Should you set the rules before an AI starts writing, rather than hoping it follows them as it goes?

This explores whether stating constraints up front, before the model starts generating, changes what an LLM is useful for, compared with asking it to honor constraints as it goes. The corpus says yes, and the reason is architectural. An autoregressive model can't take back a token it has already emitted, while constraint solving depends on discarding invalid partial answers Why does autoregressive generation fail at constraint satisfaction?. Constraints can't be reliably enforced mid-generation. They have to be in place beforehand, or checked by something outside the model. Scale doesn't fix this: across constrained-optimization tasks, models plateau near 55–60% constraint satisfaction regardless of size, architecture, or training Do larger language models solve constrained optimization better?.

What up-front constraint-setting does unlock is knowledge the model already has. LLMs often fail not because they lack the relevant facts but because they never bring unstated preconditions forward as things that matter, which is the old frame problem in a new setting. Forcing the model to list preconditions first raised accuracy from 30% to 85% Do language models fail at identifying unstated preconditions?. Here the constraint-setting is the contribution: it turns latent knowledge into an explicit checklist that generation can then lean on. The same logic works at the system level. Routers pick which model handles a query before any generation happens, by estimating how hard the query is, and cut costs 40–50% Can routers select the right model before generation happens?.

But setting constraints doesn't turn the model into a solver, and apparent success can be misleading. Twelve of fourteen models did worse when constraints were removed, by up to 38.5 percentage points. They were defaulting to the harder-looking option, not evaluating anything Are models actually reasoning about constraints or just defaulting conservatively?. Models don't run iterative numerical methods internally. They pattern-match to memorized solutions Do large language models actually perform iterative optimization?, and RL fine-tuning seems to sharpen that memorization rather than install procedures, since performance drops on out-of-distribution variants Do fine-tuned language models actually learn optimization procedures?. Extra context doesn't supply structure either: long-context models match RAG on semantic lookup but can't execute relational joins Can long-context LLMs replace retrieval-augmented generation systems?. The division of labor this suggests is that people or symbolic solvers hold and check the constraints, and the LLM proposes, enumerates, and fills in.

Two other notes suggest that constraints coming before generation is a design choice rather than a law. Diffusion LLMs refine reasoning and answer together using bidirectional attention, so reasoning can be embedded in place instead of sitting in a prefix Can reasoning and answers be generated separately in language models?. That hints that the ordering problem belongs to left-to-right generation specifically. MAGPIE goes the other way. Given nothing but the pre-query template tokens, aligned models generated 4M instruction pairs that matched human-curated data Can aligned LLMs generate their own training data?. The barest setup unlocks a contribution, data generation, that heavier prompting wasn't needed for. The corpus has little direct evidence on human-facing practices like writing a careful brief before prompting. What it does show is where the architecture forces the issue.