Multi-Hop Reasoning

The engine must identify a bridge entity—a concept, statute, or party that connects two otherwise isolated documents. In legal contexts, this often involves recognizing that a cited case in Document A is the same precedent being interpreted in Document B. Failure to resolve this coreference leads to a broken reasoning chain and a hallucinated conclusion.

Unlike single-hop retrieval, multi-hop reasoning constructs a directed acyclic graph of dependencies. The answer to sub-question 1 becomes the retrieval query for sub-question 2. Key characteristics include:

• Strict ordering: Step 2 cannot execute before Step 1 resolves.
• Fan-out risk: A single error in an early hop compounds exponentially through the chain.
• Dead-end recovery: The system must backtrack when a reasoning path yields no supporting evidence.

The core challenge is the compositional gap—the information required to answer the query is distributed across documents, and no single passage contains the full answer. The model must perform latent alignment of partial facts. For example, Document A states 'Party X breached Clause 14,' and Document B defines 'Clause 14 as requiring 30-day notice.' The conclusion that 'Party X failed to provide 30-day notice' requires composing these two facts.

To mitigate hallucination, every multi-hop conclusion must be accompanied by a complete evidence path. This is a verifiable audit trail that explicitly cites the source document and specific passage for each inferential step. A robust system outputs not just the final answer, but the full [Doc A: Para 3] -> [Doc B: Para 12] -> Conclusion chain, enabling downstream citation verification systems to validate each hop.

Iterative retrieval generates a new query for each hop based on the information gained in the previous step, dynamically interacting with the vector database. This contrasts with one-shot retrieval, which attempts to gather all potentially relevant documents upfront. Iterative methods are more precise for complex legal synthesis but introduce higher latency and require sophisticated query decomposition logic.

A critical failure mode occurs when intermediate hops retrieve contradictory information. For instance, a higher court's ruling in one document may overturn the precedent found in another. The reasoning engine must apply normative conflict resolution logic, typically prioritizing documents by temporal freshness, jurisdictional hierarchy, or explicit abrogation signals, rather than simply concatenating conflicting facts.

A prompting technique where a model drafts a response, generates a series of fact-checking questions about its own output, and then revises the initial response to correct inconsistencies. For multi-hop reasoning, CoVe is critical because it systematically interrogates each logical bridge between documents, asking questions like 'Does Document A's definition of X match Document B's usage of X?' to prevent synthesis errors.

The process of anchoring a language model's generative capabilities to a structured or unstructured knowledge base. In multi-hop legal reasoning, grounding is not a single-step lookup but a transitive process: the model must ground intermediate inferences in Document A, then use that grounded fact as a query to ground the next inference in Document B, creating a verifiable chain of provenance.

Two metrics that measure the integrity of a model's output. Citation Recall is the proportion of factual claims correctly supported by a citation. Citation Precision is the proportion of provided citations that genuinely support the associated claim. In multi-hop reasoning, a failure in either metric often signals a broken logical chain—where a model cites a source that does not contain the claimed fact or fails to cite the document that bridges two concepts.

The computational task of identifying mutually exclusive statements within a single document or across a multi-document corpus. This is a primary failure mode of multi-hop reasoning, where a model might synthesize a conclusion from Document A and Document B without recognizing that a third document, Document C, explicitly contradicts the premise. Robust contradiction detection acts as a logical safety net for the synthesis process.

The field of reverse-engineering the internal computations of a neural network into human-understandable algorithms. Applied to multi-hop reasoning, researchers aim to locate the specific attention heads and MLP layers responsible for binding information across different context windows. This allows for direct editing of faulty 'reasoning circuits' that cause a model to hallucinate when connecting facts from separate documents.