Three of our papers were accepted to the ACL 2026 main conference. They all touch the same big question: how we build structured knowledge, make it useful for models, and test reasoning when there is no single right answer on a static benchmark.
AutoSchemaKG asks: do you really need a hand drawn schema before you can build a graph? Here, LLMs pull triples from text and figure out structure as they go, like entities, events, and concepts, as a flat list of nodes. We scaled this to web scale data and built ATLAS (hundreds of millions of nodes, billions of edges). It helps multi hop QA and factuality, and the induced schemas line up surprisingly well with human designed ones.
AutoGraph R1 starts from a very common practical issue: a graph can look fine on paper and still hurt RAG, because construction was never tuned for the task. We use reinforcement learning to train the builder: the reward is how the graph actually performs inside RAG, both as facts the model reads and as something to retrieve from, not a proxy that only scores local extraction. On several QA setups, task aligned graphs beat the usual task agnostic pipeline.
The third paper is multimodal abductive reasoning. Abduction is about proposing explanations from partial clues—not one VLM picking one answer. We took inspiration from Dixit: agents play different roles (e.g. a cryptic clue tied to an image, others pick among hard distractors) so we can study both generating hypotheses and telling good ones from bad under imperfect information. We also add DixitBench for a simpler listener only setup. In practice, smaller open models can be surprisingly strong as “storytellers,” while larger proprietary models often win as “listeners”—a nice reminder that creativity and reliability do not always show up in the same model.
Papers: AutoSchemaKG · AutoGraph R1 · Dixit style multimodal abductive reasoning.
Many thanks to everyone on these projects—students, collaborators, advisors. Looking forward to the conversations in California.
Photo: Tiantian Hainanese Chicken