Pith. sign in

REVIEW 3 major objections 5 minor 45 references

LLMs are much more eager than human experts to decide that an online discussion needs facilitation, and both humans and machines are most confident when they decide not to intervene.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 00:49 UTC pith:7MUE24J3

load-bearing objection The paper's central claim—that LLMs are 'excessively eager' facilitators compared to humans—is undermined by asymmetric survey instructions; the rest is a solid empirical step worth engaging. the 3 major comments →

arxiv 2607.28643 v1 pith:7MUE24J3 submitted 2026-05-27 cs.HC cs.CL

To Facilitate or not to Facilitate: Human and LLM Facilitator Tendencies in Online Discussions

Dimitris Tsirmpas , Katerina Korre , John Pavlopoulos This is my paper
classification cs.HC cs.CL
keywords online discussion facilitationfacilitation timingLLM-as-a-judgehuman vs machine judgmentPEFK datasetModernBertpositive/negative reinforcementwhen to intervene
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper sets out to answer when a facilitator should intervene in an online discussion, a question prior work left open. It unifies the scattered facilitation datasets into PEFK, then asks 10 expert facilitators and six open LLMs to rate 1,224 six-comment excerpts. The headline result is that humans are cautious—usually choosing no intervention—while LLMs over-intervene, reading neutral narrative exchanges as needing positive reinforcement. The authors also show that deciding not to facilitate is the easy case for both humans and machines, and that traditional encoder classifiers trained on the aggregated data are more reliable than out-of-the-box LLMs, though the way 'facilitative intervention' is currently labeled caps how well any model can do.

Core claim

The paper claims that when to facilitate differs systematically between humans and LLMs: across 1,224 discussion chunks, expert human raters mostly selected 'no reinforcement,' while six open-source LLMs produced a majority-facilitate label in a far larger share of cases, and often interpreted neutral, narrative exchanges as opportunities for positive reinforcement. At the same time, both groups were more certain when they judged facilitation unnecessary than when they judged it necessary. On the predictive side, the paper claims that a fine-tuned ModernBert classifier trained on the newly aggregated PEFK corpus is more consistent and generally outperforms out-of-the-box LLMs at predicting w

What carries the argument

The central machinery is PEFK (Prosocial and Effective Facilitation in Konversations), a standardized corpus that aggregates nine prior facilitation datasets into a single schema, together with a survey instrument that asks raters to give 1–5 certainty scores for positive reinforcement, negative reinforcement, and no reinforcement. The comparison rests on the 'positive vs negative reinforcement' distinction—introduced to keep raters from defaulting to content moderation—and on the transformation of those ordinal scores into a multiclass label. PEFK is what lets the authors transfer findings from oral, professionally facilitated debates to written online settings and to train ModernBert class

Load-bearing premise

The human-versus-LLM comparison assumes that the ten experts' 1–5 ratings on six-comment excerpts, collected without inter-annotator agreement and under a prompt that warns against over-intervention, reflect how humans would actually facilitate in real discussions.

What would settle it

Re-run the survey with the same experts on full discussion threads instead of six-comment excerpts, or with prompts that do not warn against over-intervention; if human intervention rates rise to LLM levels, the 'humans are cautious' claim collapses. A behavioral experiment that tracks whether expert raters actually intervene when they say they would would directly test the self-report assumption.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Autonomous facilitation systems built on out-of-the-box LLMs would over-intervene relative to expert human judgment, potentially irritating participants in healthy discussions.
  • 'No intervention needed' is the reliable prediction; systems should be tuned conservatively, and high-certainty non-intervention can serve as a filtering decision.
  • Encoder-only classifiers trained on aggregated facilitation data are a more dependable building block for automation than generic LLM judges, especially because the decision threshold can be tuned.
  • Existing facilitation datasets, labeled by professional-facilitator speech, cap attainable accuracy; better automation requires new corpora that annotate genuinely facilitative comments (Definition 3), despite subjectivity.
  • The PEFK corpus allows cross-domain transfer from oral debates to written online facilitation, but the low ceiling warns against expecting large gains without better data.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the human-LLM gap may be partly an artifact of the survey framing, which explicitly warns that too many interventions annoy participants; a differently-framed prompt could shift human ratings toward the LLM side, so deployment caution is warranted.
  • Editorial inference: the finding that LLMs stay consistent among themselves even when humans disagree suggests a human-in-the-loop scheme—use LLM majority voting to flag 'easy' no-facilitation cases and route only ambiguous chunks to human facilitators—a testable extension not explored in the paper.
  • Editorial inference: if the Definition-2 ceiling is real, future dataset construction should shift toward Definition-3 consensus annotations, perhaps with a two-stage design where LLMs propose and experts arbitrate, which the paper's own limitations hint at.
  • Editorial inference: the paper's cross-domain transfer assumption—that oral professionally facilitated discussions teach us about written online forums—could be probed directly by measuring whether classifier errors concentrate on the written subsets; the paper does not report this breakdown explicitly.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper addresses when human and LLM facilitators decide to intervene in online discussions. It introduces PEFK, a standardized aggregation of existing facilitation datasets, and runs a survey in which 10 expert participants and six open-source LLMs rate six-comment discussion chunks on whether positive, negative, or no reinforcement is needed. The authors report that humans are generally cautious, while LLMs are excessively eager to intervene, and that both are more certain when judging that no intervention is needed. They then train ModernBERT classifiers and prompt the same LLMs on a binary 'would a facilitator intervene' task, concluding that the encoder-only classifiers are more reliable than out-of-the-box LLMs and that current Def. 2 (professional-facilitation) datasets impose a low performance ceiling.

Significance. If the central empirical claim holds, the finding that six widely used open LLMs over-intervene relative to expert human judgment is practically important for any future deployment of automated facilitation. The paper also makes a useful contribution by assembling PEFK, a multi-domain standardized dataset, and by being transparent about code release, model versions, and prompt text. The use of duplicated items to measure intra-participant consistency is a good practice. However, the headline human-versus-LLM comparison is currently vulnerable to an internal-validity confound in the survey instructions, and the quantitative comparisons in the second half of the paper lack error bars or significance tests. Both issues are fixable but are load-bearing for the main claims.

major comments (3)
  1. [§4.1, App. E.1 vs E.2.1/E.2.2] The human and LLM survey instruments are not symmetric. The human guidelines in App. E.1 explicitly instruct: 'Keep in mind that too many interventions on the moderator's part may result in an adverse reaction from the participants; no one wants to be constantly told what to say or how to say it.' The LLM prompts in App. E.2.1 and E.2.2 contain no equivalent cautionary statement, although they adopt the rest of the human guidelines. The large difference in intervention rates in Fig. 2 can therefore be explained by asymmetric instruction alone. This is an internal-validity problem for Findings 1 and 2 and the headline claim, not merely an ecological-validity caveat. I ask the authors to rerun the LLM survey with the same cautionary instruction (and/or run the human survey without it) and report whether the gap persists.
  2. [Tables 4 and 5] Finding 4 states that ModernBert classifiers are 'more reliable' and 'outperform' LLMs, but Tables 4-5 report single precision/recall/F1 point estimates with no confidence intervals, no paired tests, and no correction for evaluating six LLMs on the same test instances. For example, on the All datasets row, ModBert F1p=0.478 vs LLaMa8B F1p=0.425 is a difference that could easily be within sampling noise. I request bootstrap CIs and/or McNemar or paired bootstrap tests over the 2,000 sampled instances, plus a statement of how many random test samples were used. Without this, the reliability comparison in §5.2 and the conclusion are not quantitatively supported.
  3. [§4.1] Of 1,224 sampled chunks, only 721 remained after participants flagged excerpts as 'malformed' — a 41% drop. The paper does not report how malformed flags were distributed across datasets, participants, or eventual labels. If malformedness correlates with, e.g., oral-vs-written source, escalation, or class distribution, the survey estimates are biased. Please provide a robustness analysis, at minimum per-dataset malformed rates and a comparison of label distributions with and without malformed items, or an argument that the flagging is label-independent.
minor comments (5)
  1. [§6] 'an survey' should be 'a survey'.
  2. [App. E.2.3] Typo: 'intrevene' should be 'intervene'.
  3. [Ethical Considerations] Typo: 'unncessary' should be 'unnecessary'.
  4. [App. B.3] 'We also the data to include only messages...' is missing a verb; probably 'We also filtered/processed the data...'.
  5. [App. D.3] The label-transformation table (Table 10) is clear, but the threshold '>=3' should specify whether it applies to each of the three ordinal scales independently and what happens with ties in the original 1-5 values.

Circularity Check

0 steps flagged

No significant circularity; the main derivation chain is self-contained, with only minor non-load-bearing self-citations and a prompt-asymmetry confound that is a validity issue, not a circular reduction.

full rationale

The paper's central empirical contributions are new measurements: a 10-expert survey on facilitation timing, six LLMs run on the same survey, and ModernBERT/LLM comparisons on PEFK. These results are not derived from the claims being tested. PEFK is a compilation of external datasets listed in the authors' own prior survey (Korre et al. 2025), but the underlying data are independent and the aggregation is transparent; the self-citation is descriptive, not load-bearing. The prior claim that LLMs over-intervene (Tsirmpas et al. 2026) is cited for motivation and corroboration, but the paper reproduces the pattern with its own survey (Fig. 2) and classifier experiments (Table 4), so the citation is not the argument. Finding 5 (Def. 2 limits performance) is an empirically supported commentary on label noise, not a tautology. The most serious threat to the headline comparison is the asymmetric survey instruments: the human guidelines include a caution against over-intervention (App. E.1) that is absent from the LLM prompts (App. E.2.1/E.2.2). This is an internal-validity confound that may inflate the apparent human/LLM gap, but it is not a circular derivation: the paper does not define 'human caution' in terms of the instruction, and the LLM eagerness and classifier results stand independently. Thus no circular step is exhibited; the paper's derivation is self-contained.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

No invented entities: PEFK is an aggregation of existing datasets, and the positive/negative reinforcement axis is explicitly disclaimed as 'not a theoretical contribution' (§4.1). The free parameters are all label-construction or experimental-design choices: the >=3 binning threshold, UMOD 0.75 binarization, CMV 60th-percentile escalation cut, K=3 context window, and an unreported class weight. The load-bearing domain assumptions are the Def. 2 label, the validity of expert self-report as ground truth, the exhaustiveness of the three-definition taxonomy, and the sufficiency of truncated excerpts — the last explicitly acknowledged as a limitation.

free parameters (5)
  • Ordinal-to-multiclass binning threshold = 3 (on 1-5 scale)
    Survey labels are binarized using >=3 to flag positive/negative/no reinforcement (App. D.3, Table 10); hand-chosen, directly shapes every reported class distribution.
  • UMOD facilitation threshold = 0.75
    UMOD gradient labels are binarized at 0.75 with >50% participant agreement (App. B.1); determines which comments count as facilitative in that PEFK subset.
  • CMV-AWRY escalation percentile = 60th upper percentile
    Discussions are marked escalated when the derailment value is in the 60th upper percentile (App. B.1); arbitrary cut affecting the Def. 1 subset of PEFK.
  • Context window K = 3 preceding comments
    Every instance uses up to K=3 preceding comments along the reply chain (C.2.1); load-bearing for both the classifier and LLM intervention tasks.
  • BCE positive-class weight = not reported (blank in text)
    The class-imbalance weight in C.2.1 reads 'the positive class is weighted as' with no value given; affects trained classifier behavior and cannot be replicated as specified.
axioms (4)
  • domain assumption Definition 2: all comments written by a professional facilitator are facilitative interventions
    Adopted from Falk et al. (2021) and used as the label basis for all Section 5 experiments; the paper itself notes the risk of classifying non-facilitative comments as facilitative (§2.1, Table 1).
  • domain assumption Expert self-report on six-comment excerpts measures human facilitation tendency
    The survey replaces ground truth with individual judgment by design ('survey-based approach', §4.1); no IAA and no validation against actual intervention outcomes.
  • domain assumption The three-definition taxonomy of facilitative interventions is exhaustive
    The framing rests on the literature-derived partition into escalation, professional facilitation, and human judgment (§2.1); other possible definitions (e.g., outcome-based discussion quality) are not considered.
  • ad hoc to paper Truncated six-comment chunks preserve intervention-relevant context
    Chunks are presented without the full thread; the authors acknowledge in Limitations that 'long-term interaction dynamics, participant history, community norms, and evolving conversational context... may not be fully captured in the excerpts.'

pith-pipeline@v1.3.0-alltime-deepseek · 22797 in / 15848 out tokens · 141886 ms · 2026-08-03T00:49:24.405144+00:00 · methodology

0 comments
read the original abstract

Automating facilitation in online discussions is a long-standing social concern given the increasing time we spend on online spaces and the failure of content moderation approaches. While studies have been conducted on how to facilitate, none have answered the essential question of when to do so. A potential answer is using LLMs, which ostensibly make automated, large-scale intervention increasingly feasible. In this study, we examine when LLMs decide to facilitate by defining what facilitation is, observing when humans decide to facilitate, and comparing their decisions with those made by LLMs. To this end, we create PEFK, a corpus standardizing and aggregating all relevant facilitation datasets. We are the first to run a survey on facilitation timing, which we execute using expert facilitative participants and LLM-as-a-judge models. We discover that while humans are more cautious, LLMs are excessively eager to facilitate, although both are more certain when judging that facilitation is not needed. We then investigate whether this behavior can be corrected using alternative setups for LLMs and training ModernBert classifiers on established datasets, finding that the latter perform more reliably than the former, although current datasets impose a relatively low performance ceiling.

Figures

Figures reproduced from arXiv: 2607.28643 by Dimitris Tsirmpas, John Pavlopoulos, Katerina Korre.

Figure 1
Figure 1. Figure 1: At any point in a discussion, the facilitator [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Certainty of judgments per facilitation label. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Top 5 discussion instances where the LLM [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Percentage of comments made by professional [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Left: Comments per discussion on the PEFK dataset. Note the significant increase in comments made in transcribed datasets such as Fora and IQ2. This deviation is explained by the fact that a single human utterance is regularly broken up to many comments in order to simulate interruptions or participants talking over each other. Right: Number of words per comment in all discussions in PEFK, grouped by origi… view at source ↗
Figure 7
Figure 7. Figure 7: Most common moderator bigrams in PEFK per dataset. Function words are excluded. get, as well as scalable deployment in real-world set￾tings (Tsirmpas et al., 2026). It is also worth noting that since a part of our dataset (Fora) is not licensed, we can not risk leaking the data contained within to proprietary models. Generally, the use of local models protects the personal data of the users contained in ou… view at source ↗
Figure 8
Figure 8. Figure 8: Classifier performance (left: written datasets, right: oral datasets) when tuning the decision threshold on the test-set. The higher the threshold, the more conservative the classifier is encouraged to become when predicting a facilitator intervention. and analysis of facilitation strategies. D.3 Label Transformation To efficiently calculate intra-participant consistency, we transform the fine-grained, ord… view at source ↗
Figure 9
Figure 9. Figure 9: Label distributions of human participants per [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

45 extracted references · 9 canonical work pages

  1. [1]

    WikiConv: A Corpus of the Complete Conversational History of a Large Online Collaborative Community

    Hua, Yiqing and Danescu-Niculescu-Mizil, Cristian and Taraborelli, Dario and Thain, Nithum and Sorensen, Jeffery and Dixon, Lucas. WikiConv: A Corpus of the Complete Conversational History of a Large Online Collaborative Community. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. 2018. doi:10.18653/v1/D18-1305

    work page doi:10.18653/v1/d18-1305 2018

  2. [2]

    Moderation in the Wild: Investigating User-Driven Moderation in Online Discussions

    Falk, Neele and Vecchi, Eva and Jundi, Iman and Lapesa, Gabriella. Moderation in the Wild: Investigating User-Driven Moderation in Online Discussions. Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers). 2024

    2024

  3. [3]

    I Beg to Differ: A study of constructive disagreement in online conversations

    De Kock, Christine and Vlachos, Andreas. I Beg to Differ: A study of constructive disagreement in online conversations. Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume. 2021. doi:10.18653/v1/2021.eacl-main.173

    work page doi:10.18653/v1/2021.eacl-main.173 2021

  4. [4]

    How to disagree well: Investigating the dispute tactics used on Wikipedia

    De Kock, Christine and Stafford, Tom and Vlachos, Andreas. How to disagree well: Investigating the dispute tactics used on Wikipedia. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. 2022. doi:10.18653/v1/2022.emnlp-main.252

    work page doi:10.18653/v1/2022.emnlp-main.252 2022

  5. [5]

    Karadzhov, Georgi and Stafford, Tom and Vlachos, Andreas , title =. Proc. ACM Hum.-Comput. Interact. , month = oct, articleno =. 2023 , issue_date =. doi:10.1145/3610056 , abstract =

    doi:10.1145/3610056 2023

  6. [6]

    2025 , eprint=

    Scalable Evaluation of Online Facilitation Strategies via Synthetic Simulation of Discussions , author=. 2025 , eprint=

    2025

  7. [7]

    and Danescu-Niculescu-Mizil, Cristian

    Chang, Jonathan P. and Danescu-Niculescu-Mizil, Cristian. Trouble on the Horizon: Forecasting the Derailment of Online Conversations as they Develop. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019. doi:10.18653/v1/D19-1481

    doi:10.18653/v1/d19-1481 2019

  8. [8]

    Facilitative moderation for online participation in eRulemaking , year =

    Park, Joonsuk and Klingel, Sally and Cardie, Claire and Newhart, Mary and Farina, Cynthia and Vallb\'. Facilitative moderation for online participation in eRulemaking , year =. Proceedings of the 13th Annual International Conference on Digital Government Research , pages =. doi:10.1145/2307729.2307757 , abstract =

    arXiv

  9. [9]

    Conversational Flow in O xford-style Debates

    Zhang, Justine and Kumar, Ravi and Ravi, Sujith and Danescu-Niculescu-Mizil, Cristian. Conversational Flow in O xford-style Debates. Proceedings of the 2016 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies. 2016. doi:10.18653/v1/N16-1017

    doi:10.18653/v1/n16-1017 2016

  10. [10]

    Fora: A corpus and framework for the study of facilitated dialogue

    Schroeder, Hope and Roy, Deb and Kabbara, Jad. Fora: A corpus and framework for the study of facilitated dialogue. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024. doi:10.18653/v1/2024.acl-long.754

    doi:10.18653/v1/2024.acl-long.754 2024

  11. [11]

    WH o W : A Cross-domain Approach for Analysing Conversation Moderation

    Chen, Ming-Bin and Frermann, Lea and Lau, Jey Han. WH o W : A Cross-domain Approach for Analysing Conversation Moderation. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025. doi:10.18653/v1/2025.naacl-long.105

    work page doi:10.18653/v1/2025.naacl-long.105 2025

  12. [12]

    Evaluation and Facilitation of Online Discussions in the LLM Era: A Survey

    Korre, Katerina and Tsirmpas, Dimitris and Gkoumas, Nikos and Cabal \'e , Emma and Myrtzani, Danai and Evgeniou, Theodoros and Androutsopoulos, Ion and Pavlopoulos, John. Evaluation and Facilitation of Online Discussions in the LLM Era: A Survey. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025. doi:10.18653/v1/...

    work page doi:10.18653/v1/2025.emnlp-main.1243 2025

  13. [13]

    Moderation Matters: Measuring Conversational Moderation Impact in E nglish as a Second Language Group Discussion

    Gao, Rena and Chen, Ming-Bin and Frermann, Lea and Lau, Jey Han. Moderation Matters: Measuring Conversational Moderation Impact in E nglish as a Second Language Group Discussion. Findings of the Association for Computational Linguistics: ACL 2025. 2025. doi:10.18653/v1/2025.findings-acl.106

    work page doi:10.18653/v1/2025.findings-acl.106 2025

  14. [14]

    2026 , eprint=

    Designing Synthetic Discussion Generation Systems: A Case Study for Online Facilitation , author=. 2026 , eprint=

    2026

  15. [15]

    and Gollakota, Shyamnath

    Chen, Tuochao and Batchelder, Nicholas Scott and Liu, Alisa and Smith, Noah A. and Gollakota, Shyamnath. L lama PIE : Proactive In-Ear Conversation Assistants. Findings of the Association for Computational Linguistics: ACL 2025. 2025. doi:10.18653/v1/2025.findings-acl.710

    work page doi:10.18653/v1/2025.findings-acl.710 2025

  16. [16]

    Annotation alignment: Comparing LLM and human annotations of conversational safety

    Movva, Rajiv and Koh, Pang Wei and Pierson, Emma. Annotation alignment: Comparing LLM and human annotations of conversational safety. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024. doi:10.18653/v1/2024.emnlp-main.511

    doi:10.18653/v1/2024.emnlp-main.511 2024

  17. [17]

    Frontiers in Artificial Intelligence , VOLUME=

    Nonomura, Ryota and Mori, Hiroki , TITLE=. Frontiers in Artificial Intelligence , VOLUME=. 2025 , URL=. doi:10.3389/frai.2025.1582287 , ISSN=

    arXiv 2025

  18. [18]

    Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems , articleno =

    Schaffner, Brennan and Bhagoji, Arjun Nitin and Cheng, Siyuan and Mei, Jacqueline and Shen, Jay L and Wang, Grace and Chetty, Marshini and Feamster, Nick and Lakier, Genevieve and Tan, Chenhao , title =. Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems , articleno =. 2024 , isbn =. doi:10.1145/3613904.3642333 , abstract =

    arXiv 2024

  19. [19]

    ArXiv , year=

    Opportunities and Risks of LLMs for Scalable Deliberation with Polis , author=. ArXiv , year=

  20. [20]

    Proceedings of the International AAAI Conference on Web and Social Media , author=

    Conversation Modeling to Predict Derailment , volume=. Proceedings of the International AAAI Conference on Web and Social Media , author=. 2023 , month=. doi:10.1609/icwsm.v17i1.22200 , abstractNote=

    work page doi:10.1609/icwsm.v17i1.22200 2023

  21. [21]

    Perspective API. 2026

    2026

  22. [22]

    Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act) , author =. 2024

    2024

  23. [23]

    Communication Research Reports , volume=

    (Why) does comment presentation order matter for the effects of user comments? Assessing the role of the availability heuristic and the bandwagon heuristic , author=. Communication Research Reports , volume=. 2021 , publisher=

    2021

  24. [24]

    Government Information Quarterly , volume=

    Crowdsourcing civility: A natural experiment examining the effects of distributed moderation in online forums , author=. Government Information Quarterly , volume=. 2014 , publisher=

    2014

  25. [25]

    Journal of Educational computing research , volume=

    How habitual online practices affect the development of asynchronous discussion threads , author=. Journal of Educational computing research , volume=. 2003 , publisher=

    2003

  26. [26]

    Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems , pages=

    Supporting comment moderators in identifying high quality online news comments , author=. Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems , pages=

    2016

  27. [27]

    The Journal of Social Psychology , volume=

    Primacy versus recency in retention of information and opinion change , author=. The Journal of Social Psychology , volume=. 1977 , publisher=

    1977

  28. [28]

    International Journal of Human-Computer Studies , volume=

    The effects of interaction sequencing on user experience and preference , author=. International Journal of Human-Computer Studies , volume=. 2017 , publisher=

    2017

  29. [29]

    Smarter, Better, Faster, Longer: A Modern Bidirectional Encoder for Fast, Memory Efficient, and Long Context Finetuning and Inference

    Warner, Benjamin and Chaffin, Antoine and Clavi. Smarter, Better, Faster, Longer: A Modern Bidirectional Encoder for Fast, Memory Efficient, and Long Context Finetuning and Inference. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025. doi:10.18653/v1/2025.acl-long.127

    doi:10.18653/v1/2025.acl-long.127 2025

  30. [30]

    Political Analysis , author=

    Synthetic Replacements for Human Survey Data? The Perils of Large Language Models , volume=. Political Analysis , author=. 2024 , pages=. doi:10.1017/pan.2024.5 , number=

    doi:10.1017/pan.2024.5 2024

  31. [31]

    Proceedings of the 33rd ACM Conference on Hypertext and Social Media , pages =

    Cresci, Stefano and Trujillo, Amaury and Fagni, Tiziano , title =. Proceedings of the 33rd ACM Conference on Hypertext and Social Media , pages =. 2022 , isbn =. doi:10.1145/3511095.3536369 , abstract =

    arXiv 2022

  32. [32]

    Trujillo, Amaury and Cresci, Stefano , title =. Proc. ACM Hum.-Comput. Interact. , month = nov, articleno =. 2022 , issue_date =. doi:10.1145/3555639 , abstract =

    doi:10.1145/3555639 2022

  33. [33]

    and Lopez-Lopez, Ezequiel and Hechtlinger, Shahar and Rahwan, Zoe and Aeschbach, Samuel and Bakker, Michiel A

    Burton, Jason W. and Lopez-Lopez, Ezequiel and Hechtlinger, Shahar and Rahwan, Zoe and Aeschbach, Samuel and Bakker, Michiel A. and Becker, Joshua A. and Berditchevskaia, Aleks and Berger, Julian and Brinkmann, Levin and Flek, Lucie and Herzog, Stefan M. and Huang, Saffron and Kapoor, Sayash and Narayanan, Arvind and Nussberger, Anne-Marie and Yasseri, Ta...

  34. [34]

    Sociologica , author=

    The Problems of LLM-generated Data in Social Science Research , volume=. Sociologica , author=. 2024 , month=. doi:10.6092/issn.1971-8853/19576 , abstractNote=

    doi:10.6092/issn.1971-8853/19576 2024

  35. [35]

    Predicting Moderation of Deliberative Arguments: Is Argument Quality the Key?

    Falk, Neele and Jundi, Iman and Vecchi, Eva Maria and Lapesa, Gabriella. Predicting Moderation of Deliberative Arguments: Is Argument Quality the Key?. Proceedings of the 8th Workshop on Argument Mining. 2021. doi:10.18653/v1/2021.argmining-1.13

    work page doi:10.18653/v1/2021.argmining-1.13 2021

  36. [36]

    Seering, Joseph , title =. Proc. ACM Hum.-Comput. Interact. , month = oct, articleno =. 2020 , issue_date =. doi:10.1145/3415178 , abstract =

    doi:10.1145/3415178 2020

  37. [37]

    Social Media + Society , year=

    The Civic Labor of Volunteer Moderators Online , author=. Social Media + Society , year=

  38. [38]

    Analysis of Facilitators' Behaviors in Multi-party Conversations for Constructing a Digital Facilitator System

    Shiota, Tsukasa and Yamamura, Takashi and Shimada, Kazutaka. Analysis of Facilitators' Behaviors in Multi-party Conversations for Constructing a Digital Facilitator System. Collaboration Technologies and Social Computing. 2018

    2018

  39. [39]

    and Cicchetti, Domenic V

    Feinstein, Alvan R. and Cicchetti, Domenic V. , title =. Journal of Clinical Epidemiology , year =

  40. [40]

    Two Contrasting Data Annotation Paradigms for Subjective NLP Tasks

    R. Two Contrasting Data Annotation Paradigms for Subjective NLP Tasks. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022. doi:10.18653/v1/2022.naacl-main.13

    doi:10.18653/v1/2022.naacl-main.13 2022

  41. [41]

    Vinay Koshy and Frederick Choi and Yi. Venire:. CoRR , volume =. 2024 , url =. doi:10.48550/ARXIV.2410.23448 , eprinttype =. 2410.23448 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2410.23448 2024

  42. [42]

    arXiv preprint arXiv:2501.18645 , year=

    Layered chain-of-thought prompting for multi-agent llm systems: A comprehensive approach to explainable large language models , author=. arXiv preprint arXiv:2501.18645 , year=

    Pith/arXiv arXiv

  43. [43]

    Preprint, alphaXiv , pages=

    Chain-of-thought is not explainability , author=. Preprint, alphaXiv , pages=

  44. [44]

    Advances in Neural Information Processing Systems , volume=

    Language models don't always say what they think: Unfaithful explanations in chain-of-thought prompting , author=. Advances in Neural Information Processing Systems , volume=

  45. [45]

    plausibility: On the (un) reliability of explanations from large language models , author=

    Faithfulness vs. plausibility: On the (un) reliability of explanations from large language models , author=. arXiv preprint arXiv:2402.04614 , year=

    Pith/arXiv arXiv