Source-linked AI summary
Artificial Expert Intelligence through PAC-reasoning
Shai Shalev-Shwartz, Amnon Shashua, Gal Beniamini, Yoav Levine, Or Sharir, Noam Wies, Ido Ben-Shaul, Tomer Nussbaum, Shir Granot Peled
TL;DR
AI systems can have strong task-specific skills yet lack expert-like precision and reliable adaptation to novel, complex problems. The paper proposes PAC reasoning, using structured decomposition and empirical validation to control reasoning error during inference. Its theoretical results give high-probability approximate-correctness guarantees under stated assumptions, while reliance on domain-specific oracles and validators limits the scope described.
Problem
AI systems can excel at narrow skills, but the paper identifies a gap in precision and reliable adaptation to novel, complex problems.
Source: §1 · p. 2Method
PAC reasoning decomposes problems into functions and uses example validators and sample-based critics to control error during inference.
Source: §1 · p. 2, §2.1 · p. 4Results
The bottom-up and top-down theorems establish high-probability ε-approximate correctness for bounded computation graphs or search trees under their stated assumptions.
Source: §2.1 · p. 5, §2.2 · p. 8Takeaways & Limitations
The framework offers a way to control reasoning precision by investing inference-time compute in generating examples for learning.
Source: §1 · p. 2Reliance on domain-specific decomposition oracles and example validators challenges scaling to highly diverse or poorly understood domains.
Source: §3 · p. 8
Abstract
from arXiv · showhide
Artificial Expert Intelligence (AEI) seeks to transcend the limitations of both Artificial General Intelligence (AGI) and narrow AI by integrating domain-specific expertise with critical, precise reasoning capabilities akin to those of top human experts. Existing AI systems often excel at predefined tasks but struggle with adaptability and precision in novel problem-solving. To overcome this, AEI introduces a framework for ``Probably Approximately Correct (PAC) Reasoning". This paradigm provides robust theoretical guarantees for reliably decomposing complex problems, with a practical mechanism for controlling reasoning precision. In reference to the division of human thought into System 1 for intuitive thinking and System 2 for reflective reasoning~\citep{tversky1974judgment}, we refer to this new type of reasoning as System 3 for precise reasoning, inspired by the rigor of the scientific method. AEI thus establishes a foundation for error-bounded, inference-time learning.
1 Introduction
The paper defines Artificial Expert Intelligence (AEI) as combining AI’s domain-specific knowledge and skills with the innovative synthesis and critical judgment of leading human experts. It argues that expert-level intelligence requires precise, inference-time learning and proposes System 3 reasoning, inspired by the scientific method, to support it.
1 Introduction
- AEI combines current AI’s knowledge and skills with the innovative synthesis and critical judgment characteristic of top human experts.The authors describe this as a constructive definition and blueprint for AEI systems.Source: §1 · p. 1
- Generality is neither necessary nor sufficient for intelligence, since humanity’s greatest achievements often come from profound expertise in focused domains.Source: §1 · p. 2
- In-context learning occurs during inference, but has proved limited in solving novel, complex tasks.The paper frames inference-time skill acquisition as a measure of intelligence.Source: §1 · p. 2
- Current reasoning systems invest test-time compute in searching reasoning chains, not learning, while an incorrect claim can collapse a general-reasoning chain.Unlike games, general reasoning does not have rules that dictate valid next steps.Source: §1 · p. 2
- The proposed third type of reasoner gathers empirical observations to guarantee correctness at each step and controls ε by generating sufficient examples at inference time.The paper contrasts this approach with logical reasoners, which are limited to narrow domains, and intuitive reasoners, whose fixed stepwise error compounds over long chains.Source: §1 · p. 2
- The authors call this precise, scientific-method-inspired reasoning System 3, distinct from System 1 intuition and System 2 reflective reasoning.They describe System 3 as an invented mode of thought developed in recent centuries.Source: §1 · pp. 2–3
2 From PAC-learning to PAC-reasoning
The section develops PAC-reasoning as inference-time problem decomposition with controlled error, addressing the decay in chain accuracy caused by subproblem errors. It formalizes bottom-up and top-down approaches and gives probabilistic accuracy guarantees for both.
2 From PAC-learning to PAC-reasoning
- Classical PAC learning assumes labeled samples and a predefined hypothesis class; broad classes can be computationally infeasible, while deep learning can still fail to generalize on simple tasks.The section motivates reasoning chains as a response to these learning challenges.Source: §2 · p. 3
- A fixed average subproblem error ǫ makes intuitive reasoners’ whole-chain success decay as exp(−k ǫ), limiting reasoning-chain length.Reasoning can reduce sample complexity by decomposing a problem, but accumulated subproblem errors reduce overall precision.Source: §2 · p. 3
- PAC-reasoning formalizes inference-time decomposition and controls subproblem errors, allowing arbitrarily long reasoning chains or an explicit “I don’t know” response.The framework presents success with a probably approximately correct solution or admission of failure as its two outcomes.Source: §2 · p. 3
2.1 Bottom-up Reasoning
- Bottom-up reasoning searches for a computation graph using actor proposals, a critic that filters approximately correct proposals, and a decomposition oracle that validates the overall decomposition.The actor supplies an Example Validator for proposed functions; the decomposition oracle checks whether the graph’s specification correctly decomposes g.Source: §2.1 · pp. 4–5
- With their respective sample bounds, bottom-up outputs ǫ-approximate g with probability at least 1−δ under an (ǫ/2)-correct decomposition oracle, while top-down implementations meet the same guarantee.The bottom-up theorem bounds the computation graph size by kmax; the top-down theorem bounds the number of functions by kmax.Source: §2.1 · p. 5, §2.2 · p. 8
2.2 Top-Down Reasoning
- Top-down reasoning avoids length-specific graphs and repeated helper-function implementations, which make the merge-sort computation-graph approach inefficient and require more reasoning steps.Its motivation is that helper functions such as “merge” are called repeatedly with different inputs.Source: §2.2 · p. 6
- Top-down search implements declared functions using reference implementations for helpers, then tests each helper across all calls induced by sampled inputs from the original distribution.The critic uses these tests to identify approximately correct proposals; the approach assumes helpers are implemented after their usage.Source: §2.2 · p. 7
3 Discussion
AEI presents System 3 reasoning as a way to combine empirical validation with structured problem-solving, using bottom-up and top-down reasoning across a broad range of problems. Its reliance on domain-specific validators and decomposition oracles poses scaling challenges in highly diverse or poorly understood domains.
3 Discussion
- AEI’s System 3 combines empirical validation with structured problem-solving, and its bottom-up and top-down reasoning is presented as flexible across problem types.The discussion names algorithmic, engineering, and scientific problem-solving as examples.Source: §3 · p. 8
- Domain-specific decomposition oracles and example validators pose scaling challenges in highly diverse or poorly understood domains.The discussion identifies autonomously constructing these tools as a future research direction.Source: §3 · p. 8
- AEI emphasizes precision alongside capability and is presented as having potential to support expert-level reasoning and innovation.The discussion frames this as a potential, not an established outcome.Source: §3 · p. 8