arXiv subjects

Georgios Paraskevopoulos

Publications and source records attributed to Georgios Paraskevopoulos.

At least 19 recordsLinked to original sources

arXiv2026

Beyond Magnitude: Contrastive Routing for Modular Mixture-of-Experts

In current Mixture-of-Experts architectures, routing is performed based on representations dominated by structure shared across all tokens, limiting expert specialization. We show that contrasting each token against an Exponential Moving Average of the layer's hidden states, rather than routing on absolute magnitude, concentrates the routing signal onto a low-dimensional, highly separable subspace. Building on this, we propose the Contrastive Routing Mechanism (CoRM), which scores each expert by the gap between its affinity for the incoming token and its affinity for this shared reference state, interpreted through a distinct per-expert projection. The resulting experts have routing boundaries that align with linguistic structure significantly more than the Top-k baseline. Our experiments show that CoRM improves average zero-shot accuracy by +0.67 to +1.69 points (Top-1) and +1.38 to +1.77 points (Top-2) over standard Top-k MoE baselines on nine zero-shot reasoning benchmarks, at the minimal cost of 2.9% added parameters and 2.6% added FLOPs per token.

cs.CL↗

arXiv2026

More Capable, Less Faithful: A Multilingual Analysis of Mathematical (Un)Solvability Detection in LLMs

Solvability detection is one of the most challenging aspects of mathematical reasoning for Large Language Models (LLMs). While prior work has studied this capability extensively, these analyses have been limited to English. Consequently, it remains unclear whether multilingual failures arise from differences in internal Solvability Belief or from language-dependent failures to express it. To address this gap, we introduce the first multilingual benchmark of paired solvable and unsolvable mathematical problems, extending ReliableMath to French and Greek. Using this, we train multilingual probes predicting Solvability Belief and analyze the solvability detection capabilities of state-of-the-art LLMs behaviorally, representationally, and in terms of faithfulness. We find that Solvability Belief is encoded as a largely universal, language-agnostic feature, and that higher-resource languages such as English, despite achieving stronger mathematical reasoning performance, exhibit lower solvability-detection faithfulness.

cs.CL↗

arXiv2026

Cognitive Profiling of LRMs' Reasoning Traces Using Bloom's Taxonomy

Large Reasoning Models (LRMs) have revolutionized reasoning in LLMs, and the increasing public availability of reasoning traces creates valuable opportunities to study model behavior not only at the surface level but also at the granularity of individual reasoning steps. However, understanding the types of thinking employed during reasoning - which offers critical insights into models' reasoning patterns and enables actionable applications - remains underexplored. To address this gap, we introduce a framework for automatic annotation of reasoning steps through the lens of Bloom's Taxonomy, which classifies thinking into six cognitive levels, such as Remembering, Applying and Evaluating. Using this framework, we perform a large-scale analysis across models and datasets, revealing both similarities and differences in thinking patterns across models and tasks. Moreover, we demonstrate that thinking-type information derived from reasoning traces correlates with correctness, paving the way for improved reasoning. Our findings establish a fine-grained framework for analyzing thinking patterns in LRMs and provide actionable insights for enhancing reasoning quality.

cs.AI↗

arXiv2026

Knowledge Knows, Verbalization Tells: Disentangling Latent Directions for Mathematical Solvability in LLMs

Although LLMs have made significant progress in mathematical reasoning, determining whether a mathematical problem is solvable remains a fundamental yet challenging capability. While recent studies have probed internal representations of model solvability beliefs, verbalization has primarily been studied behaviorally rather than as an internal representation, limiting its analysis and manipulation. We address this gap by separately probing representations of solvability knowledge and verbalization, allowing us to disentangle the two within model hidden states. Across multiple LLMs, we show that knowledge and verbalization are encoded as distinct, linearly decodable representations and that fabrication is primarily associated with changes in verbalization rather than the underlying knowledge. Prompting with unsolvability cues reduces fabrication primarily by shifting verbalization, while activation steering demonstrates that these representations can be echanistically manipulated to improve model abstention.

cs.CL↗

arXiv2026

Transcoders Trace Visual Grounding and Hallucinations in Vision-Language Models

Generative Vision-Language Models (VLMs) perform well on multimodal reasoning, but how visual inputs are transformed to text remains poorly understood. Existing interpretability work on VLMs uses Sparse Autoencoders (SAEs), which decompose static residual representations and miss the functional updates that drive cross-modal interaction. We adopt a function-centric framework based on Transcoders, sparse approximations of MLP sublayers that act as a causal proxy for layer-wise computation. Applied to Gemma 3-4B-IT, the framework decomposes the model into interpretable computational pathways linking image patches to directions in token generation. Transcoder attributions produce stronger and more stable effects on visually grounded tokens under patch ablation than SAE attributions, and align better with semantically relevant image regions. A False Visual Grounding counterfactual analysis confirms that the recovered pathways are specific to vision-language interaction.Finally, we perform a structural analysis of hallucinated generations, by extracting graph-based indicators from circuit traces produced by the transcoders. A logistic classifier over these mechanistic graph features predicts hallucinations at AUC $0.68$. These results show that function-centric circuit decomposition yields interpretable and predictive accounts of multimodal computation in VLMs.

cs.LG↗

arXiv2025

VOX-KRIKRI: Unifying Speech and Language through Continuous Fusion

We present a multimodal fusion framework that bridges pre-trained decoder-based large language models (LLM) and acoustic encoder-decoder architectures such as Whisper, with the aim of building speech-enabled LLMs. Instead of directly using audio embeddings, we explore an intermediate audio-conditioned text space as a more effective mechanism for alignment. Our method operates fully in continuous text representation spaces, fusing Whisper's hidden decoder states with those of an LLM through cross-modal attention, and supports both offline and streaming modes. We introduce \textit{VoxKrikri}, the first Greek speech LLM, and show through analysis that our approach effectively aligns representations across modalities. These results highlight continuous space fusion as a promising path for multilingual and low-resource speech LLMs, while achieving state-of-the-art results for Automatic Speech Recognition in Greek, providing an average $\sim20\%$ relative improvement across benchmarks.

cs.CL↗

arXiv2025

Masked Diffusion Language Models with Frequency-Informed Training

We present a masked diffusion language modeling framework for data-efficient training for the BabyLM 2025 Challenge. Our approach applies diffusion training objectives to language modeling under strict data constraints, incorporating frequency-informed masking that prioritizes learning from rare tokens while maintaining theoretical validity. We explore multiple noise scheduling strategies, including two-mode approaches, and investigate different noise weighting schemes within the NELBO objective. We evaluate our method on the BabyLM benchmark suite, measuring linguistic competence, world knowledge, and human-likeness. Results show performance competitive to hybrid autoregressive-masked baselines, demonstrating that diffusion-based training offers a viable alternative for data-restricted language learning.

cs.CL↗

arXiv2025

Auto-Compressing Networks

Deep neural networks with short residual connections have demonstrated remarkable success across domains, but increasing depth often introduces computational redundancy without corresponding improvements in representation quality. We introduce Auto-Compressing Networks (ACNs), an architectural variant where additive long feedforward connections from each layer to the output replace traditional short residual connections. By analyzing the distinct dynamics induced by this modification, we reveal a unique property we coin as auto-compression, the ability of a network to organically compress information during training with gradient descent, through architectural design alone. Through auto-compression, information is dynamically "pushed" into early layers during training, enhancing their representational quality and revealing potential redundancy in deeper ones. We theoretically show that this property emerges from layer-wise training patterns present in ACNs, where layers are dynamically utilized during training based on task requirements. We also find that ACNs exhibit enhanced noise robustness compared to residual networks, superior performance in low-data settings, improved transfer learning capabilities, and mitigate catastrophic forgetting suggesting that they learn representations that generalize better despite using fewer parameters. Our results demonstrate up to 18% reduction in catastrophic forgetting and 30-80% architectural compression while maintaining accuracy across vision transformers, MLP-mixers, and BERT architectures. These findings establish ACNs as a practical approach to developing efficient neural architectures that automatically adapt their computational footprint to task complexity, while learning robust representations suitable for noisy real-world tasks and continual learning scenarios.

cs.LG↗

arXiv2025

MSDA: Combining Pseudo-labeling and Self-Supervision for Unsupervised Domain Adaptation in ASR

In this work, we investigate the Meta PL unsupervised domain adaptation framework for Automatic Speech Recognition (ASR). We introduce a Multi-Stage Domain Adaptation pipeline (MSDA), a sample-efficient, two-stage adaptation approach that integrates self-supervised learning with semi-supervised techniques. MSDA is designed to enhance the robustness and generalization of ASR models, making them more adaptable to diverse conditions. It is particularly effective for low-resource languages like Greek and in weakly supervised scenarios where labeled data is scarce or noisy. Through extensive experiments, we demonstrate that Meta PL can be applied effectively to ASR tasks, achieving state-of-the-art results, significantly outperforming state-of-the-art methods, and providing more robust solutions for unsupervised domain adaptation in ASR. Our ablations highlight the necessity of utilizing a cascading approach when combining self-supervision with self-training.

cs.CL↗

arXiv2025

Krikri: Advancing Open Large Language Models for Greek

We introduce Llama-Krikri-8B, a cutting-edge Large Language Model tailored for the Greek language, built on Meta's Llama 3.1-8B. Llama-Krikri-8B has been extensively trained on high-quality Greek data to ensure superior adaptation to linguistic nuances. With 8 billion parameters, it offers advanced capabilities while maintaining efficient computational performance. Llama-Krikri-8B supports both Modern Greek and English, and is also equipped to handle polytonic text and Ancient Greek. The chat version of Llama-Krikri-8B features a multi-stage post-training pipeline, utilizing both human and synthetic instruction and preference data, by applying techniques such as MAGPIE. In addition, for evaluation, we propose three novel public benchmarks for Greek. Our evaluation on existing as well as the proposed benchmarks shows notable improvements over comparable Greek and multilingual LLMs in both natural language understanding and generation as well as code generation.

cs.CL↗

arXiv2024

BloomWise: Enhancing Problem-Solving capabilities of Large Language Models using Bloom's-Taxonomy-Inspired Prompts

Despite the remarkable capabilities of large language models (LLMs) across a range of tasks, mathematical reasoning remains a challenging frontier. Motivated by the observation that humans learn more effectively when prompted not what to think but how to think, we introduce BloomWise, a cognitively-inspired prompting technique designed to enhance LLMs' performance on mathematical problem solving while making their solutions more explainable. BloomWise encourages LLMs to generate solutions - in the form of explanations - by progressing through a sequence of cognitive operations-from basic (e.g., remembering) to more advanced reasoning skills (e.g., evaluating) - mirroring how humans build understanding. The process iterates through these levels, halting early if a convergence criterion is met: specifically, if two or more consecutive levels yield the same answer, the solution from the earliest such level is output; otherwise, the process continues until all levels are completed. Through extensive experiments across five popular math reasoning datasets, we demonstrate the effectiveness of BloomWise. We also present comprehensive ablation studies to analyze the strengths of each component within our system.

cs.CL↗

arXiv2024

Y-Drop: A Conductance based Dropout for fully connected layers

In this work, we introduce Y-Drop, a regularization method that biases the dropout algorithm towards dropping more important neurons with higher probability. The backbone of our approach is neuron conductance, an interpretable measure of neuron importance that calculates the contribution of each neuron towards the end-to-end mapping of the network. We investigate the impact of the uniform dropout selection criterion on performance by assigning higher dropout probability to the more important units. We show that forcing the network to solve the task at hand in the absence of its important units yields a strong regularization effect. Further analysis indicates that Y-Drop yields solutions where more neurons are important, i.e have high conductance, and yields robust networks. In our experiments we show that the regularization effect of Y-Drop scales better than vanilla dropout w.r.t. the architecture size and consistently yields superior performance over multiple datasets and architecture combinations, with little tuning.

cs.LG↗

arXiv2024

Meltemi: The first open Large Language Model for Greek

We describe the development and capabilities of Meltemi 7B, the first open Large Language Model for the Greek language. Meltemi 7B has 7 billion parameters and is trained on a 40 billion token Greek corpus. For the development of Meltemi 7B, we adapt Mistral, by continuous pretraining on the Greek Corpus. Meltemi 7B contains up-to-date information up to September 2023. Furthermore, we have translated and curated a Greek instruction corpus, which has been used for the instruction-tuning of a chat model, named Meltemi 7B Instruct. Special care has been given to the alignment and the removal of toxic content for the Meltemi 7B Instruct. The developed models are evaluated on a broad set of collected evaluation corpora, and examples of prompts and responses are presented. Both Meltemi 7B and Meltemi 7B Instruct are available at https://huggingface.co/ilsp under the Apache 2.0 license.

cs.CL↗

arXiv2024

The Greek podcast corpus: Competitive speech models for low-resourced languages with weakly supervised data

The development of speech technologies for languages with limited digital representation poses significant challenges, primarily due to the scarcity of available data. This issue is exacerbated in the era of large, data-intensive models. Recent research has underscored the potential of leveraging weak supervision to augment the pool of available data. In this study, we compile an 800-hour corpus of Modern Greek from podcasts and employ Whisper large-v3 to generate silver transcriptions. This corpus is utilized to fine-tune our models, aiming to assess the efficacy of this approach in enhancing ASR performance. Our analysis spans 16 distinct podcast domains, alongside evaluations on established datasets for Modern Greek. The findings indicate consistent WER improvements, correlating with increases in both data volume and model size. Our study confirms that assembling large, weakly supervised corpora serves as a cost-effective strategy for advancing speech technologies in under-resourced languages.

cs.CL↗

arXiv2024

Predicting positive transfer for improved low-resource speech recognition using acoustic pseudo-tokens

While massively multilingual speech models like wav2vec 2.0 XLSR-128 can be directly fine-tuned for automatic speech recognition (ASR), downstream performance can still be relatively poor on languages that are under-represented in the pre-training data. Continued pre-training on 70-200 hours of untranscribed speech in these languages can help -- but what about languages without that much recorded data? For such cases, we show that supplementing the target language with data from a similar, higher-resource 'donor' language can help. For example, continued pre-training on only 10 hours of low-resource Punjabi supplemented with 60 hours of donor Hindi is almost as good as continued pretraining on 70 hours of Punjabi. By contrast, sourcing data from less similar donors like Bengali does not improve ASR performance. To inform donor language selection, we propose a novel similarity metric based on the sequence distribution of induced acoustic units: the Acoustic Token Distribution Similarity (ATDS). Across a set of typologically different target languages (Punjabi, Galician, Iban, Setswana), we show that the ATDS between the target language and its candidate donors precisely predicts target language ASR performance.

eess.AS↗

arXiv2023

Investigating Personalization Methods in Text to Music Generation

In this work, we investigate the personalization of text-to-music diffusion models in a few-shot setting. Motivated by recent advances in the computer vision domain, we are the first to explore the combination of pre-trained text-to-audio diffusers with two established personalization methods. We experiment with the effect of audio-specific data augmentation on the overall system performance and assess different training strategies. For evaluation, we construct a novel dataset with prompts and music clips. We consider both embedding-based and music-specific metrics for quantitative evaluation, as well as a user study for qualitative evaluation. Our analysis shows that similarity metrics are in accordance with user preferences and that current personalization approaches tend to learn rhythmic music constructs more easily than melody. The code, dataset, and example material of this study are open to the research community.

cs.SD↗

arXiv2023

Weakly-supervised forced alignment of disfluent speech using phoneme-level modeling

The study of speech disorders can benefit greatly from time-aligned data. However, audio-text mismatches in disfluent speech cause rapid performance degradation for modern speech aligners, hindering the use of automatic approaches. In this work, we propose a simple and effective modification of alignment graph construction of CTC-based models using Weighted Finite State Transducers. The proposed weakly-supervised approach alleviates the need for verbatim transcription of speech disfluencies for forced alignment. During the graph construction, we allow the modeling of common speech disfluencies, i.e. repetitions and omissions. Further, we show that by assessing the degree of audio-text mismatch through the use of Oracle Error Rate, our method can be effectively used in the wild. Our evaluation on a corrupted version of the TIMIT test set and the UCLASS dataset shows significant improvements, particularly for recall, achieving a 23-25% relative improvement over our baselines.

eess.AS↗

arXiv2023

Depression detection in social media posts using affective and social norm features

We propose a deep architecture for depression detection from social media posts. The proposed architecture builds upon BERT to extract language representations from social media posts and combines these representations using an attentive bidirectional GRU network. We incorporate affective information, by augmenting the text representations with features extracted from a pretrained emotion classifier. Motivated by psychological literature we propose to incorporate profanity and morality features of posts and words in our architecture using a late fusion scheme. Our analysis indicates that morality and profanity can be important features for depression detection. We apply our model for depression detection on Reddit posts on the Pirina dataset, and further consider the setting of detecting depressed users, given multiple posts per user, proposed in the Reddit RSDD dataset. The inclusion of the proposed features yields state-of-the-art results in both settings, namely 2.65% and 6.73% absolute improvement in F1 score respectively. Index Terms: Depression detection, BERT, Feature fusion, Emotion recognition, profanity, morality

cs.CL↗