Neuro-symbolic AI

What was covered

• Why bother. LLMs are autoregressive, so errors compound and planning suffers. They hallucinate and are sensitive to prompt format. They behave like fast, intuitive thinkers and lack the slow, analytical kind. Accuracy drops when a problem needs many logical steps or contains irrelevant information.
• Solvers. Systems such as SMT (satisfiability modulo theories) are exact and can handle millions of variables, but they are slow and rigid for real-world problems. The proposal is to combine the two.
• Ways to combine them. Build rules, such as physics or business logic, into training, as AlphaFold does for protein structure. Let the LLM translate a real-world problem into a structured form (JSON or Python) that a solver can run, working as a tool call. Or let the LLM propose a solution, code or proof while the solver certifies it, iterating until it passes.
• Help in both directions. LLMs can assist solvers by proposing streamliners, constraints that shrink the search space. In an optimisation problem about arranging groups of golfers, that cut solving time from 10 hours to 3 minutes.
• Open challenges. Translating natural language into structured graphs accurately, the lack of clear confidence metrics, and the computing cost of scaling the combination.
• Where to learn more. Material and open-source repositories from the summer school, including tools for the Model Context Protocol.

What came up

• Marketing or substance? Felipe first took the term for marketing, then concluded the team already applies the principles: ReAct agents plus Python tools for complex calculations give the language model a more solid structure.

Takeaway

Ana Maria will keep exploring integration approaches and share what she finds.