Research Interests #

My primary research interest centers on self-organizing multi-agent systems as an alternative inference approach. Work like Randazzo et al.’s “Self-classifying MNIST Digits” (Distill 2020, https://distill.pub/2020/selforg/mnist/) shows how a population of tiny “pixel agents,” each with only local information, can use learned message passing to converge on a shared classification, effectively all agreeing “we compose a 7” without any single model seeing the full image. That model-to-model communication, self-organization, and collective decision-making is the core pattern I want to generalize: swarms of specialized agents negotiating, correcting each other, and composing their partial views into a single, high-quality decision.

I want to extend this idea beyond pixels to learned multi-agent systems where the structure of interaction is itself the algorithm. Instead of one monolithic LLM or one large RL policy, I am interested in systems where many smaller models (or lightweight agents) with different roles and inductive biases communicate over time—debating, proposing tool calls, exploring strategies in parallel—and then converge on a single action, token, plan, or environmental move. That can be as small as choosing the next token or as large as selecting a long-horizon plan in a complex environment, but in all cases the emphasis is on distributed inference via communication rather than purely centralized computation.

My background in agent-based modeling (ABM) and computational social science gives me a way of thinking about these systems, but ABM is not my end goal. I treat ABM and related simulation techniques as environment and system-design tools: ways to specify local interaction rules, social structure, and stochastic dynamics that make multi-agent training and evaluation meaningful. The research focus is on learning-based agents (multi-agent RL and model-based agents, including LLM-derived agents), not on static social simulations or ABMs in isolation.

Architecturally, I am drawn to vocabulary-agnostic and flexible interfaces between agents and environments. Pagnoni et al.’s “Byte Latent Transformer: Patches Scale Better Than Tokens” (BLT, https://arxiv.org/abs/2412.09871) shows that byte-level, entropy-based “patches” can match tokenization-based LLM performance while improving robustness and efficiency. This kind of patch-based, tokenization-free interface is appealing as a way to give agents more adaptable, less brittle channels for perception and communication in multi-agent settings.

On the evaluation side, I am interested in game-based and environment-centric benchmarks that stress long-horizon, agentic behavior. Paglieri et al.’s BALROG benchmark (https://arxiv.org/abs/2411.13543) aggregates multiple RL game environments into a unified testbed for agentic LLM and VLM reasoning, including procedurally generated, highly challenging settings like the NetHack Learning Environment (NLE, https://arxiv.org/abs/2006.13760). This combination of diverse games and a hard, roguelike environment is close to the kind of long-horizon, compositional testbed I would like to use for probing multi-agent systems.

Finally, I see dictionary-learning-based interpretability not just as an AI safety tool but as a way to understand and improve reasoning in single and multi-agent systems. Anthropic’s “Towards Monosemanticity: Decomposing Language Models With Dictionary Learning” (https://transformer-circuits.pub/2023/monosemantic-features/index.html) shows how sparse autoencoders can extract semantically meaningful “features” from model activations. I am interested in using these techniques to: (a) analyze how agents specialize and divide labor in multi-agent systems, (b) diagnose how interaction protocols change which features are used, and (c) feed that understanding back into training and architecture design to enhance reasoning and coordination, not just detect pathological behavior.

All of this converges on a central program: design, train, and interpret multi-agent systems where compositional intelligence emerges from structured interaction among agents, using architectures and environments that support flexible communication, and benchmarks that actually stress long-horizon reasoning and coordination.

Problem Settings and Application Domains #

The most exciting testbeds for me are strategic games and game-like environments: settings with clear objectives, rich interaction structure, and opportunities for cooperation, competition, and negotiation. I see these as good laboratories for training and evaluating multi-agent reasoning, and for probing how systems scale with more agents, more complex rules, or more difficult coordination problems.

Beyond games, I am also interested in:

  • Multi-agent evaluation environments designed for agentic LLMs or MARL agents to interact, plan, and use tools.
  • Social dilemmas and multi-agent planning tasks where coordination, commitment, and role specialization matter.
  • Cyber-security/logs/incident-response–style environments or similar structured domains, when they are used as concrete testbeds for multi-agent reasoning rather than as purely applied analytics.

I am open to domain-heavy collaborations (social science, HCI, security, economics, etc.) as long as the core work is on multi-agent systems and reasoning rather than purely descriptive domain work. I am less interested in traditional “online platforms / social network” research that focuses solely on observational studies of user graphs, unless those insights are feeding back into the design and analysis of multi-agent systems.

I also have a secondary but related interest in agentic systems that make large-scale data analysis more accessible to non-data scientists. For example, multi-agent systems that can help a domain expert explore and analyze a corpus of 100,000+ documents with minimal bespoke engineering. This is not my primary research focus, but it is a natural application area for the multi-agent reasoning tools I want to help build.

Methods and Techniques #

Methods I actively want to deepen during the PhD, and how I would like to use them:

  • Multi-Agent Reinforcement Learning (MARL)

  • Primary method interest.

  • Training populations of agents (cooperative, competitive, mixed-motive) in rich environments, using self-play and population-based methods to study emergent strategies, coordination, and role specialization.

  • Model-based multi-agent systems and learned agents (including but not limited to LLMs)

  • I am more interested in training or fine-tuning new agents (e.g., RL or distilled models) than in purely orchestrating existing black-box APIs.

  • Using LLMs or other models as agents with tools and memory is interesting, but “prompt-engineering-plus-orchestration” is not the core; I want systems where the agent behaviors themselves are learned and improved over time.

  • Large-scale simulations and environments (ABM or otherwise)

  • Using ABM-style simulations and other rich environments as training and evaluation playgrounds for multi-agent systems.

  • Comfortable with HPC and large experiment sweeps; interested in environments that support systematic stress-testing of coordination, robustness, and scaling.

  • Representation learning and embeddings

  • Using representation learning to support coordination and communication between agents (e.g., embedding spaces that structure agent beliefs, goals, or messages).

  • Interested in approaches that relax rigid tokenization or allow more flexible interfaces between agents and environments.

  • Interpretability and diagnostics for multi-agent systems

  • Adapting dictionary-learning-style interpretability and similar tools to understand agent roles, specialization, and failure modes.

  • Emphasis on practical diagnostics: tools that help us see what a multi-agent system is actually doing and why it succeeds or fails on tasks.

  • Causal and experimental design for evaluating agents

  • Drawing on my social science background to design experiments that disentangle different mechanisms (e.g., is a system improving because of more agents, better communication protocols, or different training data?).

Scaling behavior (number of agents vs. model size vs. environment complexity) and phase transitions in multi-agent systems are topics I would like to explore further, but I do not have a fixed agenda yet. I am interested in advisors who see scaling and emergent behavior as important questions and are open to probing them empirically.

Role of ABM and Current Side Projects #

ABM and computational social science are where I started: my MA thesis used agent-based models to study protest cascades under autocracy, and I am collaborating on work using ML surrogates and sampling methods to map phase transitions in simulation parameter spaces. These projects give me a strong foundation in:

  • Designing agent-based environments.
  • Running large simulation sweeps on HPC clusters.
  • Thinking in terms of micro-level rules and macro-level emergent outcomes.

However, these are not the core of my future research interests. The surrogate modeling and phase transition detection work, for example, is valuable as a methods and tooling exercise, but I do not see it as my main research trajectory. It is better interpreted as evidence that I can handle complex simulations, build ML tooling around them, and collaborate on technical research.

Going forward, I want ABM to be:

  • A framework for environment and system design (how we structure interactions among agents), and
  • Occasionally, a comparative or supporting method, rather than the primary substantive focus on social phenomena.

Evidence of Readiness and Technical Capacity #

My previous projects are best understood as proof that I can handle large, messy, technical work, not as the exact topics I want to continue studying.

Representative examples include:

  • A thesis and subsequent work on large-scale agent-based simulations, run on university HPC infrastructure, to analyze how local interaction rules and uncertainty shape protest behavior in authoritarian regimes.
  • Leading a large-scale NLP/document-structure pipeline at the Library of Congress that processed hundreds of thousands of legal documents, involving OCR, layout analysis, and structured information extraction under strict real-world constraints.
  • Modernizing analytics and ML infrastructure in government and health-tech contexts, including dramatic performance improvements via vectorized processing and columnar data pipelines.
  • Building and operating an agentic analytics platform for security logs using LLMs, tool use, and structured interfaces, which required careful system design, API interaction, and evaluation.

Across these projects, the non-obvious capabilities I want to signal are:

  • Comfort taking on end-to-end technical responsibility for complex systems (from design and implementation to evaluation and iteration).
  • Ability to work with large, messy data and real deployment constraints, not just toy examples.
  • Willingness and ability to pivot, debug, and refactor as problems evolve.
  • Independence: I do not need close supervision to make progress on an open-ended technical problem.

I want the research system to treat these as evidence that I can thrive in labs that build and run real systems, not as evidence that I want to keep doing the same domain-specific work.

Key Papers and Inspirations #

  • Randazzo et al., “Self-classifying MNIST Digits” (Distill 2020, https://distill.pub/2020/selforg/mnist/) Shows how a neural cellular-automata-style system of “pixel agents” with only local views and neighbor communication can self-organize to correctly classify digits as a group. This is my canonical example of distributed inference via model-to-model communication and the template I want to generalize to more complex agents and tasks.

  • Pagnoni et al., “Byte Latent Transformer: Patches Scale Better Than Tokens” (https://arxiv.org/abs/2412.09871) Introduces a byte-level LLM architecture that uses entropy-based byte “patches” as units of computation, matching token-based models while improving efficiency and robustness. This suggests more vocabulary-agnostic, flexible interfaces between agents and environments, which is important for building multi-agent systems that are less brittle to tokenization details.

  • Paglieri et al., “BALROG: Benchmarking Agentic LLM and VLM Reasoning On Games” (https://arxiv.org/abs/2411.13543) Provides a benchmark and framework aggregating multiple RL game environments, including challenging, procedurally generated ones like NetHack, to evaluate long-horizon agentic reasoning in LLMs and VLMs. This aligns with my interest in game-based, long-horizon evaluations for multi-agent systems that emphasize planning, exploration, and coordination.

  • Küttler et al., “The NetHack Learning Environment” (NeurIPS 2020, https://arxiv.org/abs/2006.13760) Presents NLE as a fast but extremely rich and complex RL environment that stresses exploration, planning, skill acquisition, and generalization. I see NLE-style environments as ideal testbeds for multi-agent and agentic systems that need to operate under long horizons and sparse feedback.

  • Bricken et al., “Towards Monosemanticity: Decomposing Language Models With Dictionary Learning” (https://transformer-circuits.pub/2023/monosemantic-features/index.html) Uses sparse dictionary learning to decompose neural activations into interpretable features that correspond to concepts and behaviors. I am interested in adapting these techniques to analyze how agents specialize and how interaction protocols shape internal representations, with the goal of improving reasoning quality and coordination in both single-model and multi-agent systems.

These works collectively shape my interest in compositional intelligence, flexible agent–environment interfaces, and serious benchmarks for agentic behavior.

Advisor and Lab Preferences #

I am looking for advisors and labs that are:

  • Empirical and systems-oriented: places that build agents, run experiments, and iterate on system design, rather than primarily focusing on abstract theory or complexity proofs.
  • Working in or near multi-agent RL, agentic LLM systems, generative agents, or learning-based multi-agent systems more broadly.
  • Open to using simulations and game-like environments as training and evaluation grounds for multi-agent reasoning.

I am fully comfortable with advisors whose primary labels are “ML”, “AI”, “NLP”, “HCI”, “Information Science”, “Social Computing”, or similar, as long as the actual research intersects with:

  • Multi-agent systems / agentic LLMs / MARL.
  • Reasoning, planning, and decision-making in agents and agent collectives.
  • System-building and experimentation (not just observational studies or pure UX work).

On venues and trajectory:

  • Slight preference for labs that publish in core ML/AI venues (NeurIPS, ICML, ICLR, AAAI, etc.), especially on multi-agent or agentic topics.
  • Labs in HCI / social computing (CSCW, CHI, ICWSM) or complex systems / CSS venues are also attractive when they are doing methodologically serious work on multi-agent systems or agentic simulations, not just descriptive studies.

Regarding the spectrum “agentic LLMs vs. classical MARL/ABM”:

  • Excellent fits:

  • PIs actively working on agentic LLMs, generative agents, multi-LLM or multi-agent collaboration, or MARL in rich environments, especially those interested in training new agents and not just orchestrating opaque APIs.

  • Good fits:

  • PIs with strong backgrounds in multi-agent RL or ABM who are open to or already moving toward learning-based, model-centric agents (including LLM-based agents) in their future work.

  • Weaker fits:

  • PIs whose work is primarily static ABM without learning or purely observational social simulations with little interest in learning-based agents or systems-building.

I am more comfortable and better suited to engineering-heavy labs than strongly theory-heavy ones. I want to build systems, design experiments, and run large-scale studies, and I am willing to engage with theory and math as needed to support that, but I am not a good match for labs focused on pure theory of computation or abstract model theory without empirical components.

I am open to robotics and hardware-oriented labs when the core questions are multi-agent cooperation/competition and training in rich environments, and when the lab is situated in CS, ECE, or Information Science rather than in purely electrical-engineering-only departments.

Hard Constraints for Program and Advisor Matching #

To respect my funding constraints (e.g., CSGrad4US requirements) and my own fit:

  • Program types:

  • Must be in Computer Science, Information/Information Science/iSchool, or Computer Engineering/ECE.

  • Robotics advisors are fine as long as they are housed in CS, ECE, or an Information School, not in standalone EE-only departments.

  • Advisor/lab exclusions:

  • Avoid pure theory labs with little or no empirical systems work (e.g., purely complexity theory, purely abstract model theory, no experiments).

  • Avoid pure compilers labs and pure cryptography labs.
  • Avoid pure low-level vision labs where the work is dominated by perception benchmarks and not connected to multi-agent reasoning.

  • Prefer not to anchor my work in healthcare-only application labs unless the methods are clearly general and the multi-agent reasoning components are central, not incidental.

  • Research style:

  • Strong preference for labs that build, train, and deploy multi-agent systems, instrument them, and iterate based on empirical results.

  • Labs that combine system-building with principled evaluation and some theory are ideal; labs that are theory-only are a poor match.

This file should be interpreted as guidance for identifying advisors whose research agendas intersect with learning-based multi-agent systems, compositional intelligence, and agentic reasoning, in empirical, system-building environments within CS, Information Science, or Computer Engineering programs.