Yun-Bo Long

A shared latent geometry reveals that paralinguistic speech attributes can be precisely controlled and disentangled from linguistic content within a frozen LLM.

LLMs can now autonomously construct and reason over dynamic knowledge graphs with calibrated uncertainty, opening up complex supply chain analysis previously impossible with one-shot queries.

LLMs optimized for human-like politeness are pushovers in negotiation, but EmoDistill unlocks their inner Machiavelli by teaching them to weaponize emotions for strategic advantage.

LLMs can be harnessed to generate synthetic data that respects complex operational constraints, overcoming the limitations of existing tabular generative models.

Explicitly regulating a research agent's mental model with feedback loops beats relying on implicit LLM reasoning, leading to significant gains on complex research tasks.

LLMs can be coaxed into significantly better performance on specific tasks by distilling their own outputs, but only if you surgically isolate the relevant capability using gradient-derived subspaces.

TabGRAA flips the script on tabular data synthesis, turning static statistical replication into a dynamic, self-improving generation process.

Strategic emotional intelligence, orchestrated by a Bayesian multi-agent system, unlocks negotiation superpowers for resource-constrained language models in high-stakes scenarios.