How should dialogue state tracking change when user preferences shift mid-conversation?
This explores why the classic 'fill-in-the-slots' model of dialogue state tracking breaks when a user changes their mind partway through a conversation, and what the corpus suggests should replace it.
This reads the question as a challenge to the assumption baked into most dialogue state tracking (DST): that a user walks in with one stable set of goals and the system's job is just to fill in the blanks. The corpus suggests that assumption is exactly what fails when preferences shift mid-conversation — and points toward several repairs that don't usually get filed under 'DST.'
The sharpest critique comes from negotiation research: standard DST is a form-filling paradigm built around a single user's fixed goals, which is why it can't capture strategic moves or commitments that evolve as the conversation unfolds Why do standard dialogue systems fail at tracking negotiation agreement?. If a system can't even track two parties revising a shared agreement, it certainly can't track one person reversing themselves. The deeper fix here is treating state not as a static slot-filling snapshot but as a belief that updates turn by turn — which is what collaborative rational speech acts offer, modeling the progression from partial to shared understanding with an information-theoretic frame that token-level LLM systems lack Can dialogue systems track both speakers' beliefs across turns?.
A preference shift is, first, something you have to *notice*. The corpus is blunt that current models are bad at this: tested across health scenarios, leading LLMs only succeed once a user has a firm goal and completely miss ambivalence, resistance, or someone in the middle of changing their mind Why can't chatbots detect when users are ambivalent about change?. State tracking that can't detect wavering will happily keep optimizing toward a goal the user has already abandoned. Relatedly, persona drift research shows conversations have distinct failure modes — local drift within a turn versus global drift across the whole exchange — which is a useful lens: a genuine preference change and accidental drift look similar but should be handled oppositely Can training user simulators reduce persona drift in dialogue?.
Once a shift happens, *where* in the conversation it happened matters. Treating the dialogue as an ordered sequence rather than a bag of mentions recovers exactly this signal — modeling items and entities in the order they appear captures 'I wanted X, then changed to Y' dependencies that order-blind approaches discard Does conversation order matter for recommending items in dialogue?. And when you need to act on a shift, segment-level optimization beats both extremes: turn-level is too granular to see the change, session-level drowns it in noise, but isolating the segment around the pivot turn localizes it cleanly Does segment-level optimization work better for multi-turn dialogue alignment?.
The most counterintuitive takeaway: don't throw away the old state when preferences change. Conversational recommender work argues for keeping three preference channels — the current session, historical dialogues, and look-alike users — all conditioned on *current* intent Can conversational recommenders recover lost preference signals from history?. A mid-conversation shift, on this view, isn't a reason to wipe the slate; it's a re-weighting toward the present signal while history stays available, since today's reversal might itself reverse tomorrow.
Sources 7 notes
Standard dialogue state tracking assumes one user's goals; negotiation requires explicit agreement from both parties across multiple issues. Existing DST models, limited to form-filling paradigms, cannot capture the strategic dynamics and mutual commitments essential to genuine bilateral agreement.
CRSA integrates rate-distortion theory with RSA to enable bidirectional belief tracking across dialogue turns. Demonstrated on referential games and doctor-patient dialogues, it captures progression from partial to shared understanding, providing the information-theoretic framework that token-level LLM systems lack.
Testing three major LLMs across 25 health scenarios showed they succeed only when users have established goals but cannot detect resistance or ambivalence. Models miss relapse-prevention strategies even for users in action stages.
By inverting standard RL setups to train user simulators for consistency using three complementary metrics (prompt-to-line, line-to-line, Q&A consistency) as reward signals, persona drift decreases by over 55%. This approach captures distinct failure types: local drift within turns, global drift across conversations, and factual contradictions.
TSCR models items and entities in the order they appear in CRS dialogue, using transformers to learn dependencies between sequential mentions. This recovers information that bag-of-mentions approaches discard, improving recommendation accuracy on standard benchmarks.
SDPO identifies erroneous turns and optimizes surrounding segments, achieving simultaneous improvements in goal completion and relationship quality. Turn-level DPO is too granular; session-level introduces noise from irrelevant turns.
Current CRS systems only use the active dialogue session to infer preferences, losing item-CF and user-CF signals proven valuable in traditional recommenders. Integrating current session, historical dialogues, and look-alike users—conditioned on current intent—recovers essential user representation structure.