Conference Rankings:

• NeurIPS (#1) – World’s largest AI/ML conference (h5-index: 337)
• AAAI (#4) – Premier artificial intelligence conference (h5-index: 220)

1. Agentic Causal Graph Learning for Drug Target Discovery

Presentation: Oral

Summary: Presents a self-directed AI system that autonomously learns causal relationships in protein-protein interaction networks at the scale of STRING database (100,000+ interactions). The system uses agentic learning to identify drug targets by discovering causal mechanisms rather than mere correlations in biological networks.

Key Achievement: First autonomous system capable of causal graph learning at STRING scale for therapeutic target identification.

2. The Algorithmic Alchemist: Advanced Generative AI Methodologies for De Novo Drug Design

Presentation: Oral Presentation + Poster

Summary: Introduces advanced generative AI methods for designing novel drug molecules from scratch (de novo). The system acts as an “algorithmic alchemist,” combining multiple generative models to create drug candidates with desired properties that don’t exist in current chemical libraries.

Innovation: Integration of multiple generative paradigms for unprecedented molecular design flexibility.

3. The Acceleration of Molecular Discovery through Advanced Active Learning and Bayesian Optimization

Presentation: Poster

Summary: Develops active learning strategies combined with Bayesian optimization to dramatically accelerate the molecular discovery process. The system intelligently selects which molecules to synthesize and test, minimizing experimental costs while maximizing information gain.

Key Achievement: Reduces experimental iterations required for molecular optimization by strategically selecting most informative candidates.

4. Neural Cellular Automata and the PD-NCA AutoLab: A Differentiable Artificial Life Testbed for Autonomous Research Agents

Presentation: Poster

Summary: Presents a differentiable artificial life testbed that enables autonomous research agents to generate hypotheses, run ab initio experiments, and evaluate emergent phenomena. Introduces the PD-NCA AutoLab, a programmable, differentiable cellular-automata environment that exposes its physics for agent-driven design.

Contributions: PD-NCA AutoLab programmable environment with Python reference implementation; multi-agent research workflow with Hypothesis → Execute → Analyze → Reflect operational cycle; feasibility studies demonstrating stable BPTT gradients and discovery sparsity motivating intelligent search.

Key Innovation: Creates a “virtual laboratory” where AI agents can design and run their own scientific experiments in a differentiable physics sandbox, enabling true autonomous scientific discovery.

5. The Trajectory of Graph of Thoughts: Advancing AI for Science through Structured, Verifiable, and Causal Reasoning

Presentation: Poster

Summary: Explores how Graph of Thoughts (GoT) paradigm provides foundational architecture for autonomous scientific discovery agents. Analyzes innovations including dynamic knowledge graphs (KGoT), adaptive inference structures (AGoT), and multi-agent collaboration graphs.

Key Argument: Linear reasoning (Chain-of-Thought) and tree-based reasoning (Tree-of-Thoughts) are inadequate for scientific inquiry. Graph-structured reasoning enables: aggregation/synthesis of multiple information sources, refinement/iteration based on feedback, generation/divergence exploring multiple hypotheses simultaneously.

Specialized Extensions: HI-GoT for verifiable hypothesis validation; C-GoT for modeling dynamic causal mechanisms; KGoT integrating structured knowledge; AGoT with dynamic inference structures.

8. Neuro-Symbolic AI for Alzheimer’s Disease: Physics-Informed Biomarker Prediction and Verifiable Intervention Planning

Presentation: Poster

Summary: Combines neural networks with symbolic reasoning and physics-informed constraints for Alzheimer’s disease progression modeling. Provides verifiable intervention planning by grounding predictions in biophysical models of disease mechanisms.

Innovation: First system to combine data-driven learning with physics-based disease models for verifiable clinical decision support.

9. Glass-Box Arbitrators: An Explainable Neuro-Symbolic AI Framework for International Commercial Arbitration Proceedings

Presentation: Poster

Summary: Develops explainable neuro-symbolic architecture specifically for LCIA (London Court of International Arbitration) proceedings. Treats LCIA Rules as formal logical substrate while using LLMs for natural language parsing and explanation generation.

Architecture Layers: Rule Formalization – LCIA Rules encoded in formal logic/constraints; Case Representation – LLMs extract facts from contracts, pleadings, evidence; Symbolic Reasoning – Formal proof obligations for jurisdiction, costs, admissibility; Explanation Generation – LLMs translate proofs into natural language.

Key Principle: LLMs handle semantic parsing and drafting, but symbolic solvers provide correctness backbone—not end-to-end neural reasoning.

10. Graph of Thoughts Nanobiomaterials Assistants: Towards Logical, Tool-Augmented, and Multi-Agent Reasoning in Scientific LLMs

Presentation: Poster

Summary: Applies Graph of Thoughts reasoning framework specifically to nanobiomaterials research. Develops multi-agent systems where specialized LLM agents collaborate using graph-structured reasoning, tool augmentation, and logical verification for materials discovery.

Key Components: Logical reasoning – formal verification of material properties; Tool augmentation – integration with simulation and analysis tools; Multi-agent coordination – specialized agents for synthesis, characterization, application.

11. SOKRATES: Distilling Symbolic Knowledge into Option-Level Reasoning via Solver-Guided Preference Optimization

Presentation: Poster

Summary: Introduces SOKRATES framework that distills symbolic reasoning capabilities into LLMs through solver-guided preference optimization. Rather than teaching LLMs to imitate symbolic solvers, trains them to prefer reasoning paths that align with verified symbolic solutions.

Innovation: Uses satisfiability solvers (SAT/SMT) to generate preference data, teaching LLMs to reason at “option level”—choosing between alternative reasoning strategies based on formal correctness.

6. Regulation-Aware Neuro-Symbolic Legal World Models for Cross-Border Commerce

Presentation: Poster

Summary: Develops neuro-symbolic AI systems that combine neural language models with symbolic legal reasoning to navigate complex cross-border commerce regulations. The system maintains “world models” that represent different legal jurisdictions and can reason about regulatory compliance across borders.

Key Challenge Addressed: Automated reasoning about conflicting or inconsistent legal requirements across multiple jurisdictions in international trade.

Innovation: First neuro-symbolic system for multi-jurisdictional legal reasoning in commerce.

7. Polyculture Agents: Detecting and Mitigating Algorithmic Echo Chambers in Financial AI Workflows

Presentation: Poster

Summary: Addresses echo chambers in multi-agent financial trading systems where agents preferentially interact with similar others, leading to information homogenization and reduced alpha generation.

Key Discovery: Contrary to conventional wisdom, mild echo chambers can be beneficial for specialized tasks. Optimal performance occurs at intermediate diversity levels (Shannon entropy ≈2.3 bits).

ECHO Framework Components: Causal Detection – uses causal inference to identify harmful vs. productive echo chambers; Targeted Interventions – preserves beneficial specialization while preventing harmful isolation; Adaptive Thresholding – dynamically adjusts diversity levels based on task requirements.

Results: 73.2% echo chamber reduction; 27.4% innovation improvement; 94.1% productivity preservation (only 5.9% loss).

Key Insight: Inverted-U relationship between diversity and performance—too little OR too much diversity both harm performance.

David Scott Lewis

Cofounder & CEO
NeurIPS/AAAI/ICML published researcher
Researching AI since the 80s in Stanford University Lab.

Enrique Zueco

Cofounder & CTO
NeurIPS/AAAI published researcher
Expert in multi-agent systems