Does CoSER (arXiv 2502.09082) include a human evaluation of LLM-generated character dialogue, and was its LLM judge checked against humans?

Provisional lead, not verified from raw bytes. The pinned abstract page (v3, observed 2026-10-22) shows CoSER trains open models that sequentially portray several book characters in a scene and proposes its own evaluation protocol; the abstract does not mention human raters. A condensed read of the full HTML v3 (not quotable, no raw copy because of the batch storage stop) reports a human study in Section 5.3 with a small panel scoring LLM-simulated scenes on an overall 1-10 scale and a table comparing automatic judges with the human results. The rating criterion was not identified as character fidelity, distinguishability or voice, and no agreement statistic was seen. If confirmed from raw bytes, S52's negative narrows further: humans did rate LLM-generated character dialogue, but not as who-said-it attribution. Curator correction 2026-10-24 (discovery audit record C3-B679-01, curator-reviewed C3R-0683; excerpts re-located by independent curator GET 2026-10-24): Problem. Candidate C3-B223-S55-02 (marked provisional, from a condensed read) says the CoSER human study has a small panel, an overall 1-10 scale and no agreement statistic. Conditions. Checked against the raw PDF text of arXiv 2502.09082v3 fetched for this audit. Method. Section 5.3: 60 test samples drawn at random, 7 LLMs, three annotators who each scored 20 samples on a 1-10 scale given the background and the original conversations (so each sample has one human score, no inter-annotator figure); Table 4 gives mean scores and win rates. A separate alignment study (Table 5, appendix C.4) reports the frequency with which LLM judges agree with the human judges on model-pair preferences, excluding near ties: GCA with GPT-4o 68.6%, with DeepSeek-V3 65.1%, with DeepSeek-R1 77.5%, BLEU 75.3%, ROUGE-L 72.0%. So an agreement measure exists, it is a pairwise-preference alignment rather than a correlation or kappa, and BLEU scored higher than the GPT-4o judge. Failure modes. The earlier negative 'no agreement statistic was seen' is wrong; the rating criterion is only 'scoring' of the simulation against the original conversation, not named as character fidelity or voice attribution, so the earlier narrowing of S52 still stands. Limitations. Panel of three, 20 samples each, one scorer per sample; judge and annotators saw the original dialogue. Version. arXiv v3. Curator qualifications: CoSER does run a human scoring study of LLM-generated multi-character simulations of book characters (60 samples, 3 annotators, 20 each, 1-10 scale, with original conversations as reference). It rates overall role-play quality, not speaker attribution; annotators scored disjoint samples, so no inter-annotator agreement is reported. LLM-judge alignment with humans reaches 77.5% (DeepSeek-R1), with BLEU at 75.3%. Scope correction to C3-B217-S52-08 / C3R-0446 and T01.Q50: a human study rating LLM-generated fictional-character performances exists (CoSER v3 section 5.3), but none found tests whether readers can attribute generated lines to the right character. Evidence status: statements located in the cited sources by a correlated desk review (Claude); not executed; not independently reviewed.

{
  "citations": [
    {
      "url": "https://arxiv.org/abs/2502.09082",
      "note": "arXiv 2502.09082v3 HTML section 5.3 Human Evaluation and Table 5"
    }
  ]
}