3 x papers at NeurIPS 2025
We have 3 papers published at NeurIPS 2025:
- ProDAG: Projected Variational Inference for Directed Acyclic Graphs
- Thompson Sampling in Function Spaces via Neural Operators
- Amortized Active Generation of Pareto Sets
ProDAG: Projected Variational Inference for Directed Acyclic Graphs
When learning cause-and-effect maps from data (directed acyclic graphs, or DAGs), most methods just spit out one “best guess” map without saying how confident they are. This paper introduces a way to quantify that uncertainty properly, using a mathematical trick that reliably projects any distribution onto the space of valid, acyclic cause-and-effect maps. The result, ProDAG, is more accurate than prior methods and gives trustworthy confidence estimates alongside its predictions.
Amortized Active Generation of Pareto Sets
Many real design problems involve trade-offs between competing goals (e.g., a drug that’s both effective and low-toxicity), where there’s no single “best” answer, only a range of good compromises (the Pareto front). This paper presents a method, A-GPS, that trains a generative AI model to produce these trade-off options directly and can be steered afterward toward a user’s specific preferences without retraining. Tested on synthetic problems and protein design, it finds strong sets of options with fewer costly evaluations.
Thompson Sampling in Function Spaces via Neural Operators
Some optimization problems involve tuning an entire simulated process, like a physics simulation, rather than just a few numbers, and running the simulation is expensive. This paper adapts a classic decision-making strategy (Thompson sampling) to this setting by using “neural operators”, fast AI models that learn to mimic expensive simulators, as stand-ins for uncertainty about what the simulator would produce. The approach finds good designs for physical systems (governed by complex equations) using far fewer costly simulator runs.