Hallucination Mitigation

An architecture that injects relevant, up-to-date information from an external knowledge base into the model's prompt at inference time. Instead of relying solely on parametric memory, the model is grounded in provided context.

• Mechanism: A user query triggers a semantic search against a vector database. The top-k relevant document chunks are retrieved and prepended to the prompt.
• Key Benefit: Directly addresses knowledge cut-off dates and reduces factual drift by providing a non-parametric source of truth.
• Example: A support chatbot retrieves the latest product manual pages to answer a configuration question, rather than hallucinating a deprecated command.

A deterministic method that restricts the model's token generation process to a predefined schema or formal grammar. This prevents the model from generating structurally invalid or logically impossible outputs.

• Mechanism: During beam search or sampling, tokens that violate a specified JSON schema, regular expression, or context-free grammar are assigned a probability of zero.
• Key Benefit: Eliminates structural hallucinations in structured data extraction, ensuring the output is always parseable by downstream systems.
• Example: Forcing an LLM to output only valid ISO country codes when extracting location entities from a news article.

A prompting strategy where the model generates an initial response and then systematically plans and executes a series of independent fact-checking questions to verify its own output.

• Mechanism: (1) Draft baseline response. (2) Plan verification questions. (3) Answer questions independently, without attending to the original draft. (4) Cross-check and correct inconsistencies.
• Key Benefit: Reduces the compounding of errors by isolating the verification step from the generation step, preventing the model from justifying its own mistakes.
• Example: After writing a biography, the model asks itself "What is the exact birth date?" and "What is the official university name?" and validates against the retrieved context.

The practice of anchoring entity attributes and relationships to a deterministic, curated semantic network rather than relying on statistical token probabilities.

• Mechanism: Entities mentioned in the prompt are linked to canonical IDs in a graph database. The model is forced to use the graph's stored properties and edges as the sole source of truth for those entities.
• Key Benefit: Provides absolute factual precision for critical business entities (e.g., product specs, legal definitions) where probabilistic generation is unacceptable.
• Example: A financial analyst AI must pull the current CEO and market cap directly from a verified knowledge graph node, not from its training data.

An inference-time technique that generates multiple diverse reasoning paths for the same problem and selects the most consistent answer via marginalization.

• Mechanism: Instead of greedy decoding, the model samples a diverse set of chain-of-thought rationales using a high temperature. The final answer is chosen by a majority vote or by selecting the answer with the highest aggregate probability.
• Key Benefit: Mitigates the risk of a single erroneous reasoning path leading to a confident hallucination. It leverages the insight that there are usually many ways to reach a correct answer, but inconsistent paths lead to random noise.
• Example: For a complex math word problem, the model runs 10 different reasoning chains. If 8 result in "42" and 2 result in "43", the system outputs "42".

A supervised training approach that explicitly teaches the model to reject counterfactual queries by training on datasets containing both correct answers and deliberate refusals.

• Mechanism: The training dataset includes prompts that ask for impossible or false information. The target output is a calibrated refusal ("There is no known...") rather than a confabulated answer.
• Key Benefit: Reduces the model's sycophantic tendency to always provide an answer, teaching it a boundary between known facts and unknown gaps.
• Example: Training a model to respond "No such scientific study exists" when prompted to summarize a fabricated research paper, instead of inventing an abstract.

An architecture that grounds LLM outputs by first retrieving relevant information from an external knowledge base and providing it as context for generation. RAG is the primary defense against hallucination, ensuring the model synthesizes answers from supplied documents rather than parametric memory.

• Key Mechanism: Injects retrieved chunks into the prompt as authoritative context
• Benefit: Directly reduces factual fabrication by constraining the generation space
• Example: A support chatbot retrieves the latest product manual before answering a technical question