Stefan Ruseti

Andreea Dutulescu, Stefan Ruseti, Mihai Dascalu

Transformer-based language models often achieve strong results on mathematical reasoning benchmarks while remaining fragile on basic numerical understanding and arithmetic operations. A central limitation is that numbers are processed as symbolic tokens whose embeddings do not explicitly encode numerical value, leading to systematic errors. We introduce a value-aware numerical representation that augments standard tokenized inputs with a dedicated prefix token whose embedding is explicitly conditioned on the underlying numerical value. This mechanism injects magnitude information directly into the model's input space while remaining compatible with existing tokenizers and decoder-only Transformer architectures. Evaluation on arithmetic tasks shows that the proposed approach outperforms baselines across numerical formats, tasks, and operand lengths. These results indicate that explicitly encoding numerical value is an effective and efficient way to improve fundamental numerical robustness in language models.

Adrian Cosma, Stefan Ruseti, Emilian Radoi et al.

Despite their remarkable progress across diverse domains, Large Language Models (LLMs) consistently fail at simple character-level tasks, such as counting letters in words, due to a fundamental limitation: tokenization. In this work, we frame this limitation as a problem of low mutual information and analyze it in terms of concept emergence. Using a suite of 19 synthetic tasks that isolate character-level reasoning in a controlled setting, we show that such capabilities emerge suddenly and only late in training. We find that percolation-based models of concept emergence explain these patterns, suggesting that learning character composition is not fundamentally different from learning commonsense knowledge. To address this bottleneck, we propose a lightweight architectural modification that significantly improves character-level reasoning while preserving the inductive advantages of subword models. Together, our results bridge low-level perceptual gaps in tokenized LMs and provide a principled framework for understanding and mitigating their structural blind spots. We make our code publicly available.

Teodor-Mihai Cotet, Stefan Ruseti, Mihai Dascalu

Resources for Grammatical Error Correction (GEC) in non-English languages are scarce, while available spellcheckers in these languages are mostly limited to simple corrections and rules. In this paper we introduce a first GEC corpus for Romanian consisting of 10k pairs of sentences. In addition, the German version of ERRANT (ERRor ANnotation Toolkit) scorer was adapted for Romanian to analyze this corpus and extract edits needed for evaluation. Multiple neural models were experimented, together with pretraining strategies, which proved effective for GEC in low-resource settings. Our baseline consists of a small Transformer model trained only on the GEC dataset (F0.5 of 44.38), whereas the best performing model is produced by pretraining a larger Transformer model on artificially generated data, followed by finetuning on the actual corpus (F0.5 of 53.76). The proposed method for generating additional training examples is easily extensible and can be applied to any language, as it requires only a POS tagger

Stefan Ruseti, Teodor-Mihai Cotet, Mihai Dascalu

Diacritics restoration is a mandatory step for adequately processing Romanian texts, and not a trivial one, as you generally need context in order to properly restore a character. Most previous methods which were experimented for Romanian restoration of diacritics do not use neural networks. Among those that do, there are no solutions specifically optimized for this particular language (i.e., they were generally designed to work on many different languages). Therefore we propose a novel neural architecture based on recurrent neural networks that can attend information at different levels of abstractions in order to restore diacritics.

George-Sebastian Pîrtoacă, Traian Rebedea, Stefan Ruseti

Answering multiple-choice questions in a setting in which no supporting documents are explicitly provided continues to stand as a core problem in natural language processing. The contribution of this article is two-fold. First, it describes a method which can be used to semantically rank documents extracted from Wikipedia or similar natural language corpora. Second, we propose a model employing the semantic ranking that holds the first place in two of the most popular leaderboards for answering multiple-choice questions: ARC Easy and Challenge. To achieve this, we introduce a self-attention based neural network that latently learns to rank documents by their importance related to a given question, whilst optimizing the objective of predicting the correct answer. These documents are considered relevant contexts for the underlying question. We have published the ranked documents so that they can be used off-the-shelf to improve downstream decision models.

George-Sebastian Pirtoaca, Traian Rebedea, Stefan Ruseti

Question answering is one of the most important and difficult applications at the border of information retrieval and natural language processing, especially when we talk about complex science questions which require some form of inference to determine the correct answer. In this paper, we present a two-step method that combines information retrieval techniques optimized for question answering with deep learning models for natural language inference in order to tackle the multi-choice question answering in the science domain. For each question-answer pair, we use standard retrieval-based models to find relevant candidate contexts and decompose the main problem into two different sub-problems. First, assign correctness scores for each candidate answer based on the context using retrieval models from Lucene. Second, we use deep learning architectures to compute if a candidate answer can be inferred from some well-chosen context consisting of sentences retrieved from the knowledge base. In the end, all these solvers are combined using a simple neural network to predict the correct answer. This proposed two-step model outperforms the best retrieval-based solver by over 3% in absolute accuracy.