What happens to dialogue coherence when topic models use rigid stacks instead of flexible revisitation?
This explores what breaks in a conversation when an AI tracks topics as a rigid push-down stack — handle the newest topic, pop it, return to the last — versus letting it freely jump back to any earlier thread.
This explores what breaks in a conversation when an AI tracks topics as a rigid push-down stack — handle the newest topic, pop it, return to the last — versus letting it freely jump back to any earlier thread. The corpus's clearest answer is that stacks lose what's already been popped: once a topic is closed off, returning to it means the context is gone, because a stack only keeps the path back down, not the whole landscape of what was discussed Why do dialogue systems lose context when topics return?. Human conversation doesn't behave like a stack — we interleave, drop threads, and circle back hours later — so the structure itself fights the way people actually talk. Transformer attention sidesteps this by letting a model reach any previous turn directly, with no structural "pop" that discards it.
But swapping the stack for flexible attention doesn't automatically buy coherence — it just removes one specific failure. The corpus identifies four distinct ways dialogue coherence actually breaks: contradiction, coreference confusion (losing track of what "it" or "she" refers to), irrelevancy, and declining engagement What semantic failures break dialogue coherence most realistically?. A rigid stack maps neatly onto two of these: when a revisited topic loses its prior context, the model either contradicts what it said before or muddles the references that depended on that lost thread. So the stack's cost isn't abstract "incoherence" — it's concretely coreference breakage and self-contradiction at the seams where topics return.
There's a deeper twist, though: even models with full attention still get "lost" when topics shift gradually. Across 200,000+ conversations, every major model dropped ~39% in performance on multi-turn tasks, because they lock into a premature early guess and can't unwind it Why do language models fail in gradually revealed conversations?. That's a stack-like pathology emerging *behaviorally* even without a literal stack — the model commits early, as if it pushed an assumption it can never pop. Related work shows models treat the opening prompt as a fixed frame and can't jointly revise shared assumptions when a user pivots Can LLMs truly update shared conversational common ground?. Rigidity, in other words, isn't only an architectural choice — it can be a learned habit.
The most surprising thread is that topic flexibility may be less about structure and more about training. One study found that fine-tuning on just 1,080 dialogues with distracting turns sharply improved a model's ability to stay on topic — the gap was a missing "what to ignore" signal, not missing capacity Why do language models engage with conversational distractors?. And keeping a conversation smooth across topic shifts — repairing references, handing off cleanly — turns out to be implicit *social* work that models never learn because training rewards predicting information, not maintaining a relationship Why don't language models develop conversation maintenance skills?. So the rigid-stack-versus-flexible-revisitation question is really a doorway to a bigger one: coherence over returning topics depends not just on whether you *can* reach old context, but on whether the model was ever trained to do the quiet maintenance work that makes revisiting feel seamless.
Sources 6 notes
Research shows stack-based dialogue structures lose context when popped topics are revisited, while transformer attention enables systems to retrieve any previous turn without structural loss. Attention-based approaches naturally support the interleaved, revisiting nature of human conversation.
Research using Abstract Meaning Representation identified four distinct incoherence types: contradiction, coreference inconsistency, irrelevancy, and decreased engagement. AMR-trained classifiers detect these semantic failures while text-level manipulations alone cannot.
Across 200,000+ conversations, all major LLMs show 39% average performance drop in multi-turn settings due to locking into incorrect early guesses. Agent mitigations recover only 15-20% of this loss.
LLMs interpret all subsequent conversational turns within a fixed initial prompt frame, preventing them from symmetrically proposing updates to shared assumptions. Even when users pivot topics or contradict earlier framings, the model cannot absorb revisions into jointly held background—making the user the sole maintainer of conversational scoreboard.
Fine-tuning on just 1,080 synthetic dialogues with distractor turns significantly improves topic resilience, revealing that the gap is not model capacity but absent training signal. Models learn to follow what-to-do instructions but not what-to-ignore instructions.
Humans keep conversations smooth through implicit techniques like reference repair and topic hand-off that sustain relational interaction, not convey information. Language models don't develop these because training signals reward information prediction, not relational work.