💬 NLP

Social Meaning in Large Language Models: Structure, Magnitude, and Pragmatic Prompting

This paper demonstrates that while large language models qualitatively replicate the structural patterns of human social inferences, they often distort the magnitude of these inferences, a limitation that can be partially mitigated by prompting strategies grounded in pragmatic theory, specifically those addressing speaker knowledge and motives.

Original authors: Roland Mühlenbernd

Published 2026-04-06

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Roland Mühlenbernd

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are teaching a robot to understand human conversation. You want it to not just hear the words, but to "get" the hidden meaning behind them—like knowing that if someone says, "The meeting is around 3 PM," they might be a bit disorganized, but if they say, "The meeting is at 3:02 PM," they are likely a perfectionist.

This paper is about testing three super-smart AI robots (Large Language Models, or LLMs) to see if they can do this kind of "social reading" as well as real humans, and if we can teach them to do it better using specific instructions.

Here is the breakdown of the study using simple analogies:

1. The Problem: The Robot Gets the "Direction" Right, but the "Volume" Wrong

The researchers set up a test based on a human experiment. They asked people to judge how "competent" or "helpful" a speaker sounded when they gave a precise number (e.g., "$500") versus an approximate one (e.g., "about $500") in different situations (like talking to an insurance agent vs. a friend).

The Finding:
The AI robots were surprisingly good at getting the direction right.

Human: "If I say 'about $500' to an insurance agent, I sound less competent."
AI: "Yes, I agree. The score for competence goes down."

However, the robots were terrible at getting the magnitude (the strength) right.

Human: "The score drops a little bit."
AI: "The score drops by a huge amount!"

The Analogy:
Imagine a volume knob on a stereo. The robots know exactly which way to turn the knob (up or down) to match human feelings. But they often turn the volume knob way too far, making the sound either deafeningly loud or barely a whisper. They understand the trend, but they exaggerate the intensity.

2. The New Tools: Measuring "Exaggeration"

Before this paper, scientists mostly checked if the robot said "Yes" or "No." This paper introduced two new "rulers" to measure exactly how much the robot exaggerated:

The "Exaggeration Meter" (Effect Size Ratio): This measures if the robot's reaction is 1x, 2x, or 10x stronger than a human's.
The "Calibration Score" (Calibration Deviation): This gives a single number showing how far off the robot is from the human average.

3. The Experiment: Can We "Prompt" the Robots to Calibrate?

The researchers tried four different ways of talking to the robots (called "prompting") to see if they could fix the volume knob. They used two ideas from human language theory:

Thinking about Alternatives: "What else could the speaker have said?" (e.g., "Why didn't they say the exact number?")
Thinking about Motives: "What is the speaker thinking or feeling?" (e.g., "Are they nervous? Do they not know the exact number?")

The Results of the "Training":

The "Alternative" Trick (Thinking about other words):
- Result: Mixed. For some robots, it helped a little. For others, it made them worse. It was like telling a nervous actor, "Think about all the other lines you could have said!" and they ended up overacting even more.
The "Motive" Trick (Thinking about the speaker's mind):
- Result: Very helpful! When the robots were told to imagine the speaker's knowledge and feelings, they stopped exaggerating so much. They became more like humans.
- Analogy: It's like telling a judge, "Before you decide, imagine you are the person on trial. What were they thinking?" This makes the judgment more balanced.
The "Combo" Trick (Thinking about both):
- Result: This was the best overall strategy. When the robots were asked to think about both the alternative words AND the speaker's motives, they performed the most consistently well across all three different robot models.

4. The Big Takeaway

The paper concludes that these AI robots are like brilliant but dramatic actors.

They know the script perfectly (they understand the social rules).
But they lack the subtle "human touch" to know how strongly to feel those rules. They tend to overact.

While we can give them instructions to tone it down (especially by asking them to think about the speaker's mind), we haven't quite solved the problem completely. Some robots are naturally better at this than others, and even the best instructions only get them "mostly" human-like.

In a nutshell:
AI is getting really good at understanding what we mean socially, but it still struggles to match how strongly we feel about it. We can help them improve by asking them to "put themselves in the speaker's shoes," but they still need more practice to get the volume just right.

1. Problem Statement

While Large Language Models (LLMs) have demonstrated an ability to reproduce human-like pragmatic and social reasoning qualitatively (e.g., identifying correct conversational implicatures), there is a critical gap in understanding their quantitative performance. Existing evaluations often focus on directional agreement (does the model get the "right" answer?) or categorical rankings, ignoring the graded strength of social inferences.

The paper addresses two core questions:

Quantitative Fidelity: Do LLMs approximate human social meaning not just in direction but in magnitude? (i.e., Do they exaggerate or attenuate the strength of social judgments compared to humans?)
Intervention: Can prompting strategies grounded in pragmatic theory improve this quantitative approximation?

The authors hypothesize a "structure–magnitude dissociation," where models correctly identify the direction of an effect (e.g., precise language implies competence) but systematically distort the magnitude of that effect.

2. Methodology

Experimental Design

Task: The study uses a case study on numerical (im)precision. It evaluates how the choice between precise (e.g., "$500") and approximate (e.g., "about $500") numerical expressions influences social attributions (competence, knowledgeability, helpfulness, etc.).
Contexts: The study crosses utterance form with contextual precision requirements:
- High-Precision (HP): Formal contexts (e.g., insurance claims) where precision is expected.
- Low-Precision (LP): Casual contexts (e.g., chatting with a friend) where approximation is acceptable.
Human Benchmark: The evaluation is grounded in Experiment 1 of Solt et al. (2025), which established a quantitative baseline of human social inferences across 10 significant effects (5 main effects of form, 5 form × context interactions).

Models and Protocols

Models: Three frontier LLMs were evaluated via API: GPT-4o-mini, Claude-sonnet-4, and Gemini-2.5-pro.
Prompting Conditions: Four distinct prompting regimes were tested to probe pragmatic reasoning mechanisms:
1. Minimal (MIN): Baseline instructions mirroring the human experiment.
2. Alternative-Aware (ALT): Includes a Chain-of-Thought (CoT) exemplar forcing the model to reason over linguistic alternatives (Gricean principle).
3. Knowledge-and-Motives-Aware (KMA): Instructs the model to consider the speaker's potential knowledge states and communicative motives before rating.
4. Combined (COM): Integrates both ALT and KMA components.

Evaluation Metrics

The paper introduces two novel metrics to distinguish structural fidelity from magnitude calibration:

Effect Size Ratio (ESR): $ESR = |\Delta_M| / |\Delta_H|$ . It measures the ratio of the model's effect size to the human effect size. $ESR=1$ is perfect; $ESR>1$ indicates exaggeration; $ESR<1$ indicates attenuation.
Calibration Deviation Score (CDS): The mean absolute deviation of ESR from 1 across all significant effects. Lower scores indicate better magnitude calibration.

Supporting Metrics: Spearman rank correlation ( $\rho$ ), Concordance Correlation Coefficient (CCC), and Root Mean Square Error (RMSE).

3. Key Contributions

New Metrics: The introduction of ESR and CDS provides a principled way to quantify the "structure–magnitude dissociation" in LLMs, moving beyond simple accuracy or ranking metrics.
Theoretical Framework: The study operationalizes two core pragmatic assumptions (Reasoning over Alternatives and Speaker Knowledge/Motives) into specific prompting interventions to test if they can modulate model behavior.
Empirical Evidence of Dissociation: It provides robust evidence that LLMs reliably learn the direction of social inferences but fail to capture the graded strength of those inferences, often inflating effect magnitudes by factors of 2–10.

4. Results

Structural vs. Magnitude Alignment

Structural Fidelity: All models achieved perfect directional agreement (DAS = 1.0, ISS = 1.0) and high rank correlations ( $\rho \approx 0.83–0.95$ ). They correctly identified which inferences to make.
Magnitude Calibration: Significant failures were observed.
- Gemini: Showed severe magnitude inflation (RMSE 1.07–1.42), often exaggerating effects by 2–3x the human baseline.
- Claude: Showed moderate but uneven miscalibration, highly sensitive to prompting.
- GPT: Performed best overall, with near-calibrated magnitudes (RMSE 0.55–0.81).

Impact of Prompting Strategies

Knowledge-and-Motives-Aware (KMA): This was the most consistently effective intervention for reducing magnitude deviation in models prone to exaggeration (Claude and Gemini). It reduced the Calibration Deviation Score (CDS) significantly, suggesting that explicitly modeling speaker uncertainty moderates social inferences.
Alternative-Aware (ALT): Produced mixed or adverse effects. While it helped GPT slightly, it amplified magnitude inflation in Gemini and worsened overall calibration in Claude. This suggests that forcing models to consider alternatives without grounding them in speaker epistemic states can exacerbate contrast effects.
Combined (COM): This was the only condition that improved all calibration-sensitive metrics (CCC, RMSE, CDS) across all three models simultaneously. However, it did not fully resolve the calibration gap, particularly for Gemini, indicating that the two pragmatic components are not fully additive and may interfere with each other in specific architectures.

5. Significance and Conclusion

Theoretical Implication: The results support the Rational Speech Act (RSA) framework, suggesting that human-like pragmatic inference requires the simultaneous operation of reasoning over alternatives and inferring speaker states. Isolating these components (via prompting) often leads to suboptimal or distorted reasoning.
Evaluation Shift: The paper argues that evaluating LLMs for social reasoning must move beyond "did it get the right answer?" to "how strongly does it believe the answer?" The proposed ESR and CDS metrics are essential for diagnosing this gap.
Practical Limitations: While pragmatic prompting offers a partial solution, it is architecture-dependent and incomplete. The study highlights that current LLMs capture the inferential structure of social meaning but variably distort the inferential strength, limiting their reliability as perfect simulators of human social judgment in consequential decision-making contexts.

In summary, the paper demonstrates that while LLMs have mastered the logic of social inference, they struggle with the calibration of social strength, and that pragmatic theory provides a useful, though currently incomplete, toolkit for mitigating these errors.