Current Practices for Building LLM-Powered Reasoning Tools Are Ad Hoc—and We Can Do Better
summary
Proposes a Neurosymbolic Transition System that pairs a standard small-step transition system (states, nondeterministic steps) with a domain of intuition and two operators—combine/infer operators—so each initial state and each transition carries an LM-updatable intuition. Runtime runs symbols and intuition in lockstep: LMs act as the infer operator to bias nondeterministic grammar productions, verification predicates both use and emit intuition, and logic-programming search is guided but still backed by the symbolic transition system. Prototype is not implemented—currently vaporware.
AI-written summary · about
chapters (7)
- 0:00Neurosymbolic reasoning tools
- 4:04Sequential architectures limitations
- 5:48Neurosymbolic transition systems
- 8:00Model properties and expressivity
- 9:04NSTS programming language
- 11:16Conclusion and future work
- 11:59Q&A
related talks
53mSemantics-based program manipulation in the age of LLMs
22mTowards Verified Code Reasoning by LLMs
30mToward Programming Languages for Reasoning -- Humans, Symbolic Systems, and AI ...
33mLeveraging LLM Reasoning with Dual Horn Programs
12mNeuroSync: Intent-Aware Code-Based Problem Solving via Direct LLM Understanding Modification
01h 00mFrom LLM Theory to Practical Agentic Implementations
26mLanguage Model Agents Enable Semi-Formal Programming
44mJoint Keynote Address: Accelerating Software Development: The LLM (R)evolution