I believe AI agents will radically transform how research is conducted over the next decade: enabling investigations across disciplines, accelerating cycles of hypothesis and experiment, and expanding the scale of scientific work. The opportunity is not only to automate existing tasks, but to make new kinds of research possible.
My research focuses on the reasoning, decision-making and learning capabilities that agents need to conduct sustained scientific investigations: generating hypotheses, designing and interpreting experiments, using tools and models of the world, learning from evidence, and deciding what to do next.
My long-term ambition extends from AI co-scientists to autonomous research teams and research factories: systems that sustain investigations over long horizons, learn from success and failure, and produce rigorously validated discoveries. I am particularly interested in applying these ideas to biomedical and therapeutic discovery.
How can agents conduct coherent scientific investigations over extended periods of time? I study planning, reasoning, memory, tool use and multi-agent coordination, as well as how agents allocate computation and decide when to reason, retrieve information, use a tool, delegate or stop.
How should agents generate hypotheses and determine what evidence is needed to distinguish between them? I study hypothesis search, experiment selection and scientific verification, including the use of predictive and world models to simulate interventions and guide decisions before expensive real-world experiments.
How can scientific agents improve through experience? I study how experimental outcomes—including failures—should update hypotheses, memory, reasoning strategies and future actions, and how agents can improve their own research processes without reinforcing errors or drifting away from evidence.
What makes scientific discovery computationally possible, and how should progress be measured? I study the structure and search of hypothesis spaces, the relationship between evidence and scientific value, and the computational, data and experimental resources required for discovery. I am also interested in evaluating whether agentic systems produce genuinely novel, well-supported and useful scientific results.
My research papers examine what scientific agents contribute, how their ideas should be evaluated and what governs the complexity of discovery, through controlled comparisons, model diagnostics and computational analysis. Working with limited compute made careful scoping central to these papers: identifying novel, promising ideas and designing informative experiments that were feasible within the available budget.
The current manuscripts may include revisions made after acceptance to improve readability and clarity.
When does an AI agent add value to experimental decision-making? Across retrospective CRISPR screens, experimental feedback improves LLM-guided hit discovery, but a simple numerical policy using the same feedback finds more hits on average. Results vary by screen and initialization. The study separates the value of feedback from the value of repeated LLM decisions, providing a controlled way to test where scientific agents genuinely help.
Combining biomedical LLMs can appear beneficial because of how their answers are counted, not simply because the models are diverse. Holding responses fixed reveals that discarding answer ranks can weaken the single-model comparison and change the apparent benefit of mixing. Under equal call budgets, a selected single model outperforms a mixed-model policy on MedXpertQA. The findings show why model allocation and answer aggregation must be evaluated together.
Which scientific problems become achievable as AI improves, and when do new capabilities lead to validated discoveries? This position paper proposes a computational framework linking a discovery's resource and information requirements to available methods and the time needed for validation. Mathematical examples and retrospective tests expose why more compute or better benchmark performance need not translate directly into scientific progress, outlining a research program rather than a validated forecast of breakthroughs.
Is a protein-design model learning from surface geometry, or from information about the sequence it is meant to predict? Audits of two released pipelines identify surface features derived from the target sequence. In a released SurfPro model, changing those features redirects predictions toward a specified alternative sequence. The paper develops tests that distinguish useful representations from target-information leakage, while leaving open how much genuinely backbone-derived surfaces help protein design.
An AI scientist's apparent contribution can change when its menu of methods is rewritten, even without adding a new executable strategy. Experiments in protein landscapes and knapsack optimization show that ordering and duplicating options alter choices and can reverse conclusions about the agent's added value. Grouping methods by their behavior on fixed tests makes evaluation more stable, without establishing a general performance gain: a step toward measuring scientific judgment rather than menu design.
What should an AI system learn to value when selecting research ideas? An audit of 3,443 ICLR submissions finds that technical and empirical novelty/significance scores have different relationships with later citation uptake. At a fixed reading budget, the empirical score selects more high-uptake papers in the studied cohort; a separate analysis bounds the effect of missing outcomes. The results motivate testing selection criteria against their intended consequences, without equating citations with scientific value.
Does a high discovery-benchmark score reflect scientific reasoning, or recovery of a published conclusion already present in the input? Audits show how revealing or swapping source material changes the recoverability of reference conclusions, and why matching those conclusions does not establish scientific validity. The paper proposes measurement contracts that specify the available evidence, target answer, scoring rule and intended scientific claim, making clearer what a benchmark can actually demonstrate.
When does letting a language model write programs and recursively call other models justify the extra cost? Where the required operation is known, task-matched alternatives generally match or outperform recursive execution more cheaply. In open search, advantages over stronger alternatives do not hold consistently across budget-matched repeats. Counting every attempted query, including failures, can reverse apparent winners, offering a practical framework for judging when flexible agent architectures are worth their cost.
I wrote these essays after noticing that discussions of AI, including talks and writing by leading AI researchers, sometimes blur important conceptual distinctions. The underlying ideas may be understood, yet the way they are communicated can obscure what follows from them. These essays make those distinctions explicit and examine their consequences for reasoning, learning, evaluation and scientific progress.
When continued experience teaches reusable ideas rather than merely improving recall of past examples.
PDFHow recognizing good answers can unlock capability, and why flawed tests can reward the wrong results.
PDFWhy the promise of visual world models in robotics does not establish comparable gains in conceptual reasoning, and why complementing LLMs is more plausible than replacing them.
PDFHow retrieval, search and verification can support useful intelligence without copying the brain, and where weak integration becomes a limit.
PDFWhy expanding a working method can be more predictable than inventing one, and how compute can help discover better algorithms.
PDFWhich obstacles advances in language processing remove in biology and robotics, and which still require different evidence or skills.
PDFJEPA learns representations before all their later uses are known. How much useful information survives when the task changes, and when does a new task require further learning or observation?
PDFCould language coordinate increasingly diverse AI capabilities without making their underlying computations linguistic?
PDFAI could help build the tools and working practices needed for its own usefulness, making more projects feasible.
PDFHow to measure which demands of scientific discovery AI can meet as its methods scale and help improve their own successors.
PDFTracing technical gains through adoption, checking and changes in work to understand where economic value appears and who receives it.
PDFHow AI as artificial memory could make room for first principles without replacing the practice needed for independent thinking.
PDFA 2035 scenario, not a prediction: how directing research agents could reshape expertise, education and the scale of scientific inquiry.
PDFI created these lecture series because I have not found any public ones dedicated to formalizing scientific discovery and connecting it with the latest methods in advanced reasoning and AI agents in the depth and combination I wanted to study. They bring together two complementary questions: how scientific agents should reason, and what a computational theory of discovery should explain.
In addition, even excellent courses I encountered often explained ideas without showing how they arose. These lectures aim to reconstruct that creative process, from the initial question to the design choices and evidence, through three questions:
What observation, opportunity or dissatisfaction starts the inquiry? Which assumption becomes questionable, and what analogy or reframing makes a new direction conceivable?
How are candidate solutions constructed? Why pursue one design rather than credible alternatives, and what information could change that choice?
What mechanism or derivation explains the method? What evidence supports it, which assumptions does it depend on, and where does it fail?
Even strong lectures often focus only on the third question. My aim is to emphasize the first two as well. Explaining an idea is not the same as explaining the creative process behind it. Funnily enough, AI models will probably learn to innovate more effectively this way too.
Reasoning & AI Agents for Scientific DiscoveryAn 18-lecture synthesis of advanced reasoning and AI agent methods for scientific discovery: search and test-time computation, latent and diffusion reasoning, hierarchy, reinforcement learning, causal reasoning, memory and multi-agent systems. Worked examples and failure analyses connect these methods to tool use and autonomous experimentation.| No. | Lecture | Focus | Slides |
|---|---|---|---|
| 01 | Scientific discovery as a computational processDiscovery as a computational process, with an explicit account of the evidence needed to support a claim. | Discovery as a computational process, with an explicit account of the evidence needed to support a claim. | |
| 02 | Patterns of discovery: analogy, recombination, anomaly, mechanismAnalogy, recombination, anomalies and mechanisms as routes to discovery and reasons to change a representation. | Analogy, recombination, anomalies and mechanisms as routes to discovery and reasons to change a representation. | |
| 03 | Hypothesis generation and conceptual noveltyConstructing competing causal hypotheses, distinguishing verbal from predictive diversity and choosing interventions that separate them. | Constructing competing causal hypotheses, distinguishing verbal from predictive diversity and choosing interventions that separate them. | |
| 04 | Searching hypothesis spacesSearching candidate hypotheses with external verification and feedback, including the example of AlphaEvolve. | Searching candidate hypotheses with external verification and feedback, including the example of AlphaEvolve. | |
| 05 | Experiment selection and information gainChoosing experiments by information gain and decision value, including when biased simulations are worth less than direct measurements. | Choosing experiments by information gain and decision value, including when biased simulations are worth less than direct measurements. | |
| 06 | Search, verifiers, and test-time computeScientific generation as search: comparing retraining, conditioning, ranking and intermediate steering while separating scores from scientific validity. | Scientific generation as search: comparing retraining, conditioning, ranking and intermediate steering while separating scores from scientific validity. | |
| 07 | Latent and continuous reasoningReasoning in latent or continuous representations and the relationship between internal computation and scientific decisions. | Reasoning in latent or continuous representations and the relationship between internal computation and scientific decisions. | |
| 08 | Recurrent, diffusion, and hybrid reasoningRecurrent, diffusion and hybrid reasoning, including particle steering, reweighting and the need for correction and independent evaluation. | Recurrent, diffusion and hybrid reasoning, including particle steering, reweighting and the need for correction and independent evaluation. | |
| 09 | Hierarchical reasoning and abstractionBuilding and reusing abstractions while preserving the assumptions that make a reasoning step valid. | Building and reusing abstractions while preserving the assumptions that make a reasoning step valid. | |
| 10 | Hierarchical planning and subgoal discoveryDecomposing investigations into subgoals, with a distinction between checking a plan and validating its scientific interpretation. | Decomposing investigations into subgoals, with a distinction between checking a plan and validating its scientific interpretation. | |
| 11 | Learning reasoning strategies with RLLearning reasoning strategies through reinforcement learning, including what the reward encourages and what it leaves untested. | Learning reasoning strategies through reinforcement learning, including what the reward encourages and what it leaves untested. | |
| 12 | Causal, mechanistic, and counterfactual reasoningConstructing causal alternatives and separating treatment effects, mechanisms and transfer through explicit interventions and assumptions. | Constructing causal alternatives and separating treatment effects, mechanisms and transfer through explicit interventions and assumptions. | |
| 13 | Scientific evidence and tool useUsing literature, databases and tools without confusing successful execution or a supporting citation with sufficient evidence. | Using literature, databases and tools without confusing successful execution or a supporting citation with sufficient evidence. | |
| 14 | Long-horizon research, memory, and self-improvementMemory and self-improvement across long research trajectories, and how persistent changes affect evaluation. | Memory and self-improvement across long research trajectories, and how persistent changes affect evaluation. | |
| 15 | Multi-agent scientific reasoningWhen collaboration between agents adds value beyond a single system with matched information and resources. | When collaboration between agents adds value beyond a single system with matched information and resources. | |
| 16 | Biomedical discovery systemsBiomedical discovery workflows that distinguish simulator predictions from measurements, and treatment effects from mechanisms and transfer. | Biomedical discovery workflows that distinguish simulator predictions from measurements, and treatment effects from mechanisms and transfer. | |
| 17 | Closed-loop experimentation and hypothesis revisionRevising competing hypotheses through noisy experimental results, with explicit likelihood updates, dependence checks and model-expansion limits. | Revising competing hypotheses through noisy experimental results, with explicit likelihood updates, dependence checks and model-expansion limits. | |
| 18 | Can AI Actually Discover? Novelty, Evidence and VerificationEvaluating whether AI systems make scientific discoveries: novelty, evidence and verification, with model accuracy, research behavior and scientific validity assessed separately. | Evaluating whether AI systems make scientific discoveries: novelty, evidence and verification, with model accuracy, research behavior and scientific validity assessed separately. |
The Nature and Scaling of Scientific DiscoveryA 10-lecture investigation of discovery as a computational process: hypothesis spaces, identifiability, experimental feedback, scaling laws and resource trade-offs. Examples from mathematical construction, program synthesis and molecular design examine what a predictive theory of scientific progress would need to explain.| No. | Lecture | Focus | Slides |
|---|---|---|---|
| 01 | What Is a Scientific Discovery?What distinguishes discovery from solving a fixed task, and how interventions separate prediction, causal effects and competing mechanisms. | What distinguishes discovery from solving a fixed task, and how interventions separate prediction, causal effects and competing mechanisms. | |
| 02 | What Makes a Discovery Problem Complex?Separating computational difficulty, missing information and identifiability: when more computation cannot resolve a question. | Separating computational difficulty, missing information and identifiability: when more computation cannot resolve a question. | |
| 03 | Representations and the Geometry of DiscoveryHow representations reshape discovery, with symmetry motivating alternatives based on augmentation, invariant features and equivariant models. | How representations reshape discovery, with symmetry motivating alternatives based on augmentation, invariant features and equivariant models. | |
| 04 | Search, Experiments, and the Value of FeedbackTurning verbal hypotheses into explicit models, choosing experiments that distinguish them and updating beliefs from noisy evidence. | Turning verbal hypotheses into explicit models, choosing experiments that distinguish them and updating beliefs from noisy evidence. | |
| 05 | Sequential and Parallel DiscoveryWork, depth and adaptive dependencies: which parts of discovery can run in parallel and which must wait for evidence. | Work, depth and adaptive dependencies: which parts of discovery can run in parallel and which must wait for evidence. | |
| 06 | Scaling Laws and Resource Trade-offsTrading off data, computation and experiments while distinguishing predictive loss, success probability and validated discovery yield. | Trading off data, computation and experiments while distinguishing predictive loss, success probability and validated discovery yield. | |
| 07 | When Are Discovery Problems Equivalent or Transferable?When insights transfer between problems, and the different assumptions behind reductions, generalization and causal transport. | When insights transfer between problems, and the different assumptions behind reductions, generalization and causal transport. | |
| 08 | Proofs, Programs, and Molecules: Comparing Discovery Across DomainsComparing proofs, programs and molecules, including how Coscientist's tool-error repair differs from chemical measurement and scientific discovery. | Comparing proofs, programs and molecules, including how Coscientist's tool-error repair differs from chemical measurement and scientific discovery. | |
| 09 | Cumulative Discovery and Moving FrontiersHow discoveries accumulate and move the frontier, and why repeated claims or shared errors do not amount to independent corroboration. | How discoveries accumulate and move the frontier, and why repeated claims or shared errors do not amount to independent corroboration. | |
| 10 | Toward a Predictive Theory of Scientific ProgressWhat a predictive theory of scientific progress could claim, how to test it and where its predictions must remain conditional. | What a predictive theory of scientific progress could claim, how to test it and where its predictions must remain conditional. |
Masters in Data Science
Top-4 CS university in the US
NLP, AI agents, biomedical AI
Bachelors/Masters in Computer Science
Top CS university in Greece
AI, robotics, mathematics
Perfect scores in selected UC Berkeley courses.
Entrance scholarship based on prior academic performance.
Offered an entrance scholarship based on prior academic performance (chose UC Berkeley over Columbia).
Top academic performance in mathematics courses.
Top academic performance in mathematics courses.
Top 0.1% (top 40 among 100,000 candidates) in Greece's national university admissions exams.
Selected for the national reserve squad.