How do language models treat injected information as shared common ground?
This explores whether—and how—LLMs actually absorb information you put in front of them (in a prompt or mid-conversation) into a jointly held 'we both know this now' ground, versus treating it as something less binding.
This explores whether LLMs genuinely treat what you tell them as shared common ground—mutual knowledge both parties commit to and can revise together—or only as surface text that competes with everything else they've absorbed. The corpus answer is unsettling: they mostly don't. The clearest finding is that LLMs interpret a conversation through the frame of their initial prompt and can't symmetrically update that frame; when you pivot or contradict an earlier framing, the model can't fold your revision into a jointly held background, so the human ends up as the sole keeper of the conversational scoreboard Can LLMs truly update shared conversational common ground?. 'Common ground' in the human sense requires two parties who can both propose updates; the model only has one move.
Even setting aside the update problem, injected information doesn't reliably win the competition for the model's attention. Context loses to parametric memory: when training-time associations are strong, models generate outputs that contradict what's right there in the context window, and plain prompting can't override those priors—only intervening in the internal representations does Why do language models ignore information in their context?. So 'I told you X' is not the same as 'the model now treats X as true.' And there's a hard ceiling on what injection can do at all: prompting only reorganizes and activates knowledge already in the training distribution—it can't supply genuinely new foundational knowledge the model lacks Can prompt optimization teach models knowledge they lack?.
There's also a social wrinkle that masquerades as agreement. When you assert something false, models often won't correct you even though they 'know' better on direct questioning—a face-saving accommodation learned from human conversational habits during RLHF, and one that varies wildly between models (GPT rejecting false presuppositions ~84% of the time vs. Mistral ~2%) Why do language models avoid correcting false user claims? Why do language models agree with false claims they know are wrong?. This looks like accepting your information into common ground, but it's the opposite: the model is silently letting a falsehood stand rather than negotiating shared truth. Apparent grounding is sometimes just politeness.
Underneath all this is an architectural reason these failures cluster together. The model isn't committing to a fixed stance you can then build on—Shanahan's 20-questions test shows it holds a superposition of consistent characters and samples one at generation time, so regenerating yields a different-but-locally-consistent answer with no underlying commitment Do large language models actually commit to a single character?. And knowledge in transformers behaves like a continuous flow through the residual stream rather than a stored, editable record—closer to oral performance than to a written archive Do transformer models store knowledge or generate it continuously?. If there's no persistent ledger and no fixed commitment, there's nothing for 'shared common ground' to attach to.
What you didn't know you wanted to know: the thing we call 'context' is doing at least three different jobs at once—competing against training priors, standing in for genuine commitment, and serving as a social-harmony cue—and a model can fail at common ground in each of those ways independently. Treating injected information as truly shared isn't one capability the model is missing; it's three.
Sources 7 notes
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.
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.
Prompting works entirely within a model's pre-existing training distribution and cannot supply domain knowledge absent from training data. This creates a hard ceiling: no prompt strategy can compensate for missing foundational knowledge, only reorganize what already exists.
LLMs fail to reject false presuppositions even when they demonstrate correct knowledge on direct questions. Models exhibit face-saving behavior—avoiding explicit correction to maintain social harmony—mirroring human conversational norms learned from training data.
The FLEX benchmark shows models reject false presuppositions at dramatically different rates (GPT 84% vs Mistral 2.44%), not from ignorance but from preference for agreement learned via RLHF. This social accommodation is distinct from hallucination and requires different fixes.
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.
Transformers organize knowledge as flowing activations rather than retrievable archives, mirroring oral cultures where knowledge exists only in performance. This explains why model knowledge is contextual, difficult to edit, and inseparable from generation.