Multi-Hop Reasoning

Analyzes complex legal documents by connecting clauses across hundreds of pages. A system must link a definition clause on page 4 with a limitation of liability on page 78 and a governing law clause on page 102 to assess total risk exposure.

• Entity Chaining: Identifies defined terms and traces their usage through the document graph.
• Conflict Detection: Flags logical contradictions between separate but related clauses.
• Precedent Synthesis: Connects a clause to external case law databases to validate enforceability.

Synthesizes patient data from unstructured physician notes, structured lab results, and genomic reports to identify non-obvious diagnoses. The system must connect a symptom mentioned in a radiology report with a rare genetic marker found in a separate sequencing file.

• Temporal Reasoning: Orders events across different timestamped documents to establish causality.
• Evidence Linking: Connects a specific medication from the pharmacy record to a side effect noted weeks later in a specialist's note.
• Differential Generation: Produces a ranked list of possible conditions by combining findings from multiple independent sources.

Detects sophisticated money laundering rings by connecting seemingly innocuous transactions across multiple accounts, geographies, and time zones. A single hop might link a wire transfer to an account; the second hop connects that account to a shell company; the third reveals a common beneficial owner.

• Graph Traversal: Walks a knowledge graph of entities and transactions to find hidden paths.
• Pattern Matching: Identifies circular fund flows where money returns to its origin through a chain of intermediaries.
• Anomaly Correlation: Links a flagged transaction in one system to a suspicious phone call log in another.

Answers a scientific research question by aggregating findings from multiple papers. A query like 'What is the downstream effect of Drug A on Protein C?' requires finding Paper 1 (Drug A inhibits Gene B), Paper 2 (Gene B upregulates Protein C), and synthesizing the transitive relationship.

• Cross-Document Coreference: Resolves that 'the inhibitor' in Paper 2 refers to Drug A from Paper 1.
• Contradiction Resolution: Identifies and weighs conflicting evidence from different studies to provide a nuanced answer.
• Hypothesis Generation: Proposes novel connections by bridging two previously unlinked bodies of literature.

Traces the cascading impact of a single event, like a port closure, through a multi-tier supply network. The system must hop from the closed port to specific delayed shipments, then to affected manufacturing lines, and finally to customer order fulfillment delays.

• Bill of Materials Explosion: Recursively traverses component dependencies to find all affected finished goods.
• Supplier Mapping: Links a Tier-1 supplier's delay to a Tier-3 raw material provider's geographic risk.
• Inventory Rebalancing: Calculates the optimal re-routing of inventory across the network by analyzing connected node capacities.

Resolves complex technical issues by linking a user's vague problem description to multiple internal knowledge base articles, past support tickets, and product specification sheets. The system must infer that 'the screen is flickering' on Model X, combined with a recent driver update log, points to a specific known firmware conflict.

• Symptom Clustering: Groups disparate user-reported symptoms to identify a common root cause.
• Temporal Correlation: Links a spike in similar tickets to a recent product change or external event.
• Resolution Path Synthesis: Generates a step-by-step fix by merging instructions from two separate articles.

The foundational technique for eliciting intermediate reasoning steps from a language model. By providing examples that include a reasoning trace (few-shot) or a simple trigger phrase (zero-shot), the model is guided to decompose a complex problem into a sequence of explicit logical steps before stating a final answer. This process makes the latent multi-hop connections visible and auditable.

A reasoning trace that accurately reflects the true causal process by which the model arrived at its answer. This is a critical distinction from a merely plausible-sounding explanation. A faithful chain-of-thought is free from post-hoc rationalization, where a model generates a convincing but inaccurate justification after the fact, masking the heuristics it actually used.

A training methodology that rewards a model for the correctness of each intermediate step, not just the final outcome. A Process Reward Model (PRM) is trained to evaluate these steps, providing granular feedback that directly incentivizes logically sound multi-hop reasoning. This contrasts with outcome supervision, which only verifies the final answer and can inadvertently reward faulty logic that coincidentally leads to a correct result.

An extension of chain-of-thought that explores multiple reasoning paths simultaneously in a tree structure. At each step, the model generates several possible next thoughts, evaluates their promise, and can look ahead and backtrack. This framework is designed for complex multi-hop problems where a single linear chain is insufficient, enabling deliberate planning and exploration of different document-connection strategies.

A prompting paradigm that interleaves Reasoning traces and Action steps. For multi-hop reasoning, this allows a model to not only think about what it needs to know but also to take an action—like querying a search engine or a database—to retrieve the specific external document required for the next logical hop. This grounds the reasoning chain in retrieved facts.

A critical failure mode in multi-hop reasoning where an initial factual error in a reasoning chain causes a cascade of subsequent errors. The model builds further logic on an incorrect premise, leading to a completely fabricated conclusion. This is a primary risk in multi-hop tasks, as the model must connect information across multiple sources, and a single broken link corrupts the entire chain.