Making language models more reliable and more data-efficient — from agentic systems and memory management for multi-agent setups, to alignment that needs less human feedback, to retrieval and multimodal reasoning over text, image, and tabular data.
The first principled flow diffusion for graph-based retrieval-augmented generation, with online query-aware edge reweighting, exponential convergence guarantees, and state-of-the-art results across multiple benchmarks. Open sourced, with code and models released.
Combines partial-match rewards with verbal reinforcement learning so text-to-SQL models get useful learning signal from near-correct queries rather than only exact matches.
Defines an Uncertainty-aware Reward Model (URM) for preference training of LLMs, using a Bayesian approximation to quantify the uncertainty of paired responses. URM-based fine-tuning boosts instruction-following capability and policy optimization objectives, surpassing existing methods by a large margin on benchmarks such as Vicuna and MT-Bench.
Agent design and memory management for multi-agent systems (MAS), addressing how agents retain, share, and retrieve context across long-running collaborative tasks.
Alignment and fine-tuning techniques that reach strong instruction-following performance with substantially less human feedback, built on uncertainty-aware reward modeling.
RAG architectures for knowledge-intensive tasks, including graph-based retrieval with formal guarantees on which evidence a query recovers.
Reinforcement learning for structured query generation, using partial-match rewards and verbal feedback to make training signal denser and more informative.
Language models tailored to text, image, and tabular data, supporting business and engineering workflows where the inputs are rarely text alone.