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.

Hyperspectral images contain a large volume of source data that exhibits high correlations along neighboring spectral bands. This makes it necessary to select the most informative features among correlated groups of features to effectively solve various machine learning problems. A method of feature importance evaluation for hyperspectral image data is proposed. This method combines iterative training of Decision Tree classifiers based on spectral features with matrix factorization to overcome sparsity. Decision trees provide intrinsic feature selection mechanism but only a small number of features are usually taken into account by the CART algorithm for training a single decision tree classifier instance. Furthermore when features are highly correlated (e.g., Pearson $\rho > 0.8$), tree-based methods like Random Forest or XGBoost arbitrarily assign importance to one feature while suppressing others, as they redundantly capture the same signal. To overcome this problem, an additional balancing term was incorporated into the optimization function used to obtain the matrix factorization. The considered method of feature importance evaluation is compared with such model-specific tree-based methods as vanilla Gini impurity decrease and more complicated Boruta algorithm. Classification accuracy is tested using a Random Forest classifier on significant features. Selecting features with higher importance scores yields models boasting greater training accuracy.