Research

Multi-agent AI systems — where multiple LLM-based agents perceive, reason, and act together — concentrate a peculiar set of difficulties that single-agent setups largely sidestep. Problems of interest include:

• Structuring collaboration. Agents may cooperate, compete, or do both at once (coopetition), under centralized, decentralized, or hierarchical topologies, governed by rule-, role-, or model-based strategies — and no single combination dominates across tasks.
• Governance and collective decision-making. There is still no agreed framework for distributing tasks, assigning roles, and aggregating heterogeneous agent outputs into accurate joint decisions — without collapsing the useful disagreement that motivated using multiple agents in the first place.
• Scalability. Communication overhead, context length, and orchestration complexity grow quickly with agent count, and the field's cleaner demos rarely survive that growth intact.
• Reliability, safety, and accountability. LLM hallucinations, bias, and vulnerabilities do not stay isolated — they propagate through inter-agent dialogue, raising the stakes of misalignment, prompt-injection, and unclear responsibility for joint outcomes.

Multimodal understanding — getting a single model to perceive, align, and reason across text, images, video, 3D, and audio — has become the dominant theme of recent vision-language research. Problems of interest include:

• Cross-modal alignment and fusion. Designing interfaces — cross-attention, Q-Formers, projectors, prompt and adapter tuning — that let heterogeneous encoders share a usable representation without erasing the structure each modality contributes.
• Multimodal reasoning. Moving past alignment to composition: treating each modality as a partial projection of a shared world-knowledge manifold and unrolling evidence across modalities before producing an answer.
• Robustness and adversarial safety. Modern VLMs are "semantically strong but spatially fragile" — low-severity geometric and resampling perturbations often degrade performance — while the continuous, high-dimensional nature of images also leaves them open to imperceptible adversarial perturbations that transfer across models and bypass safety mechanisms.

Controllable image and video generation — where a model must obey the user's intent down to the level of pose, layout, identity, or motion — concentrates difficulties that unconditional generation sidesteps. Problems of interest include:

• Multi-signal conditioning. Combining text, sketches, depth, pose, masks, references, and audio into a single coherent generation, when the signals often conflict or under-specify.
• Fidelity and compositionality. Honouring complex specs without sacrificing visual quality — long-form coherent composition and generation from short-form videos is particularly challenging.
• Temporal and identity consistency. Keeping subjects, objects, and physics stable across frames or repeated generations, as small per-step errors compound.
• Evaluation and efficiency. Evaluation metrics for evaluating the controllability of the generated output is still an open problem.