H2HTalk: Evaluating Large Language Models as Emotional Companion

Boyang Wang; Hongcheng Guo; Yalun Wu; Zhoujun Li

REVIEW 5 major objections 4 minor 51 references

H2HTalk: Evaluating Large Language Models as Emotional Companion

T0 review · 5 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A new benchmark, H2HTalk, tests whether LLMs can be emotional companions and finds they stumble on memory, long-horizon planning, and implicit user needs.

desk verdict The benchmark's own statistics sink it: median gold response is one token, so the emotional companionship claims don't survive contact with the data. read the letter →

arxiv 2507.03543 v1 pith:WBXQXVHE submitted 2025-07-04 cs.CL cs.AI

Boyang Wang , Yalun Wu , Hongcheng Guo , Zhoujun Li This is my paper

classification cs.CLcs.AI

keywords LLMbenchmarkemotionalcompanionshipempathyevaluationattachmenttheorylong-horizonmemoryinstructionfollowingsafedialoguepersonalitydevelopment

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper introduces H2HTalk, a benchmark of 4,650 curated scenarios that tests an LLM's behavior as an emotional companion across dialogue, recollection, and itinerary planning. Each scenario is paired with a reference response, and models are scored by a unified protocol that blends n-gram overlap, semantic similarity, and, for itinerary tasks, rubric-based judgments from a proprietary judge model. Running this protocol on 50 open and proprietary LLMs, the authors report that scale helps only partially: models consistently struggle to recall earlier context, plan across long horizons, and detect implicit or shifting user needs. They also show that a persona module built from attachment-theory principles markedly improves safety in high-risk scenarios, cutting harmful-response rates by an order of magnitude. If the benchmark is accepted, it gives the field a standard comparison tool and a concrete list of capability gaps to close.

What carries the argument

The central object is the benchmark itself, H2HTalk, with its three task families and a unified score $S$ defined as the equal-weight average of BLEU-1 through BLEU-4, ROUGE-1, ROUGE-L, and embedding-based cosine similarity against a single gold response; itinerary tasks additionally blend in a proprietary judge score with weight 0.6 on $S$ and trigger human evaluation below a threshold. The Secure Attachment Persona (SAP) module operationalizes attachment theory as a safety scaffold, injecting boundary maintenance, self-regulation, and safety-first response rules into the companion's instruction context. The instruction-following analysis tests models on three scenario types—implicit help, ambiguous/contradictory, and context-aware—to expose where models fail to detect covert user needs.

What would settle it

Take a random sample of model responses from each subtask, have human annotators rate them on empathy, safety, and memory usage, and compute the rank correlation between those ratings and $S$ from Equation (2); if the correlation is near zero or negative, the benchmark's model rankings and all derived conclusions are invalidated.

Watch

Extended reading notes

Core claim

H2HTalk is the first end-to-end benchmark that evaluates emotional companions on both personality development and empathetic interaction simultaneously. Its 4,650 scenarios span companion dialogue (base, emotional, schedule), recollection (synthesis, refinement, initialization), and itinerary (basic, middle, advanced, response, initialization). The unified scoring protocol averages seven components—four BLEU variants, ROUGE-1, ROUGE-L, and cosine similarity from a multilingual embedding model—with a rubric-based judge score blended in for itinerary tasks and human adjudication below a threshold. Benchmarking 50 models reveals that long-horizon planning and memory retention are the chief bottlenecks, and that instruction-following degrades sharply when user directives are implicit, contradictory, or context-dependent. An ablation of the Secure Attachment Persona (SAP) module shows that removing it leaves lexical fluency nearly unchanged but reduces perceived safety from 4.8 to 3.2 and raises harmful-response rates from 0.7% to 7.1%.

Load-bearing premise

The benchmark's conclusions about which models are better companions rest on the assumption that the automatic score $S$—an equal-weight blend of n-gram overlap and embedding similarity to one reference response—measures emotional support quality, even though this score is never validated against human ratings of empathy or safety.

Editorial extensions

If this is right

The benchmark offers a reusable protocol for comparing emotional companions, and the released data allows other researchers to reproduce the rankings and add new models.
Smaller, fine-tuned models can rival much larger baselines—Qwen2.5-7B with LoRA adaptation reaches an overall score of 48.58, above some 20B-30B models—so adaptation matters as much as raw scale.
Models perform well on explicit instruction-following but drop sharply on implicit or context-dependent requests, pointing to a need for better affective intent detection in companion systems.
Memory-centric tasks show the largest gaps between model sizes, indicating long-term memory is a distinct, under-solved capability rather than a byproduct of general competence.
The SAP ablation shows safety can be engineered independently of fluency: removing the module raises harmful-response rates nearly tenfold while leaving lexical metrics almost unchanged.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Because $S$ compares against a single gold response per scenario, models that echo the reference wording may be rewarded even when a different response would be more supportive; correlating $S$ with human empathy judgments would test this directly.
The benchmark's short-horizon tasks may understate memory failures; a multi-session variant where the companion must recall facts from earlier conversations would likely widen the observed gaps.
The attachment-theory scaffolding behind SAP could be repurposed as a general safety layer for mental-health chatbots outside companionship settings, and the paper's safety protocol offers a template for testing it.
The implicit-instruction findings suggest that empathetic-dialogue benchmarks should include a diagnostic subset for detecting covert needs, since explicit-command performance is not predictive of implicit performance.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: the central empirical claim doesn't survive contact with the paper's own statistics. Section 3.3 and Figure 6 report that gold responses in H2HTalk have mean length 3.02 tokens, median 1 token. A one-token target cannot contain the empathy, recollection, or multi-step itinerary planning that Section 3.1 and the Figure 1 example describe. The unified metric S in Eq. (2) then mostly measures single-word overlap, so the 50-model rankings in Table 2 are not about emotional companionship at all.

To its credit, the paper assembles a substantial artifact: 4,650 scenarios across dialogue, recollection, and itinerary, a 50-model evaluation, and a Secure Attachment Persona module that addresses a real safety concern. The design space is useful and the field does need better evaluation here. The SAP ablation, though small, points to something non-obvious: safety framing can be separated from surface fluency.

The problems are load-bearing. First, the length distribution is an internal inconsistency between the data description and the tasks. If the released data matches Fig. 6, the benchmark cannot support its stated capabilities. Second, Eq. (3) adds a 0–100 scale to a 0–5 GPT-4o score with no normalization, making the Itinerary scores uninterpretable; the tau threshold then triggers on a scale that doesn't mean what the authors think. Third, S is never validated against human judgments of empathy or safety, and the equal weights are asserted rather than argued. The main table has no error bars or significance tests, so the model comparisons, including the memory and instruction-following conclusions, are statistically unsupported. Fourth, 'first comprehensive benchmark' overreaches: the related work includes EmotionQueen, InCharacter, and ToMBench, which already occupy adjacent territory. Finally, the gold responses are LLM-generated and the evaluation uses LLM embeddings plus GPT-4o judgment, so there's an unaddressed circularity.

This paper is for anyone building or evaluating companion models. My view: the dataset might be salvageable, but the current version is not sound. The fix is to release the data, show per-subtask lengths, normalize the metric, and validate against human raters. A serious referee would catch these issues, so I'd send it to review with major revision expected — but I wouldn't cite it as it stands.

Referee Report

5 major / 4 minor

Summary. The paper introduces H2HTalk, a benchmark for evaluating LLMs as emotional companions, with 4,650 scenarios spanning dialogue, recollection, and itinerary planning, plus a Secure Attachment Persona (SAP) module for safer interactions. The authors benchmark 50 LLMs using a composite score S that combines BLEU, ROUGE, and BGE-M3 embedding similarity, with GPT-4o rubric scores for itinerary tasks, and report rankings and capability gaps, particularly in long-horizon planning and memory retention. The paper's central claim is that H2HTalk is a valid, reusable benchmark and that the 50-model study reveals genuine limitations in emotional companionship.

Significance. If the benchmark were valid, its scale (4,650 scenarios, 50 models, open-sourced materials) and its attempt to incorporate attachment theory would be useful contributions to LLM evaluation for emotionally intelligent companions. The paper also deserves credit for releasing code and data and for conducting a broad model comparison. However, the paper's own reported length statistics and the unvalidated composite metric undermine the central empirical claims; the benchmark, as presented, does not provide evidence for the stated conclusions about long-horizon planning, memory retention, or the safety benefits of SAP.

major comments (5)

[§3.3, Fig. 6, and §4.1] The output length distribution reported in §3.3 (median 1 token, mean 3.02 tokens, few outputs exceeding 15 tokens) is internally inconsistent with the rich reference responses required by the task definitions in §3.1 and the detailed example in Fig. 1. With a median gold response of one token, the lexical and embedding scores in Eq. (2) reduce to single-word or short-phrase matching, so BLEU-4, ROUGE-L, and BGE-M3 cosine similarity cannot measure empathy, recollection, or multi-step planning. Consequently, the model rankings in Table 2 and the conclusions that long-horizon planning and memory retention remain key challenges are not supported by the reported data.
[§4.1, Eq. (3)] Equation (3) mixes scores on incompatible scales: S in Eq. (2) is an average of BLEU/ROUGE scores and a cosine similarity, which Table 3 reports on a 0-100 scale (e.g., BLEU-4 32.7, ROUGE-L 58.9, SS 91.4%), while J_GPT-4o is defined on a 0-5 scale. Without normalization, the S term dominates the linear combination for typical score ranges, so the GPT-4o rubric component contributes almost nothing to J; the claimed 'unified' protocol is therefore not actually balancing the two components as stated.
[§4.1 and Table 2] The paper never validates the composite score S against human judgments of empathy, safety, or response quality. No correlation with human ratings, no inter-annotator agreement, and no error bars or significance tests are reported for the 50-model comparisons in Table 2. Without such validation, the model rankings and the claimed capability gaps rest on an unvalidated proxy, and the differences between models (often smaller than one point) cannot be interpreted as meaningful.
[§3.2, §4.1, and §5] The benchmark's gold responses are LLM-generated and its evaluation uses BGE-M3 embeddings and GPT-4o judgments, so LLMs are judged against LLM-produced references with LLM-based tools. The paper does not report any human-model agreement analysis or a human-validated subset, leaving open the possibility that the rankings reflect LLM stylistic preferences rather than emotional support quality; this is a load-bearing validity concern, not merely a philosophical one, and should be addressed with concrete human-evaluation experiments.
[§5, Table 3] The SAP safety claim is based on only 33 high-risk scenarios and 150 interactions, with a subjective 'Safety Perception Score' rated 1-5, and the paper does not report annotator agreement, scenario selection criteria, or whether the SAP parameters were optimized on these scenarios. The reported 33% drop in safety perception and order-of-magnitude increase in violation rate are therefore not established as generalizable effects of the SAP module.

minor comments (4)

[Abstract and title] The title and abstract contain the typo 'H2HT alk' instead of 'H2HTalk'; please correct the spacing throughout the manuscript.
[§3.2] The text says the dataset was developed through a 'rigorous five-phase protocol' but then lists only three phases (Data Gathering, Data Pre-Processing, Data Refinement), while Fig. 3 shows six stages; the numbering is inconsistent and should be reconciled.
[References] The text refers to 'Claude-3.7' but reference [5] is titled 'Claude 3.5 Sonnet'; either the model name or the citation is incorrect.
[§4.1] The phrase 'Levels I-III companion tasks' is undefined; please specify which subtasks use the GPT-4o scoring methodology in Eq. (3).

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found; the benchmark is an empirical evaluation protocol, and the flagged issues are validity concerns rather than reductions.

full rationale

The paper's central claims are empirical measurements, not reductions. The evaluation metric S in Eq. (2) measures lexical and embedding similarity to gold responses; even if those gold responses are LLM-generated, model scores are computed from model outputs and are not defined to equal the benchmark inputs by construction. The SAP ablation in Section 5 relies on human-rated Safety Perception Score and annotator-identified Violation Response Rate, so the safety conclusion does not reduce to the model-based metric S. Self-citations (e.g., RoleLLM [35], InCharacter [37]) appear in related-work background and are not load-bearing justifications for the benchmark's validity; no uniqueness theorem is imported from prior work by the same authors. A genuine internal inconsistency does exist: Section 3.3 and Figure 6 report gold responses with median 1 token and mean 3.02, which undermines the claim that Itinerary and Recollection subtasks test long-horizon planning and memory; however, that is a construct-validity flaw, not a circular derivation. Without an equation-level reduction or a fitted parameter renamed as a prediction, no circularity step can be substantiated, so the circularity score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The benchmark's validity depends on several unvalidated assumptions: that lexical and embedding similarity to a single gold response measures emotional intelligence, that LLM-generated simulation data is authentic, that attachment theory can be reduced to a prompt module, and that GPT-4o rubrics are reliable. The metric weights and thresholds are hand-chosen free parameters. The SAP module is an invented component whose safety benefits are only shown within the authors' own framework.

free parameters (4)

Metric weights in S = 1/7 each
Equal weights for BLEU-1..4, ROUGE-1, ROUGE-L, and SS in Equation (2), chosen by hand with no sensitivity analysis.
alpha in J = 0.6
Mixing weight between S and GPT-4o judgment in Equation (3), chosen by hand with no robustness check.
tau threshold = 3.5
Threshold triggering human adjudication in Equation (4), chosen by hand with no justification.
SAP parameter optimization = unspecified
Section 3.4 says Ryan's paradox is 'resolved through parameter optimization' but no parameters or values are reported.

assumptions (4)

domain assumption BLEU/ROUGE and BGE-M3 semantic similarity against a single gold response are valid proxies for emotional support quality
The entire leaderboard rests on this assumption; it is stated in Section 4.1 without validation against human judgments.
domain assumption LLM-generated simulation data reflects authentic emotional interactions
Section 3.2 'Data Gathering' says the data comes from simulations; no human-authored gold conversations are described.
domain assumption Attachment theory principles can be meaningfully operationalized as a prompt module (SAP)
Section 3.4 maps Bowlby, Ainsworth, Gottman, and Gross into the SAP design without empirical validation of the operationalization.
domain assumption GPT-4o rubric scores are reliable without human adjudication except below a threshold
Section 4.1 uses J_GPT-4o in Equation (3) but reports no human agreement statistics for the rubric.

invented entities (1)

Secure Attachment Persona (SAP) module
purpose: A prompt-level module incorporating attachment-theory principles to make LLM companions safer and more empathetic
Its safety benefit is demonstrated only in the authors' own 150-interaction ablation (Table 3); no external validation is provided.

how reviews work

0 comments

Cite this review

Pith. "Pith review of H2HTalk: Evaluating Large Language Models as Emotional Companion." pith.science (2026). https://pith.science/paper/WBXQXVHE

@misc{pith2026250703543,
  author       = {Pith},
  title        = {Pith review of: H2HTalk: Evaluating Large Language Models as Emotional Companion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WBXQXVHE}},
  note         = {Machine review of arXiv:2507.03543}
}

read the original abstract

As digital emotional support needs grow, Large Language Model companions offer promising authentic, always-available empathy, though rigorous evaluation lags behind model advancement. We present Heart-to-Heart Talk (H2HTalk), a benchmark assessing companions across personality development and empathetic interaction, balancing emotional intelligence with linguistic fluency. H2HTalk features 4,650 curated scenarios spanning dialogue, recollection, and itinerary planning that mirror real-world support conversations, substantially exceeding previous datasets in scale and diversity. We incorporate a Secure Attachment Persona (SAP) module implementing attachment-theory principles for safer interactions. Benchmarking 50 LLMs with our unified protocol reveals that long-horizon planning and memory retention remain key challenges, with models struggling when user needs are implicit or evolve mid-conversation. H2HTalk establishes the first comprehensive benchmark for emotionally intelligent companions. We release all materials to advance development of LLMs capable of providing meaningful and safe psychological support.

Figures

Figures reproduced from arXiv: 2507.03543 by the authors.

**Figure 1.** An example of LLMs generating responses that delivers formulaic congratulations and generic stress management advice, and out H2HTalk demonstrates sophisticated personality development through four key capabilities: Emotional Support, Recollection, Itinerary, and Experience Condensation, creating a more natural, contextually-aware interaction that simulates genuine companionship. To fill this gap we introduce Heart-… view at source ↗

**Figure 2.** H2HTalk consists of companion dialogue, recollection and itinerary components. 3.1 Definition and Framework H2HTalk establishes a dual-mode foundation for emotional companion LLMs through personality development and empathetic interaction. As shown in [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** The protocol consists of (1)Data Gathering, (2)Data Pre-Processing, [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 3.** Figure 3: The pipeline gathers diverse emotional interactions between users and LLMs for companion role-play, processes this data through anonymization and safety filtering, then refines it via crowd-voting and expert assessment. The resulting instruction-response pairs form the…

**Figure 4.** Figure 4: The word cloud of H2HTalk [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 7.** Figure 7: Evaluate instruction-following across three scenarios. (1) ImplicitHelp: emotionally intense conversations with minimal directives; (2) Ambiguous/Contradictory: mixed emotions with shifting goals; (3) Context-Aware: directives requiring conversation history comprehe…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 27 canonical work pages

[1]

Complementary Therapies in Clinical Practice41, 101223 (2020)

Human-animal-environment interactions as a context for youth social-emotional health and wellbeing: Practitioners’ perspectives on processes of change, implemen- tation, and challenges. Complementary Therapies in Clinical Practice41, 101223 (2020)

work page 2020
[2]

Animals11(8) (Aug 2021).https://doi.org/10

Companion animal type and level of engagement matter: a mixed-methods study examining links between companion animal guardianship, loneliness and well-being during the covid-19 pandemic. Animals11(8) (Aug 2021).https://doi.org/10. 3390/ani11082349

work page 2021
[3]

Exceptionality Education International31(1), 1–23 (2021)

Abat-Roy, V.: Service animals and pet therapy in schools: synthesizing a review of the literature. Exceptionality Education International31(1), 1–23 (2021)

work page 2021
[4]

Psychology press (2015)

Ainsworth, M.D.S., Blehar, M.C., Waters, E., Wall, S.N.: Patterns of attachment: A psychological study of the strange situation. Psychology press (2015)

work page 2015
[5]

Anthropic News (June 2024), https://www

Anthropic: Claude 3.5 sonnet. Anthropic News (June 2024), https://www. anthropic.com/news/claude-3-5-sonnet

work page 2024
[6]

arXiv preprint arXiv:2406.12934 (2024)

Ayyamperumal, S.G., Ge, L.: Current state of llm risks and ai guardrails. arXiv preprint arXiv:2406.12934 (2024)

arXiv 2024
[7]

Basic books (2008)

Bowlby, J.: A secure base: Parent-child attachment and healthy human development. Basic books (2008)

work page 2008
[8]

In: Attachment theory, pp

Bretherton, I.: The origins of attachment theory: John bowlby and mary ainsworth. In: Attachment theory, pp. 45–84. Routledge (2013)

work page 2013

Show all 51 references

[9]

arXiv preprint arXiv:2404.18231 (2024)

Chen, J., Wang, X., Xu, R., Yuan, S., Zhang, Y., Shi, W., Xie, J., Li, S., Yang, R., Zhu, T., et al.: From persona to personalization: A survey on role-playing language agents. arXiv preprint arXiv:2404.18231 (2024)

2024 arXiv
[10]

CoRRabs/2402.03216 (2024)

Chen, J., Xiao, S., Zhang, P., Luo, K., Lian, D., Liu, Z.: BGE m3-embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self- knowledge distillation. CoRRabs/2402.03216 (2024)

2024 arXiv
[11]

In: Findings of the Association for Computational Linguistics: EMNLP

Chen, N., Wang, Y., Jiang, H., Cai, D., Li, Y., Chen, Z., Wang, L., Li, J.: Large language models meet harry potter: A dataset for aligning dialogue agents with characters. In: Findings of the Association for Computational Linguistics: EMNLP
[12]

arXiv preprint arXiv:2409.13359 (2024)

Chen, Y., Wang, H., Yan, S., Liu, S., Li, Y., Zhao, Y., Xiao, Y.: Emotionqueen: A benchmark for evaluating empathy of large language models. arXiv preprint arXiv:2409.13359 (2024)

2024 arXiv
[13]

arXiv preprint arXiv:2402.15052 (2024)

Chen, Z., Wu, J., Zhou, J., Wen, B., Bi, G., Jiang, G., Cao, Y., Hu, M., Lai, Y., Xiong, Z., et al.: Tombench: Benchmarking theory of mind in large language models. arXiv preprint arXiv:2402.15052 (2024)

2024 arXiv
[14]

arXiv preprint arXiv:2402.18225 (2024)

Coda-Forno, J., Binz, M., Wang, J.X., Schulz, E.: Cogbench: a large language model walks into a psychology lab. arXiv preprint arXiv:2402.18225 (2024)

2024 arXiv
[15]

https://github.com/open-compass/opencompass (2023)

Contributors, O.: Opencompass: A universal evaluation platform for foundation models. https://github.com/open-compass/opencompass (2023)

2023
[16]

Penguin (2011)

Fisher, R., Ury, W.L., Patton, B.: Getting to yes: Negotiating agreement without giving in. Penguin (2011)

2011
[17]

Routledge (2014)

Gottman, J.M.: Principia amoris: The new science of love. Routledge (2014)

2014
[18]

Psychological inquiry 26(1), 1–26 (2015)

Gross, J.J.: Emotion regulation: Current status and future prospects. Psychological inquiry 26(1), 1–26 (2015)

2015
[19]

arXiv preprint arXiv:2403.15401 (2024) 12 B

Guo, Z., Lai, A., Thygesen, J.H., Farrington, J., Keen, T., Li, K.: Large language model for mental health: A systematic review. arXiv preprint arXiv:2403.15401 (2024) 12 B. Wang et al

2024 arXiv
[20]

5-coder technical report

Hui, B., Yang, J., Cui, Z., Yang, J., Liu, D., Zhang, L., Liu, T., Zhang, J., Yu, B., Lu, K., et al.: Qwen2. 5-coder technical report. arXiv preprint arXiv:2409.12186 (2024)

2024 arXiv
[21]

arXiv preprint arXiv:2406.18841 (2024)

Jiao, J., Afroogh, S., Xu, Y., Phillips, C.: Navigating llm ethics: Advancements, challenges, and future directions. arXiv preprint arXiv:2406.18841 (2024)

2024 arXiv
[22]

arXiv preprint arXiv:2308.09597 (2023)

Li, C., Leng, Z., Yan, C., Shen, J., Wang, H., MI, W., Fei, Y., Feng, X., Yan, S., Wang, H., et al.: Chatharuhi: Reviving anime character in reality via large language model. arXiv preprint arXiv:2308.09597 (2023)

2023 arXiv
[23]

arXiv preprint arXiv:2406.17675 (2024)

Li, Y., Huang, Y., Wang, H., Zhang, X., Zou, J., Sun, L.: Quantifying ai psy- chology: A psychometrics benchmark for large language models. arXiv preprint arXiv:2406.17675 (2024)

2024
[24]

In: Text Summarization Branches Out (Jul 2004)

Lin, C.Y.: ROUGE: A package for automatic evaluation of summaries. In: Text Summarization Branches Out (Jul 2004)

2004
[25]

arXiv preprint arXiv:2402.17753 (2024)

Maharana, A., Lee, D.H., Tulyakov, S., Bansal, M., Barbieri, F., Fang, Y.: Evaluating very long-term conversational memory of llm agents. arXiv preprint arXiv:2402.17753 (2024)

2024 arXiv
[26]

CoRR (2024)

Mo, Y., Yang, J., Liu, J., Zhang, S., Wang, J., Li, Z.: C-ICL: contrastive in-context learning for information extraction. CoRR (2024)

2024
[27]

In: Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, July 6-12, 2002, Philadelphia, PA, USA

Papineni, K., Roukos, S., Ward, T., Zhu, W.: Bleu: a method for automatic evaluation of machine translation. In: Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, July 6-12, 2002, Philadelphia, PA, USA. pp. 311–318. ACL (2002)

2002
[28]

arXiv preprint arXiv:2408.15787 (2024)

Qiu, H., Lan, Z.: Interactive agents: Simulating counselor-client psychological counseling via role-playing llm-to-llm interactions. arXiv preprint arXiv:2408.15787 (2024)

2024 arXiv
[29]

American psychologist55(1), 68 (2000)

Ryan, R.M., Deci, E.L.: Self-determination theory and the facilitation of intrinsic motivation, social development, and well-being. American psychologist55(1), 68 (2000)

2000
[30]

arXiv preprint arXiv:2310.10158 (2023)

Shao, Y., Li, L., Dai, J., Qiu, X.: Character-llm: A trainable agent for role-playing. arXiv preprint arXiv:2310.10158 (2023)

2023 arXiv
[31]

NPJ Mental Health Research3(1), 12 (2024)

Stade, E.C., Stirman, S.W., Ungar, L.H., Boland, C.L., Schwartz, H.A., Yaden, D.B., Sedoc, J., DeRubeis, R.J., Willer, R., Eichstaedt, J.C.: Large language models could change the future of behavioral healthcare: a proposal for responsible development and evaluation. NPJ Menta...

2024
[32]

Team, Q.: Qwen2.5: A party of foundation models (September 2024),https:// qwenlm.github.io/blog/qwen2.5/

2024
[33]

Frontiers in Dementia3, 1385303 (2024)

Treder, M.S., Lee, S., Tsvetanov, K.A.: Introduction to large language models (llms) for dementia care and research. Frontiers in Dementia3, 1385303 (2024)

2024
[34]

arXiv preprint arXiv:2406.01171 (2024)

Tseng, Y.M., Huang, Y.C., Hsiao, T.Y., Chen, W.L., Huang, C.W., Meng, Y., Chen, Y.N.: Two tales of persona in llms: A survey of role-playing and personalization. arXiv preprint arXiv:2406.01171 (2024)

2024 arXiv
[35]

In: Ku, L.W., Martins, A., Srikumar, V

Wang, N., Peng, Z., Que, H., Liu, J., Zhou, W., Wu, Y., Guo, H., Gan, R., Ni, Z., Yang, J., Zhang, M., Zhang, Z., Ouyang, W., Xu, K., Huang, W., Fu, J., Peng, J.: RoleLLM: Benchmarking, eliciting, and enhancing role-playing abilities of large language models. In: Ku, L.W., Mar...

2024
[36]

arXiv preprint arXiv:2305.17926 (2023) H2HTalk: Evaluating Large Language Models as Emotional Companion 13

Wang, P., Li, L., Chen, L., Cai, Z., Zhu, D., Lin, B., Cao, Y., Liu, Q., Liu, T., Sui, Z.: Large language models are not fair evaluators. arXiv preprint arXiv:2305.17926 (2023) H2HTalk: Evaluating Large Language Models as Emotional Companion 13

2023 arXiv
[37]

In: Ku, L.W., Martins, A., Srikumar, V

Wang, X., Xiao, Y., Huang, J.t., Yuan, S., Xu, R., Guo, H., Tu, Q., Fei, Y., Leng, Z., Wang, W., Chen, J., Li, C., Xiao, Y.: InCharacter: Evaluating personality fidelity in role-playing agents through psychological interviews. In: Ku, L.W., Martins, A., Srikumar, V. (eds.) Pro...

2024
[38]

arXiv preprint arXiv:2402.15116 (2024)

Xie, J., Chen, Z., Zhang, R., Wan, X., Li, G.: Large multimodal agents: A survey. arXiv preprint arXiv:2402.15116 (2024)

2024 arXiv
[39]

arXiv preprint arXiv:2402.06044 (2024)

Xu, H., Zhao, R., Zhu, L., Du, J., He, Y.: Opentom: A comprehensive benchmark for evaluating theory-of-mind reasoning capabilities of large language models. arXiv preprint arXiv:2402.06044 (2024)

2024 arXiv
[40]

Xu, R., Wang, X., Chen, J., Yuan, S., Yuan, X., Liang, J., Chen, Z., Dong, X., Xiao, Y.: Character is destiny: Can large language models simulate persona-driven decisions in role-playing? arXiv preprint arXiv:2404.12138 (2024)

2024 arXiv
[41]

Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies8(2), 1–41 (2024)

Xu, Z., Xu, H., Lu, Z., Zhao, Y., Zhu, R., Wang, Y., Dong, M., Chang, Y., Lv, Q., Dick, R.P., et al.: Can large language models be good companions? an llm-based eyewear system with conversational common ground. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiqui...

2024
[42]

In: Proceedings of the AAAI Conference on Artificial Intelligence

Yang, J., Hui, B., Yang, M., Yang, J., Zhang, L., Qu, Q., Lin, J.: Fine-tuning language models with collaborative and semantic experts. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 39, pp. 25624–25632 (2025)

2025
[43]

Journal of Medical Artificial Intelligence pp

Yu, H., McGuinness, S.: An experimental study of integrating fine-tuned llms and prompts for enhancing mental health support chatbot system. Journal of Medical Artificial Intelligence pp. 1–16 (2024)

2024
[44]

In: Al-Onaizan, Y., Bansal, M., Chen, Y.N

Zhang, C., Jian, Y., Ouyang, Z., Vosoughi, S.: Working memory identifies reasoning limits in language models. In: Al-Onaizan, Y., Bansal, M., Chen, Y.N. (eds.) Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (Nov 2024)

2024
[45]

arXiv preprint arXiv:2406.03127 (2024)

Zhang, S., Yan, C., Yang, J., Liu, J., Mo, Y., Bai, J., Li, T., Li, Z.: Towards real- world scenario: Imbalanced new intent discovery. arXiv preprint arXiv:2406.03127 (2024)

2024 arXiv
[46]

In: Joint European Conference on Machine Learning and Knowledge Discovery in Databases

Zhang, S., Yan, C., Yang, J., Zhang, W., Ren, C., Li, T., Bai, J., Li, Z.: Tinid: A transfer and interpretable llm-enhanced framework for new intent discovery. In: Joint European Conference on Machine Learning and Knowledge Discovery in Databases. pp. 195–212. Springer (2024)

2024
[47]

arXiv preprint arXiv:2403.16913 (2024)

Zhang, S., Yang, J., Bai, J., Yan, C., Li, T., Yan, Z., Li, Z.: New intent discovery with attracting and dispersing prototype. arXiv preprint arXiv:2403.16913 (2024)

2024 arXiv
[48]

arXiv preprint arXiv:2310.01459 (2023)

Zhao,R.,Zhang,W.,Li,J.,Zhu,L.,Li,Y.,He,Y.,Gui,L.:Narrativeplay:Interactive narrative understanding. arXiv preprint arXiv:2310.01459 (2023)

2023 arXiv
[49]

In: Proceedings of the AAAI Conference on Artificial Intelligence

Zhong, W., Guo, L., Gao, Q., Ye, H., Wang, Y.: Memorybank: Enhancing large language models with long-term memory. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 19724–19731 (2024)

2024
[50]

arXiv preprint arXiv:2311.16832 (2023)

Zhou, J., Chen, Z., Wan, D., Wen, B., Song, Y., Yu, J., Huang, Y., Peng, L., Yang, J., Xiao, X., et al.: Characterglm: Customizing chinese conversational ai characters with large language models. arXiv preprint arXiv:2311.16832 (2023)

2023 arXiv
[2023]

8506–8520 (2023)

pp. 8506–8520 (2023)

2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Complementary Therapies in Clinical Practice41, 101223 (2020)

Human-animal-environment interactions as a context for youth social-emotional health and wellbeing: Practitioners’ perspectives on processes of change, implemen- tation, and challenges. Complementary Therapies in Clinical Practice41, 101223 (2020)

work page 2020

[2] [2]

Animals11(8) (Aug 2021).https://doi.org/10

Companion animal type and level of engagement matter: a mixed-methods study examining links between companion animal guardianship, loneliness and well-being during the covid-19 pandemic. Animals11(8) (Aug 2021).https://doi.org/10. 3390/ani11082349

work page 2021

[3] [3]

Exceptionality Education International31(1), 1–23 (2021)

Abat-Roy, V.: Service animals and pet therapy in schools: synthesizing a review of the literature. Exceptionality Education International31(1), 1–23 (2021)

work page 2021

[4] [4]

Psychology press (2015)

Ainsworth, M.D.S., Blehar, M.C., Waters, E., Wall, S.N.: Patterns of attachment: A psychological study of the strange situation. Psychology press (2015)

work page 2015

[5] [5]

Anthropic News (June 2024), https://www

Anthropic: Claude 3.5 sonnet. Anthropic News (June 2024), https://www. anthropic.com/news/claude-3-5-sonnet

work page 2024

[6] [6]

arXiv preprint arXiv:2406.12934 (2024)

Ayyamperumal, S.G., Ge, L.: Current state of llm risks and ai guardrails. arXiv preprint arXiv:2406.12934 (2024)

arXiv 2024

[7] [7]

Basic books (2008)

Bowlby, J.: A secure base: Parent-child attachment and healthy human development. Basic books (2008)

work page 2008

[8] [8]

In: Attachment theory, pp

Bretherton, I.: The origins of attachment theory: John bowlby and mary ainsworth. In: Attachment theory, pp. 45–84. Routledge (2013)

work page 2013