ACHT accelerates hypothesis generation through Bayesian optimization and causal discovery. VAP verifies autonomous pipelines through self-falsification and physics-informed surrogates. Together, they form a complete discovery-to-validation framework.
A domain-agnostic framework that fuses Bayesian optimization, generative AI, and causal discovery to compress years of hypothesis exploration into days. Validated across drug design, materials science, financial modeling, and legal reasoning.
Protein-protein interaction networks provide the causal scaffold for biological hypothesis testing. ACHT leverages STRING's curated interaction database to constrain the generative search space and validate discovered causal relationships.
Gaussian process surrogates with expected improvement acquisition functions. Adaptively samples the hypothesis space, concentrating evaluation budget on the most promising regions. Supports multi-objective optimization with Pareto frontier tracking.
Diffusion models and variational autoencoders trained on domain-specific representations. Generates novel candidates — molecules, materials, financial instruments, legal arguments — that satisfy learned structural constraints while maximizing novelty.
Structure learning algorithms (PC, GES, NOTEARS) combined with interventional data to infer directed acyclic graphs. Distinguishes correlation from causation, enabling mechanistic understanding and robust out-of-distribution predictions.
The same ACHT pipeline adapts to four distinct scientific domains through domain-specific encoders and decoders, while sharing the core optimization and causal inference machinery.
A multi-agent verification framework that prevents AI echo chambers, validates hypotheses against physical constraints, and ensures autonomous research pipelines produce trustworthy results. Built on self-falsification principles.
Diverse agent populations with heterogeneous training, architectures, and reasoning strategies. Prevents monoculture consensus failures by ensuring genuine intellectual diversity in the verification panel. Each agent is a domain specialist with a unique epistemological stance.
A dedicated adversarial agent actively attempts to disprove each hypothesis before it can advance. Applies Popperian epistemology programmatically: claims survive only if they withstand systematic attempts at refutation, counter-example generation, and edge-case stress testing.
Neural network surrogates trained with physics-informed loss functions enforce conservation laws, symmetries, and domain constraints. Provides fast approximate validation while guaranteeing physical plausibility. Catches violations that pure data-driven models miss.
VAP continuously monitors inter-agent agreement patterns. When agent outputs converge suspiciously fast or exhibit correlated reasoning chains, the echo chamber detector triggers diversity injection: introducing new agents with orthogonal training data, swapping reasoning strategies, or escalating to human review. This mechanism has prevented 23% of false-positive validations in benchmark evaluations, catching cases where homogeneous agents would have unanimously endorsed flawed hypotheses.
Explore our full publication record across AAAI, NeurIPS, ELLIS, Stanford, and more. Every claim is peer-reviewed. Every result is reproducible.