Pith. sign in

REVIEW 4 major objections 4 minor 28 references

ZifaMem: Structured Memory for Persona, Preference, and Emotional Continuity in AI Companions

T0 review · 4 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read Structured memory, not injected emotion state, is what gives AI companions their sense of emotional continuity.

desk verdict An unusually honest empirical paper whose headline emotional-intelligence gain rests on a single unvalidated LLM judge—worth engaging, but treat the A5 numbers as conditional. read the letter →

arxiv 2607.17564 v1 pith:OSIZBKGW submitted 2026-07-20 cs.AI cs.LG

Jingzhe Fang , Guozhi Xu , Yunfan Cui , Xiaochen Yang , Zhangyu Hua This is my paper
classification cs.AIcs.LG
keywords AIcompanionmemoryemotionalcontinuitystructuredpersonagroundingpreferencefollowingLLM-as-a-judgeaffectstateestimationepisodic
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 claims that for AI companions, emotional continuity—remembering who the user is, what they prefer, and how past interactions felt—is carried by consolidated, affect-tagged structured memory, not by separately injected emotion signals. Testing against a deliberately strong comparator that receives the full raw dialogue history, structured memory improved judged emotional-intelligence scores by 11.4% pooled across four backbones (95% CI 6.3% to 17.1%), with persona and preference gains concentrated where the underlying model left headroom. The paper also reports honest nulls: an explicit affect state machine failed its primary comparison, and injecting an affect block added nothing once structured memory was present. A preregistered same-protocol comparison found the system statistically equivalent to an existing open memory system within ±5 percentage points on the primary preference endpoint, while simple verbatim retrieval was a strong baseline for short-history preference tasks. A sympathetic reader would care because the paper sharply separates 'what to store' from 'what to compute live,' with evidence that live affect dynamics earn no measurable response value.

What carries the argument

The carrying mechanism is the layered memory substrate: L1 recent turns, L2 session summaries, L3 episodic memories, L4 a consolidated user model, with a Boolean promotion gate (importance ≥ 0.5 or an emotional peak recorded) moving sessions into episodic memory, and a reranker score of 0.35 semantic similarity + 0.30 strength + 0.35 importance. Affect tags enter through this substrate rather than through a separate module. The counterpoint studied is the legacy affect path—a companion-mood appraisal updating a VAD state with damped second-order dynamics—and a confidence-adaptive online estimator of user affect with state, trend, and change-point handling. The experiments isolate the value o

What would settle it

Have human raters score the same paired raw-history versus structured-memory responses on emotional continuity, and check whether the pooled +0.42 A5 delta survives human judgment; and run the knockout ablation that removes affect tags from consolidation and retrieval to test whether the gain is actually attributable to the affect tagging.

Watch

Extended reading notes

Core claim

The central claim is that a layered memory representation—recent-turn buffer, session summaries, episodic memories, and a consolidated user model, with affect tags participating in consolidation and retrieval—rather than a raw transcript or an explicit emotion module, is what produces emotionally continuous AI companion responses. Against a comparator that supplies the full raw dialogue history, the structured representation raised judged emotional intelligence (a 1–5 rubric score) by +0.42 pooled across four verified backbones, an 11.4% relative gain; persona grounding improved on all four backbones, and the largest capability gains appeared where the backbone's raw-history baseline left he

Load-bearing premise

All headline gains rest on a single LLM judge's scores, with no demonstrated human agreement; if that judge rewards structured formatting or longer context rather than genuine emotional continuity, the 11.4% improvement would be an artifact.

Editorial extensions

If this is right

  • Replacing raw dialogue history with any tested memory system improves companion responses consistently; raw history is the dominated deployment option.
  • Gains are workload- and backbone-specific: persona and preference improvements concentrate where the underlying model leaves headroom, and emotional intelligence improves directionally on every verified backbone.
  • Emotional threads beat snapshots: multi-turn affect context wins a +39% net preference over a single-turn snapshot, but a latent dynamics layer adds no measurable value beyond recent observations.
  • For short-history preference tasks, verbatim retrieval is a cheap, strong baseline that exact wording summarization can lose.
  • The long-history advantage is directionally positive but unproven at current power; temporal-reasoning questions regress under summarization.

Reading between the lines

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

  • Editorial inference: if the affect-tag ablation (removing affect metadata from consolidation and retrieval) were run, the A5 gain might shrink or vanish, which would relocate the benefit from emotional tagging to generic structure; this is a direct test the paper leaves open.
  • Editorial inference: the temporal-reasoning regression at full history and the failed latent-state contrast jointly suggest a concrete design change—keep timestamped, verbatim-anchored episodic threads alongside summaries—that would likely improve both retrieval and reasoning on long histories.
  • Editorial inference: the backbone-dependent sign of affect injection implies future memory evaluations should report per-backbone, per-workload maps rather than pooled averages, since pooled gains can hide real costs, as shown by the Gemini A2/A3 regressions.
  • Editorial inference: the equivalence result on the primary endpoint, combined with the strong verbatim-retrieval baseline, suggests the near-term deployment choice among memory systems may hinge on integration cost rather than measured quality, until better-powered A5 evidence accumulates.
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.

Referee Report

4 major / 4 minor

Summary. The paper proposes ZifaMem, a structured memory system for AI companions that consolidates dialogue into session summaries, episodic memories, and a user model with affect tags. It evaluates the system on ALPSBench Task 4 (dev split) under a fixed DeepSeek-v4 LLM-as-a-judge protocol, comparing structured memory against full raw dialogue history across four backbones, and in a preregistered comparison against Mem0 and simple filtered retrieval. The headline result is a pooled +0.42 (11.4%) gain on the A5 emotional-intelligence endpoint, with persona/preference gains concentrated on Claude. The paper also reports a paired affect-injection matrix, a failed DYAFFECT primary, a null effect of an emotion state machine, and an underpowered long-history experiment. The manuscript is notable for disclosing failures, route audits, and preregistration tiers.

Significance. If the A5 endpoint were human-validated, the paper would provide a carefully controlled demonstration that structured memory improves judged emotional continuity over a deployment-honest raw-history comparator, plus a clean negative result on explicit affect state machines. The evaluation discipline—route audits, frozen preregistrations, reporting failed primaries, separating exploratory and preregistered evidence—is exemplary and worth emulating. The decision-map framing (when to use structured memory vs. verbatim retrieval) is practical. However, the central construct is measured only by a single unvalidated LLM judge, and the pooled A5 analysis may have a statistical clustering issue; the significance is therefore conditional.

major comments (4)
  1. [§3.2, Table 1; Limitations (§8)] The headline pooled A5 gain (11.4%, +0.42) rests entirely on a single DeepSeek-v4 LLM judge with three repeats, and the judge bridge in Appendix A compares two LLM judges, not judges to humans. The paper's own Limitations state 'Human evaluation of longitudinal companion responses remains necessary.' If the judge systematically rewards longer or more structured responses, the A5 gain would be a stylistic artifact rather than emotional continuity. This is load-bearing because A5 is the only endpoint that directly supports the title's 'emotional continuity' claim. Please provide human-alignment evidence on a sample, or explicitly reframe the abstract as 'judged emotional intelligence by an LLM judge' with the human-validation caveat.
  2. [§3.2, Table 1 (Pooled n=208)] The A5 instances are the same 52 benchmark dialogues for each of the four backbones, yet the pooled estimate reports n=208. If the paired-instance bootstrap resamples the 208 backbone-instance pairs independently, it ignores the clustering of four responses per dialogue; the effective independent unit is 52 dialogues, not 208. This could materially narrow the headline 95% CI [+0.25,+0.60]. Please report a cluster bootstrap resampling dialogues, or clarify that the pooled interval already accounts for this correlation.
  3. [§3.4, §5.1, Table 6] The preregistered prior-system comparison reuses the ZifaMem and raw-history columns from pilot cohorts run five to six days earlier than the Mem0 and simple-RAG columns. The primary contrast ZifaMem−Mem0 therefore is not a single concurrent batch, and uncontrolled batch effects (model drift, endpoint state) are not eliminated by the disclosed same-prompt assertion. Additionally, the ±5pp TOST margin was chosen post hoc, matching the judge-noise band rather than being frozen in the preregistration. These two facts weaken the abstract's phrase 'statistically equivalent within ±5 points'; the evidence supports a weaker claim such as 'no detected difference, with a post-hoc equivalence bound.'
  4. [§3.1, Eq. (1)–(2), §8 Limitations] All main results come from the ALPSBench Task 4 dev split, with test gold withheld by the maintainers. The reranker weights in Eq. (2) and the promotion threshold in Eq. (1) are presented as fixed constants, but the paper does not state whether any component was tuned on this dev split. If the memory system's parameters were selected using the dev endpoints, the measured structured-memory advantage may be optimistic. Please disclose the tuning history, or validate on an unseen split once available, to support generalizability.
minor comments (4)
  1. [§3.2] The judge endpoint is described as 'temperature zero' with 'three repeats.' At temperature zero, repeats are expected to be identical unless there is hardware non-determinism; the paper does not report repeat agreement. Please clarify what variance the three repeats expose, or use sampling temperature > 0 for judge noise estimation.
  2. [§4.3, Table 4] The C4 vs C2 comparison is reported as a win (+0.395) despite answer-order discordance of 18.4%, above the frozen 15% limit. The text carefully labels it exploratory, but the section title 'Emotional threads beat snapshots' may overstate what the evidence supports; consider rephrasing to 'directionally favor multi-turn context.'
  3. [§3.1] The paper says 'A4 was not executed' but does not explain why. A one-sentence clarification would help readers assess the completeness of the capability map.
  4. [Table 3 caption] The statement about 'third decimal' differences across bootstrap seeds is fine, but consider moving it to the main text for clarity since Table 3 is cited for precise deltas.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical comparisons against external benchmark and baselines; post-hoc choices disclosed.

full rationale

This paper is an empirical benchmark study rather than a derivation chain. The central claims are measured paired comparisons against external baselines: ZifaMem versus full raw dialogue history on AlpsBench Task 4, and ZifaMem versus Mem0 and filtered verbatim retrieval under a preregistered protocol. The outcome numbers are observed deltas under a fixed judge protocol, not quantities derived from the model equations. The memory scoring and affect equations (Eqs. 1-7) are component mechanics, and they are not fitted to the A5/A2 endpoints; in fact the affect-dynamics mechanics fail their own preregistered validation gates (Section 5.6), which is the opposite of circular forcing. There are no load-bearing self-citations: the references are external literature, and no uniqueness theorem or prior author result is invoked to constrain the design. The potentially concerning post-hoc elements are explicitly disclosed as post-hoc in the manuscript: the 40/30/30 projection weights were chosen after outcome review and labeled illustrative; the TOST margin was chosen post hoc and labeled as such; the equivalence claim is stated on the preregistered primary endpoint with a post-hoc margin. These are inferential choices, not fitted parameters renamed as predictions. The main validity risk identified by the reader—the A5 endpoint relies on a single DeepSeek-v4 LLM judge with no human-alignment evidence, acknowledged in the Limitations section ('Human evaluation of longitudinal companion responses remains necessary')—is a measurement-validity concern, not circularity. Moreover, the paper reports numerous null or failed primary comparisons (DyAffect's primary comparison against C3, the long-history H1 at L3, affect-layer nulls under memory), which is inconsistent with a story in which results are forced by construction. No circular step can be exhibited from the paper's own text, so the appropriate circularity score is 0.

Assumptions & free parameters 5 free parameters · 5 assumptions · 2 invented entities

The memory system's retrieval weights and promotion threshold are hand-set; the affect estimator has many undisclosed hyperparameters; and the paper's own evidence rejects the latent state machine, so it contributes no independent support. The post-hoc projection weights and TOST margin are additional analysis-level parameters chosen after outcome review.

free parameters (5)
  • Retrieval reranker weights (0.35 semantic, 0.30 strength, 0.35 importance) = 0.35 / 0.30 / 0.35
    Hand-set weights in Eq. (2) directly shape which memories are surfaced; no fitting procedure or ablation is reported.
  • Session-promotion threshold theta=0.5 = 0.5
    Hand-set threshold in Eq. (1) controlling which sessions become episodic memories.
  • DyAffect estimator constants (gmin, gmax, beta, gamma, delta, eta, omega_t) = not reported
    The confidence-adaptive gain and trend-update hyperparameters in Eqs. (5)-(7) are tunable constants; values and tuning procedure are not disclosed.
  • Post-hoc workload weights for Figure 3 projection = 40/30/30
    Persona 40%, content preference 30%, style preference 30%; chosen after outcome review and explicitly labeled post-hoc.
  • TOST equivalence margin for primary endpoint = ±0.05
    The ±5pp equivalence margin was chosen post hoc to match the judge-noise band, not prespecified in the preregistration.
assumptions (5)
  • domain assumption DeepSeek-v4 LLM judge scores on ALPSBench A5 measure emotional intelligence/continuity well enough to support effect claims.
    All headline A5 results and the pooled 11.4% gain depend on this judge. Appendix A bridges two LLM judges only; no human-alignment evidence is provided.
  • domain assumption The public ALPSBench Task 4 dev split is representative of AI-companion workloads and its labels are correct.
    The paper uses only the dev split with test-set gold withheld; results may not transfer to the hidden test set.
  • domain assumption Route-audit records unambiguously identify the generating backbone for every response.
    The evaluation architecture depends on route provenance; a routing error in any admitted cohort would invalidate that cohort's comparisons. The contaminated Gemini cohort shows such risks are real.
  • domain assumption The preregistrations were frozen before data collection and the stated stopping rules were actually followed.
    Preregistration IDs are cited, but the files are not independently verifiable from the paper alone; the paper asserts compliance.
  • standard math Paired bootstrap confidence intervals are valid at the reported sample sizes and across multiple endpoints.
    Used throughout with 10,000 resamples; reasonable, but multiple comparisons and small A5 n=52 are not formally adjusted for.
invented entities (2)
  • DyAffect latent user-affect state (s_t, h_t, omega_t)
    purpose: Estimate a latent emotional state and trend from noisy confidence-labeled observations, to improve response quality over snapshots.
    The paper's own evidence rejects it: C4 loses to the C3 recent-observations baseline (net preference -0.133), so the latent state has no external validation.
  • Legacy companion VAD mood state e(t)
    purpose: Track the companion's appraised valence-arousal-dominance state and inject it into prompts; evaluated but not adopted.
    The deterministic validation suite fails on multiple gates, and the response-level benefit is backbone-dependent and null under memory.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ZifaMem: Structured Memory for Persona, Preference, and Emotional Continuity in AI Companions." pith.science (2026). https://pith.science/paper/OSIZBKGW

@misc{pith2026260717564,
  author       = {Pith},
  title        = {Pith review of: ZifaMem: Structured Memory for Persona, Preference, and Emotional Continuity in AI Companions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OSIZBKGW}},
  note         = {Machine review of arXiv:2607.17564}
}
read the original abstract

AI companions are judged not only by single-turn fluency but by whether they sustain emotional continuity: remembering who the companion is, what the user prefers, and how the relationship has felt. We present ZifaMem, a structured memory system that organizes dialogue into session summaries, episodic memories, and a consolidated user model. Against a deployment-honest comparator that supplies the full raw dialogue history, and under a fixed LLM-as-a-judge protocol with route audits, structured memory raises pooled four-backbone emotional-intelligence scores by 11.4% (95% CI 6.3% to 17.1%), and persona grounding improves on all four backbones (Claude +42% relative). Multi-turn affect context wins a +39% net preference over a single-turn snapshot (exploratory), whereas an additional emotion state machine yields no measurable gain on any of five endpoints. Under an identical preregistered protocol, three memory systems (ZifaMem, Mem0, and filtered verbatim retrieval) each improve significantly over raw-history deployment, and ZifaMem and Mem0 are statistically equivalent within +/-5 points on the preregistered primary preference endpoint. The ZifaMem SDK, CLI, and portable Agent Skills are open-sourced at https://github.com/zifacorp/zifamem.

Figures

Figures reproduced from arXiv: 2607.17564 by the authors.

Figure 1
Figure 1. System decomposition of structured memory and affect continuity. Dialogue branches into [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Legacy affect dynamics under a synthetic event sequence. The audited numerical integrator [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Observed Claude pass-rate gains and an illustrative target-workload projection. Solid [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Paired simple effects of adding the legacy affect layer to raw-history and ZifaMem context [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 6 linked inside Pith

  1. [1]

    My chatbot companion — a study of human-chatbot relationships.International Journal of Human- Computer Studies, 149:102601, 2021

    Marita Skjuve, Asbjørn Følstad, Knut Inge Fostervold, and Petter Bae Brandtzaeg. My chatbot companion — a study of human-chatbot relationships.International Journal of Human- Computer Studies, 149:102601, 2021

    2021

  2. [2]

    My AI friend: How users of a social chatbot understand their human-AI friendship.Human Communication Research, 48(3): 404–429, 2022

    Petter Bae Brandtzaeg, Marita Skjuve, and Asbjørn Følstad. My AI friend: How users of a social chatbot understand their human-AI friendship.Human Communication Research, 48(3): 404–429, 2022

    2022

  3. [3]

    Patil, Ion Stoica, and Joseph E

    Charles Packer, Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G. Patil, Ion Stoica, and Joseph E. Gonzalez. MemGPT: Towards LLMs as operating systems.arXiv preprint arXiv:2310.08560, 2023

    arXiv 2023

  4. [4]

    MemoryBank: Enhancing large language models with long-term memory.arXiv preprint arXiv:2305.10250, 2023

    Wanjun Zhong, Lianghong Guo, Qiqi Gao, He Ye, and Yanlin Wang. MemoryBank: Enhancing large language models with long-term memory.arXiv preprint arXiv:2305.10250, 2023

    arXiv 2023

  5. [5]

    Mem0: Building production-ready AI agents with scalable long-term memory.arXiv preprint arXiv:2504.19413, 2025

    Prateek Chhikara, Dev Khant, Saket Aryan, Taranjeet Singh, and Deshraj Yadav. Mem0: Building production-ready AI agents with scalable long-term memory.arXiv preprint arXiv:2504.19413, 2025

    arXiv 2025

  6. [6]

    Evaluating very long-term conversational memory of LLM agents

    Adyasha Maharana, Dong-Ho Lee, Sergey Tulyakov, Mohit Bansal, Francesco Barbieri, and Yuwei Fang. Evaluating very long-term conversational memory of LLM agents. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, 2024

    2024

  7. [7]

    Long- MemEval: Benchmarking chat assistants on long-term interactive memory

    Di Wu, Hongwei Wang, Wenhao Yu, Yuwei Zhang, Kai-Wei Chang, and Dong Yu. Long- MemEval: Benchmarking chat assistants on long-term interactive memory. InInternational Conference on Learning Representations, 2025

    2025

  8. [8]

    Feelings change: Accounting for individual differences in the temporal dynamics of affect.Journal of Personality and Social Psychology, 99(6):1042–1060, 2010

    Peter Kuppens, Zita Oravecz, and Francis Tuerlinckx. Feelings change: Accounting for individual differences in the temporal dynamics of affect.Journal of Personality and Social Psychology, 99(6):1042–1060, 2010

    2010

Show all 28 references
  1. [9]

    Allen, and Lisa B

    Peter Kuppens, Nicholas B. Allen, and Lisa B. Sheeber. Emotional inertia and psychological maladjustment.Psychological Science, 21(7):984–991, 2010

    2010

  2. [10]

    Clore, and Allan Collins.The Cognitive Structure of Emotions

    Andrew Ortony, Gerald L. Clore, and Allan Collins.The Cognitive Structure of Emotions. Cambridge University Press, 1988

    1988

  3. [11]

    Marsella and Jonathan Gratch

    Stacy C. Marsella and Jonathan Gratch. EMA: A process model of appraisal dynamics.Cognitive Systems Research, 10(1):70–90, 2009

    2009

  4. [12]

    Alpsbench: An LLM personal- ization benchmark for real-dialogue memorization and preference alignment.arXiv preprint arXiv:2603.26680, 2026

    Jianfei Xiao, Xiang Yu, Chengbing Wang, Wuqiang Zheng, Xinyu Lin, Kaining Liu, Hongxun Ding, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He. Alpsbench: An LLM personal- ization benchmark for real-dialogue memorization and preference alignment.arXiv preprint arXiv:2603.266...

    2026 arXiv

  5. [13]

    Wild- chat: 1M chatGPT interaction logs in the wild

    Wenting Zhao, Xiang Ren, Jack Hessel, Claire Cardie, Yejin Choi, and Yuntian Deng. Wild- chat: 1M chatGPT interaction logs in the wild. InInternational Conference on Learning Representations (ICLR), 2024

    2024

  6. [14]

    A-MEM: Agentic memory for LLM agents.arXiv preprint arXiv:2502.12110, 2025

    Wujiang Xu, Zujie Liang, Kai Mei, Hang Gao, Juntao Tan, and Yongfeng Zhang. A-MEM: Agentic memory for LLM agents.arXiv preprint arXiv:2502.12110, 2025

    2025 arXiv

  7. [15]

    HippoRAG: Neurobiologically inspired long-term memory for large language models

    Bernal Jiménez Gutiérrez, Yiheng Shu, Yu Gu, Michihiro Yasunaga, and Yu Su. HippoRAG: Neurobiologically inspired long-term memory for large language models. InAdvances in Neural Information Processing Systems, 2024

    2024

  8. [16]

    Personalizing dialogue agents: I have a dog, do you have pets too? InProceedings of the 56th Annual Meeting of the Association for Computational Linguistics, 2018

    Saizheng Zhang, Emily Dinan, Jack Urbanek, Arthur Szlam, Douwe Kiela, and Jason Weston. Personalizing dialogue agents: I have a dog, do you have pets too? InProceedings of the 56th Annual Meeting of the Association for Computational Linguistics, 2018

    2018

  9. [17]

    RoleLLM: Benchmarking, eliciting, and enhancing role-playing abilities of large language models

    Zekun Moore Wang, Zhongyuan Peng, Haoran Que, Jiaheng Liu, Wangchunshu Zhou, Yuhan Wu, Hongcheng Guo, Ruitong Gan, Zehao Ni, Jian Yang, et al. RoleLLM: Benchmarking, eliciting, and enhancing role-playing abilities of large language models. InFindings of the Association for Com...

    2024

  10. [18]

    CharacterEval: A Chinese benchmark for role-playing conversational agent evaluation

    Quan Tu, Shilong Fan, Zihang Tian, and Rui Yan. CharacterEval: A Chinese benchmark for role-playing conversational agent evaluation. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, 2024

    2024

  11. [19]

    Towards empathetic open-domain conversation models: A new benchmark and dataset

    Hannah Rashkin, Eric Michael Smith, Margaret Li, and Y-Lan Boureau. Towards empathetic open-domain conversation models: A new benchmark and dataset. InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics, 2019

    2019

  12. [20]

    Towards emotional support dialog systems

    Siyang Liu, Chujie Zheng, Orianna Demasi, Sahand Sabour, Yu Li, Zhou Yu, Yong Jiang, and Minlie Huang. Towards emotional support dialog systems. InProceedings of the 59th Annual Meeting of the Association for Computational Linguistics, 2021

    2021

  13. [21]

    James A. Russell. A circumplex model of affect.Journal of Personality and Social Psychology, 39(6):1161–1178, 1980

    1980

  14. [22]

    Pleasure-arousal-dominance: A general framework for describing and measuring individual differences in temperament.Current Psychology, 14:261–292, 1996

    Albert Mehrabian. Pleasure-arousal-dominance: A general framework for describing and measuring individual differences in temperament.Current Psychology, 14:261–292, 1996

    1996

  15. [23]

    Large language models understand and can be enhanced by emotional stimuli.arXiv preprint arXiv:2307.11760, 2023

    Cheng Li, Jindong Wang, Yixuan Zhang, Kaijie Zhu, Wenxin Hou, Jianxun Lian, Fang Luo, Qiang Yang, and Xing Xie. Large language models understand and can be enhanced by emotional stimuli.arXiv preprint arXiv:2307.11760, 2023

    2023 arXiv

  16. [24]

    Xing, Hao Zhang, Joseph E

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging LLM-as-a-judge with MT-Bench and Chatbot Arena. InAdvances in Neural Information Processi...

    2023

  17. [25]

    G- Eval: NLG evaluation using GPT-4 with better human alignment

    Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. G- Eval: NLG evaluation using GPT-4 with better human alignment. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 2023

    2023

  18. [26]

    Large language models are not fair evaluators

    Peiyi Wang, Lei Li, Liang Chen, Zefan Cai, Dawei Zhu, Binghuai Lin, Yunbo Cao, Qi Liu, Tianyu Liu, and Zhifang Sui. Large language models are not fair evaluators. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, 2024

    2024

  19. [27]

    Too human and not human enough: A grounded theory analysis of mental health harms from emotional dependence on the social chatbot Replika.New Media & Society, 2022

    Linnea Laestadius, Andrea Bishop, Michael Gonzalez, Diana Illen ˇcík, and Celeste Campos- Castillo. Too human and not human enough: A grounded theory analysis of mental health harms from emotional dependence on the social chatbot Replika.New Media & Society, 2022

    2022

  20. [28]

    mood mismatch

    Iryna Pentina, Tyler Hancock, and Tian Xie. Exploring relationship development with social chatbots: A mixed-method study of Replika.Computers in Human Behavior, 140:107600, 2023. 17 A Judge Bridge We re-judged 4,096 rows from the independently routed comparison cohort using t...

    2023

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.