reasoning-Intensive Retrieval
Jamie Holdcroft, ๐๐ฏ๐ฑ๐ฒ๐น๐ฟ๐ฎ๐ต๐บ๐ฎ๐ป ๐๐ฏ๐ฑ๐ฎ๐น๐น๐ฎ๐ต, Adam Jatowt
๐ Coming soon ยท ๐ป Coming soon
Findings
๐ REGREACT: Self-Correcting Multi-Agent Pipelines for Structured Regulatory Information Extraction
Mohammed Ali, ๐๐ฏ๐ฑ๐ฒ๐น๐ฟ๐ฎ๐ต๐บ๐ฎ๐ป ๐๐ฏ๐ฑ๐ฎ๐น๐น๐ฎ๐ต, Adam Jatowt
๐ arxiv.org/abs/2604.12054 ยท ๐ป 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. ๐
1/๐๏ธ 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)
- ๐ฆ 1024-dim head, 4.5ร smaller index
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.
๐ arxiv.org/abs/2606.04300
๐จ๐ฝโ๐ป github.com/DataScienceUIBK/Aโฆ
Abdelrahman Abdallah retweeted
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.
๐ arxiv.org/abs/2606.04300
๐จ๐ฝโ๐ป github.com/DataScienceUIBK/Aโฆ
Abdelrahman Abdallah retweeted
another late interaction w
[1/3] tiny 32M reasoning Colber retriever, just dropped. ๐ค
Reason-mxbai-colbert-v0-32m a late-interaction ColBERT fine-tuned for BRIGHT-style retrieval.
19.0 avg nDCG@10 on BRIGHT, ~5x smaller than Reason-ModernColBERT.
huggingface.co/DataScience-Uโฆ
[1/3] tiny 32M reasoning Colber retriever, just dropped. ๐ค
Reason-mxbai-colbert-v0-32m a late-interaction ColBERT fine-tuned for BRIGHT-style retrieval.
19.0 avg nDCG@10 on BRIGHT, ~5x smaller than Reason-ModernColBERT.
huggingface.co/DataScience-Uโฆ
[1/3] Built on @mixedbreadai's mxbai-edge-colbert-v0-32m with one key change: we widened the projection head from 64โ128 dim using small-random init (zero-pad = dead neurons under L2 MaxSim, classic gotcha).