Multi-Hop Reasoning

The defining logical structure where the answer to one sub-question is a prerequisite for formulating the next. The system cannot retrieve the final answer in a single step; it must follow a chain of evidence.

• Bridge Entity Resolution: Step 1 identifies an intermediate entity (e.g., a person's name), which becomes the search term for Step 2.
• Contrast with Parallel Retrieval: Unlike simple factoid QA where all keywords are known upfront, sequential dependency requires dynamic query reformulation based on intermediate results.

The process of collecting and fusing information fragments from multiple distinct source documents to construct a complete answer. No single passage contains the full answer.

• Redundancy Handling: The system must identify when two passages provide the same fact versus complementary facts.
• Conflict Resolution: When sources contradict each other, the reasoning engine must weigh source authority and recency to select the ground truth.

The capability to connect two entities or facts that are never explicitly linked in any single text. The system must infer the relationship by recognizing that Entity A is connected to Entity B, and Entity B is connected to Entity C.

• Transitive Reasoning: Applying the logical property of transitivity across text spans.
• Graph Completion: Effectively predicting a missing edge in a latent knowledge graph constructed dynamically from unstructured text.

The initial planning step where a complex natural language question is parsed into a structured set of simpler, answerable sub-questions. This decomposition defines the reasoning topology.

• Decomposition Types: Can be sequential (linear chain), parallel (independent branches), or a directed acyclic graph (DAG) of operations.
• Tool Assignment: Each sub-question may be routed to a different tool, such as a vector search, a SQL executor, or a calculator.

A self-correcting mechanism where the system evaluates the validity of a reasoning chain and discards dead ends. If an intermediate step yields an entity that doesn't exist or a logical contradiction, the system backtracks to explore alternative paths.

• Search Algorithms: Implements strategies like beam search or Monte Carlo Tree Search (MCTS) over the space of possible reasoning traces.
• Hallucination Mitigation: Prevents the model from committing to a false premise early in the chain by verifying each hop against the retrieval corpus.

A reasoning pattern requiring the system to retrieve attributes of multiple entities and perform a comparative operation (e.g., greater than, less than, closest to) to determine the answer.

• Example: 'Which company had higher revenue, the one founded by A or the one founded by B?' requires retrieving revenue for both companies and comparing the numerical values.
• Numerical Reasoning: Often requires the integration of a symbolic solver or code interpreter to perform the final arithmetic comparison accurately.

The technique of breaking a complex, multi-faceted user query into a set of simpler, independently answerable sub-questions that can be solved sequentially or in parallel.

• Sequential Decomposition: Sub-question B depends on the answer to A
• Parallel Decomposition: Sub-questions are independent and can be executed concurrently
• Example: "Which company acquired the startup that invented the transformer architecture?" decomposes into: (1) Who invented the transformer? (2) Who acquired that company?

An intermediate, often unmentioned entity that must be identified to connect two pieces of information across different documents.

• Serves as the critical logical link in a multi-hop reasoning path
• Requires the model to infer an implicit connection rather than match explicit keywords
• Example: To answer "Where was the CEO of DeepMind born?", the bridge entity is the CEO's name, which must first be resolved from one document before querying their birthplace in another

A dynamic search process where the initial query is repeatedly reformulated based on newly retrieved information.

• Each retrieval round gathers additional context required to resolve the original complex question
• Contrasts with single-shot retrieval that attempts to gather all evidence in one pass
• Key advantage: Allows the system to discover information needs that were not apparent in the original query formulation

A reasoning paradigm where the model generates intermediate rationales and retrieves supporting evidence for each step.

• Interleaves retrieval with the generation of a logical path to the final answer
• Each reasoning step can trigger a new retrieval call to ground the next inference
• IRCoT (Interleaving Retrieval with Chain-of-Thought) is a specific implementation that uses generated rationale sentences as search queries

The algorithmic process of navigating a structured semantic network by following relationships from a starting entity across multiple hops.

• Uses explicit edge types (e.g., "founded_by", "acquired", "located_in") to constrain reasoning paths
• Provides deterministic grounding compared to probabilistic text-based retrieval
• Example: Entity A → [acquired_by] → Entity B → [headquartered_in] → Location C

A prompting framework that interleaves discrete reasoning traces with tool-use actions.

• Enables a language model to dynamically plan, execute, and update its strategy based on external feedback
• Reasoning traces: "I need to find X, so I should search for Y"
• Actions: API calls to search engines, calculators, or databases
• The observation from each action feeds back into the next reasoning step, creating a closed loop