Dmytro Vitel
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Dmytro Vitel has a total of 7 co-authors. Here are the top co-authors by number of shared publications:
- Alessio Gaspar
University of South Florida, Tampa, USA
3 shared publications - Kok Cheng Tan
University of South Florida, Tampa, USA
2 shared publications - R. Paul Wiegand
Winthrop University, Rock Hill, South Carolina, USA
2 shared publications - Anshuman Chhabra
University of South Florida
1 shared publication - Daniel Zantedeschi
Muma College of Business, School of Information Systems, University of South Florida, Tampa, Florida 33620
1 shared publication - John Licato
University of South Florida
1 shared publication
Latest Publications
- Open Access
First is Not Really Better Than Last: Evaluating Layer Choice and Aggregation Strategies in Language Model Data Influence Estimation
Identifying how training samples influence/impact Large Language Model (LLM) decision-making is essential for effectively interpreting model decisions and auditing large-scale datasets. Current training sample influence estimation methods (also known as influence functions) undertake this goal by utilizing information flow through the model via its first-order and higher-order gradient terms. However, owing to the large model sizes of today consisting of billions of parameters, these influence computations are often restricted to some subset of model layers to ensure computational feasibility. Prior seminal work by Yeh et al. (2022) in assessing which layers are best suited for computing language data influence concluded that the first (embedding) layers are the most informative for this purpose, using a hypothesis based on influence scores canceling out (i.e., the cancellation effect). In this work, we propose theoretical and empirical evidence demonstrating how the cancellation effect is unreliable, and that middle attention layers are better estimators for influence. Furthermore, we address the broader challenge of aggregating influence scores across layers, and showcase how alternatives to standard averaging (such as ranking and vote-based methods) can lead to significantly improved performance. Finally, we propose better methods for evaluating influence score efficacy in LLMs without undertaking model retraining, and propose a new metric known as the Noise Detection Rate (NDR) that exhibits strong predictive capability compared to the cancellation effect. Through extensive experiments across LLMs of varying types and scales, we concretely determine that the first (layers) are not necessarily better than the last (layers) for LLM influence estimation, contrasting with prior knowledge in the field.
Open AccessFirst is Not Really Better Than Last: Evaluating Layer Choice and Aggregation Strategies in Language Model Data Influence Estimation
Identifying how training samples influence/impact Large Language Model (LLM) decision-making is essential for effectively interpreting model decisions and auditing large-scale datasets. Current training sample influence estimation methods (also known as influence functions) undertake this goal by utilizing information flow through the model via its first-order and higher-order gradient terms. However, owing to the large model sizes of today consisting of billions of parameters, these influence computations are often restricted to some subset of model layers to ensure computational feasibility. Prior seminal work by Yeh et al. (2022) in assessing which layers are best suited for computing language data influence concluded that the first (embedding) layers are the most informative for this purpose, using a hypothesis based on influence scores canceling out (i.e., the cancellation effect). In this work, we propose theoretical and empirical evidence demonstrating how the cancellation effect is unreliable, and that middle attention layers are better estimators for influence. Furthermore, we address the broader challenge of aggregating influence scores across layers, and showcase how alternatives to standard averaging (such as ranking and vote-based methods) can lead to significantly improved performance. Finally, we propose better methods for evaluating influence score efficacy in LLMs without undertaking model retraining, and propose a new metric known as the Noise Detection Rate (NDR) that exhibits strong predictive capability compared to the cancellation effect. Through extensive experiments across LLMs of varying types and scales, we concretely determine that the first (layers) are not necessarily better than the last (layers) for LLM influence estimation, contrasting with prior knowledge in the field.
- Open Access
- 14 July 2025
- Published by Association for Computing Machinery (ACM)
- p. 911-914
Adaptive Denoising of Sequential Data with Multi-Objective Genetic Algorithms
- 14 July 2025
- Published by Association for Computing Machinery (ACM)
- p. 911-914
Open AccessAdaptive Denoising of Sequential Data with Multi-Objective Genetic Algorithms
- Open Access
- 13 July 2025
- Published by Association for Computing Machinery (ACM)
- p. 1081-1089
Coordinate System Extraction as the Search Driver in Test-Based Genetic Programming
- 13 July 2025
- Published by Association for Computing Machinery (ACM)
- p. 1081-1089
Open AccessCoordinate System Extraction as the Search Driver in Test-Based Genetic Programming
- 17 April 2025
- Published by Springer Nature
- p. 541-556
Robust Search for the Underlying Objectives in Black-Box Games with Binary Outcomes
- 17 April 2025
- Published by Springer Nature
- p. 541-556
Robust Search for the Underlying Objectives in Black-Box Games with Binary Outcomes
- Open Access
- 8 May 2023
- Published by University of Florida George A Smathers Libraries in The International FLAIRS Conference Proceedings
- Vol. 36
Enforcing Grammar in Code Synthesis with Transformers
Even more so than natural language, code is extremely sensitive to syntax; a small error could make an entire snippet invalid. It is therefore important to explore methods for ensuring syntactic correctness in generated code. Existing methods to resolve this issue often rely on the complex architecture of syntax-guided decoders. In this work, we present the grammar enforcement method, which introduces a separate layer that constrains the decisions of the transformer during fine-tuning according to syntactic constructs present both in the target language grammar and the given training set. We experiment with the Hearthstone dataset to study its effects on result programs and compare it with the existing state-of-art syntax-guided decoders. We demonstrate a statistically significant positive effect of grammar enforcement on the quality of generated programs in terms of exact match accuracy and grammatically correct percent of samples. At the same time, we observe lower values for text-based metrics, chrF, and BLEU, potentially indicating their inability to represent the quality of generated abstract syntax sequences.
- 8 May 2023
- Published by University of Florida George A Smathers Libraries in The International FLAIRS Conference Proceedings
- Vol. 36
Open AccessEnforcing Grammar in Code Synthesis with Transformers
Even more so than natural language, code is extremely sensitive to syntax; a small error could make an entire snippet invalid. It is therefore important to explore methods for ensuring syntactic correctness in generated code. Existing methods to resolve this issue often rely on the complex architecture of syntax-guided decoders. In this work, we present the grammar enforcement method, which introduces a separate layer that constrains the decisions of the transformer during fine-tuning according to syntactic constructs present both in the target language grammar and the given training set. We experiment with the Hearthstone dataset to study its effects on result programs and compare it with the existing state-of-art syntax-guided decoders. We demonstrate a statistically significant positive effect of grammar enforcement on the quality of generated programs in terms of exact match accuracy and grammatically correct percent of samples. At the same time, we observe lower values for text-based metrics, chrF, and BLEU, potentially indicating their inability to represent the quality of generated abstract syntax sequences.