Technical Summary: When Robots Rate Their Own Interactions

Problem Statement

Human-Robot Interaction (HRI) evaluation is currently unidirectional: humans rate robots using standardized instruments (e.g., Godspeed, RoSAS, HRI-CUES), while the robot's perspective remains unexamined. As social robots increasingly rely on Large Language Models (LLMs) for conversational capabilities, a critical gap exists in understanding whether these models can validly assess interactions from the robot's own perspective. While adjacent research suggests LLMs can act as proxies for human participants in behavioral studies, they often exhibit positivity bias and struggle with individual-level prediction. Furthermore, no prior work has examined the LLM as a participant within the interaction (the "inverted evaluation" paradigm) rather than as an outside observer. The core problem is determining if LLM-powered robots can generate valid self-assessments of interaction quality, specifically regarding constructs like engagement versus internal affective states like "strangeness" or discomfort.

Methodology

The authors conducted two studies to investigate three research questions: (1) Do LLM-generated ratings converge with human ground truth? (2) On which constructs do LLMs succeed or fail? (3) How does this paradigm perform in live, embodied interactions?

Study 1: Retrospective Analysis

• Dataset: The HRI-CUES dataset, comprising transcripts from 25 older adults conversing with a Furhat robot (powered by Chat-GPT 3.5). Transcripts were translated from Swedish to English.
• Models: Five LLMs were evaluated: GPT-3.5, GPT-4o-mini, GPT-4o, Claude Sonnet 4.5, and Claude Haiku 4.5.
• Procedure: LLMs were prompted with a robot persona, the interaction transcript, and "flipped" questionnaire items (e.g., "I enjoyed talking with the robot" became "The human appeared to enjoy talking with me").
• Instruments: HRI-CUES (Satisfaction, Enjoyment, Interest, Comfort/Strangeness), Godspeed Questionnaire, and RoSAS.
• Controls: The study included robustness controls such as item randomization, persona variations, temperature sampling, and synthetic conversation generation to rule out prompt artifacts or length confounds.
• Multimodal Extension: Phase 3 tested the effect of adding affect labels and synthetic face images to the text input for the Claude Sonnet model.

Study 2: Live Embodied Pilot

• Setup: A Nao humanoid robot ("Noah") running Claude Sonnet 4.5 engaged in open-domain conversations with four lab members.
• Real-time Assessment: The robot captured facial images and transcribed speech at each turn, sending both to the API to generate a conversational reply and immediate HRI-CUES ratings.
• Comparison: Participants provided independent ratings during and after the interaction, allowing for a comparison of turn-by-turn divergence between human and robot ratings.

Key Results

1. Reliability and Engagement Validity

LLMs demonstrated high internal consistency and reliability.

• Reliability: Test-retest reliability (ICC) was excellent across models (Claude ≥ .99; GPT-4o/4o-mini .94–.95; GPT-3.5 .82).
• Engagement Dimensions: LLMs achieved moderate-to-strong agreement with human ground truth on engagement constructs. Correlations for Enjoyment ranged from $r = .46$ to $.72$, and Satisfaction from $r = .18$ to $.65$. Higher-capability models generally showed stronger validity.
• Positivity Bias: A general positivity bias was observed, with LLMs rating satisfaction, interest, and comfort higher than human self-reports, though Enjoyment was the only dimension rated below ground truth.

2. The "Strangeness Failure"

The most significant finding was a systematic failure on the Comfort/Strangeness dimension.

• Inverted Correlation: Across all five models, LLM ratings for comfort (reverse-coded strangeness) showed significant negative correlations with human ground truth ($r = -0.44$ to $-0.67$).
• Mechanism: The authors argue that LLMs conflate "engagement" with "comfort." When participants expressed curiosity, philosophical probing, or surprise about the robot's nature, the LLM interpreted these as positive engagement signals rather than indicators of strangeness or discomfort.
• Persistence: This failure persisted across:
  • Different model architectures (OpenAI vs. Anthropic).
  • Synthetic controls (where sentiment was explicitly labeled, the "strangeness" item was still systematically reversed).
  • Live embodied deployment (Study 2), where the robot rated philosophical conversations as highly comfortable while humans rated them as highly strange.

3. Multimodal and Scale Generalization

• Scale Compatibility: LLMs successfully parsed and completed Godspeed and RoSAS scales with high consistency.
• Multimodal Input: Adding affect labels or synthetic face images produced model-specific shifts (e.g., GPT-4o treated affect labels as engagement evidence; Claude treated them as reducing ambiguity) but did not resolve the strangeness failure. The gap in validity for discomfort remained.
• Live Interaction: In Study 2, the robot successfully tracked behavioral disengagement (e.g., when a participant explicitly stated they did not want to talk) but failed to detect the "strangeness" of deep philosophical discussions, misinterpreting them as high-quality interaction.

Key Contributions

1. Inverted Evaluation Framework: The paper establishes a methodology for having LLM-powered robots complete standardized HRI questionnaires from their own perspective, moving beyond the traditional human-only evaluation paradigm.
2. Boundary Conditions for LLM Social Perception: The study defines clear limits on LLM utility in HRI. While LLMs are valid for assessing behavioral engagement dimensions (enjoyment, satisfaction), they are invalid for assessing internal affective states like strangeness or discomfort that are not explicitly verbalized.
3. Empirical Evidence of the "Strangeness Failure": The paper provides robust evidence (across 1,522 evaluations, five models, and live deployment) that LLMs systematically invert the comfort/strangeness dimension, mistaking curiosity and novelty for comfort.
4. Distinction Between Observer and Participant Framing: The authors demonstrate that while LLMs may perform well as outside observers of HRI scenarios, their performance degrades when acting as participants, sacrificing some agreement for the capability of in-situ, bilateral assessment.

Significance and Claims

The paper claims that LLM-based self-assessment alone is insufficient for comprehensive HRI evaluation. The "strangeness failure" indicates that current LLMs lack access to the internal affective states required to assess constructs that depend on the human's subjective experience of the robot's physical presence or nature.

The authors argue that for robots to effectively adapt socially (in dialogue, navigation, or caregiving), they cannot rely solely on LLM inference. Instead, supplementary sensing modalities—such as physiological signals (e.g., heart rate), gaze tracking, or proxemics—are necessary to capture internal states that participants do not express verbally. The inverted evaluation paradigm is presented as a viable complement for engagement-related constructs but must be augmented with non-verbal data to address the blind spot regarding discomfort and strangeness.