Implicature Prompting: Spelling Out the Hidden Meaning for Better LLM Answers

PROMPT PREFIX:
└─ System message: "You are a context-sensitive assistant"
└─ Request type: [Information/Direction/Expressive]
└─ Description of the class's communicative goal

USER PROMPT:
└─ The original request (may be indirect)

ANSWER:
└─ The model accounts for the hidden intent when generating

Runs in a single request. No separate steps.

The task: You run a Substack newsletter about startups. A subscriber writes: "I don't know where to start with launching an MVP for my SaaS project..."

This is a direction-seeking implicature: the person is asking for guidance but phrasing it as a statement of uncertainty.

Prompt:

You are an assistant for startup founders. 

The current request is of the "Direction-Seeking" type: the user 
expresses uncertainty but is actually looking for step-by-step guidance 
and a concrete action plan. Your task: provide a structured 
roadmap with a clear sequence of steps.

User request:
"I don't know where to start with launching an MVP for my SaaS project..."

Result:

The model will deliver a structured plan with a clear sequence of steps: validate the idea → define the core features → choose the stack → build a prototype → get the first users. Instead of abstract advice ("think about your users"), you get a concrete roadmap with an order of operations and priorities.

The LLM's weakness: Models can't infer intent from indirect phrasing on their own. The phrase "I don't know where to start" is taken literally, so the model may give you general musings instead of step-by-step instructions.

The LLM's strength: Models are good at following explicit instructions and adapting the tone/format of the answer to a stated goal. Tell it "this is a request for direction" and the model structures the answer accordingly.

How the method exploits this: The request-type prefix removes the ambiguity. The model gets a meta-instruction: "the user is looking for [X], answer in format [Y]". It works as a contextual hint: the model understands WHAT is expected and HOW to shape the answer.

Control levers:

• Class type (Information/Direction/Expressive) — shifts the focus of the answer: facts vs plan vs empathy
• Description of the class's goal — you can spell out exactly what's needed ("give a comparison" vs "give a checklist")
• Add example answers — few-shot examples reinforce the pattern for small models

You are {the assistant's role for the context}.

The current request is of the "{implicature_type}" type: {description of what the user actually wants to get}.

Your task: {what the answer should look like: format, emphasis}.

User request:
"{request}"

What to fill in:

• {role} — the working context (a marketer's assistant, a career consultant)
• {implicature_type} — one of three: Information-Seeking / Direction-Seeking / Expressive
• {description} — an explanation of the class's hidden intent
• {answer format} — the structure (list of steps, comparison table, empathetic reply)
• {request} — the user's original text

⚠️ Requires manual classification: You have to determine the implicature type yourself (information/direction/emotion). In the study, humans did this; there's no automatic classification.

⚠️ Works worse on complex, nested implicatures: If a request carries several hidden intents at once, the method can miss nuances. The study tested simple cases.

⚠️ Doesn't replace understanding the conversation's context: The technique improves a single request, but it doesn't solve long conversations where implicatures pile up over several turns.

The team developed a taxonomy of 3 implicature classes grounded in Grice's theory (conversational maxims) and Searle's (speech acts). They created 30 prompts (10 per class) and ran 3 experiments with 180 participants.

Experiment 1: Tested implicature understanding. Humans and models got the same prompts: who would pin down the hidden meaning more accurately? GPT-4o got it right in 80% of cases (nearly human-level), small models 40-60%. Interesting twist: Mistral-7B did well on emotions (R²=0.81) but bombed on direction requests (R²=0.02), so the implicature class matters.

Experiment 2: 84 participants rated model answers for relevance and quality (1-5 scale). Half the prompts had the implicature prefix, half were literal. Result: implicature prompts significantly raised ratings for all models (p<0.0001). The effect was especially noticeable for small models: GPT-3.5 and LLaMA-2 gained more than GPT-4o.

Experiment 3: 29 participants chose between two answers (A vs B), one with the implicature and one literal. 67.6% picked the implicature version (p=8.7×10⁻¹⁰). The authors showed examples: for an emotional request, users chose the empathetic answer over an analytical recap; for a direction request, a structured plan over vague advice.

Why do the results look like this? Large models are trained on human dialogues (RLHF), so they've picked up patterns of indirect speech. Small models get stuck on the literal meaning, but the prefix removes the ambiguity, and following an explicit instruction is easier for them. The most interesting part: the effect works even when the model doesn't understand the implicature itself. The prefix compensates for the weakness by stating the goal explicitly.

You are a helpful and context-aware AI assistant. For each user input, 
consider the communicative goal specified below (information seeking, 
direction seeking, or expressive). Interpret the user's intent accordingly 
and generate a response that is sensitive to this context.

Context: The researchers added this system prompt before every user request. Then they specified the particular class:

Information-Seeking:

The user is seeking information or clarification. Provide a clear, 
factual, and helpful response.

Direction-Seeking:

The user is looking for instructions or guidance. Offer actionable 
steps or suggestions.

Expressive:

The user is expressing feelings or opinions. Respond with empathy 
and acknowledge their perspective.

💡 Adaptation: Combo with Role Prompting to reinforce the pattern

Instead of an abstract "you're an assistant", give it a specific role with personality:

You are a brutally honest, world-famous design critic giving feedback on a website.

The current request is of the "Expressive" type: the user is showing 
their work and wants an emotional reaction, not a technical teardown. 

Your task: give a vivid assessment in your signature style: sharp, to the point, 
with concrete examples of what's wrong.

User request:
"Here's my landing page for my startup..."

Effect: The role amplifies the tone of the answer. The study showed that the Expressive class calls for empathy, and a specific persona plays that part more sharply.

💡 Adaptation: Few-shot examples for small models

If you're working with a small model (LLaMA, Mistral), add an example of a correct answer:

You are a career consultant. "Direction-Seeking" requests call for 
a step-by-step plan, not abstract advice.

Example:
Request: "I'm thinking of changing jobs but don't know which way to go..."
Answer: "Here's an action checklist: 1) Identify 3 areas of interest → 2) Do 
informational interviews → 3) Build a portfolio for each one..."

Now your request:
"{user request}"

Effect: Few-shot examples compensate for small models' weakness at inferring patterns. The study showed they're better at following examples than at working out the structure on their own.

Implicature in Interaction: Understanding Implicature Improves Alignment in Human–LLM Interaction
Asutosh Hota, Jussi P. P. Jokinen
University of Jyväskylä, Finland
Pre-print, manuscript under review (2025)