Technical Summary: Implicit vs. Explicit Prompting Strategies for LVLMs in Referential Communication

Problem Statement

Recent literature presents contradictory findings regarding the ability of Large Vision-Language Models (LVLMs) to exhibit human-like collaborative behavior in referential communication tasks. Specifically, two studies reached opposing conclusions on whether AI agents can develop "conceptual pacts"—shared, efficient referring expressions that emerge dynamically through interaction.

• Zeng et al. (2026) found that AI agents failed to shorten descriptions over time, persisting in verbose communication and showing decreased accuracy, suggesting a lack of common ground utilization.
• Jones et al. (2026) found that AI agents successfully shortened referring expressions and improved accuracy over time, mimicking human lexical entrainment.

This paper investigates whether these divergent results stem from differences in model architecture, task design, or prompting strategies. The authors hypothesize that the discrepancy is driven primarily by the prompting style (implicit vs. explicit) rather than the models or tasks themselves.

Methodology

The authors conducted a controlled experiment using an open-source pipeline based on Zeng et al. (2026), adapted to test specific variables.

Experimental Design

• Task: A multi-round, multi-turn collaborative object-matching task involving 12 baskets (non-lexicalized objects).
• Participants: AI-AI pairs where both the "Director" (speaker) and "Matcher" (listener) roles were played by the same model instance.
• Models: Two versions were tested: GPT-5.2 (used in Zeng et al., 2026) and GPT-5.5 (the latest model at the time of writing).
• Conditions: The study crossed two prompting strategies with two model versions:
  1. Implicit Prompting: Replicates Zeng et al. (2026). Instructions rely on Gricean maxims (be concise but informative) and collaborative grounding norms. The model is not explicitly told to shorten phrases or reuse words; any entrainment must emerge spontaneously.
  2. Explicit Prompting: Adapts Jones et al. (2026). The Director is explicitly instructed to reduce description length, use only 1-2 words in later rounds, and establish conventions (e.g., "SERIOUSLY–in later rounds just 1-2 words").

Technical Refinements

To ensure a fair comparison, the authors corrected the visual context injection pipeline used in Zeng et al. (2026):

• Chronological Alignment: Past chat histories and visual feedback images are paired chronologically at the start of the context.
• Dynamic Injection: The current round's image grid is injected dynamically using explicit round boundary markers, ensuring the LVLM sees the freshest visual frame immediately preceding its action.
• Unified Visual Input: The Matcher's active state (12-slot sequence) and candidate pool (18 baskets) are rendered in a single composite image with structured text reminders, grounding all spatial tracking in one temporally aligned visual representation.

Key Results

The study evaluated 40 complete AI-AI runs (200 round-level observations), measuring task accuracy, dialogue length, turn count, referring expression (RE) length, and lexical overlap.

1. Task Accuracy: All conditions achieved high task accuracy, indicating that both prompting styles allow models to complete the core task.
2. Communicative Efficiency (Implicit Prompt): Under implicit prompting, models remained verbose.
   • GPT-5.2 averaged 1,250.7 words/round; GPT-5.5 averaged 710.4 words/round.
   • Reductions across rounds were modest. Models reused visual descriptions but did not spontaneously treat accumulated history as a license to be brief.
3. Communicative Efficiency (Explicit Prompt): Under explicit prompting, models drastically compressed wording.
   • GPT-5.2 reduced length by 62.8%; GPT-5.5 reduced length by 75.6% compared to their implicit counterparts.
   • GPT-5.5 (Explicit): Showed the strongest "conceptual pact" behavior. RE length dropped from 58.8 words (Round 1) to 32.7 words (Round 5). Lexical overlap reached 1.00, and accuracy remained high (97.5%).
   • GPT-5.2 (Explicit): Also shortened descriptions, but accuracy dropped to 92.5% in Round 5, suggesting an accuracy-brevity trade-off.
4. Transcript Analysis: Explicit prompting changed the form of interaction. Models introduced compact, discriminative labels (e.g., "rect picnic," "tall cylinder") and pruned them into stable, telegraphic descriptions. In contrast, implicit prompting models retained full descriptive captions and confirmation routines despite reusing some lexical material.

Key Contributions

• Reconciliation of Conflicting Findings: The paper demonstrates that the divergence between Zeng et al. (2026) and Jones et al. (2026) is attributable to prompting style, not model version or task difficulty.
• Demonstration of Prompt Dependence: LVLMs can produce short, stable referring expressions when explicitly forced to do so, but they fail to infer the need for communicative efficiency from implicit pragmatic principles alone.
• Distinction Between Form and Process: The study highlights that while LVLMs can reproduce the surface pattern of human lexical entrainment (shortening expressions), this behavior does not emerge from the same underlying "common ground" mechanisms observed in human dialogue.

Significance and Claims

The paper argues that entrainment is not simply compression. In human dialogue, shortened expressions are evidence of a coordinated perspective and shared conceptual pact. Current LVLMs can be guided to mimic this outward form through heavy-handed prompting, but the effect is entirely prompt-dependent.

The authors caution against attributing this behavior to the same underlying cognitive processes as humans. The findings suggest that without explicit instruction to compress, LVLMs do not naturally develop the collaborative efficiency seen in human partners. This implies that current models may lack the intrinsic ability to dynamically update shared perspectives to reduce communicative effort, relying instead on explicit constraints to simulate such behavior.

Limitations

The authors note several constraints:

• The study was conducted only in English with non-lexicalized objects (baskets).
• It utilized only two specific LVLM versions (GPT-5.2/5.5).
• The experiment compared pairs of the same model, which may align more readily than pairs of different models or humans.
• The study does not evaluate scenarios where initial perspectives are misaligned, requiring models to pragmatically produce efficiency based on partner feedback rather than explicit instructions.