Each extra turn stacks recognition and understanding error

What it is

Ordering a pizza: turn one hears “thick crust” as “thin.” Later turns that ask toppings and address treat thin as an established fact. The user thinks they are filling later slots; they are signing an order that is already wrong. Turn-wise error accumulation is this: each turn’s recognition or intent error becomes the next turn’s context. More turns are not merely more chances to mishear — an early error contaminates slots that have not been spoken yet. That is not the same claim as “short commands are better.” Even a goal that ought to take several turns will forward error along the dialogue.

Why it happens

Slot filling writes filled items into common ground. The next turn’s language model, anaphora, and defaults all read from that ground. A substitution on turn one (thick/thin), once treated as grounded, never re-enters the repair window if later confirmations only echo the new slot. Understanding errors travel the same way: “no cilantro” parsed as “cilantro” leaves the negation out of the state machine when the system later asks “what else.” People rarely audit turn one while answering turn three; working memory holds the current question, not a ledger of the whole form. Completion therefore falls with turn count often not because the last turn was noisy, but because the first mistake already locked later questions onto the wrong branch.

Studying it

In multi-turn slot filling, inject an error at a designated turn: substitute one slot on turn 1, inject nothing later, and score downstream slot accuracy, the turn at which the user notices and rolls back, and whether the final submit still carries that seed error. Control: the same task as a one-shot packed utterance. Do not stop at word error: score seed-error survival — is the first mistake still present at task end.

On live logs, histogram sessions by “turn of first system misunderstanding,” then overlay abandonment and wrong submits. If the mass of completed sessions still carries the slot value from that early failure, the problem is accumulation, not per-turn noise. Lab scripts that force an explicit confirmation of every filled slot each turn wipe the accumulation by construction and will not see the spread.

Where it stops holding

Confirming every filled slot every turn truncates accumulation, at the cost of turning the dialogue into a read-aloud. Short, closed, high-confidence slots (on/off, today/tomorrow) already have tiny per-turn error, so the accumulation term is near zero. When the user is watching a filled form on screen, visual store does the audit and spoken accumulation is no longer the only discovery path. The reverse also holds: silent adoption at low confidence makes accumulation worse than “misheard but asked immediately” — turns are not toxic by themselves; ungrounded assumptions treated as fact are.

Applying it

• For slots that will write the later agenda (the few that decide what gets asked next), confirm or highlight that item before it enters common ground, rather than a full read-back only at the end.
• Let the user name and change an earlier slot on any turn, and recompute downstream questions that slot pulls; do not keep asking down the old branch.
• Split “turn of first error” and “was that item still right at submit” into two metrics. Task success alone counts a seed-wrong completion as a win.
• How to check: inject a turn-one substitution in the test set, run the flow, read the submit. If the seed error is still there and the user was given no rollback point, the skill is accumulating.

Related

• Same group: M1.03.2 Every turn must make clear whose floor it is · M1.03.3 If the turn budget is spent, hand the task to another channel
• Nearby: M1.01 When Voice-First Is Appropriate · M1.06 Repair Strategies · C7.03 Types of Recognition Errors
• Search terms: turn-wise error accumulation · grounding failure cascade · slot error propagation