Preslav Nakov · page 3
Papers
- Atlas-Chat: Adapting Large Language Models for Low-Resource Moroccan Arabic Dialect (2025 · 18 citations)
Arabic-focused LLMs such as Jais, AceGPT and ALLaM mostly balance English and Modern Standard Arabic and largely ignore spoken dialects. - Aristotle: Mastering Logical Reasoning with A Logic-Complete Decompose-Search-Resolve Framework (2025 · 16 citations)
Large language models still struggle with logical reasoning. General search-based methods such as ToT and GoT ignore the formal structure of logic problems: they split tasks along surface wording, depend on unreliable LLM evaluators to choose paths, and make logical slips when solving sub-steps. The authors present… - BALSAM: A Platform for Benchmarking Arabic Large Language Models (2025 · 14 citations)
Arabic LLMs still lag behind English ones, and the gap is hard to measure. Most Arabic benchmarks use static public data and cover only a narrow set of capabilities. They also lack a hosted platform with hidden test sets, so training-data contamination is likely and progress is difficult to track. A large consortium… - Unconditional Truthfulness: Learning Unconditional Uncertainty of Large Language Models (2025 · 10 citations)
LLMs score each new token assuming everything generated so far is correct. If an early token is wrong, the tokens that follow can still look highly confident, so uncertainty estimates built on raw probabilities understate the risk. The authors argue that what matters is a token's unconditional chance of being… - DialectalArabicMMLU: Benchmarking Dialectal Capabilities in Arabic and Multilingual Language Models (2026 · 9 citations)
Most Arabic LLM benchmarks test Modern Standard Arabic (MSA), but everyday Arabic is spoken and written in regional dialects that differ in lexicon, morphology and syntax. - Nile-Chat: Egyptian Language Models for Arabic and Latin Scripts (2025 · 9 citations)
Egyptian Arabic is often written in two ways: in Arabic script and in a Latin-based romanization called Arabizi or Franco-Arabic. - FinCARDS: Card-Based Analyst Reranking for Financial Document Question Answering (2026 · 8 citations)
Answering questions over long SEC filings (10-K/10-Q) depends on finding passages that match the right company, financial metric and fiscal period, and that often contain a concrete number. - Exploring the Limitations of Detecting Machine-Generated Text (2025 · 7 citations)
Detectors of machine-generated text (MGT) often report strong accuracy, but they are usually tested on data resembling their training data. - IITD at WANLP 2022 Shared Task: Multilingual Multi-Granularity Network for Propaganda Detection (2022 · 7 citations)
The paper describes a second-place system for the WANLP 2022 shared task on detecting propaganda in Arabic tweets. - Can LLMs Automate Fact-Checking Article Writing? (2026 · 5 citations)
Research on automated fact-checking usually ends at spotting claims, retrieving evidence, assigning a verdict and, at most, producing a short justification. - FAID: Fine-grained AI-generated Text Detection using Multi-task Auxiliary and Multi-level Contrastive Learning (2026 · 3 citations)
Detecting machine-written text is harder when humans and LLMs co-write, and most detectors only output a human vs. - COMMUNITYNOTES: A Dataset for Exploring the Helpfulness of Fact-Checking Explanations (2026 · 2 citations)
Platforms such as X, Meta and TikTok increasingly rely on crowd-written notes rather than professional fact-checkers to explain why a post may mislead. - Instruction-Guided Poetry Generation in Arabic and Its Dialects (2026 · 2 citations)
Research on Arabic poetry with LLMs has mostly covered analysis, such as meter detection or poet attribution. Controllable generation, where a user sets constraints like meter, rhyme, era or dialect, has received little attention. The authors treat Arabic poetry writing as an instruction-following problem. - Beyond the Crowd: LLM-Augmented Community Notes for Governing Health Misinformation (2026 · 1 citation)
Community Notes on X lets volunteers attach context to misleading posts, but the process is slow. The authors studied 30.8K English health-related notes and found that corrections usually arrive after public attention has peaked. Most notes never get enough votes to be shown at all. To speed this up, they propose… - Can Dialects Be Steered Like Languages? Sparse Neurons and Distributed Directions in Arabic LLMs (2026 · 1 citation)
Arabic LLMs are trained mostly on Modern Standard Arabic (MSA), so they tend to answer in MSA even when a user writes in a dialect. - From Multiple-Choice to Extractive QA: A Case Study for English and Arabic (2025 · 1 citation)
Annotated data for task-specific NLP is scarce outside a few major languages. The authors ask whether a dataset built for one QA format can be turned into another. They take part of BELEBELE, a multiple-choice reading-comprehension benchmark covering 122 language variants, and convert it into extractive QA (EQA)… - A Multi-View Media Profiling Suite: Resources, Evaluation, and Analysis (2026)
Automatically judging news outlets for political bias and factuality is a scalable alternative to checking individual claims or articles. - Beyond Memorization: Extending Reasoning Depth with Recurrence, Memory and Test-Time Compute Scaling (2026)
The paper asks whether multi-step reasoning in neural models reflects real generalization or memorization, and how difficulty grows with the number of required inference steps. - Curriculum Learning and Pseudo-Labeling Improve the Generalization of Multi-Label Arabic Dialect Identification Models (2026)
Arabic Dialect Identification has usually been treated as assigning one country label per sentence. In reality, many utterances sound natural in several dialects. The multi-label variant (MLADI) still has no large multi-label training resource: only small evaluation sets carry multiple labels. The authors trained… - Schützen: Evaluating LLM Safety in Bulgarian and German Contexts (2026)
Most LLM safety benchmarks are built for English or Chinese and usually test other languages one at a time. - MuDRiC: Multi-Dialect Reasoning for Arabic Commonsense Validation (2025)
Arabic commonsense benchmarks have mostly covered Modern Standard Arabic (MSA), while everyday communication happens in regional dialects. MSA-trained models transfer poorly to them. The authors build MuDRiC, which they describe as the first multi-dialect Arabic commonsense validation resource.
This page in Markdown → https://md.research.elamur.ai/free/en/authors/preslav-nakov/page/3/