KBevo at COLM 2026: Co-Evolving Knowledge Base Construction and Reasoning in LLMs
yoavartzi · x · 2026-10-07
A Cornell-led team (Yoav Artzi, Kilian Weinberger, Jennifer Sun, et al.) is presenting Co-Evolving Structured Knowledge and Reasoning in Language Models (arXiv:2608.26386) at COLM 2026, introducing the KBevo framework.
- Key idea: Retrieval over unstructured text often injects irrelevant context and offers little control, while structured knowledge bases are costly to build and brittle to reason over. KBevo jointly learns knowledge base construction and reasoning over it, optimized end-to-end with QA-outcome rewards — so successful reasoning directly improves what gets stored.
- Results: Larger, better-connected knowledge structures with higher answer reachability, plus improved compositional factual reasoning and controllability over standard retrieval baselines.
- Accepted at COLM 2026; code and blog are available, with a poster session (#137) presentation.
More from Research
- Anders Sandberg: preprints are often more honest than final publications — anderssandberg · 2026-10-07
- Two counterexamples to the Shafarevich conjecture spotted, littmath jokes about double-dipping — littmath · 2026-10-07
- Anders Sandberg: co-writing a paper with Claude makes pre-registering hypotheses painless — anderssandberg · 2026-10-07
- Faster High-Assurance AEGIS Implementation in Jasmin for x86 AESNI — jedisct1 · 2026-10-07
- New COLM 2026 paper: flipped answers under swapped demographics don't prove LLM bias — yoavgo · 2026-10-07
- IR4RL turns intermediate render progress into RL rewards, new SOTA for image-to-code — phillip_isola · 2026-10-07