Selected research
Co-VLN studies whether independently navigating agents can benefit from one another's observations. It detects overlap between their topological maps and fuses shared perceptual memory, expanding each agent's effective view without additional exploration.
Read the preprint
This work shows how persistent memory can redirect an LLM agent's tool-use control flow long after an injection occurs. MEMFLOW generates memory control-flow attacks and audits their effects across agent models, memory mechanisms, and tool frameworks.
Read the preprint
MAPA develops a multi-turn jailbreak evaluation method for large vision-language models. It alternates text and vision attack actions within each turn, then refines the attack trajectory across turns to progressively test model safety.
Read the preprint
FedHydra addresses model and data heterogeneity in one-shot federated learning without accessing client data. It stratifies client models by capability, generates synthetic data, and distils their knowledge into one global model.
Read the paper
FLOW represents multi-agent workflows as graphs and revises task assignments as execution unfolds. Its modular design reduces unnecessary dependencies, enables parallel work, and supports efficient recovery when a subtask fails.
Read the paper
This work reframes machine-generated text detection as a relative hypothesis test: is a text distribution closer to human or machine references? A learned deep kernel improves test power while reducing false positives when human writing shifts from the seen distribution.
Read the paper
SeqCV addresses inter-agent misalignment when workflow dependencies cannot be treated as conditionally independent. It executes subtasks sequentially, validates response segments across models, and recursively splits inconsistent subtasks.
Read the paper
FedInverse evaluates whether a federated global model leaks participant data through model inversion. Its HSIC-based regularisation diversifies the attack generator, revealing privacy risks that may remain hidden behind apparently benign clients.
Read the paper
MMD-MP targets the variance inflation caused by machine-generated text from multiple LLM populations. Its multi-population-aware objective learns a more stable kernel for paragraph- and sentence-level detection.
Read the paper
This paper introduces regional concept drift detection and density-synchronised adaptation. It identifies where drift occurs and updates the learner with relevant samples, avoiding the overestimation and unnecessary retraining caused by global adaptation.
Read the paper
SEGA introduces a sequentially updated drift gradient that scores which historical segments remain most relevant as each new instance arrives. It reduces adaptation delay by updating predictors with selected segments instead of waiting for a complete batch.
Read the paper
DAR learns from streams that exhibit both changing distributions and temporal dependence. It reconstructs the feature space and uses LDD+ to remove outdated instances, supporting robust online regression under two simultaneous forms of non-stationarity.
Read the paper
FUZZ-CARE uses fuzzy clustering to identify latent patterns in a drifting stream and assign membership degrees to individual instances. It fits pattern-aware regression models so predictions can follow recurring or overlapping concepts.
Read the paperFor the complete publication record, visit my Google Scholar profile