How do I set how strictly an AI follows my instructions — and tell if it actually changed?

October 24, 2026 · 6 min read

SentX SI

The short answer is that you still cannot turn a dial, because no mainstream assistant exposes one — but a paper posted to arXiv on October 22, 2026 makes the missing dial precise enough that you can build a workable substitute today. It is called DIAL: Steerable Language Models with Distilled Instruction-Adherence Levels, from Thanh Q. Tran and colleagues, including Daniela Rus and Bryan Kian Hsiang Low. The authors start from the frustration behind your question: users say things like "keep it short," the model guesses how short, and prompting does not reliably provide fine-grained control over how strictly an instruction is followed. Their answer is to make adherence a number with a target, and to show that the number moves when you turn it. That second half — showing the change landed — is the part you can copy without waiting for the technology.

Source: DIAL: Steerable Language Models with Distilled Instruction-Adherence Levels

Why plain prompting falls short is easy to see once you name it. Text carries intent, not a scale. "Be concise" has no units, so the model resolves it differently on every run, and the same wording lands differently across models and contexts. You can ratchet up the pressure — "must be under 80 words," "at most three bullets" — and checkable constraints like those genuinely help, but nothing guarantees the response moves in proportion to how hard you pushed. Adherence stays something you sample by eye, one reply at a time.

Source: DIAL: Steerable Language Models with Distilled Instruction-Adherence Levels

DIAL's mechanism, as the abstract describes it, has three parts. First, it takes a frozen LLM — weights left untouched — and uses classifier-free guidance, a technique that contrasts the model's behavior with and without the instruction, to produce training signals for distilling adherence at several levels. In plain terms, the gap between instructed and uninstructed behavior becomes the teaching material for a smaller controller. Second, that controller is a lightweight adapter that reads the model's internal state and applies a small correction at each generated token. The correction is closed-form — computed directly from a formula rather than searched for case by case — and its magnitude is set by the target adherence level, adjusted for the geometry of the internal state so responses stay coherent. Third, the controller is conditioned on the instruction text itself, so one shared adapter handles new instructions without per-instruction fitting, and training needs only instruction and request text. Together that is the dial: a numerical level that scales how strongly the model honors the instruction.

Source: DIAL: Steerable Language Models with Distilled Instruction-Adherence Levels

What the abstract establishes, with its hedges intact: evaluation across multiple LLM families (unnamed in the abstract) shows a competitive balance between concept expression and response fluency, with concept scores rising monotonically as the adherence level rises — the knob pulls in one direction, which is the property a control needs before it means anything. The authors also introduce SteerIF, an instruction-only dataset of 98,625 concept, persona, verifiable, and descriptive instructions, and report that DIAL trained on SteerIF provides fine-grained control over personas and, when combined with prompting, improves verifiable instruction following. Three boundaries matter. "Competitive balance" is not a state-of-the-art claim. The verifiable-following gain is reported in the SteerIF-trained setting together with prompting, not as a standalone replacement for it. And this is an abstract-level read: specific metrics, family names, and effect sizes are not established from the listing.

Source: DIAL: Steerable Language Models with Distilled Instruction-Adherence Levels

The distance between that framework and your chat box is real. No consumer product currently exposes a numerical adherence knob, so you cannot set DIAL's level on any assistant you use this week. What transfers is the design principle: adherence as a measured quantity with a target, and a change that counts only if the measurement moves in the expected direction.

Source: DIAL: Steerable Language Models with Distilled Instruction-Adherence Levels

Here is a working rule, offered as an editorial suggestion rather than a finding from the paper. Sort your tasks by how you could check compliance. If the instruction is soft — tone, style, rough length — and the cost of drifting is low, coarse prompting suffices: state it once, look at the result, move on. If the instruction is checkable — an exact phrase that must appear, a word or item cap, a schema, a count, a detail that must survive across messages — you are in the regime the paper says prompting alone handles unreliably. There, two habits do the work a dial would do: put the constraint in the current message every time it matters, and score the output against it. The litmus test is simple. If you can write a yes/no check for compliance, you need the measurement loop.

Source: DIAL: Steerable Language Models with Distilled Instruction-Adherence Levels

For the second half of your question — did the change actually land? — keep a fixed probe. A handful of representative inputs, the same instruction each time, scored against your yes/no checks, run before and after whatever you touched: a reworded prompt, a new setting, a model update. Compare pass rates. If the rate does not move, the change did not reach behavior, however plausible it sounded. This is the reader-scale version of the monotonicity check the authors used to validate their own knob, adapted here as a method, not a result they report for end users.

Source: DIAL: Steerable Language Models with Distilled Instruction-Adherence Levels

A concrete illustration, with invented numbers. Your standing instruction is "answer in at most 80 words," and you keep five test questions. Baseline run: four of five answers fit the cap — an 80 percent pass rate. You tighten the prompt to "Answer in at most 80 words. Count your words before sending." Same five questions again: five of five fit — 100 percent. The jump from 80 to 100 is the evidence that the change worked; "it feels tighter now" is not. The probe catches regressions just as cleanly: a model update that silently abandons the cap shows up as a falling pass rate even when individual answers still look fine. Every figure in this paragraph is illustrative arithmetic, not an observed result from any product.

The same discipline already appears in how SentX SI documents its own assistant. Victoria's public model card — a first-party disclosure, not an independent certification — notes that recall is selective and exact retention is not guaranteed, and its known-limits section directs users to repeat instructions and facts that must remain exact. Read next to DIAL, that direction is the manual version of the dial: the constraint travels with the message, and you verify the output against it. One boundary to hold when building probes: the privacy policy states that submissions are not confidential and may enter training or shared memory, so compose your test set from material you are comfortable submitting.

Source: Victoria AI Model Card | Capabilities, Evidence and Limits | SentX SI; Privacy Policy | SentX SI

Until a real adherence dial ships, the verifiable-check habit is the control you actually have. The paper's core lesson is that a control only earns the name if turning it moves a score in the direction you asked for — and that lesson costs nothing to adopt today.