Multi-Turn Reasoning

The structured backbone of multi-turn reasoning that maintains a running representation of user goals, intents, and slot values across turns. DST estimates a probability distribution over possible dialogue states at each step, enabling the system to remember that a user asking 'What about tomorrow?' refers to a previously discussed flight booking. Modern approaches use generative DST where the full state is expressed as a natural language string rather than a fixed ontology, allowing open-domain tracking without predefined slot schemas.

The NLP task of linking pronouns, definite descriptions, and other anaphoric expressions to their antecedents across turns. Without coreference resolution, a system cannot understand that 'it,' 'that one,' and 'the cheaper option' all refer to the same entity mentioned three turns ago. End-to-end neural coreference models using span-ranking architectures now achieve near-human performance, resolving both identity and bridging references that are critical for maintaining topic continuity in extended dialogues.

The engineering discipline of fitting relevant conversation history within the model's finite token budget. Techniques include:

• Sliding window attention that drops oldest turns first
• Conversation summarization that compresses prior exchanges into dense abstracts
• Retrieval-augmented memory that fetches only semantically relevant history chunks Effective management prevents context collapse, where the model loses track of early constraints or user preferences established at the conversation's start.

The process of resolving uncertainty when a user utterance maps to multiple potential interpretations. In multi-turn settings, disambiguation leverages accumulated context to narrow possibilities—a query like 'Book it' could mean a flight, hotel, or restaurant reservation depending on prior turns. When confidence remains low, systems issue clarification questions targeting the specific ambiguity, such as 'Did you mean the 3 PM or 7 PM showing?' This proactive resolution prevents cascading errors in downstream task execution.

Extending Chain-of-Thought (CoT) prompting beyond single-turn reasoning to maintain explicit reasoning traces throughout a dialogue. The model generates intermediate reasoning steps that reference prior conclusions, creating an auditable logical chain. For example, when a user asks a follow-up math question, the system can reference its previously derived intermediate values rather than recalculating from scratch. This technique significantly improves accuracy on multi-hop reasoning tasks where answers depend on chaining facts across multiple exchanges.

A retrieval optimization that applies a cross-encoder model to reorder candidate documents based on their relevance to the full dialogue history, not just the current query. Standard retrieval treats each turn independently, but conversational reranking evaluates whether a document answers the user's implicit need given what has already been discussed. This prevents the system from retrieving redundant information already covered or missing documents that resolve a dependency established three turns prior.

The process of maintaining a structured representation of user goals, intents, and slots across multiple turns. DST acts as the working memory for multi-turn reasoning, updating a belief state as new information arrives. For example, in a booking system, DST tracks the destination, date, and time as the user reveals them incrementally, handling corrections like 'change that to Friday' without losing previously captured slots.

The system architecture component responsible for storing, summarizing, and recalling dialogue history to maintain stateful context. Multi-turn reasoning depends on short-term memory (the current session buffer) and long-term memory (vector-stored past interactions). Effective memory management prevents context window overflow by summarizing older turns while preserving critical entities and decisions for logical coherence.

The NLP task of identifying all expressions that refer to the same real-world entity. This is essential for multi-turn reasoning because users naturally use pronouns ('it', 'they') and definite references ('the first option') in follow-up turns. Without resolving 'it' to 'the Q4 financial report' from turn one, the reasoning chain breaks and the AI loses the thread of the conversation.

A prompting strategy that instructs the model to generate intermediate reasoning steps before arriving at a final answer. In multi-turn scenarios, CoT enables the model to externalize its logic across turns, allowing users to see and correct faulty assumptions. This transforms opaque reasoning into a collaborative, verifiable process where each step can be challenged or refined in subsequent exchanges.

The process of applying a cross-encoder model to reorder retrieved documents based on their relevance to the full conversational context, not just the current query. Multi-turn reasoning often requires information from documents retrieved in earlier turns. Reranking ensures that the most contextually relevant passages—considering the entire dialogue history—are prioritized for the final grounded generation step.

The process of resolving uncertainty when a user query maps to multiple potential intents. In multi-turn reasoning, ambiguity compounds across turns. A clarification question like 'Did you mean delete the draft or the published version?' is a reasoning act that prevents cascading errors. Effective disambiguation maintains the logical coherence of the entire interaction by establishing a shared, unambiguous ground truth before proceeding.