Induction and Inquiry via Probabilistic Reasoning over Language and Code
A new computational model combines symbolic knowledge (mental programs blending natural language and source code) with LLM-guided Bayesian learning to explain how humans acquire abstract knowledge from sparse, noisy data The model satisfies three key desiderata: data/compute efficiency, uncertainty quantification for intelligent inquiry, and representational flexibility for diverse concepts It successfully reproduces quantitative signatures of human inductive learning and active inquiry, includi
Analysis
TL;DR
- A new computational model combines symbolic knowledge (mental programs blending natural language and source code) with LLM-guided Bayesian learning to explain how humans acquire abstract knowledge from sparse, noisy data
- The model satisfies three key desiderata: data/compute efficiency, uncertainty quantification for intelligent inquiry, and representational flexibility for diverse concepts
- It successfully reproduces quantitative signatures of human inductive learning and active inquiry, including anchoring and garden-pathing effects
- Pure LLMs and classic Bayesian models fail to match human behavior or require exorbitant computational cost to approximate the same results
- The framework proposes that a bottom-up neural mechanism (LLM) makes sequential hypothesis revision tractable and learnable, bridging symbolic and subsymbolic approaches
Why It Matters
This work offers a compelling hybrid architecture that could inform the next generation of AI systems seeking human-like learning efficiency and reasoning flexibility. By demonstrating that LLM-guided Bayesian inference over program-like representations can reproduce nuanced human cognitive phenomena, it provides a concrete blueprint for building systems that are both computationally efficient and capable of active inquiry. For AI practitioners, this suggests a viable path beyond either pure scaling of neural networks or purely symbolic approaches.
Technical Details
- Mental Program Representation: Symbolic knowledge is encoded as programs that combine natural language descriptions with executable source code, enabling flexible representation of concepts spanning language-like and program-like structures
- LLM-Guided Bayesian Learning: Sequential inference over mental programs uses LLMs as a bottom-up neural mechanism to propose and revise hypotheses, approximating Bayesian updates in a computationally tractable manner
- Behavioral Validation: The model was tested across a range of behavioral studies, quantitatively reproducing human inductive learning signatures such as anchoring effects and garden-pathing phenomena
- Comparative Baselines: Pure LLMs and classic Bayesian models were evaluated as baselines; both failed to simultaneously achieve human-like behavior and computational efficiency, with classic Bayesian approaches requiring exorbitant computational cost
- Three Desiderata Framework: The model is explicitly designed to satisfy (1) data and compute efficiency, (2) gradations of uncertainty for intelligent information gathering, and (3) representational flexibility for learning an open-ended range of concepts
Industry Insight
- Hybrid neuro-symbolic architectures that combine LLMs with structured probabilistic reasoning over program-like representations may represent a critical direction for achieving human-level sample efficiency in learning systems
- The demonstrated ability to reproduce anchoring and garden-pathing effects suggests these frameworks can capture subtle cognitive biases, which is valuable for building AI systems with more predictable and interpretable reasoning behaviors
- Researchers and engineers should explore LLM-guided sequential hypothesis revision as a practical approach to making Bayesian inference tractable at scale, potentially unlocking more efficient and flexible AI learning systems
Disclaimer: The above content is generated by AI and is for reference only.