I am an Electrical and Computer Engineering PhD student at Princeton University, working with Jonathan D. Cohen and Taylor Webb. I study reasoning and learning in foundation models, drawing inspiration from cognitive science and neuroscience (read more about what shaped my research).
My current work focuses on sequential visual reasoning in VLMs and efficient long-context reasoning in LLMs (read more about my recent and current research).
I earned my B.Sc. in Biomedical Engineering from the Department of Electronic and Telecommunication Engineering at the University of Moratuwa, Sri Lanka, in 2022, with First Class honors.
Outside academia, I enjoy chess, music, and playing instruments.
I study how biological and artificial systems learn, reason, and process information over time, combining ideas from cognitive science, neuroscience, and machine learning.
My current research focuses on understanding and improving visual reasoning in vision-language models (VLMs). In my recent work, I investigated the binding problem in VLMs, showing how spatially grounded reasoning induces serial attention routines that reduce interference between object representations and improve generalization. I am extending this work to develop recurrent and video-generative VLM architectures capable of more general forms of sequential visual processing.
In parallel, I investigate efficient long-context reasoning in foundation models with in-place test-time training, episodic memory, and KV-cache compaction.
Stanford University
January – June 2024
Visiting Student Researcher · Department of Biology
I worked in Professor Mark Schnitzer’s lab on building machine learning models that replicate activity recorded from the mouse brain (i.e., digital twin brains) during visual decision-making tasks.
Further, using dynamical systems theory, I analyzed how these learned neural activity patterns evolve over time to achieve specific behaviors and interact with other brain regions.
Harvard University
July 2022 – December 2023
Post-Baccalaureate Fellow · Faculty of Arts and Sciences
I worked on understanding learning dynamics in autoregressive machine learning models (e.g., recurrent neural networks), and how biologically inspired neural computation allows them to have stable convergence dynamics.
I also developed computational models of high-throughput imaging systems combining machine learning, signal processing, optics, and low-power electronics.
Recent & Current Research
1. The Need for Sequential Processing in VLMs: Insights from the Binding Problem
One recent example is my work on the binding problem in vision-language models (Haputhanthri et al., 2026, NeurIPS). The binding problem refers to a system’s difficulty keeping track of which features belong to which objects when several objects must be processed at once. A long-standing idea in cognitive science, supported by theoretical work, is that this reflects a tradeoff between generalization and interference: compositional representations allow features to be reused across objects, supporting generalization, but that same reuse can cause similar objects to interfere with one another. In this work, we first found evidence for this same representational structure and interference in VLMs, helping explain their capacity limits on multi-object visual reasoning. We then brought together two observations. First, evidence from theoretical and empirical work in cognitive science suggests that humans address similar binding limits through serial attention, processing objects one at a time. Second, recent VLMs often benefit from reasoning traces that explicitly ground references to object locations. This motivated us to ask whether position-grounded reasoning in VLMs could induce a human-like serial visual-search process. In models trained to point, we found exactly this: an emergent search routine supported by specific attention heads, processing objects one at a time, reducing interference between competing representations, and therefore preserving the correct object-feature bindings. Finally, we showed that training models to point while reasoning largely resolves binding failures and enables strong out-of-distribution generalization, whereas variants without sequential processing (i.e., a serial attention routine) or without spatial grounding fail to generalize in the same way. Together, these results provide a mechanistic account of why pointing-based, visually grounded reasoning can improve VLM performance, and connect a long-standing cognitive account of capacity limits with the internal computations that emerge in modern VLMs.
2. Towards More Sophisticated Sequential Processing in VLMs via Latent Reasoning
I am currently extending this work to develop foundation model architectures capable of more general forms of sequential visual processing. My previous work showed that training VLMs to point can induce serial attention routines, allowing them to solve tasks such as visual search and counting. While this provides a relatively straightforward way to improve visual reasoning in existing VLMs through chain-of-thought supervision, it is largely limited to tasks whose intermediate computations can be expressed as discrete textual reasoning steps. Human visual cognition, however, is not subject to this restriction. Motivated by this distinction and inspired by recent models with continuous latent representations, I am exploring how models can carry out more general-purpose sequential processing, going beyond simple serial-attention-related sequential processing, which could be easily achieved through textual-CoT supervision. My broader goal is to investigate the hypothesis that such general-purpose/sophisticated sequential processing methods allow models to generalize out of distribution.
One could argue that such continuous latent representations (visual or textual), which replace multiple discrete reasoning tokens, act as an indirect form of KV-cache compaction. Another direction of my research addresses the KV-cache compaction problem more directly by combining cognitively inspired concepts (e.g., episodic memory) with recent developments in ML (e.g., test-time training).
3. Understanding Latent Reasoning through Dynamical Systems
My interest in recurrent reasoning builds on my previous work on understanding how recurrent networks develop their internal dynamics to memorize and manipulate information. In this series of works (Haputhanthri et al., 2024, ICML MI; Haputhanthri et al., 2025, preprint), we studied RNNs learning short-term memory tasks and asked how these networks gain new capabilities for solving the task (or parts of the task). We found that solving the task happens in a stage-like manner, with abrupt improvements associated with geometric restructuring in the networks’ attractor manifolds (i.e., bifurcations). Also, such attractor manifolds with slow dynamics act as memory that can retain information. This led us to ask whether we could help these structures form rather than wait for task training to discover them. We developed temporal consistency regularization, which encourages a subset of neurons to change slowly, allowing networks to go through bifurcations and create memory subspaces faster, therefore reducing the time they spend in each ability-acquisition phase, significantly improving the stability of convergence, and reducing the time needed to learn the full task. Building on this work, I am now exploring dynamical-systems approaches to understanding continuous latent reasoning in language models, studying how their reasoning can be understood as representations evolve over time in state space.
H. Arguello, J. Bacca, H. Kariyawasam, E. Vargas, M. Marquez, H. Garcia, R. Hettiarachchi, K. Herath, U. Haputhanthri, B. S. Ahluwalia, P. So, D. N. Wadduwage, C. Edussooriya
How do the brain’s circuits give rise to complex behavior, and how could we replicate that? My research has followed this question through engineering 🤖, imaging 🔬🦠, neuroscience 🧠, and cognitive science 🧠💭.
At Harvard, I worked with Dr. Dushan Wadduwage on computational microscopy 🔬🦠, combining my machine learning experience and biomedical background to develop new imaging systems. Building these systems made me think more deeply about what it would take to observe the brain well enough to recreate its circuits and behavior, and how imaging limitations constrain our understanding. In parallel, I worked with Dr. Hidenori Tanaka on NeuroAI 🧠🤖, exploring what we could learn from the limited knowledge and data we already had about the brain. I drew on biological learning and population dynamics to understand stability and convergence issues in artificial neural networks, and ways to mitigate them.
This work made me want to study those dynamics in actual brain recordings 🧠. At Stanford, I joined Dr. Mark Schnitzer’s lab to build artificial neural networks that model mouse neural activity and reproduce behavioral patterns (i.e., digital twin brains). Using dynamical systems theory, I studied how these learned activity patterns evolve over time and relate to behavior 🤖 → 🧠.
Having worked across machine learning, imaging, NeuroAI, and computational neuroscience, I came to think that AI and neuroscience had drifted somewhat apart, despite their history of driving one another forward. When I began my PhD at Princeton, I wanted to work where the two could inform each other: taking inspiration from the biological brain to advance AI 🧠 → 🤖, and using AI to understand the brain better 🤖 → 🧠. This led me to work with Jonathan Cohen and Taylor Webb. My current work draws on cognitive science 🧠💭 and neuroscience 🧠 to explain how machine learning works and build more human-like, trustworthy ML systems.
