Dialogue Systems and Conversational State

A model handles an instruction well when it gets the whole thing at once. Split the same instruction into shards and release them one turn at a time, and it comes apart. That is the experiment: 200,000+ simulated conversations, six generation tasks, 15 LLMs. Not one model held up. 90% on the single-turn version of the instruction, 65% when the same information arrived in pieces — an average drop of 39%.

The decomposition is the part that matters for design. Aptitude, the level a model reaches at its best, fell by an average of 16%. Unreliability, the spread between its good runs and its bad ones on the same task, rose by an average of 112%. The models had not become much less able. They had become far less able to stay on one track once they had left it. “In simpler terms, we discover that when LLMs take a wrong turn in a conversation, they get lost and do not recover,” write Laban and colleagues, in LLMs Get Lost In Multi-Turn Conversation, published at ICLR 2026.

Every turn in such a conversation can be fluent, relevant and well formed. The reply is not wrong as a sentence. It is wrong as a move, because the commitment underneath it was made early, from an underspecified turn, and was never taken back.

Task-oriented dialogue completes a bounded workflow with explicit slots and actions. Retrieval dialogue selects a response from an approved collection and carries provenance. A generative assistant composes from context, evidence and policy, and pays for that flexibility in support and safety burden. Open-domain conversation maintains interaction without one narrow task, and has the hardest success definition of the four. Many products combine these rather than fitting one label.

The first column gets expensive at scale. Google's Schema-Guided Dialogue dataset, published in 2020, holds over 16k multi-domain conversations spanning 16 domains. Sixteen domains will not each hold still long enough for a hand-built, static ontology of intents and slots, and a system that hard-codes one per service cannot be extended to a new service without retraining. So the schema moves into the input: “Along the same lines, we present a schema-guided paradigm for task-oriented dialogue, in which predictions are made over a dynamic set of intents and slots, provided as input, using their natural language descriptions.” That is the abstract's own summary of the paradigm.

That design decision is testable rather than decorative, because the evaluation is built to break it. Google Research's own public release of the dataset states that “the dataset contains unseen domains and services in the evaluation set to quantify the performance in zero-shot or few-shot settings” — the state tracker is scored on APIs it never saw in training. Choosing a family is choosing what your system owes: a completed workflow, a citable source, a composed answer, or a conversation worth continuing.

Flattening the conversation into one summary can erase alternatives and corrections. The layers are kept apart for that reason. Raw turn history: exact messages, speakers, timestamps, attachments, channel events. Entities and candidate references, including the alternatives still under consideration. Goals, slots and constraints, including the values the user changed. Evidence and policy state, with freshness, permissions and unresolved conflicts. Action and repair state, naming the proposed action, its confirmation, and the responsible human or service.

Whether a committed state actually matches the turns it claims to summarise is a measurable question, and on the field's standard corpus the answer was no. MultiWOZ is the collection on which dialogue state trackers are scored. Eric and colleagues put its annotations back in front of crowd workers and re-checked them against the dialogues themselves. Their report, in 2020: “This correction process results in changes to over 32% of state annotations across 40% of the dialogue turns.” The repaired version is MultiWOZ 2.1.

One re-annotation did not exhaust the problem. An independent Google team noted in turn that “recent works have reported presence of substantial noise in the dialogue state annotations”, and went on to “identify and fix dialogue state annotation errors across 17.3% of the utterances on top of MultiWOZ 2.1”. That is MultiWOZ 2.2, later the same year.

These are the gold labels: state written down by careful annotators with the whole transcript in front of them and no time pressure. A third of them disagreed with the turns they were supposed to summarise, and it took two successive passes by two independent teams to find it. A running system, committing state under uncertainty one turn at a time, has no reason to do better. That is the argument for keeping the raw turns, the candidates and the commitments in separate, revisable layers: the commitment is a claim about the evidence, and the claim can be wrong while the evidence is still there to check it against.

A system can wait for the user, ask for missing information, offer options, or proactively warn about a conflict. Initiative affects task length, accessibility, user burden, and error exposure.

Use expected information gain, consequence, user preference, and available evidence rather than maximizing question count or reducing turns at all costs. A question asked early, while a wrong commitment is still cheap to retract, is worth several asked after the action has been taken.

Errors and ambiguity are expected events rather than exceptional failures, and in human conversation the rate has been counted. Dingemanse and colleagues coded spontaneous conversation in 12 languages from 8 language families, and counted every occasion on which one speaker signalled trouble with what another had just said. Their Results section reports: “In the 48.5 exhaustively sampled hours of conversation we find 2053 cases, meaning that there is a repair initiation about once every 1.4 minutes across all languages.” The abstract states the same finding as a property of the system rather than of any one culture — it is “frequently used (on average about once per 1.4 minutes in any language)”. That was published in PLOS ONE in 2015.

The structure is as regular as the rate. Three functionally distinct repair-initiator types account for a mean 92% of cases, with a standard deviation of 4.5%, and 95% of independent repair initiations fall within 4.13 minutes of the last. Repair is therefore not an error path bolted onto conversation. It is a small, closed, recurring set of moves that competent speakers of unrelated languages run roughly every minute and a half. A dialogue system that files repair under exception handling has built the abnormal path for the most normal event in the medium.

Put it on the main line in five steps. Detect a mismatch: conflicting values, low confidence, a failed tool result, a user correction. Surface the current understanding: briefly state the relevant date, entity, goal, or action under consideration. Ask a discriminating question: offer clear alternatives or request the missing evidence. Update and invalidate state: record the correction and retire the dependent assumptions and any action resting on them. Escalate with context: transfer history, evidence, unresolved issues and user consent to a human.

Intent accuracy, response relevance, and state-slot scores help localize failures. Conversation-level evaluation must also include task completion, unnecessary turns, successful repair, abandonment, handoff quality, user effort, and downstream outcome.

Use scenario suites with interruptions, corrections, topic changes, unavailable tools, and conflicting evidence. Human review should separate clarity, support, policy compliance, and conversational appropriateness.

The corpus most of these measurements are taken on arrived with the same layered structure. MultiWOZ is “a fully-labeled collection of human-human written conversations spanning over multiple domains and topics”. Budzianowski and colleagues released it at EMNLP in 2018. At 10k dialogues, they note, “it is at least one order of magnitude larger than all previous annotated task-oriented corpora.” Its baselines were reported not as one score but as three: the paper reports “a set of benchmark results of belief tracking, dialogue act and response generation”. That is the state, the act and the response measured apart.

The top of the ladder has its own benchmark. τ-bench, published at ICLR 2025, does not judge a customer-service agent's replies at all. It compares the database state at the end of the conversation against an annotated goal state — what was actually booked, cancelled or refunded. Then it adds pass^k, which requires all k independent trials of the same task to succeed. In the abstract, Yao and colleagues report: “Our experiments show that even state-of-the-art function calling agents (like gpt-4o) succeed on <50% of the tasks, and are quite inconsistent (pass^8 <25% in retail).”

Those two numbers are the whole argument for climbing. A system that finishes fewer than half the tasks it is given, and finishes the same task eight times running less than a quarter of the time, can still score well on relevance judged turn by turn. Only an end-state metric sees the first gap, and only a repeat-run metric sees the second.

Define states for issue identification, evidence collection, proposed resolution, confirmation, execution, failure, and human handoff. For each transition, list required evidence, allowed system actions, user-visible state, and rollback.

Run five conversations with corrections and ambiguous references, delivering the requirements a shard at a time rather than all at once. Check whether an old value can survive after its supporting assumption changes. Then run the same five a second and a third time, and compare the end states, not the transcripts. A scenario that passes once and fails on repetition is the unreliability finding reproduced on your own system.