projects | Pasan Dissanayake

Accelerating Multi-Property Molecular Design via Counterfactual Explanations

Proposed FINDER, a unified framework for composite molecular inverse design that efficiently finds molecular candidates satisfying multiple target properties simultaneously using entropic-risk-based counterfactual explanations and projected gradient descent.

Counterfactual Explanations and Model Reconstruction

Explored how the additional information provided by counterfactual explanations can be exploited for reconstructing a surrogate model. The work was later extended to multiple directions including retrieving counterfactual explanations in private and data-efficient LLM distillation using counterfactual explanations.

Quantifying Knowledge Distillation Using Partial Information Decomposition

Proposed information-theoretic metrics based on Partial Information Decomposition (PID) to quantify and explain knowledge transfer in distillation. This led to the Redundant Information Distillation (RID) framework, which filters task-irrelevant information and improves distillation under nuisance teachers.

MAPLEBench — A Benchmark for Multi-Agent Privacy in Long-Horizon Executions

MAPLEBench evaluates privacy leakage in realistic long-horizon multi-agent workflows. Built from 51 ALE benchmark tasks with injected sensitive information, sandboxed tool environments, adversarial third-party agents, and a dual detection framework (Disclosure Detector + Leakage Estimator).

Single-spiked complex Wishart matrices

Studied the distributions of two test statistics of centralized complex Wishart matrices with a single-spiked covariance structure