Showcase Your
AI Innovations

We propose a scalable reward modeling framework that incorporates fine-grained human preference data to achieve better alignment in large language models, demonstrating a 23% improvement in human preference metrics.

We introduce MemAgent, a framework for persistent episodic memory in LLM-based agents that enables coherent multi-session interactions without context window limitations.

A novel cross-modal grounding approach that enables language-guided visual reasoning in embodied environments, achieving state-of-the-art performance on ScanQA and EmbodiedScan benchmarks.

A comprehensive framework for systematically red-teaming frontier language models across 12 risk categories, with automated discovery of novel attack vectors using adversarial agents.

We derive new empirical scaling laws governing multi-task learning in transformer models, demonstrating that task diversity can substitute for model scale at inference time under certain conditions.

A sparse attention mechanism specifically designed for temporal reasoning tasks, achieving 3.2x inference speedup with equivalent accuracy on benchmarks including TimeQA and TempReason.

We introduce LoRA-Merge, a gradient-aware composition algorithm that combines multiple LoRA adapters trained for different tasks without the catastrophic interference observed in naive weight averaging.

A family of lightweight vision transformer architectures for real-time object detection, achieving 45 FPS on mobile hardware while matching ResNet-50 baselines on MS-COCO.

A multi-agent deep RL approach for adaptive resource allocation in distributed computing systems, demonstrating 31% reduction in latency and 18% increase in throughput over heuristic baselines.

A regularised constraint-based algorithm for causal discovery in high-dimensional time series, with theoretical guarantees under non-stationarity and practical applications in financial markets.