Can prompt engineering improve reasoning or only move requests into denser regions?
This explores whether prompting genuinely makes a model reason better, or whether it just relocates a request into a richer part of what the model already learned — the 'denser regions' framing.
This explores whether prompting genuinely improves reasoning or just moves a request into a richer part of the model's existing training distribution. The corpus suggests the honest answer is: both are true, and which one you get depends on whether the knowledge and reasoning paths already exist inside the model. The cleanest statement of the ceiling comes from work showing that prompt optimization retrieves and reorganizes what's already there but cannot inject knowledge the model never learned Can prompt optimization teach models knowledge they lack?. By that account, prompting is exactly 'moving requests into denser regions' — you can steer toward better-populated parts of the distribution, but you can't conjure new territory.
There's an even sharper version of the 'denser regions' worry: prompting may be steering toward what the *user* expects rather than what's true. One line of work frames prompt engineering as divergence minimization between the model's output and the user's own anticipated answer — meaning iterative refinement quietly bends generation toward your priors, making the output a co-production of model and user rather than independent reasoning How much does the user shape what a model generates?. So 'denser region' can mean 'closer to what you already believed,' which is not the same as 'more correct.'
But the corpus also shows prompting doing real reasoning work, not just relocation. Structured argument prompts that force a model to name its warrants and backing catch logical gaps that ordinary chain-of-thought skips over — that's the prompt changing *how* the model reasons, not just where it lands Can structured argument prompts make LLM reasoning more rigorous?. The catch is that this only helps for the right problems: saliency analysis shows step-by-step prompting actually *hurts* simple questions, where direct question-to-answer flow wins, so the gain is conditional on question type rather than universal Why do some questions perform better without step-by-step reasoning?. The same conditionality shows up across model tiers — rephrasing and background-knowledge prompts lift cheap models while step-by-step reasoning *reduces* accuracy in strong ones Do prompt techniques work the same across all LLM tiers?.
The deeper twist is that 'reasoning' itself may not be the bottleneck prompting can move. Reasoning-tuned models don't reliably beat standard models on constraint-bound numerical tasks — extended thinking produces more *text*, not more *computation*, so no amount of prompting more reasoning helps when the real limit is a numeric procedure Do reasoning models actually beat standard models on optimization?. And even when good solution paths exist inside the model, it abandons them prematurely — wandering and underthinking are structural failures that prompting alone doesn't fix; decoding-level interventions do Why do reasoning models abandon promising solution paths?.
So the resolution is less either/or than it first appears. Prompting can't add knowledge or invent reasoning capacity, and it can drift you toward your own priors — that's the 'denser regions' case. But it *can* activate latent reasoning structure that the model otherwise skips, provided you match the prompt to the problem and don't optimize it in isolation. Tellingly, prompts tuned without knowledge of the inference strategy (best-of-N, majority voting) systematically misfire, and jointly optimizing prompt and inference yields up to 50% gains — evidence that the lever is real but only when you pull it in coordination with everything else Does prompt optimization without inference strategy fail?. The thing you didn't know you wanted to know: prompt quality is itself a structured, measurable space — six dimensions grounded in communication theory, where improving one cascades to others — which means 'prompt engineering' is less a bag of tricks than navigation through that space Can we measure prompt quality independent of model outputs?.
Sources 9 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.
Foundation Priors research shows prompt engineering as divergence minimization between synthetic output and user priors. The refinement process systematically steers generation toward what users already expect, making outputs co-productions of model and user subjectivity.
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.
Saliency analysis reveals that CoT prompting fails when question information doesn't aggregate into the prompt structure before reasoning begins. For simple questions, direct question-to-answer flow outperforms step-by-step reasoning, showing the optimal prompt depends on question type, not just task category.
A 23-prompt benchmark across 12 LLMs shows rephrasing and background-knowledge prompts boost cheap models, while step-by-step reasoning reduces accuracy in high-performance models. Task structure, not generic best practices, determines which prompts help.
Reasoning variants with extended CoT show no consistent advantage over standard models on constraint-bound numerical tasks like optimal power flow. Extended thinking produces more text, not more iterative computation, suggesting the bottleneck is numeric procedure rather than reasoning steps.
Reasoning LLMs exhibit two reinforcing failures: wandering (invalid exploration) and underthinking (premature path-switching). Decoding-level interventions like thought-switching penalties improve accuracy without fine-tuning, suggesting viable solutions exist but are abandoned prematurely.
Prompts optimized without knowledge of the inference strategy (best-of-N, majority voting) systematically underperform. Joint optimization of both prompt and inference strategy yields up to 50% improvement across reasoning and generation tasks.
Research identifies six evaluable dimensions—Communication, Cognition, Instruction, Logic, Hallucination, and Responsibility—with 20 sub-criteria based on Grice, cognitive load theory, and instructional design. Improvements in one dimension cascade to others, revealing prompt quality as a structured space rather than a flat checklist.