{"id":{"repo_id":"mit","oai_identifier":"oai:dspace.mit.edu:1721.1/156748"},"canonical_url":"https://search.dev.ndltd.org/etd/mit/oai:dspace.mit.edu:1721.1/156748","repository":{"repo_id":"mit","name":"MIT","base_url":"https://dspace.mit.edu/oai/request"},"display":{"title":"Mechanistic Interpretability for Progress Towards Quantitative AI Safety","abstract":"In this thesis, we conduct a detailed investigation into the dynamics of neural networks, focusing on two key areas: inference stages in large language models (LLMs) and novel program synthesis methods using mechanistic interpretability. We explore the robustness of LLMs through layer-level interventions such as zero-ablations and layer swapping, revealing that these models maintain high accuracy despite perturbations. As a result, we hypothesize the stages of inference in LLMs. This work suggests implications for LLM dataset curation, model optimization, and quantization. Subsequently, we introduce MIPS, an innovative method for program synthesis that distills the operational logic of neural networks into executable Python code. By transforming an RNN into a finite state machine and applying symbolic regression, MIPS successfully addresses 32 out of 62 algorithmic tasks, outperforming GPT-4 in 13 unique challenges. The work intends to take a step forward in enhancing the interpretability and reliability of AI systems, promising significant advances in our understanding and utilization of current and future AI capabilities. Together, these studies highlight the importance of comprehending the inferential behaviors of neural networks to foster more interpretable and efficient AI.","abstract_html":"In this thesis, we conduct a detailed investigation into the dynamics of neural networks, focusing on two key areas: inference stages in large language models (LLMs) and novel program synthesis methods using mechanistic interpretability. We explore the robustness of LLMs through layer-level interventions such as zero-ablations and layer swapping, revealing that these models maintain high accuracy despite perturbations. As a result, we hypothesize the stages of inference in LLMs. This work suggests implications for LLM dataset curation, model optimization, and quantization. Subsequently, we introduce MIPS, an innovative method for program synthesis that distills the operational logic of neural networks into executable Python code. By transforming an RNN into a finite state machine and applying symbolic regression, MIPS successfully addresses 32 out of 62 algorithmic tasks, outperforming GPT-4 in 13 unique challenges. The work intends to take a step forward in enhancing the interpretability and reliability of AI systems, promising significant advances in our understanding and utilization of current and future AI capabilities. Together, these studies highlight the importance of comprehending the inferential behaviors of neural networks to foster more interpretable and efficient AI.","abstract_has_math":false,"creators":["Lad, Vedang K."],"institution":"Massachusetts Institute of Technology","degree_name":"Master","degree_level":null,"degree_discipline":null,"degree_department":"Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science","school":null,"contributors":[],"advisors":["Tegmark, Max"],"committee_chairs":[],"committee_members":[],"year":2024,"date_issued":"2024-05","date_published":"2024-05","updated_at":"2026-07-22T22:21:44Z","subjects":[],"languages":[],"rights":["Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0)","Copyright retained by author(s)"],"rights_urls":["https://creativecommons.org/licenses/by-nc-nd/4.0/"],"identifier_entries":[]},"links":{"outbound_url":"https://hdl.handle.net/1721.1/156748","outbound_label":"Handle","outbound_source":"dc:identifier.uri"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:contributor.advisor","label":"Advisor","values":["Tegmark, Max"]},{"key":"dc:contributor.department","label":"Department","values":["Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science"]},{"key":"dc:creator","label":"Author","values":["Lad, Vedang K."]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date.accessioned","label":"Dc Date Accessioned","values":["2024-09-16T13:46:46Z"]},{"key":"dc:date.available","label":"Dc Date Available","values":["2024-09-16T13:46:46Z"]},{"key":"dc:date.issued","label":"Date","values":["2024-05"]},{"key":"dc:publisher","label":"Institution","values":["Massachusetts Institute of Technology"]},{"key":"dc:type","label":"Dc Type","values":["Thesis"]},{"key":"thesis:degree_name","label":"Degree Name","values":["Master","Master of Engineering in Electrical Engineering and Computer Science"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:rights","label":"Dc Rights","values":["Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0)","Copyright retained by author(s)"]},{"key":"dc:rights.uri","label":"Rights URI","values":["https://creativecommons.org/licenses/by-nc-nd/4.0/"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier.uri","label":"Identifier URI","values":["https://hdl.handle.net/1721.1/156748"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description.abstract","label":"Abstract","values":["In this thesis, we conduct a detailed investigation into the dynamics of neural networks, focusing on two key areas: inference stages in large language models (LLMs) and novel program synthesis methods using mechanistic interpretability. We explore the robustness of LLMs through layer-level interventions such as zero-ablations and layer swapping, revealing that these models maintain high accuracy despite perturbations. As a result, we hypothesize the stages of inference in LLMs. This work suggests implications for LLM dataset curation, model optimization, and quantization. Subsequently, we introduce MIPS, an innovative method for program synthesis that distills the operational logic of neural networks into executable Python code. By transforming an RNN into a finite state machine and applying symbolic regression, MIPS successfully addresses 32 out of 62 algorithmic tasks, outperforming GPT-4 in 13 unique challenges. The work intends to take a step forward in enhancing the interpretability and reliability of AI systems, promising significant advances in our understanding and utilization of current and future AI capabilities. Together, these studies highlight the importance of comprehending the inferential behaviors of neural networks to foster more interpretable and efficient AI."]},{"key":"dc:description.degree","label":"Dc Description Degree","values":["M.Eng."]},{"key":"dc:title","label":"Title","values":["Mechanistic Interpretability for Progress Towards Quantitative AI Safety"]}]}],"canonical_facts":{"dc:contributor.advisor":["Tegmark, Max"],"dc:contributor.department":["Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science"],"dc:creator":["Lad, Vedang K."],"dc:date.accessioned":["2024-09-16T13:46:46Z"],"dc:date.available":["2024-09-16T13:46:46Z"],"dc:date.issued":["2024-05"],"dc:description.abstract":["In this thesis, we conduct a detailed investigation into the dynamics of neural networks, focusing on two key areas: inference stages in large language models (LLMs) and novel program synthesis methods using mechanistic interpretability. We explore the robustness of LLMs through layer-level interventions such as zero-ablations and layer swapping, revealing that these models maintain high accuracy despite perturbations. As a result, we hypothesize the stages of inference in LLMs. This work suggests implications for LLM dataset curation, model optimization, and quantization. Subsequently, we introduce MIPS, an innovative method for program synthesis that distills the operational logic of neural networks into executable Python code. By transforming an RNN into a finite state machine and applying symbolic regression, MIPS successfully addresses 32 out of 62 algorithmic tasks, outperforming GPT-4 in 13 unique challenges. The work intends to take a step forward in enhancing the interpretability and reliability of AI systems, promising significant advances in our understanding and utilization of current and future AI capabilities. Together, these studies highlight the importance of comprehending the inferential behaviors of neural networks to foster more interpretable and efficient AI."],"dc:description.degree":["M.Eng."],"dc:identifier.uri":["https://hdl.handle.net/1721.1/156748"],"dc:publisher":["Massachusetts Institute of Technology"],"dc:rights":["Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0)","Copyright retained by author(s)"],"dc:rights.uri":["https://creativecommons.org/licenses/by-nc-nd/4.0/"],"dc:title":["Mechanistic Interpretability for Progress Towards Quantitative AI Safety"],"dc:type":["Thesis"],"thesis:degree_name":["Master","Master of Engineering in Electrical Engineering and Computer Science"]},"updated_at":"2026-07-22T22:21:44Z"}