Implicit vs. Explicit Prompting Strategies for LVLMs in Referential Communication

2026-06-18 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Social Sciences & Behavioral Studies · Depth: Expert, long

Summary

A recent study investigated conflicting findings regarding large vision-language models' (LVLMs) ability to coordinate efficient referring expressions, specifically comparing implicit and explicit prompting styles. Replicating prior work, the research found that LVLMs, including GPT-5.2 and GPT-5.5, can achieve human-like communicative efficiency when explicitly instructed. For instance, GPT-5.5's referring expression length dropped from 58.8 words to 32.7 words over five rounds with explicit prompts, maintaining 97.5% accuracy and 1.00 lexical overlap. Conversely, models given implicit, pragmatically informed prompts remained verbose (GPT-5.2 averaged 1250.7 words/round, GPT-5.5 710.4 words/round), failing to spontaneously shorten expressions. This indicates that prompting style, rather than model version or task differences, drives the observed divergence in LVLM communicative behavior.

Key takeaway

For prompt engineers designing LVLM interactions for efficient communication, you must prioritize explicit, direct instructions over general pragmatic principles. Your models, such as GPT-5.5, will only reliably shorten referring expressions and achieve human-like entrainment when forcefully commanded to do so. Be mindful that extreme brevity, as seen with GPT-5.2, can sometimes lead to a slight drop in accuracy, so balance conciseness with task performance.

Key insights

LVLMs require explicit, forceful prompting to achieve human-like communicative efficiency and lexical entrainment, not implicit pragmatic cues.

Principles

• Explicit prompts are critical for LVLM brevity.
• Implicit pragmatic instructions do not induce entrainment.
• LVLM "conceptual pacts" differ from human common ground.

Method

This study used a multi-round, multi-turn collaborative object-matching task, comparing explicit and implicit prompt designs for GPT-5.2 and GPT-5.5 in AI-AI pairs, based on Zeng et al. (2026)'s pipeline.

In practice

• Use direct commands for LVLM brevity.
• Monitor accuracy-brevity tradeoffs carefully.
• Do not rely on implicit pragmatic cues.

Topics

• Large Vision-Language Models
• Prompt Engineering
• Referential Communication
• Communicative Efficiency
• Lexical Entrainment
• Human-AI Interaction

Best for: AI Engineer, Machine Learning Engineer, NLP Engineer, AI Scientist, Research Scientist, Prompt Engineer

Related on AIssential

• Implicit vs. Explicit Prompting Strategies for LVLMs in Referential Communication — LVLMs require explicit prompts for efficient referential communication, unlike humans.
• LVLMs and Humans Ground Differently in Referential Communication — LVLMs fail to build common ground and adapt communication in multi-turn referential tasks, unlike humans.
• As Easy as Rocket Science: Assessing the Ability of Large Language Models to Interpret Negation in Figurative Language — LLMs struggle with combined negation and figurative language, with prompt style significantly impacting performance.
• You Might Want to Re-Learn Prompting... — ChatGPT 3.5 models perform better with outcome-first, shorter prompts that define the desired result.
• As Easy as Rocket Science: Assessing the Ability of Large Language Models to Interpret Negation in Figurative Language — LLMs struggle with combined negation and figurative language, with prompt style significantly impacting interpretation accuracy.
• Stop Being Nice to Your Chatbot: The Surprising Science of Prompt Politeness — Polite prompts reduce LLM accuracy and increase computational cost, while direct or emotionally-charged prompts improve performance.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.