Can prompting alone inject new domain knowledge into a model?
This explores whether clever prompting can give a model knowledge it never learned during training — or whether prompts can only rearrange what's already there.
This reads the question as asking whether prompting *injects* new domain knowledge versus merely *activating* knowledge the model already holds — and the corpus is unusually direct here: it can't. Prompt optimization works entirely inside a model's pre-existing training distribution, so it reorganizes and surfaces what's latent but cannot supply facts that were never learned Can prompt optimization teach models knowledge they lack?. That creates a hard ceiling no clever phrasing can climb past. The failure isn't even just about missing facts — even when the missing knowledge is sitting right there in the context window, models routinely ignore it: when parametric associations from training are strong, in-context information gets overridden, and textual prompting alone can't break the prior. Fixing that requires intervening in the model's internal representations, not rewording the prompt Why do language models ignore information in their context?. You can even watch this resistance directly in the personality work, where most open models stubbornly retain their trained defaults no matter what persona you prompt them to adopt Can open language models adopt different personalities through prompting?.
So what *does* prompting do well? It sharpens reasoning over knowledge the model already has. Forcing a model through explicit argument-checking steps — identify the warrant, check the backing — catches reasoning failures that ordinary chain-of-thought glosses over Can structured argument prompts make LLM reasoning more rigorous?. There's an even sharper version of this: a single internal 'reasoning' feature can be steered to match chain-of-thought performance without any reasoning prompt at all, which suggests the capability was always there and prompting is just one switch among several for turning it on Can we trigger reasoning without explicit chain-of-thought prompts?. That's the tell — prompting is an activation lever, not a deposit slot.
If you actually need *new* domain knowledge, the corpus points to two different routes. One is to bring the knowledge in at inference time without touching weights: retrieval-augmented generation, where the model reasons over external documents — though the research stresses retrieval has to adapt dynamically and couple tightly to reasoning, not bolt on as a fixed lookup How should systems retrieve and reason with external knowledge?. The other is to actually train it in, and here the surprising finding is that *how* you structure the knowledge beats *how much* you feed in. StructTuning hits 50% of full-corpus performance using 0.3% of the data by organizing chunks into a domain taxonomy so the model learns where a fact sits in a conceptual map, the way a student learns from a textbook rather than from raw pages Can organizing knowledge structures beat raw training data volume?. A knowledge-graph curriculum pushes this further, turning graph paths into reasoning tasks that build genuine domain expertise Can knowledge graphs teach models deep domain expertise?.
The non-obvious payoff is that the *training method* changes what gets absorbed. Reinforcement learning from augmented generation embeds knowledge more durably than ordinary supervised fine-tuning, because it rewards coherent explanation rather than just token-level correctness — the model internalizes a knowledge structure instead of memorizing surface strings Can reinforcement learning embed domain knowledge more effectively than supervised fine-tuning?. But none of these are free: every adaptation method has a domain-specific sweet spot, and visible gains often hide quiet costs in reasoning faithfulness, transfer to other tasks, and format flexibility How do domain training techniques actually reshape model behavior?. And there's a deeper argument that systems learning *only* from data — never given explicit structured knowledge — end up uninterpretable and brittle outside their training distribution, which is exactly the failure prompting can't rescue you from Does refusing explicit knowledge harm AI system performance?.
The short version: prompting is a flashlight, not a library. It illuminates what's already on the shelves brilliantly, but if the book isn't in the building, you either carry it in at read-time (RAG) or you build a new wing (structured fine-tuning, graph curricula, RL) — and the smart move is investing in the shelving system, not the volume of books.
Sources 11 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.
Applying Toulmin's argument model as explicit prompting steps (CQoT) improves LLM reasoning by forcing models to identify warrants and backing rather than skipping implicit premises. The method catches failures that standard chain-of-thought prompting allows.
SAE-identified reasoning features can be directly steered to match or exceed chain-of-thought performance across six model families. This reasoning mode activates early in generation and overrides surface-level instructions, suggesting latent reasoning is a fundamental capability independent of explicit prompting.
Research shows retrieval should adapt dynamically rather than follow fixed patterns, reasoning and retrieval must integrate closely, and embedding-based retrieval has fundamental limits requiring architectural alternatives.
StructTuning achieves 50% of full-corpus performance using only 0.3% of training data by organizing chunks into auto-generated domain taxonomies. The model learns knowledge position within conceptual structures rather than raw text patterns, matching how students learn from textbooks.
Fine-tuning a 32B model on 24,000 reasoning tasks derived from medical knowledge graph paths produces state-of-the-art performance across 15 medical domains, demonstrating that structured knowledge composition matters more than scale.
RLAG rewards both answer accuracy and explanation rationality by cycling between augmented and unaugmented generation, progressively internalizing coherent knowledge structures. This outperforms SFT because it prioritizes reasoning quality over token-level correctness.
Research shows every adaptation method—from parameter-efficient tuning to knowledge graph curricula—has optimal conditions tied to specific domains. The key finding: visible benefits like performance gains often come with hidden degradation in reasoning faithfulness, capability transfer, and format flexibility.
AI systems that learn exclusively from data produce uninterpretable representations, inherit statistical biases uncorrected by normative rules, and fail to generalize beyond training distributions. Structured knowledge injection at minimal corpus cost substantially improves performance.