56 Harvard Stem Cell Institute Internship Program Effect of DNMT1-inhibiting RNA Aptamer in MDS Mice Models

Authors:

Deborah Wells, Lucrezia Rinaldi, Annalisa Di Ruscio

Date Created:

2025-01-01

Course Title:

Professor:

Not specified

About Paper:

Myelodysplastic syndromes (MDS) are a category of bone marrow Tet2 and Asxl1 loss of function mutations, commonly found in disorders that result from dysplastic hematopoietic stem cells MDS patients. Flow cytometry data was collected periodically failing to differentiate into healthy myeloid cells.In 30-40% to monitor the engraftments using peripheral blood from the 15 of cases MDS can progress to acute myeloid leukemia (AML). mice. The data analyzed after 9 weeks showed active engrafted Aberrant methylation in tumor suppressor gene promoters is a cell populations in all the mice, and at 12 weeks there was a clear prominent mechanism that contributes to the progression of MDS. expansion of myeloid cell counts in the engrafted populations. The Current hypomethylation treatments, such as azacytidine, have morphology of cell dysplasia in the mice was analyzed with a serious side effects as they are toxic and non-specific. This peripheral blood smear at 12 weeks. Three week treatments will project looks to test the effectiveness of AptaDiR, an RNA-based then be implemented with PBS as a control, azacytidine, and the therapeutic, for MDS treatment. AptaDiR inhibits DNMT1, a AptaDiR with two different lipid nanoparticles (LNPs), as carriers. DNA methyltransferase protein that methylates new strands of ThiswillallowforanalysisontheimpactAptaDiRhasonengrafted DNA during replication, mirroring the methylation on the templatecells modeling MDS, and its effectiveness compared to existing strand. This low toxicity therapeutic consists of an aptamer, a hypomethylation agents. Understanding how RNA-therapeutics single-stranded piece of RNA that can competitively bind and such as AptaDiR can affect methylation in MDS could provide inhibit DNMT1. To model MDS, 15 congenic mice including one important insight into treatment for other diseases also fueled by control, were used with engrafted bone marrow from mice with excess DNA methylation, such as cancer. Kempner Research in Artificial and Natural Intelligence for Undergraduates with Mentorship 60 Kempner Research in Artificial and Natural Intelligence for Undergraduates with Mentorship Infant Eye-Gaze and the Role of Collaborative Understanding in Human Intelligence Victoria Chen, Hannah Hok Kim, Ashley Thomas Harvard College | Leverett House | Psychology | 2028 Collaboration — our ability to pool knowledge and resources rich understanding of 1) infants see people as agents and attribute towards attaining more complex goals — is a defining aspect goals to the actions they perform, 2) collaborators engaging in of human intelligence. Prior research has shown it to be complementary actions likely have a shared goal , and 3) that effort uniquely adaptive, socially embedded, and powerful from early andefficiencyarekeymetricsingoalattainment. Thissupportsthe in development. But if infants don’t possess the adequate verbal results of our initial pilot study, where infants recognize and show and motor skills to engage in goal-directed actions, how do a violation of expectation — characterized by prolonged look- they develop this understanding of task difficulty, complementary time — when shown the unexpected event of fewer collaborators actions, and the relationship between effort and performance at leading to better outcomes in the building task. Additionally, such an early age? This study addresses how infants can still eye gaze is directed mainly at the single builder, showing that quickly acquire initial expectations of collaboration and its relatedthe surprise of seeing a single builder overachieving surpasses effects. We showed 16 infants videos of two task-completion that of two collaborators underachieving. Ultimately, we hope scenarios — one with a solo builder and the other with two that studying infants’ ability to recognize collaboration as an collaborators. This was followed by an analysis of their looking enhancer of task performance will provide a basis for future behavior using iCatcher+, a machine learning-based program that research surrounding humans’ rapid behavioral learning, and early codes for infants’ gaze direction and duration. Previous literature tendencies to be social agents. has established that, by the age of 16-20 months, infants possess a Diffusion Models and their Relational Compositions Hannah Kim, Binxu Wang Harvard College | Currier House | Computer Science | 2026 Recent advances in text-to-image machine learning models like accuracy, we analyzed cross-attention maps that revealed how text Stable Diffusion and DALL-E have enabled the generation of tokens influenced different parts of the image during generation. highly realistic images from natural language prompts. However, The controlled setting allowed us to isolate how relational these models often struggle to understand relational prompts, like understanding emerges or fails within the model. Preliminary “cat on top of iguana” and the causes of these errors remain results suggest that the model can learn simple spatial relations unclear. To better understand this failure mode, we investigated such as left-right positioning, but not reliably. There are instances PixArt-alpha, a transformer-based text-to-image diffusion model where small variations, such as the existence of articles in the that gained popularity because of its ability to produce high qualitytraining text prompts, can surprisingly modulate the accuracy of images with low training costs. We constructed a synthetic dataset generation, suggestinghowthefailuremodeisnotsolelyattributed of images with simple objects and colors paired with structured to model architecture. Our work contributes to broader efforts to relational captions like “Object A [relation] Object B” and trained interpret the internal mechanisms of diffusion models and offers a the PixArt model on it. In addition to evaluating generation framework for diagnosing and improving their reasoning abilities. Examining User Gender-Based Variations in LLM Response Trends Pei Yao Simon Ma, Naomi Saphra, Isabel Papadimitriou (in collaboration with Laasya Nagumalli, Kempner Fel- low) Harvard College | Pforzheimer House | Statistics | 2026 A large language model (LLM), upon being prompted with a sought further clarification. We generated 590 responses for each question, should respond with follow-up questions when the query and applied uniform manifold approximation and projection model has insufficient information to answer the user query. (UMAP) to reduce the dimensionality of the data, followed by This paper investigates whether desirable behaviors like follow- clustering to investigate patterns. Preliminary results show the up questions are subject to systematic biases. To this end, model is more likely to make assumptions about the user when the we analyze LLM responses from Cohere Command R+ and prompt contains a female persona and more likely to ask clarifying Llama 3.1-8B to twenty demographic-neutral queries about food, questionswhenthepersonaismale. Furthermore,responsestomen health, clothing, and more. Each prompt was prefixed with included more epistemic modal verbs like “can” while responses a gendered persona. These prompts were intentionally under- to women included more deontic modal verbs like “should.” If specified, requiring further clarification from the LLM for a full these preliminary results hold, these trends imply that female users response to the request. We classified the responses into two receive lower quality, less personalized responses from LLMs than categories: thosethatweredirect repliesto thequery andthose that male users. Sample Efficient Generalization to Very Different New Actions in Reinforcement Learning Ian Moore, Susan Murphy Harvard College | Quincy House | Applied Math | 2026 Insettingswheredataisexpensiveorhardtocollect,generalization new actions. In some settings such as digital advertising it may of reinforcement learning models is very useful as it allows us to be reasonable to assume the new actions will be similar to or adapt to a new RL environment while using the data we already have dynamics that are a mixture of that of the existing actions. have, reducing the need to collect new data. We consider the However in settings such as RL for digital health interventions and case where the RL model must generalize to new actions that robotics it is likely that some new actions will be very different are introduced. Existing work attempts to solve this problem, from the existing actions because they result from advances in but we show empirical results indicating these methods perform technology or science that enable the RL model to interact with poorly when the new actions are very different than the existing the environment in entirely new ways not previously seen or even actions we have data on — even in a very simple environment conceivable. Thus, this methodology will make sample efficient with discrete states and actions. This points to the need for a newaction generalization possible in these settings, enabling more methodology for sample efficient adaptation to the very different applications and capability for reinforcement learning. 62 Kempner Research in Artificial and Natural Intelligence for Undergraduates with Mentorship Examining User Gender-Based Variations in LLM Response Trends Laasya Nagumalli, Isabel Papadimitriou, Naomi Saphra (in collaboration with Pei Yao Simon Ma, Kempner Fel- low) Harvard College | Currier House | Mathematics | 2028 A large language model (LLM), upon being prompted with a sought further clarification. We generated 590 responses for each question, should respond with follow-up questions when the query and applied uniform manifold approximation and projection model has insufficient information to answer the user query. (UMAP) to reduce the dimensionality of the data, followed by This paper investigates whether desirable behaviors like follow- clustering to investigate patterns. Preliminary results show the up questions are subject to systematic biases. To this end, model is more likely to make assumptions about the user when the we analyze LLM responses from Cohere Command R+ and prompt contains a female persona and more likely to ask clarifying Llama 3.1-8B to twenty demographic-neutral queries about food, questionswhenthepersonaismale. Furthermore,responsestomen health, clothing, and more. Each prompt was prefixed with included more epistemic modal verbs like “can” while responses a gendered persona. These prompts were intentionally under- to women included more deontic modal verbs like “should.” If specified, requiring further clarification from the LLM for a full these preliminary results hold, these trends imply that female users response to the request. We classified the responses into two receive lower quality, less personalized responses from LLMs than categories: thosethatweredirect repliesto thequery andthose that male users. Interpreting and Manipulating Language Model Embeddings via Concept Manifolds Carl Scandelius, Pranav Misra, Haim Sompolinsky Harvard College | Winthrop House | Mathematics | 2027 Representations in language models (LMs) remain largely each transformer layer; (2) analyse the interpretability of the basis uninterpretable. We propose a geometric framework for analysing LMrepresentationsusingconceptmanifolds—pointcloudsformed vectors of these manifolds; (3) use this framework to guide causal intervention on LM embeddings to generate predictably altered by the final-token residual-stream embeddings of sentences output. representing a single dominant concept. These manifolds enable the definition of theoretically grounded geometric measures that Insights from this project hope to inform safety-minded efforts directly relate to downstream task performance. This framework like discovering misaligned latent knowledge, as well as fine- builds on successful applications to vision and vision-language tuning strategies or perturbation-based methods to mitigate against models. misaligned model behaviour. Our project has three main stages: (1) characterise the internal structure and inter-concept alignments of the concept manifolds at Testing In-Context Generalisation of Language Models on Anti-Unification Tasks Yiding Song, Christopher Bates, Kazuki Irie, Samuel Gershman Harvard College | Quincy House | Computer Science | 2028 Despite their underlying neural architectures, modern language operation ‘repeat k times’ via anti-unification, which enables models can display strong systematic generalisation, skills that faster learning of a new instance, say ‘repeat 5 times’. More Fodor and Pylyshyn have traditionally argued to rest on symbolic generally, anti-unification is also an important skill in science, reasoning. Previous work has demonstrated that Transformers can wheremultipletheoriesmaybeabstractedintoasingleframework, be trained using meta-learning to exhibit systematic generalisation from which new theories can subsequently be drawn to explain in certain kinds of tasks involving compositional concepts. Here new phenomena. Building on the work of Lake and Baroni we apply the same meta-learning setup to study a more complex (2023), we propose a meta-learning approach to teaching neural but important kind of generalisation, specifically ‘least-general networks to anti-unify in-context through a diverse curriculum of generalisation’ (or ‘anti-unification’): the ability to inductively training samples. We generate a novel dataset inspired by reverse form new concepts based on an abstraction over a family of engineering Boolean circuits, and test it on Transformer-based related concepts. For example, a model that learns to ‘repeat 2 architectures, reporting the anti-unification capability of modern times’ and ‘repeat 3 times’ can generalise to the more abstract language models after specialised finetuning. The Effects of Pretraining on the Downstream Solution Learned by Transformers Alexandru-Raul Todoran, Bingbin Liu, Samy Jelassi Harvard College | Cabot House | Computer Science | 2028 The widespread adoption of Large Language Models (LLMs) has We first replicate prior findings showing that when the pretraining made improving their practical performance a central research task shares substructure with the downstream objective, such as focus. However, given the complexity of the reasoning tasks they parity and S3 state tracking, the model often learns a shortcut are learning to solve, multiple local minima of varying quality can that impairs generalization outside the training distribution. To be found, and standard metrics often fall short of distinguishing between them. Therefore, recent research has aimed to understand probe the cause of this behavior, we compare pretraining tasks that how different elements of the training pipeline affect the internal yieldsimilarsolutionsbutdifferinrelationtothedownstreamtask. For instance, pretraining on modulo-5 addition before fine-tuning characteristics of the solutions these models learn. This project on S3 state tracking does not induce a shortcut, suggesting that focuses on pretraining, and specifically, how it can bias the model the structure of the learned solution alone may be insufficient to toward different types of internal mechanisms for solving complex explain shortcut formation. reasoning tasks. Ongoing experiments aim to isolate other aspects of the To study reasoning abilities in a more controlled setting, we model pretraining-fine-tuning relationship. For example, we are them through synthetic tasks such as modular addition and state tracking. We select a series of task pairs for pretraining and investigating whether subgroup relationships between tasks fine-tuning, and analyze the mechanistic structure of the learned naturally encourage shortcut formation. This research ultimately seeks to inform principled pretraining dataset design. solutions. For example, we classify solutions as requiring a constant, logarithmic, or linear number of layers in sequence length, based on their signatures under two interpretability methods: linear probing and prefix patching. 64 Kempner Research in Artificial and Natural Intelligence for Undergraduates with Mentorship RepresentationalAlignmentbetweenNaturalSmellsandLLM-GeneratedOdor Perceptual Embeddings Eric Xu, Farhad Pashakhanloo, Venkatesh Murthy Harvard College | Mather House | Computer Science | 2028 Predicting how humans perceive smells has mainly focused on the relationship between different odorant representations through monomolecular odors, but inferring the perception of natural correlations between pairwise distances of samples. smells, which are complex combinations of molecules, remains a As a control, we used the Gemini text-embedding-004 model challenge. Gaining a stronger understanding of the extent to which to obtain embeddings for raw-text names of natural smells. different representations of olfactory information agree, including chemical compositions, human ratings for perceptual qualities, Using the same model, we also obtained embeddings of LLM- and textual descriptions, could improve performance of models on generated perceptual descriptions for each natural smell. Through representational similarity analysis, we computed pairwise tasks that involve perception of natural smells. distancematricesforeachrepresentationusingwell-suiteddistance This project aims to understand the mapping from complex metrics. Then, we calculated the Pearson correlation coefficient of mixtures to perception through the analysis of different odorant pairwise distances between matching elements in these matrices. representations, obtained from a natural smell dataset and the Preliminary results using the Jaccard distance metric for chemical outputs of text embedding models and Large Language Models composition representations and Euclidean distance metric for (LLMs). We used the Volatile Compounds in Food Dataset, raw-text embeddings yield correlations from 0.21 to 0.43 with p- which contains binary encodings for the presence of over 8000 values from 1.2▯10 ▯19 to 0.023. We are currently investigating chemical compounds within over 1300 foods. Due to the scarcity the relationship between odorant chemical composition and LLM- of human-rated perception datasets for natural smells, we used generated perceptual representations using other distance metrics an LLM to obtain proxies for human odor perceptions. Using and improving prompting techniques to obtain more informative representational similarity analysis, we quantified the strength ofodor perceptual descriptions.

Abstract:

Myelodysplastic syndromes (MDS) are a category of bone marrow Tet2 and Asxl1 loss of function mutations, commonly found in disorders that result from dysplastic hematopoietic stem cells MDS patients. Flow cytometry data was collected periodically failing to differentiate into healthy myeloid cells.In 30-40% to monitor the engraftments using peripheral blood from the 15 of cases MDS can progress to acute myeloid leukemia (AML). mice. The data analyzed after 9 weeks showed active engrafted Aberrant methylation in tumor suppressor gene promoters is a cell populations in all the mice, and at 12 weeks there was a clear prominent mechanism that contributes to the progression of MDS. expansion of myeloid cell counts in the engrafted populations. The Current hypomethylation treatments, such as azacytidine, have morphology of cell dysplasia in the mice was analyzed with a serious side effects as they are toxic and non-specific. This peripheral blood smear at 12 weeks. Three week treatments will project looks to test the effectiveness of AptaDiR, an RNA-based then be implemented with PBS as a control, azacytidine, and the therapeutic, for MDS treatment. AptaDiR inhibits DNMT1, a AptaDiR with two different lipid nanoparticles (LNPs), as carriers. DNA methyltransferase protein that methylates new strands of ThiswillallowforanalysisontheimpactAptaDiRhasonengrafted DNA during replication, mirroring the methylation on the templatecells modeling MDS, and its effectiveness compared to existing strand. This low toxicity therapeutic consists of an aptamer, a hypomethylation agents. Understanding how RNA-therapeutics single-stranded piece of RNA that can competitively bind and such as AptaDiR can affect methylation in MDS could provide inhibit DNMT1. To model MDS, 15 congenic mice including one important insight into treatment for other diseases also fueled by control, were used with engrafted bone marrow from mice with excess DNA methylation, such as cancer. Kempner Research in Artificial and Natural Intelligence for Undergraduates with Mentorship 60 Kempner Research in Artificial and Natural Intelligence for Undergraduates with Mentorship Infant Eye-Gaze and the Role of Collaborative Understanding in Human Intelligence Victoria Chen, Hannah Hok Kim, Ashley Thomas Harvard College | Leverett House | Psychology | 2028 Collaboration — our ability to pool knowledge and resources rich understanding of 1) infants see people as agents and attribute towards attaining more complex goals — is a defining aspect goals to the actions they perform, 2) collaborators engaging in of human intelligence. Prior research has shown it to be complementary actions likely have a shared goal , and 3) that effort uniquely adaptive, socially embedded, and powerful from early andefficiencyarekeymetricsingoalattainment. Thissupportsthe in development. But if infants don’t possess the adequate verbal results of our initial pilot study, where infants recognize and show and motor skills to engage in goal-directed actions, how do a violation of expectation — characterized by prolonged look- they develop this understanding of task difficulty, complementary time — when shown the unexpected event of fewer collaborators actions, and the relationship between effort and performance at leading to better outcomes in the building task. Additionally, such an early age? This study addresses how infants can still eye gaze is directed mainly at the single builder, showing that quickly acquire initial expectations of collaboration and its relatedthe surprise of seeing a single builder overachieving surpasses effects. We showed 16 infants videos of two task-completion that of two collaborators underachieving. Ultimately, we hope scenarios — one with a solo builder and the other with two that studying infants’ ability to recognize collaboration as an collaborators. This was followed by an analysis of their looking enhancer of task performance will provide a basis for future behavior using iCatcher+, a machine learning-based program that research surrounding humans’ rapid behavioral learning, and early codes for infants’ gaze direction and duration. Previous literature tendencies to be social agents. has established that, by the age of 16-20 months, infants possess a Diffusion Models and their Relational Compositions Hannah Kim, Binxu Wang Harvard College | Currier House | Computer Science | 2026 Recent advances in text-to-image machine learning models like accuracy, we analyzed cross-attention maps that revealed how text Stable Diffusion and DALL-E have enabled the generation of tokens influenced different parts of the image during generation. highly realistic images from natural language prompts. However, The controlled setting allowed us to isolate how relational these models often struggle to understand relational prompts, like understanding emerges or fails within the model. Preliminary “cat on top of iguana” and the causes of these errors remain results suggest that the model can learn simple spatial relations unclear. To better understand this failure mode, we investigated such as left-right positioning, but not reliably. There are instances PixArt-alpha, a transformer-based text-to-image diffusion model where small variations, such as the existence of articles in the that gained popularity because of its ability to produce high qualitytraining text prompts, can surprisingly modulate the accuracy of images with low training costs. We constructed a synthetic dataset generation, suggestinghowthefailuremodeisnotsolelyattributed of images with simple objects and colors paired with structured to model architecture. Our work contributes to broader efforts to relational captions like “Object A [relation] Object B” and trained interpret the internal mechanisms of diffusion models and offers a the PixArt model on it. In addition to evaluating generation framework for diagnosing and improving their reasoning abilities. Examining User Gender-Based Variations in LLM Response Trends Pei Yao Simon Ma, Naomi Saphra, Isabel Papadimitriou (in collaboration with Laasya Nagumalli, Kempner Fel- low) Harvard College | Pforzheimer House | Statistics | 2026 A large language model (LLM), upon being prompted with a sought further clarification. We generated 590 responses for each question, should respond with follow-up questions when the query and applied uniform manifold approximation and projection model has insufficient information to answer the user query. (UMAP) to reduce the dimensionality of the data, followed by This paper investigates whether desirable behaviors like follow- clustering to investigate patterns. Preliminary results show the up questions are subject to systematic biases. To this end, model is more likely to make assumptions about the user when the we analyze LLM responses from Cohere Command R+ and prompt contains a female persona and more likely to ask clarifying Llama 3.1-8B to twenty demographic-neutral queries about food, questionswhenthepersonaismale. Furthermore,responsestomen health, clothing, and more. Each prompt was prefixed with included more epistemic modal verbs like “can” while responses a gendered persona. These prompts were intentionally under- to women included more deontic modal verbs like “should.” If specified, requiring further clarification from the LLM for a full these preliminary results hold, these trends imply that female users response to the request. We classified the responses into two receive lower quality, less personalized responses from LLMs than categories: thosethatweredirect repliesto thequery andthose that male users. Sample Efficient Generalization to Very Different New Actions in Reinforcement Learning Ian Moore, Susan Murphy Harvard College | Quincy House | Applied Math | 2026 Insettingswheredataisexpensiveorhardtocollect,generalization new actions. In some settings such as digital advertising it may of reinforcement learning models is very useful as it allows us to be reasonable to assume the new actions will be similar to or adapt to a new RL environment while using the data we already have dynamics that are a mixture of that of the existing actions. have, reducing the need to collect new data. We consider the However in settings such as RL for digital health interventions and case where the RL model must generalize to new actions that robotics it is likely that some new actions will be very different are introduced. Existing work attempts to solve this problem, from the existing actions because they result from advances in but we show empirical results indicating these methods perform technology or science that enable the RL model to interact with poorly when the new actions are very different than the existing the environment in entirely new ways not previously seen or even actions we have data on — even in a very simple environment conceivable. Thus, this methodology will make sample efficient with discrete states and actions. This points to the need for a newaction generalization possible in these settings, enabling more methodology for sample efficient adaptation to the very different applications and capability for reinforcement learning. 62 Kempner Research in Artificial and Natural Intelligence for Undergraduates with Mentorship Examining User Gender-Based Variations in LLM Response Trends Laasya Nagumalli, Isabel Papadimitriou, Naomi Saphra (in collaboration with Pei Yao Simon Ma, Kempner Fel- low) Harvard College | Currier House | Mathematics | 2028 A large language model (LLM), upon being prompted with a sought further clarification. We generated 590 responses for each question, should respond with follow-up questions when the query and applied uniform manifold approximation and projection model has insufficient information to answer the user query. (UMAP) to reduce the dimensionality of the data, followed by This paper investigates whether desirable behaviors like follow- clustering to investigate patterns. Preliminary results show the up questions are subject to systematic biases. To this end, model is more likely to make assumptions about the user when the we analyze LLM responses from Cohere Command R+ and prompt contains a female persona and more likely to ask clarifying Llama 3.1-8B to twenty demographic-neutral queries about food, questionswhenthepersonaismale. Furthermore,responsestomen health, clothing, and more. Each prompt was prefixed with included more epistemic modal verbs like “can” while responses a gendered persona. These prompts were intentionally under- to women included more deontic modal verbs like “should.” If specified, requiring further clarification from the LLM for a full these preliminary results hold, these trends imply that female users response to the request. We classified the responses into two receive lower quality, less personalized responses from LLMs than categories: thosethatweredirect repliesto thequery andthose that male users. Interpreting and Manipulating Language Model Embeddings via Concept Manifolds Carl Scandelius, Pranav Misra, Haim Sompolinsky Harvard College | Winthrop House | Mathematics | 2027 Representations in language models (LMs) remain largely each transformer layer; (2) analyse the interpretability of the basis uninterpretable. We propose a geometric framework for analysing LMrepresentationsusingconceptmanifolds—pointcloudsformed vectors of these manifolds; (3) use this framework to guide causal intervention on LM embeddings to generate predictably altered by the final-token residual-stream embeddings of sentences output. representing a single dominant concept. These manifolds enable the definition of theoretically grounded geometric measures that Insights from this project hope to inform safety-minded efforts directly relate to downstream task performance. This framework like discovering misaligned latent knowledge, as well as fine- builds on successful applications to vision and vision-language tuning strategies or perturbation-based methods to mitigate against models. misaligned model behaviour. Our project has three main stages: (1) characterise the internal structure and inter-concept alignments of the concept manifolds at Testing In-Context Generalisation of Language Models on Anti-Unification Tasks Yiding Song, Christopher Bates, Kazuki Irie, Samuel Gershman Harvard College | Quincy House | Computer Science | 2028 Despite their underlying neural architectures, modern language operation ‘repeat k times’ via anti-unification, which enables models can display strong systematic generalisation, skills that faster learning of a new instance, say ‘repeat 5 times’. More Fodor and Pylyshyn have traditionally argued to rest on symbolic generally, anti-unification is also an important skill in science, reasoning. Previous work has demonstrated that Transformers can wheremultipletheoriesmaybeabstractedintoasingleframework, be trained using meta-learning to exhibit systematic generalisation from which new theories can subsequently be drawn to explain in certain kinds of tasks involving compositional concepts. Here new phenomena. Building on the work of Lake and Baroni we apply the same meta-learning setup to study a more complex (2023), we propose a meta-learning approach to teaching neural but important kind of generalisation, specifically ‘least-general networks to anti-unify in-context through a diverse curriculum of generalisation’ (or ‘anti-unification’): the ability to inductively training samples. We generate a novel dataset inspired by reverse form new concepts based on an abstraction over a family of engineering Boolean circuits, and test it on Transformer-based related concepts. For example, a model that learns to ‘repeat 2 architectures, reporting the anti-unification capability of modern times’ and ‘repeat 3 times’ can generalise to the more abstract language models after specialised finetuning. The Effects of Pretraining on the Downstream Solution Learned by Transformers Alexandru-Raul Todoran, Bingbin Liu, Samy Jelassi Harvard College | Cabot House | Computer Science | 2028 The widespread adoption of Large Language Models (LLMs) has We first replicate prior findings showing that when the pretraining made improving their practical performance a central research task shares substructure with the downstream objective, such as focus. However, given the complexity of the reasoning tasks they parity and S3 state tracking, the model often learns a shortcut are learning to solve, multiple local minima of varying quality can that impairs generalization outside the training distribution. To be found, and standard metrics often fall short of distinguishing between them. Therefore, recent research has aimed to understand probe the cause of this behavior, we compare pretraining tasks that how different elements of the training pipeline affect the internal yieldsimilarsolutionsbutdifferinrelationtothedownstreamtask. For instance, pretraining on modulo-5 addition before fine-tuning characteristics of the solutions these models learn. This project on S3 state tracking does not induce a shortcut, suggesting that focuses on pretraining, and specifically, how it can bias the model the structure of the learned solution alone may be insufficient to toward different types of internal mechanisms for solving complex explain shortcut formation. reasoning tasks. Ongoing experiments aim to isolate other aspects of the To study reasoning abilities in a more controlled setting, we model pretraining-fine-tuning relationship. For example, we are them through synthetic tasks such as modular addition and state tracking. We select a series of task pairs for pretraining and investigating whether subgroup relationships between tasks fine-tuning, and analyze the mechanistic structure of the learned naturally encourage shortcut formation. This research ultimately seeks to inform principled pretraining dataset design. solutions. For example, we classify solutions as requiring a constant, logarithmic, or linear number of layers in sequence length, based on their signatures under two interpretability methods: linear probing and prefix patching. 64 Kempner Research in Artificial and Natural Intelligence for Undergraduates with Mentorship RepresentationalAlignmentbetweenNaturalSmellsandLLM-GeneratedOdor Perceptual Embeddings Eric Xu, Farhad Pashakhanloo, Venkatesh Murthy Harvard College | Mather House | Computer Science | 2028 Predicting how humans perceive smells has mainly focused on the relationship between different odorant representations through monomolecular odors, but inferring the perception of natural correlations between pairwise distances of samples. smells, which are complex combinations of molecules, remains a As a control, we used the Gemini text-embedding-004 model challenge. Gaining a stronger understanding of the extent to which to obtain embeddings for raw-text names of natural smells. different representations of olfactory information agree, including chemical compositions, human ratings for perceptual qualities, Using the same model, we also obtained embeddings of LLM- and textual descriptions, could improve performance of models on generated perceptual descriptions for each natural smell. Through representational similarity analysis, we computed pairwise tasks that involve perception of natural smells. distancematricesforeachrepresentationusingwell-suiteddistance This project aims to understand the mapping from complex metrics. Then, we calculated the Pearson correlation coefficient of mixtures to perception through the analysis of different odorant pairwise distances between matching elements in these matrices. representations, obtained from a natural smell dataset and the Preliminary results using the Jaccard distance metric for chemical outputs of text embedding models and Large Language Models composition representations and Euclidean distance metric for (LLMs). We used the Volatile Compounds in Food Dataset, raw-text embeddings yield correlations from 0.21 to 0.43 with p- which contains binary encodings for the presence of over 8000 values from 1.2▯10 ▯19 to 0.023. We are currently investigating chemical compounds within over 1300 foods. Due to the scarcity the relationship between odorant chemical composition and LLM- of human-rated perception datasets for natural smells, we used generated perceptual representations using other distance metrics an LLM to obtain proxies for human odor perceptions. Using and improving prompting techniques to obtain more informative representational similarity analysis, we quantified the strength ofodor perceptual descriptions.

Source:

Harvard / Johns Hopkins University | Neuroscience | 2028 / 2025

Topics:

model, task, new, action, response, different, natural, llm, like, language, user, mds

Co-authors:

@deborahwells205 , @lucreziarinaldi206 , @annalisadiruscio207