"I tried deduction, induction, and abduction. When none of them worked, I tried seduction. The data still refused to talk."
A Syllogism That Skipped the Minor Premise
Overview
Science does not advance by observation alone. Raw data, no matter how abundant, remains inert until a reasoning process transforms it into explanations, predictions, and testable hypotheses. This chapter equips you with the computational machinery of reasoning: from the classical syllogism to the modern causal calculus, from simple pattern matching to the structured analogy engines that let AI systems transfer knowledge across domains.
We begin with the five forms of reasoning that underpin scientific thought: deduction (deriving necessary consequences), induction (generalizing from instances), abduction (inferring the best explanation), analogy (mapping structure across domains), and causal reasoning (distinguishing correlation from intervention). Each form has a precise computational realization, and each plays a distinct role in the discovery pipeline we will build throughout this book.
The chapter then turns to how large language models reason, or fail to reason, examining chain-of-thought prompting, scratchpad methods, and the emerging class of reasoning models that allocate extra computation at inference time. We formalize causal reasoning through Pearl's do-calculus, giving you the mathematical tools to distinguish "patients who take the drug recover more often" from "the drug causes recovery." Finally, we assemble a complete reasoning pipeline that proposes abductive explanations for observed phenomena and ranks them by plausibility and testability.
The reasoning machinery introduced here returns in Chapter 29 (Reasoning Models for Discovery), Chapter 31 (Causal Discovery and Causal Inference), and Chapter 39 (Hypothesis Generation), where each thread is developed to production depth.
Prerequisites
This chapter assumes you have read Chapter 1: Discovery as Search (for search-space vocabulary), Chapter 2: Scientific Discovery and Knowledge Creation (for the hypothesis-experiment cycle), and Chapter 3: Knowledge Representation (for graph and logic formalisms). You should be comfortable with Python, basic propositional logic, and elementary probability. Familiarity with directed graphs is helpful but not required.
What You Will Learn
- Distinguish five forms of reasoning and identify which form applies in a given scientific scenario
- Implement deductive inference with resolution and unification in Python
- Build structure-mapping engines for analogical transfer using Gentner's framework
- Formalize causal questions using Pearl's do-calculus and the three rules of causal inference
- Prompt language models for chain-of-thought and scratchpad reasoning, and diagnose common failure modes
- Construct a reasoning pipeline that generates abductive hypotheses and ranks them by plausibility and testability
Chapter Sections
4.1 Forms of Reasoning
Deductive, inductive, abductive, analogical, and causal reasoning. Formal definitions, computational realizations, and their roles in scientific discovery. Resolution-based theorem proving and Gentner's structure mapping.
4.2 Reasoning in Language Models
Chain-of-thought prompting, scratchpad computation, tool-augmented reasoning, and reasoning models. AlphaGeometry as a case study. Systematic reasoning failures and mitigation strategies.
4.3 Causal Reasoning and the Do-Calculus
Interventions versus observations, counterfactuals, $P(Y \mid do(X))$, the three rules of do-calculus, and backdoor/frontdoor criteria. Computational identification of causal effects from observational data.
4.4 Building a Reasoning Pipeline
Recipe: a pipeline that proposes abductive explanations for observed data, scores them by plausibility and testability, and outputs ranked hypotheses ready for experimental design.
What's Next
With reasoning machinery in hand, Chapter 5: Discovery Through Data, Models, and Simulation turns to the raw materials that reasoning operates on: datasets, statistical and machine-learning models, and computational simulations. Where this chapter asks "how do we draw conclusions?", Chapter 5 asks "from what evidence, and with what computational tools?" Together, the two chapters complete the intellectual toolkit that Chapter 6 assembles into a coherent Discovery Workbench architecture.
Bibliography
Foundational Papers
The original formulation of structure-mapping theory, which defines analogy as the transfer of relational structure (not surface features) between domains. The foundation for all computational analogy engines discussed in Section 4.1.
Introduced resolution refutation, the workhorse of automated theorem proving. Our deductive reasoning implementation in Section 4.1 builds directly on this method.
Demonstrated that prompting models to show intermediate steps dramatically improves performance on arithmetic, commonsense, and symbolic reasoning tasks. Central to Section 4.2.
AlphaGeometry combines a neural language model with a symbolic deduction engine to solve olympiad-level geometry problems. A key case study in Section 4.2.
Books
The definitive treatment of causal inference from a graphical-model perspective. Introduces do-calculus, backdoor and frontdoor criteria, and counterfactual reasoning. Required reading for Section 4.3.
Peirce's original articulation of abduction as "inference to the best explanation," the reasoning form most central to scientific discovery. The philosophical backbone of Section 4.1 and Section 4.4.
A computational theory of scientific revolutions showing how explanatory coherence drives theory change. Informs our plausibility scoring in Section 4.4.
Tools & Libraries
Open-source symbolic computation library used throughout this chapter for logical inference and algebraic manipulation.
Graph library used for causal DAGs, structure mapping, and reasoning graph construction in Sections 4.1 and 4.3.
Reference for the Claude API used in chain-of-thought prompting and the reasoning pipeline recipe in Sections 4.2 and 4.4.
Tutorials & Surveys
Comprehensive survey of reasoning capabilities and limitations in LLMs, covering deductive, inductive, abductive, and mathematical reasoning benchmarks.
Evaluates LLM performance on causal reasoning tasks including causal discovery, identification, and counterfactual reasoning. Context for Section 4.3.