M. L. Cummings
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Research Interests
M. L. Cummings has a total of 63 co-authors. Here are the top co-authors by number of shared publications:
- Benjamin Bauchwitz
Duke University, Durham, NC, USA
3 shared publications Songpo Li
Duke University, USA
3 shared publicationsWanyi Chen
Duke University Durham NC
3 shared publicationsPaul W. Quimby
Massachusetts Institute of Technology, Cambridge, MA, USA
2 shared publicationsA. Mkrtchyan
Massachusetts Institute of Technology, Massachusetts Avenue 33-311, Cambridge, MA, UK
1 shared publicationA.S. Clare
Aeronautics & Astronautics, Massachusetts Institute of Technology, 77 Massachusetts Ave, Cambridge, MA 02139, USA
1 shared publication
Latest Publications
- Open Access
- 29 June 2026
- Published by Elsevier in SSRN Electronic Journal
Breaking Bad Intentions: Identifying Semantic Drift in Multi-Turn Language Model Inputs
Multi-turn interaction fundamentally changes the safety landscape for large language models, as potentially harmful intent can be distributed across multiple individually benign queries. Existing defenses struggle to detect coordinated misuse that emerges at the level of conversational trajectories. This paper addresses this gap by framing deceptive questioning as a problem of semantic drift in embedding space. Each multi-turn user interaction is represented as a set of embedded questions anchored to a fixed reference vector, from which a compact set of geometric and similarity-based features is extracted to capture angular coverage, distance structure, linearity, and internal semantic coherence. Using these features, we trained a lightweight neural classifier to distinguish non-deceptive from deceptively coordinated question sets attepting to illicit information about developed methamphetamine in a fully black-box, transcript- only setting. The approach is evaluated across four experimental conditions. Results show strong performance across experimental settings. We also conducted a cross-model generalization analysis using an external dataset constructed around a different harmful task. Despite this shift in domain, the proposed features retain meaningful discriminative power, indicating that they capture structural properties of coordinated deceptive questioning rather than topic-specific artifacts. Detailed error analyses further reveal that false negatives arise primarily from geometric overlap near the decision boundary, while false positives remain partially sensi- tive to specific lexical cues. Together, the results demonstrate that multi-turn geometric structure provides a scalable foundation for deception detection, while underscoring the importance of jointly considering geometric and lexical effects in robust LLM safety monitoring
- 29 June 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessBreaking Bad Intentions: Identifying Semantic Drift in Multi-Turn Language Model Inputs
Multi-turn interaction fundamentally changes the safety landscape for large language models, as potentially harmful intent can be distributed across multiple individually benign queries. Existing defenses struggle to detect coordinated misuse that emerges at the level of conversational trajectories. This paper addresses this gap by framing deceptive questioning as a problem of semantic drift in embedding space. Each multi-turn user interaction is represented as a set of embedded questions anchored to a fixed reference vector, from which a compact set of geometric and similarity-based features is extracted to capture angular coverage, distance structure, linearity, and internal semantic coherence. Using these features, we trained a lightweight neural classifier to distinguish non-deceptive from deceptively coordinated question sets attepting to illicit information about developed methamphetamine in a fully black-box, transcript- only setting. The approach is evaluated across four experimental conditions. Results show strong performance across experimental settings. We also conducted a cross-model generalization analysis using an external dataset constructed around a different harmful task. Despite this shift in domain, the proposed features retain meaningful discriminative power, indicating that they capture structural properties of coordinated deceptive questioning rather than topic-specific artifacts. Detailed error analyses further reveal that false negatives arise primarily from geometric overlap near the decision boundary, while false positives remain partially sensi- tive to specific lexical cues. Together, the results demonstrate that multi-turn geometric structure provides a scalable foundation for deception detection, while underscoring the importance of jointly considering geometric and lexical effects in robust LLM safety monitoring
- Open Access
- 19 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 9778-9787
Modeling Image Segmentation Label Noise Through Concave Region Removal
The effects of label noise geometry on image segmentation model training are under-studied, in part because prevailing segmentation metrics such as the Intersection over Union are based on pixel counts and cannot describe spatial properties of label noise. It is well established that the degree of label noise randomness affects training, however, the effects of non-random geometric label noise have not been evaluated. To this end, we describe a new class of non-random geometric label noise characterized by removal of concave regions from connected components in segmentation masks. This class of label noise geometry is motivated by the observation that both human and automated annotations have vulnerabilities that disproportionately affect concave regions. We validate this class of label noise by demonstrating its prevalence in a dataset containing noisy crowdsourced labels for the PASCAL, BDD100K, and COCO benchmarks. We then show that concave region label noise affects model training differently than uniformly applied geometric transformations across a range of model architectures and image domains. These results motivate the study of new segmentation metrics better tuned to detect label noise geometries most detrimental to model training.
- 19 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 9778-9787
Open AccessModeling Image Segmentation Label Noise Through Concave Region Removal
The effects of label noise geometry on image segmentation model training are under-studied, in part because prevailing segmentation metrics such as the Intersection over Union are based on pixel counts and cannot describe spatial properties of label noise. It is well established that the degree of label noise randomness affects training, however, the effects of non-random geometric label noise have not been evaluated. To this end, we describe a new class of non-random geometric label noise characterized by removal of concave regions from connected components in segmentation masks. This class of label noise geometry is motivated by the observation that both human and automated annotations have vulnerabilities that disproportionately affect concave regions. We validate this class of label noise by demonstrating its prevalence in a dataset containing noisy crowdsourced labels for the PASCAL, BDD100K, and COCO benchmarks. We then show that concave region label noise affects model training differently than uniformly applied geometric transformations across a range of model architectures and image domains. These results motivate the study of new segmentation metrics better tuned to detect label noise geometries most detrimental to model training.
- 16 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Systems, Man, and Cybernetics Magazine
- Vol. 12 (1) , 118-127
R4 Trustworthy Human–Artificial Intelligence Symbiosis [Conference Reports]
Provides society information that may include news, reviews or technical notes that should be of interest to practitioners and researchers.
- 16 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Systems, Man, and Cybernetics Magazine
- Vol. 12 (1) , 118-127
R4 Trustworthy Human–Artificial Intelligence Symbiosis [Conference Reports]
Provides society information that may include news, reviews or technical notes that should be of interest to practitioners and researchers.
- Open Access
Assessing LLM code generation quality through path planning tasks
As LLM-generated code grows in popularity, more evaluation is needed to assess the risks of using such tools, especially for safety-critical applications such as path planning. Existing coding benchmarks are insufficient as they do not reflect the context and complexity of safety-critical applications. To this end, we assessed six LLMs' abilities to generate the code for three different path-planning algorithms and tested them on three maps of various difficulties. Our results suggest that LLM-generated code presents serious hazards for path planning applications and should not be applied in safety-critical contexts without rigorous testing.
Open AccessAssessing LLM code generation quality through path planning tasks
As LLM-generated code grows in popularity, more evaluation is needed to assess the risks of using such tools, especially for safety-critical applications such as path planning. Existing coding benchmarks are insufficient as they do not reflect the context and complexity of safety-critical applications. To this end, we assessed six LLMs' abilities to generate the code for three different path-planning algorithms and tested them on three maps of various difficulties. Our results suggest that LLM-generated code presents serious hazards for path planning applications and should not be applied in safety-critical contexts without rigorous testing.
- Open Access
- 17 March 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 54338-54349
Identifying AI Hazards and Responsibility Gaps
With the expansion of artificial intelligence (AI) in general, but especially in self-driving cars in the United States, there is an increasing need for new hazard analysis approaches to determine if and how AI contributes to accidents. The original Swiss Cheese model widely used for hazard analyses focuses uniquely on human activities that lead to accidents, but cannot address accidents where AI is a possible causal factor. To this end, the Taxonomy for AI Hazard Analysis (TAIHA) is proposed that introduces layers focusing on the oversight, design, maintenance and testing of AI. Three case studies demonstrate how TAIHA can be used to better understand how AI contributes to accidents. Resulting TAIHA analyses suggest responsibility gaps are emerging in the self-driving car community in the US where humans, either unknowingly or on purpose, side step accountability, leading to serious consequences. Recommendations for mitigating these gaps include more regulatory involvement, as well as the need for company-appointed senior lead test engineers, who should become the new first-line actors in any autonomous system operation.
- 17 March 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 54338-54349
Open AccessIdentifying AI Hazards and Responsibility Gaps
With the expansion of artificial intelligence (AI) in general, but especially in self-driving cars in the United States, there is an increasing need for new hazard analysis approaches to determine if and how AI contributes to accidents. The original Swiss Cheese model widely used for hazard analyses focuses uniquely on human activities that lead to accidents, but cannot address accidents where AI is a possible causal factor. To this end, the Taxonomy for AI Hazard Analysis (TAIHA) is proposed that introduces layers focusing on the oversight, design, maintenance and testing of AI. Three case studies demonstrate how TAIHA can be used to better understand how AI contributes to accidents. Resulting TAIHA analyses suggest responsibility gaps are emerging in the self-driving car community in the US where humans, either unknowingly or on purpose, side step accountability, leading to serious consequences. Recommendations for mitigating these gaps include more regulatory involvement, as well as the need for company-appointed senior lead test engineers, who should become the new first-line actors in any autonomous system operation.
- Open Access
- 4 December 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Vehicles
- Vol. 10 (11) , 4903-4912
Identifying Research Gaps Through Self-Driving Car Data Analysis
There are currently around thirty companies testing self-driving cars in San Francisco, CA, effectively creating a living laboratory. Of these companies, only Waymo is engaged in commercial operations, while Zoox conducts routine driverless testing operations in San Francisco. Despite these successes, federal investigations have been opened into both companies for safety concerns, and Cruise is attempting to reinstate its permit after a near-fatal pedestrian crash. An analysis of these three companies' crash data from required reporting illustrates that many areas of self-driving need improvement. The most significant crash type for Waymo and Zoox are struck-from-behind events, while Cruise struggled most with unexpected actions by others. Computer vision systems are very brittle and likely play an outsized role in crashes. Self-driving cars also struggle to reason under uncertainty, and simulations are not effectively bridging the physical-to-real-world testing gap. This analysis underscores that research is lacking, especially for artificial intelligence involving computer vision and reasoning under uncertainty.
- 4 December 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Vehicles
- Vol. 10 (11) , 4903-4912
Open AccessIdentifying Research Gaps Through Self-Driving Car Data Analysis
There are currently around thirty companies testing self-driving cars in San Francisco, CA, effectively creating a living laboratory. Of these companies, only Waymo is engaged in commercial operations, while Zoox conducts routine driverless testing operations in San Francisco. Despite these successes, federal investigations have been opened into both companies for safety concerns, and Cruise is attempting to reinstate its permit after a near-fatal pedestrian crash. An analysis of these three companies' crash data from required reporting illustrates that many areas of self-driving need improvement. The most significant crash type for Waymo and Zoox are struck-from-behind events, while Cruise struggled most with unexpected actions by others. Computer vision systems are very brittle and likely play an outsized role in crashes. Self-driving cars also struggle to reason under uncertainty, and simulations are not effectively bridging the physical-to-real-world testing gap. This analysis underscores that research is lacking, especially for artificial intelligence involving computer vision and reasoning under uncertainty.
- Open Access
- 16 October 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in Computer
- Vol. 57 (11) , 31-40
A Root Cause Analysis of a Self-Driving Car Dragging a Pedestrian
Using the Taxonomy for Artificial Intelligence Hazard Analysis to examine the recent accident where a pedestrian was dragged by a self-driving car, major themes include lacking consideration for remote operations issues, significant post-collision assessment deficiencies, and a lack of periodic design and test protocol reviews.
- 16 October 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in Computer
- Vol. 57 (11) , 31-40
Open AccessA Root Cause Analysis of a Self-Driving Car Dragging a Pedestrian
Using the Taxonomy for Artificial Intelligence Hazard Analysis to examine the recent accident where a pedestrian was dragged by a self-driving car, major themes include lacking consideration for remote operations issues, significant post-collision assessment deficiencies, and a lack of periodic design and test protocol reviews.
- Open Access
- 20 May 2024
- Published by Association for the Advancement of Artificial Intelligence (AAAI) in Proceedings of the AAAI Symposium Series
- Vol. 3 (1) , 22-29
Subjectivity in Unsupervised Machine Learning Model Selection
Model selection is a necessary step in unsupervised machine learning. Despite numerous criteria and metrics, model selection remains subjective. A high degree of subjectivity may lead to questions about repeatability and reproducibility of various machine learning studies and doubts about the robustness of models deployed in the real world. Yet, the impact of modelers' preferences on model selection outcomes remains largely unexplored. This study uses the Hidden Markov Model as an example to investigate the subjectivity involved in model selection. We asked 33 participants and three Large Language Models (LLMs) to make model selections in three scenarios. Results revealed variability and inconsistencies in both the participants’ and the LLMs' choices, especially when different criteria and metrics disagree. Sources of subjectivity include varying opinions on the importance of different criteria and metrics, differing views on how parsimonious a model should be, and how the size of a dataset should influence model selection. The results underscore the importance of developing a more standardized way to document subjective choices made in model selection processes.
- 20 May 2024
- Published by Association for the Advancement of Artificial Intelligence (AAAI) in Proceedings of the AAAI Symposium Series
- Vol. 3 (1) , 22-29
Open AccessSubjectivity in Unsupervised Machine Learning Model Selection
Model selection is a necessary step in unsupervised machine learning. Despite numerous criteria and metrics, model selection remains subjective. A high degree of subjectivity may lead to questions about repeatability and reproducibility of various machine learning studies and doubts about the robustness of models deployed in the real world. Yet, the impact of modelers' preferences on model selection outcomes remains largely unexplored. This study uses the Hidden Markov Model as an example to investigate the subjectivity involved in model selection. We asked 33 participants and three Large Language Models (LLMs) to make model selections in three scenarios. Results revealed variability and inconsistencies in both the participants’ and the LLMs' choices, especially when different criteria and metrics disagree. Sources of subjectivity include varying opinions on the importance of different criteria and metrics, differing views on how parsimonious a model should be, and how the size of a dataset should influence model selection. The results underscore the importance of developing a more standardized way to document subjective choices made in model selection processes.
- 9 January 2024
- Published by SAGE Publications in Journal of Cognitive Engineering and Decision Making
- Vol. 18 (4) , 327-332
A Taxonomy for AI Hazard Analysis
With the rise of artificial intelligence in safety-critical systems like surface transportation, there is a commensurate need for new hazard analysis approaches to determine if and how AI contributes to accidents, which are also increasing in number and severity. The original Swiss Cheese model widely used for hazard analyses focuses uniquely on human activities that lead to accidents, but cannot address accidents where AI is a possible causal factor. To this end, the Taxonomy for AI Hazard Analysis (TAIHA) is proposed that introduces layers focusing on the oversight, design, maintenance, and testing of AI. TAIHA is illustrated with real-world accidents. TAIHA does not replace the traditional Swiss Cheese model, which should be used in concert when a joint human-AI system exists, such as when people are driving a car with AI-based advanced driving assist features.
- 9 January 2024
- Published by SAGE Publications in Journal of Cognitive Engineering and Decision Making
- Vol. 18 (4) , 327-332
A Taxonomy for AI Hazard Analysis
With the rise of artificial intelligence in safety-critical systems like surface transportation, there is a commensurate need for new hazard analysis approaches to determine if and how AI contributes to accidents, which are also increasing in number and severity. The original Swiss Cheese model widely used for hazard analyses focuses uniquely on human activities that lead to accidents, but cannot address accidents where AI is a possible causal factor. To this end, the Taxonomy for AI Hazard Analysis (TAIHA) is proposed that introduces layers focusing on the oversight, design, maintenance, and testing of AI. TAIHA is illustrated with real-world accidents. TAIHA does not replace the traditional Swiss Cheese model, which should be used in concert when a joint human-AI system exists, such as when people are driving a car with AI-based advanced driving assist features.
- Open Access
Subjectivity in Unsupervised Machine Learning Model Selection
Model selection is a necessary step in unsupervised machine learning. Despite numerous criteria and metrics, model selection remains subjective. A high degree of subjectivity may lead to questions about repeatability and reproducibility of various machine learning studies and doubts about the robustness of models deployed in the real world. Yet, the impact of modelers' preferences on model selection outcomes remains largely unexplored. This study uses the Hidden Markov Model as an example to investigate the subjectivity involved in model selection. We asked 33 participants and three Large Language Models (LLMs) to make model selections in three scenarios. Results revealed variability and inconsistencies in both the participants' and the LLMs' choices, especially when different criteria and metrics disagree. Sources of subjectivity include varying opinions on the importance of different criteria and metrics, differing views on how parsimonious a model should be, and how the size of a dataset should influence model selection. The results underscore the importance of developing a more standardized way to document subjective choices made in model selection processes.
Open AccessSubjectivity in Unsupervised Machine Learning Model Selection
Model selection is a necessary step in unsupervised machine learning. Despite numerous criteria and metrics, model selection remains subjective. A high degree of subjectivity may lead to questions about repeatability and reproducibility of various machine learning studies and doubts about the robustness of models deployed in the real world. Yet, the impact of modelers' preferences on model selection outcomes remains largely unexplored. This study uses the Hidden Markov Model as an example to investigate the subjectivity involved in model selection. We asked 33 participants and three Large Language Models (LLMs) to make model selections in three scenarios. Results revealed variability and inconsistencies in both the participants' and the LLMs' choices, especially when different criteria and metrics disagree. Sources of subjectivity include varying opinions on the importance of different criteria and metrics, differing views on how parsimonious a model should be, and how the size of a dataset should influence model selection. The results underscore the importance of developing a more standardized way to document subjective choices made in model selection processes.