Can prompting inject new knowledge into already-trained AI models?
This explores whether clever prompting can add knowledge a model never learned in training — or whether prompts can only rearrange and surface what's already inside.
This explores whether prompting can add genuinely new knowledge to a finished model, and the corpus answers with a fairly hard 'no' — prompts move existing knowledge around, they don't import any. The cleanest statement of the ceiling is that prompt optimization retrieves and reorganizes what the model already learned but cannot supply domain facts that were absent from training Can prompt optimization teach models knowledge they lack?. No prompt trick compensates for a missing foundation; it only finds better paths through the foundation that's there.
The more surprising part is *why* this ceiling holds even when you do hand the model new information in the prompt. You'd think pasting a fact into context counts as injecting knowledge — but models often ignore context when it conflicts with strong training-time associations, generating answers that follow their parametric priors instead of the text in front of them Why do language models ignore information in their context?. Overriding those priors isn't a wording problem; it can require intervening directly in the model's internal representations rather than through prompting at all. The same stubbornness shows up in personality: most open models retain their trained default disposition and resist prompts asking them to be someone else Can open language models adopt different personalities through prompting?.
So if prompts don't add knowledge, what *do* they do? They act as selectors and activators. A growing line of work finds that base models already contain latent capabilities — reasoning, in particular — and that post-training and prompting mostly *elicit* what's present rather than create anything new Do base models already contain hidden reasoning ability?. This reframes the whole question: the bottleneck is elicitation, not acquisition. Even the strongest theoretical result points the same way — a single transformer is Turing-complete and can be 'programmed' by prompt to compute almost anything Can a single transformer become universally programmable through prompts? — but that's reorganizing existing computational machinery, not teaching the model facts it lacks.
The knowledge-vs-skill distinction sharpens this. Factual recall depends on narrow, document-specific memorization — the fact has to have been seen — while reasoning leans on broad procedural knowledge spread across pretraining Does procedural knowledge drive reasoning more than factual retrieval?. That's why prompting can dramatically improve *how* a model reasons (procedures generalize and can be triggered) but can't conjure a *fact* the model never encountered. It also explains why structured prompting can mimic multi-agent systems or compose personas Can branching prompts replicate what multi-agent systems do?: those gains come from rearranging latent skills, not from new information entering the model.
The quietly useful takeaway: a prompt is better understood as a static program over frozen weights than as a teaching channel How do prompts reshape the role of context in AI conversation?. If you actually need new knowledge in the model — not just better access to old knowledge — prompting is the wrong tool, and the corpus points instead toward methods that touch weights or representations: inference-time expert composition Can models dynamically activate expert skills at inference time? or self-play that manufactures its own training signal Can language models learn skills without human supervision?. Prompting is for unlocking, not for filling.
Sources 10 notes
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.
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.
Research shows most open models fail to adopt prompted personalities, stubbornly retaining their trained ENFJ-like defaults. Only a few flexible models succeed. Combining role and personality conditioning improves results but doesn't fully overcome resistance.
Five independent mechanisms—RL steering, critique fine-tuning, decoding changes, SAE feature steering, and RLVR—all elicit reasoning already present in base model activations. Post-training selects rather than creates reasoning; the bottleneck is elicitation, not capability acquisition.
Research proves a single finite-size transformer exists that can compute any computable function given the right prompt, achieving complexity bounds nearly matching unbounded models. However, standard training rarely produces models that learn to implement arbitrary programs this way.
Analysis of 5 million pretraining documents shows reasoning relies on broad, transferable procedural knowledge from diverse sources, unlike factual recall which depends on narrow, document-specific memorization of target facts.
Research shows single LLMs using dynamic persona simulation achieve multi-agent cognitive synergy without multiple model instances. Solo Performance Prompting validates that structured prompting techniques map directly to multi-agent debate architectures, enabling equivalent outcomes through structural equivalence.
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.
Transformer2 demonstrates that tuning only singular values within weight matrices produces composable expert vectors that dynamically mix at inference without interference, outperforming LoRA with fewer parameters and enabling continual specialization.
Ctx2Skill's three-role self-play loop manufactures missing feedback through internal signals: the Challenger escalates difficulty as curriculum, the Judge gives binary verdicts as reward, and both sides evolve via natural-language skill edits. Success requires balancing adversarial pressure against a generalization safeguard to prevent collapse.