Published Papers

Large language models (LLMs) have proven themselves to be powerful tools for many natural language tasks — from being a high-quality text classifiers to acting as agents in complex retrieval-augmented generation (RAG) systems. However, from early beggining they suffer from a major limitation: hallucinations, i.e. confidently generating incorrect or misleading information that can also slightly correlate with the given task. This issue is critical in error-sensitive domains such as finance, medicine, and law, where even small inaccuracies can cause significant harm and detriment. In this study we address the early detection of hallucinating answers based on user input (prompt), answer by the LLM, and which is more important — token-level probabilty signals that can also be extracted from the LLM during its inference time. We constructed a dataset that combines textual information with sequences of token log-probabilities and their statistics (mean, min, variance, percentiles, etc.), labeled the answers whether they are hallucinations or not. We trained a lightweight classifier that outputs the probability that a given response is a hallucination. We evaluate the classifier and perform ablation studies to quantify the contribution of token-level signals versus text-only features. The intended use of the trained model is to be a standalone output guard agent in multi-agent system that rejects the answer of LLM-generator if its hallucination probability is above acceptance threshold and protects the users of it from having incorrect or misleading answer by making the whole system regenerate such answer or confirm that it cannot give the faithfull reply.

Modern automated resume screening systems are typically based on neural text classification models that encode a resume as a feature representation and predict a discrete label corresponding to candidate category, suitability level, or job role. Such models commonly produce class logits parameterized by model weights, which are converted into class probabilities via the softmax function over the target classes. These models are typically trained using cross-entropy loss and deployed as the first stage of automated candidate filtering. Despite their effectiveness, resume classifiers may encode implicit bias through correlations between predictions and non-job-related or proxy textual features. To study this effect, we analyze feature influence using Integrated Gradients, which assign an attribution score to each input feature based on the path integral of partial derivatives between a baseline representation and the actual input. This analysis reveals systematic dependencies on features that should be irrelevant to candidate evaluation. Building on these observations, we evaluate multiple debiasing techniques and propose an interpretability-guided framework for bias mitigation. We compare six methods spanning in-processing approaches (GroupDRO, Focal Loss, Label Smoothing, Adversarial debiasing) and attribution-based approaches (Data Scrubbing, Attention Regularization) that leverage the interpretability findings directly. This allows explainable analysis to guide the development of fairer resume screening models.