Why does coreference resolution become implicit in full-transcript prompting?
This explores why, when you hand an LLM an entire conversation transcript at once, it resolves references like 'it,' 'that one,' or 'the second option' as a side-effect of predicting the next token — rather than running coreference as a separate, explicit step the way classical NLP pipelines did.
This explores why coreference resolution disappears into the prompt rather than happening as its own visible step. The short version: when the whole transcript is the input, there's nowhere else for reference resolution to live. A prompt isn't just an utterance — it simultaneously carries the utterance, the shared context, and the role assignment, all bundled into one static frame the model reads in a single pass How do prompts reshape the role of context in AI conversation?. Because the antecedent of every pronoun is already sitting in that same window, the model never needs to *retrieve* a referent; it just continues the most probable text given everything visible. Coreference becomes implicit because it's been absorbed into ordinary next-token prediction over a context that already contains its own answers.
The interesting part is what that absorption hides. Resolving 'it' isn't a deterministic lookup — the model is sampling a consistent reading from a superposition of possible ones. Shanahan's 20-questions test shows an LLM doesn't commit to a single referent so much as hold a distribution of consistent continuations and draw from it at generation time; regenerate the same turn and you can get a different resolution, each internally coherent with the prior text Do large language models actually commit to a single character?. So 'implicit' here doesn't mean 'reliable and invisible' — it means 'never pinned down.' The reference is resolved fresh, probabilistically, every time the transcript is reprocessed.
That's also where the failure modes hide, and they're the thing worth knowing. When resolution is implicit, an early wrong guess about what 'it' or 'that' points to gets baked into the transcript and carried forward — models lock into premature assumptions in underspecified conversations and rarely recover, with a 39% average performance drop across multi-turn settings Why do language models fail in gradually revealed conversations?. And because the model can't renegotiate the framing mid-stream the way humans cooperatively repair a misunderstanding, fixing a botched reference requires you to explicitly re-prompt rather than the model implicitly adjusting How do prompts reshape the role of context in AI conversation?.
There's a deeper reason implicit resolution can quietly go wrong: it competes with the model's training priors. Even when the correct antecedent is right there in context, strong parametric associations can override what the transcript actually says, so the model resolves toward what it 'expects' rather than what was written Why do language models ignore information in their context?. Add to that the erosion of grounding behavior under preference optimization — clarifying questions and understanding-checks get trained out by roughly 77.5% below human levels — and the model is both less likely to ask 'which one do you mean?' and more likely to silently resolve a reference the wrong way Does preference optimization harm conversational understanding?.
The takeaway a curious reader might not expect: making coreference implicit is exactly what makes full-transcript prompting feel effortless *and* what makes it brittle. The same mechanism that lets a model track 'it' across twenty turns without any explicit machinery also means there's no checkpoint where a misresolution can be caught, no commitment to audit, and no built-in repair loop — the resolution is re-sampled, prior-biased, and unrecoverable all at once.
Sources 5 notes
LLM prompts bundle utterance, context assignment, and role specification into a single static frame the model cannot renegotiate, unlike human dialogue where context evolves cooperatively. This makes mid-conversation pivots require explicit re-prompting rather than implicit adjustment.
Shanahan's 20-questions test shows LLMs maintain a superposition of consistent objects or characters and sample from that distribution at generation time. Regenerating the same response yields different outputs, each consistent with prior context, proving no fixed commitment exists.
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.
Research demonstrates that LMs generate outputs inconsistent with their context because parametric knowledge from training dominates over in-context information. Textual prompting alone cannot override strong priors; causal intervention in representations is required.
RLHF optimizes models for single-turn helpfulness by rewarding confident responses over clarifying questions and understanding checks. This preference alignment systematically reduces grounding acts by 77.5% below human levels, creating an alignment tax where models appear helpful but fail silently in multi-turn contexts.