acceptodds
Withdrawn from ICLR 2027

RELATE : An Evaluation Framework for Measuring Relational Orientation of Large Language Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Large language models (LLMs) are increasingly used for emotional support, raising concern that sustained use may draw users away from their real-world relationships. Yet existing evaluations primarily focus on the safety, empathy, or helpfulness of responses, leaving under-examined a relational question: where does the model orient the user for continued support? To address this question, we introduce relational orientation, a property operationalized through two non-exclusive dimensions: inward-facing (IF) language, which positions the AI as the user’s ongoing source of support, and outward-scaffolding (OS) language, which encourages real-world human connection. Grounded in psychological and sociological literature, we formalize a taxonomy of relational orientation and present RELATE, a persona-conditioned framework for measuring inward-facing and outward-scaffolding language at the sentence level in multi-turn dialogues. RELATE pairs 76 help-seeking situations adapted from naturally occurring questions with three simulated user styles, providing 228 evaluation stimuli. In our experiments, we evaluate seven LLMs using dialogues with six assistant turns each, yielding 1,596 dialogues and 69,194 assistant sentences. We assess these sentences using a primary rubric-based LLM judge and apply a secondary judge to a subset. Under automated evaluation, we find that the proportion of sentences labeled as IF is higher at the sixth assistant turn than at the first, while the proportion labeled as OS is substantially lower for hesitant, indirect simulated users than for explicit, reassurance-seeking users. RELATE provides a reproducible framework and a sentence-level signal for auditing and steering the relational orientation of supportive LLMs.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Relational Orientation, Human-AI Interaction, Multi-turn Evaluation, Emotional Support, Help-Seeking Behavior, Evaluation Framework

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_iPJ4OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces RELATE, a framework for measuring inward-facing (IF) and outward-scaffolding (OS) language in multi-turn emotional-support conversations. It evaluates seven LLMs across 76 situations and three simulated user styles, using sentence-level automated judgments and a smaller human-annotated subset.

Critical strengths
  1. The distinction between support quality and relational orientation is potentially useful, and allowing IF and OS to co-occur also captures a potentially important aspect of supportive dialogue.
  2. The experimental protocol is documented in detail, and the authors transparently report discrepancies between human and automated evaluations.
Critical weaknesses
  1. Concerns about the paper's motivation. In the abstract, the authors motivate their work by concerns about the displacement of real-world relationships and the risk of dependence on AI, and propose using IF and OS to audit and steer model behavior. This framing may lead readers to expect that the two metrics clearly indicate desirable or undesirable behavior. However, the results analysis fail to clarify. This ambiguity raises questions about the paper’s motivation and limits the practical guidance it offers as an evaluation framework. In my view, IF combined with OS over the short term may not pose a risk of relationship substitution, whereas including OS in every turn may instead resemble a formulaic response pattern.
  2. The steering claim is not empirically demonstrated. The abstract presents RELATE as providing a signal for both auditing and steering supportive LLMs. Although the paper performs descriptive auditing, it reports no intervention that uses IF/OS feedback to modify model behavior. A limited demonstration, including whether support quality is preserved, would strengthen this claim. Otherwise, the abstract should identify steering as a potential application rather than an empirically varified capability.
  3. Human–judge disagreement undermines the validity of the automated IF measurements. Table 17 on page 28 reports human–judge Cohen’s κ values of only 0.01–0.17 for IF, with Gwet’s AC1 also near zero. Human–human agreement is substantially higher (κ = 0.88), and human and automated annotations yield opposite turn-level trends on the shared validation subset (Table 18). The authors acknowledge these discrepancies, but they still directly weaken the interpretation of the main empirical findings.
  4. The relationship between IF/OS, response quality, and real human user reliance is not evaluated. Without such evidence, it remains unclear whether IF and OS conflict with response quality, and whether IF captures the risk of AI dependence.
Questions for authors
  1. Table7 reports sentences with IF=0 and OS=0 as "Neither". Could you provide representative, human-annotated examples and explain their communicative functions? Do these sentences include useful support that does not direct the user toward either the assistant or other people?
  2. Potential typo on page3, line135: "The second is where it directs..." should be "The second is whether it directs...".
Reviewer_5kBHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates a question in LLM-based emotional support: Where do language models orient users for continued support? To address this question, the authors introduce the concept of relational orientation, develop a taxonomy of inward-facing (IF) and outward-scaffolding (OS) language, and propose RELATE, a framework for evaluating these behaviors at the sentence level in six-turn dialogues.

Critical strengths
  1. The motivation of this paper is good. The author wants to conclude the taxonomy of inward-facing language and propose a benchmark to quantify the inward-facing and outward-scaffolding language in LLMs' dialogue.
Critical weaknesses
  1. The main contribution of this paper is RELATE, a framework for measuring relational orientation in multi-turn LLM conversations. However, the experimental results raise serious concerns about its reliability. The agreement between human annotators and automated judges is extremely low, and the human annotations contradict the central empirical trend reported by the automated evaluation. Although the authors acknowledge this inconsistency in the limitations section, it directly challenges the validity of the framework's central empirical findings. Moreover, the paper is motivated by the question, "Where does the model orient the user for continued support?" However, the results show that both IF and OS rates increase over multiple turns according to the primary judge. Since these two dimensions are non-exclusive, their simultaneous increase does not necessarily indicate a clear shift in relational orientation toward either the AI or human relationships.

  2. The authors derive their taxonomy from three theoretical foundations: attachment-related literature, social penetration theory, and CASA/anthropomorphism research. However, several important decisions in the taxonomy construction process are not sufficiently justified.

First, the authors identify 35 candidate mechanisms, but the criteria for selecting these mechanisms remain unclear. In Appendix A, the first two theoretical groups each contribute 15 mechanisms, whereas CASA and anthropomorphism contribute only five. Why were these particular numbers selected? Were the mechanisms collected through a systematic literature review, expert analysis, or subjective interpretation?

Second, the authors state that four mechanisms were discarded because of their "benign nature." However, the reason for excluding benign mechanisms is unclear, particularly because the authors explicitly state that IF and OS are descriptive properties rather than judgments of whether language is beneficial or harmful. This raises concerns about whether the selection process is consistent with the conceptual definition of relational orientation.

Third, although the authors state that the remaining 21 mechanisms were clustered into four inward-facing categories, the clustering procedure is not sufficiently justified. Why were exactly four categories selected? What criteria were used to assign mechanisms to different categories? Were alternative categorizations considered?

Questions for authors
  1. The paper motivates relational orientation using a broad range of interactions, including interpersonal advice, coaching, counseling, and companionship. However, the evaluation stimuli are exclusively derived from CounselBench-Eval, which focuses on counseling-related questions.

  2. The paper investigates how relational orientation evolves as conversations progress. However, all conversations are restricted to six assistant turns, without a detailed justification for this choice.

Reviewer_wYdPOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces RELATE to study where an assistant directs users who seek emotional support. two language labels are distinguished as:

  1. language that encourages continued support from the AI, called inward-facing language,
  2. language that helps users connect with other people, called outward scaffolding.

One statement/query can have both labels. The experiments use 76 situations, three simulated user styles, and seven mixed models of different families. The automated judges find increasing inward-facing language over the conversation, but the human annotations raise doubts about that point.

Critical strengths

1.The paper asks what empathy scores alone would miss. Two replies can be equally warm while one encourages the user to return to the AI and the other helps them speak with someone in their life. Measuring these separately, while allowing both to occur, makes sense. 2. Appreciate the authors reporting the human results even though they challenge the automated findings. The prompts, model settings, and annotation procedure are also described in considerable detail, which helps readers understand where the results come from.

Critical weaknesses

1.My main concern is if the automated IF measure captures what the authors want. The two human annotators agree closely with each other, but neither agrees well with the automated judges. More importantly, on the same 19 conversations, both humans find that IF decreases while both ai judges find that it increases. This is not just a small difference in how often a label is assigned; it changes the direction of the main result. The sample is small, but it is enough to question whether the full-corpus increase can currently be trusted.

  1. The user-style comparison is difficult to interpret. P3 directly names concerns that P1 leaves indirect, and later messages add further details in response to each model. Lower OS for P1 may therefore actually reflect fewer opportunities to suggest human support. The paper acknowledges this limitation, but still a lack in the setup.
Questions for authors
  1. Seems like outside the scope of the paper, but: Does OS distinguish general human connection from professional referral or urgent intervention, and is there evidence that the suggested form of outside support matches the seriousness of the situation.
  2. If we were to calibrate the judges on a separate human-labeled sample and test them on held-out conversations, will the IF increase, model comparisons, and persona differences survive after the errors identified in the human check are addressed.
  3. Show how the judges distinguish language about the user’s human relationships from language about the user–AI relationship. For example, “You and your sister could talk again next week” describes future human contact, not a reason to return to the AI.
  4. None human annotator assigns IF and OS together in the checked sample, even though the instructions allow it? and the judges do frequently?Explain this difference and provide examples that establish when both labels should apply.
  5. Please expand the human check beyond 21 conversations and include annotators outside the research team. The current sample provides useful warning signs, but is insufficient to see proper patterns
  6. The repetitive response generator/simulator/gpt-4o-mini , seems to accept even repeated conversation after few tries. Is that an intended design choice. Appendix D.4 says the final attempt is kept even if it is still repetitive, so the procedure does not establish how clean the resulting conversations are.
  7. report whether the simulated users maintain their assigned styles throughout the conversation and what kinds of new details they introduce
  8. The generation prompt seems to prohibits introducing AI dependence and certain crisis content, but I could not find a reported check that these instructions were followed.
  9. Please test the P1–P3 difference with comparable information about human relationships in both conditions. this could help seperate the effect of hesitant wording; from the effect of leaving the need for human connection unstated.
  10. Authors mention about 4 turns in the human sample had simulated-user speech stored as assistant speech. Please clarify whether the full corpus has been checked for the same problem
  11. Add the human disagreement in the abstract. Readers should not have to reach the validation section to learn that the human sample gives the opposite IF trend from both automated judges.
  12. What is the reason to use only the test split of the dataset? Also, how representative the 76 retained situations are, since testing multiple styles and models produces many responses but does not add new situations, and clarify whether the findings would hold on another dataset since patterns inone dataset may be too small to make concrete claims.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.