ICML 2026
A geometry-based steganography scheme that hides secrets in fine-tuned LLM outputs via embedding-space hyperplanes, together with a linear-probe detector that exposes it more reliably than traditional steganalysis.
TMLR 2026
The Transfer Entropy Redundancy Criterion (TERC): an information-theoretic test that provably drops state variables with no effect on agent performance, improving sample efficiency across Q-learning, Actor-Critic, and PPO.
SIGKDD 2025
An information-theoretic feature selection method (FSNID) that drops uninformative inputs for network intrusion detection while preserving classifier performance.
AISTATS 2025
PIDF replaces a single feature-importance score with three: how much information a feature shares with the target, how much arises only in combination with others (synergy), and how much is redundant with what other features already carry.
arXiv preprint, 2026
Linear-probe detectors for LLM steganography can be evaded by adversarial fine-tuning (58–79% covert recovery preserved). A theory-guided recontextualization intervention restores detection where activation-only methods fail.
arXiv preprint, 2025
Recasts the Information Bottleneck through synergy — information that only appears when features are processed jointly — yielding interpretable compression phases in ReLU networks, CNNs, and Transformers where standard IB struggles.
arXiv preprint, 2024
A structured pruning method that keeps the nodes carrying mutual information between adjacent layers, with a guarantee that the pruned upstream activations can still be mapped to the downstream layer — so the network remains retrainable.