acceptodds
Under review as a conference paper at ICLR 2027

AnthroDial: Benchmarking LLM Anthropomorphism in Autonomous Social Interaction

Abstract

Large language models (LLMs) are increasingly deployed as social agents, yet credible human-like interaction requires more than fluent responses or persona consistency. Agents must autonomously decide whether, when, and how to communicate while adapting to evolving contexts, goals, and relationships. Existing research, however, lacks a unified approach to enabling, evaluating, and improving such capabilities in continuous, open-ended interaction. We introduce AnthroDial, a unified framework for developing anthropomorphic social agents from three complementary aspects: MindFlow, a lightweight interaction harness that enables autonomous, asynchronous, and adaptive communication through a dynamic Mind Buffer; CAPS-Eval, a theory-grounded framework for evaluating cognitive, affective, and behavioral dimensions of anthropomorphic interaction; and a scalable training paradigm that combines SEEDS for environment expansion with DiAPO for adaptive capability optimization. We further construct evaluation datasets covering everyday communication, game interaction, and long-horizon character interaction. Extensive experiments across diverse models and scenarios demonstrate improved interaction autonomy and naturalness, validate CAPS-Eval's reliability, discriminativeness, and agreement with human rankings, and confirm the effectiveness of our training paradigm. Together, these components provide a unified framework for developing credible human-like social agents in open-ended interaction. Code is available at https://anonymous.4open.science/r/AnthroDial-2650.

Primary area:
datasets and benchmarks
Keywords:
Large Language Models, Anthropomorphism, Benchmarking, Social Interaction

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_NW6EOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces AnthroDial, a framework for developing and evaluating anthropomorphic social agents in continuous, open-ended interaction. It contains four main components: (1) MindFlow, an event-driven interaction harness with a private Mind Buffer that allows agents to autonomously schedule, revise, cancel, and send messages rather than following strict turn-taking; (2) CAPS-Eval, a theory-motivated hierarchical evaluation framework covering validity constraints, per-turn behavior, and holistic interaction quality; (3) SEEDS, a persona–scenario expansion pipeline for synthesizing training environments; and (4) DiAPO, an adaptive policy optimization method that reweights reward dimensions according to estimated capability deficits and task–capability matching. The authors evaluate several frontier and open-weight models on Everyday Chat, Game Interaction, and Long-Horizon Character scenarios, and train Qwen3.5-9B using SEEDS-based SFT followed by DiAPO. The paper reports substantial improvements under CAPS-Eval, human preference for MindFlow over conventional turn-based interaction, and reasonable agreement between automatic judges and human rankings.

Critical strengths
  1. MindFlow provides a useful interaction abstraction beyond strict turn-taking. The paper correctly identifies that realistic social interaction involves not only deciding what to say, but also whether and when to communicate. The private draft buffer and event-driven virtual clock provide a simple and interpretable mechanism for delayed responses, proactive follow-ups, message revision/cancellation, and autonomous conversational closure. I found this formulation useful, and the human comparison against a turn-based harness provides some evidence that the interaction scaffold matters.
  2. The paper is unusually complete as an end-to-end system and provides extensive diagnostics. The work connects interaction generation, multidimensional evaluation, data synthesis, SFT, and RL, and evaluates several model families across three settings. The authors also include detailed rubrics, ablations of DiAPO components, and judge-robustness checks. These make the empirical artifact more useful than a benchmark consisting only of aggregate leaderboard scores.
Critical weaknesses
  1. I do not yet see a sufficiently well-isolated scientific contribution beyond the integration of several reasonable components. The paper presents MindFlow, CAPS-Eval, SEEDS, and DiAPO as a unified framework, but it is difficult to identify the central scientific result that should be taken away from the work. MindFlow is a plausible event-driven social interaction harness; CAPS-Eval is an expert-designed evaluation rubric; SEEDS is a synthetic persona/scenario expansion pipeline; and DiAPO is an adaptive reward-weighting/curriculum mechanism based on current capability estimates. Each component is reasonable, but the current experiments mainly establish that the complete system can be built and optimized, rather than revealing a new phenomenon about LLM social behavior or establishing a substantially new learning principle. In particular, I would have found the paper considerably more compelling if MindFlow were used as an experimental instrument to establish a clear result such as: once models must autonomously decide whether and when to communicate, frontier LLMs exhibit systematic failure modes that are hidden by turn-based evaluation. The current paper instead moves quickly from introducing the interaction scaffold to building an evaluation and training pipeline around it. Similarly, for DiAPO, the paper does not sufficiently establish why the proposed ZPD-inspired weighting is fundamentally preferable to simpler adaptive weighting, hard-example mining, curriculum learning, or inverse-performance weighting. As a result, the work currently reads more as a well-engineered unified system than as a focused ICLR contribution.
  2. The training and evaluation objectives are too tightly coupled to support the strongest claims about improved anthropomorphic capability. CAPS-Eval is not only the final evaluation framework: the same evaluation dimensions are used to filter synthesized trajectories for SFT, and they directly provide the dimension-level reward signal for DiAPO. The appendix further states that the expert-designed evaluation dimensions are shared across construction, trajectory filtering, and downstream evaluation. This creates a substantial risk that the very large post-training gains primarily demonstrate optimization toward CAPS-Eval rather than a general improvement in human-like social interaction. This concern is especially important because the reported gains are extremely large (e.g., the Qwen3.5-9B Long-Horizon Character Score rises from 0.212 to 0.728 after SFT and to 0.993 after DiAPO). Judge agreement and the MindFlow-vs-turn-based human study are useful sanity checks, but they do not provide an independent evaluation of whether the trained model generalizes to human-likeness criteria that were not involved in data filtering or reward construction. For a submission whose primary area is datasets and benchmarks, I would expect substantially stronger evidence of construct validity and out-of-distribution validity, such as blinded human evaluation of the trained checkpoints, evaluation on independently designed third-party benchmarks/rubrics, or transfer to real human interaction data not generated or selected through CAPS-Eval.
Questions for authors
  1. Independent evaluation of the trained models. Can the authors evaluate the Base, SFT, GRPO, and DiAPO checkpoints using a genuinely independent evaluation process whose criteria were not used for SFT filtering or RL reward construction? For example, this could be a sufficiently powered blinded human evaluation, an independently designed rubric, or an external social/anthropomorphism benchmark. A result showing that the large DiAPO gains persist under such an evaluation would materially change my assessment.
  2. What scientific phenomenon does MindFlow reveal that cannot be observed under turn-based interaction? Can the authors provide controlled experiments that separately manipulate autonomous initiation, response delay, follow-up without an incoming message, draft revision/cancellation, and autonomous termination? I would particularly like to know whether allowing models to control these dimensions exposes systematic failure modes of frontier models, rather than only improving perceived naturalness.
  3. What part of DiAPO requires the proposed ZPD formulation? Can the authors compare DiAPO against strong simpler adaptive baselines, such as inverse-current-performance weighting, uncertainty-based weighting, hard-example mining, or standard curriculum sampling, under the same rollout and compute budget? If the ZPD matching term provides a robust advantage over these alternatives, that would strengthen the algorithmic contribution considerably.
8 comments
AuthorsOpenReview ↗

Response to Reviewer NW6E (1/5)

We thank the reviewer for the detailed and constructive feedback. We address each concern below.

W1: I do not yet see a sufficiently well-isolated scientific ... as a focused ICLR contribution.

Response:

Regarding the concern that the scientific contribution of each component is not sufficiently isolated, we describe the design rationale and the validation evidence for MindFlow and DiAPO separately.

1. MindFlow: The main text already provides targeted experimental analyses of MindFlow: (i). Basic capabilities that small models lack. §5.2 shows that for weaker models such as Qwen3.5-9B, a large per-turn–holistic gap indicates that structural failures such as repetition, malformed output, and AI-identity leakage remain a major bottleneck. (ii). Difficulties shared by large models. As shown in §5.2 and Figure 3, under MindFlow the problems concentrate on four dimensions: context consistency (E1, down to 0.770.77 for Qwen3.5-397B on Everyday Chat and 0.590.59 on Long-Horizon Character), scenario–relationship fit (E2, also degrading on Long-Horizon Character), unpunctuated chat style (D11, at most 0.540.54 even for the strongest models), and fragmentation and immediacy (E8, markedly lower on Game Interaction).

To show more fully that MindFlow exposes model problems, we sample four models and measure the failure rate of every dimension (4 models ×\times 50 scenarios ×\times both roles A/B = 400 cases in theory, 397 valid cases after removing L0-fatal cases; a case counts as one failure on a dimension if any checklist item of that dimension is not satisfied; the models are GPT-5.5, DeepSeek-V4-Flash, Gemini-3.5 Flash, and Qwen3.5-397B-A17B).

Comparing MindFlow with the turn-based baseline, MindFlow substantially lowers the failure rate of most dimensions, in particular persona expression (D3, 46.10% →\rightarrow 3.53%), segmentation and rhythm (D5, 30.98% →\rightarrow 2.27%), chat naturalness (D1, 26.45% →\rightarrow 1.26%), and knowledge and action boundaries (D4, 22.67% →\rightarrow 0.76%). The mechanism requires the model itself to take responsibility for turn-taking, pacing, and persona consistency instead of letting a fixed turn-taking protocol absorb them, so failures caused by forced immediate replies are largely eliminated.

We nevertheless find that models remain weak on D2 and E1 under MindFlow.

D2 requires replies not to be overly redundant, yet models still over-interpret or over-explain, and their replies are not concise enough or have too low an information density; this is especially visible under successive follow-up questions, where models tend to re-explain earlier speech acts.

E1 concerns context consistency: models confuse earlier events, produce self-contradictory facts, or repeat meaningless dialogue in order to avoid errors. Notably, the failure rate of E1 rises from 14.36% under the turn-based baseline to 20.40% under MindFlow: once a model may revise a draft, postpone, or cancel a message, consistency between what has been sent and what is planned also becomes a critical bottleneck.

These are the aspects in which we hope anthropomorphic behavior can be further improved. We also note that CAPS-Eval and SEEDS are conventional in form — an expert rubric and an environment-expansion pipeline — and we do not claim either as a novel mechanism: they are the instruments that make the failure modes measurable and the training environments diverse.

Table 1: Case-level failure rate of each dimension under MindFlow (4 models ×\times 50 scenarios ×\times both roles A/B; Qwen3.5-397B-A17B judge).

RankDimensionCase failure rate
1D2 Information density23.93%
2E1 Context consistency20.40%
3D3 Persona expression3.53%
4D5 Rhythm and segmentation2.27%
5E3 Social-emotional fit2.02%
6E4 Initiative1.51%
7D1 Chat naturalness1.26%
8D4 Knowledge and action boundaries0.76%
9E2 Scenario and relationship fit0.76%
10E5 Dialogue development0.50%

Table 2: Case-level failure rate of each dimension under the turn-based baseline (same cases as Table 1).

RankDimensionCase failure rate
1D3 Persona expression46.10%
2D5 Rhythm and segmentation30.98%
3D2 Information density29.22%
4D1 Chat naturalness26.45%
5D4 Knowledge and action boundaries22.67%
6E1 Context consistency14.36%
7E2 Scenario and relationship fit2.52%
8E4 Initiative1.01%
9E5 Dialogue development1.01%
10E3 Social-emotional fit0.76%
AuthorsOpenReview ↗

Response to Reviewer NW6E (2/5)

2. DiAPO: DiAPO consists of two core components, CDT and ZPD. CDT measures the deficit of a dimension from its historical capability estimate θk\theta_k, so that the reward weight is proportional to its remaining headroom (ck=ϵ+(1−θk)pc_k=\epsilon+(1-\theta_k)^p, where ϵ\epsilon preserves signals for mastered dimensions). ZPD follows the Zone of Proximal Development principle to estimate the range in which the model learns most effectively, and gates the weight by how close the rollout-group mean sˉg,k\bar{s}_{g,k} is to the center of that range, θk−δ\theta_k-\delta, while asymmetrically penalizing overly easy and difficult samples.

CDT itself is a simple form of adaptive weighting under hard-example mining; the core of this paper is the introduction of ZPD on top of it, so that the model attends not only to hard cases but also, progressively, to the region just beyond its current capability. Multiplying the two terms (wg,k∝αkckωg,kw_{g,k}\propto\alpha_k c_k\omega_{g,k}) confines the weight to dimensions that are both underdeveloped and within reach, which yields the balanced multidimensional improvement targeted in this paper.

To validate the proposed method, we additionally compare two baselines: Fast-RL [1] and MAESTRO [2]. The ablation and baseline results verify that DiAPO outperforms previous adaptive weighting algorithms. The core reason is that previous adaptive weighting algorithms emphasize difficulty-adaptive learning but neglect the optimal learning range of the model's own capability, a gap that ZPD fills.

Table 3: Comparison with adaptive-weighting baselines on Qwen3.5-9B (Score and ACC@95 in %; Qwen3.5-397B-A17B judge).

Method (Qwen3.5-9B)Everyday ScoreEveryday ACC@95Game ScoreGame ACC@95CLAM ScoreCLAM ACC@95
+ GRPO97.3172.0098.7988.1498.2074.00
+ Fast-RL97.3568.0098.5986.4497.0666.00
+ MAESTRO97.3466.0098.0784.7597.6974.00
+ CDT only97.5868.0098.9989.8397.7570.00
+ ZPD only97.5170.0099.0694.9298.1482.00
+ DiAPO98.0176.0099.2696.6199.3092.00

[1] Li J, Zhang H, Zhang F, Chang TW, Kuang K, Chen L, Zhou J. Optimizing language models with fair and stable reward composition in reinforcement learning. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 2024 Nov (pp. 10122-10140).

[2] Zhao Y, Wang H, Ding X, Ouyang Y, Cai B, Xiong K, Gao J, Sun Z, Du L, Qin B, Liu T. Maestro: Meta-learning adaptive estimation of scalarization trade-offs for reward optimization. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2026 Jul (pp. 22267-22283).

AuthorsOpenReview ↗

Response to Reviewer NW6E (3/5)

W2: The training and evaluation objectives are too tightly ... generated or selected through CAPS-Eval.

Response: We thank the reviewer for raising this important concern about construct validity. We address it from three complementary angles: (i) the rubric itself is theoretically grounded and empirically aligned with human annotation; (ii) real human conversations provide an external anchor for the scale; and (iii) the gains transfer to third-party rubrics whose dimensions were fixed before our training and played no role in data filtering or reward construction.

(i) The rubric is grounded in psychology and aligns with human annotation. Each dimension instantiates one CAPS component—persona (SS), situational understanding (MtM_t), affective adaptation (AtA_t), and goal-directed participation (GtG_t)—so the checklist covers the anthropomorphism construct rather than ad-hoc criteria. Figure 7 confirms the alignment: human annotators and the two automatic judges give the nearly same ordering at both ends.

(ii) Real human game conversations anchor the scale from outside. We score the in-game conversations collected from the game company—the source material of Game Interaction as below. Real conversations rank first on Score, on per-turn evidence and on every pass rate, and their holistic score stays at the level of the strongest model.

Table 4: Models (Table 1, Game Interaction) and company-collected real conversations under the same rubric (best per column in bold).

Model (Game Interaction)ScorePer-TurnHolisticACC@85ACC@90ACC@95
Real conversations0.91500.91700.912967.80%40.68%25.42%
Claude-4.6 Thinking0.82900.74540.913523.73%5.08%0.00%
Gemini-3.5 Flash0.84900.77820.920028.81%13.56%3.39%
GPT-5.50.70100.58110.82010.00%0.00%0.00%
DeepSeek-V4-Flash0.75200.62530.87820.00%0.00%0.00%
Qwen3.5-397B-A17B0.83600.76490.907425.42%6.78%0.00%
Qwen3.6-35B-A3B0.79700.73450.858710.17%1.69%0.00%
Qwen3.5-9B0.66200.59140.73200.00%0.00%0.00%

(iii) The gains transfer to third-party rubrics. Following HumanLLM [3] and SimBench [4], we extract their rubric into our evaluation and score the same checkpoints with it as below. This provides strong evidence that the improvements from training under our evaluation framework generalize effectively and remain compatible with other evaluation dimensions.

Table 5: The same four checkpoints under CAPS-Eval (Everyday Chat) and under the HumanLLM and SimBench rubrics.

Checkpoint (Qwen3.5-9B)CAPS ScoreCAPS Per-TurnCAPS HolisticCAPS ACC@85CAPS ACC@90CAPS ACC@95HumanLLM ScoreHumanLLM Per-TurnHumanLLM HolisticHumanLLM ACC@85HumanLLM ACC@90HumanLLM ACC@95SimBench ScoreSimBench ACC@60SimBench ACC@65SimBench ACC@70
Base0.64100.72610.555812.00%4.00%0.00%67.6177.2857.9410.00%6.00%4.00%43.1128.00%24.00%20.00%
+ SFT0.92100.96770.873780.00%52.00%26.00%93.6892.0595.3188.00%70.00%36.00%68.6670.00%58.00%44.00%
+ SFT + GRPO0.97310.99620.950198.00%88.00%72.00%95.1394.9195.3596.00%72.00%28.00%74.6492.00%88.00%72.00%
+ SFT + DiAPO0.98010.99660.9636100.00%98.00%76.00%95.5594.7696.35100.00%86.00%30.00%75.5592.00%90.00%84.00%

[3] Wang X, Yang J, Li W, Xie R, Huang JT, Gao J, Huang S, Kang Y, Guo Y, Feng H, Xiao Y. HumanLLM: Benchmarking and Improving LLM Anthropomorphism via Human Cognitive Patterns. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2026 Jul (pp. 38486-38517).

[4] Hu T, Baumann J, Lupo L, Collier N, Hovy D, Röttger P. SimBench: Benchmarking the ability of large language models to simulate human behaviors. In International Conference on Learning Representations, 2026 Apr (Vol. 2026, pp. 28290-28341).

Reviewer_ytNcOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how LLMs can engage in credible human-like social interaction beyond persona consistency and conventional turn-taking. The authors introduce AnthroDial, which combines MindFlow for autonomous and asynchronous communication, CAPS-Eval for evaluating anthropomorphic interaction, and SEEDS with DiAPO for environment synthesis and adaptive policy optimization. Experiments across everyday chat, long-horizon character interaction, and game interaction report improved anthropomorphic evaluation scores after training, while human evaluation shows that MindFlow produces more human-like dialogues than conventional turn-based interaction in everyday chat. The paper argues that human-like social agents require not only persona consistency but also autonomy, appropriate timing, and adaptation to evolving contexts and relationships.

Critical strengths
  1. The paper addresses an important research problem by extending LLM anthropomorphism beyond persona consistency toward autonomous and continuous social interaction. MindFlow introduces a lightweight interaction framework that supports proactive communication, delayed responses, and continuous dialogue through a dynamic Mind Buffer. The comparison with conventional turn-based interaction in Figure 6 reports higher human evaluation scores across six models on Everyday Chat, providing empirical support for the proposed interaction framework.

  2. The paper presents a unified framework integrating interaction, evaluation, and training. CAPS-Eval provides a theory-grounded, multidimensional benchmark that reveals distinct capability gaps across models and interaction scenarios. SEEDS and DiAPO further support diverse environment synthesis and adaptive capability optimization. Experiments across three scenarios show substantial improvements after training, while comparisons with GRPO and component ablations provide additional evidence for the effectiveness of the proposed learning pipeline.

Critical weaknesses
  1. The benchmark does not sufficiently establish whether high CAPS-Eval scores reflect genuinely human-like interaction. The same evaluation criteria are used for SFT data filtering, RL rewards, and final evaluation, so improvements may mainly reflect better alignment with predefined rubrics. Human validation is limited to Everyday Chat, and the reported judge agreement primarily concerns model rankings rather than individual dialogues. Without stronger comparisons against real human conversations, the benchmark's ability to measure behavioral realism remains unclear.

  2. The methodological novelty and individual contributions of the proposed components remain insufficiently established. MindFlow combines existing ideas in proactive dialogue, memory, and message scheduling, while DiAPO is closely related to adaptive reward weighting and curriculum learning. Figure 6 compares the complete MindFlow framework with turn-based interaction but does not isolate the effects of its key mechanisms. Although DiAPO includes ablations of CDT and ZPD, these experiments are limited to Everyday Chat. The substantial gains from SFT also make it difficult to distinguish the contribution of SEEDS from the effects of supervised training itself.

Questions for authors
  1. How does CAPS-Eval distinguish genuinely human-like social interaction from compliance with expert-designed rubrics? Could the authors provide a few representative case studies comparing model-generated dialogues with real human conversations to illustrate whether higher CAPS-Eval scores correspond to more natural and realistic interaction patterns?

  2. Could the authors clarify the human evaluation protocol, including the number of annotators, annotations per dialogue, and inter-annotator agreement? Given that the current validation focuses on Everyday Chat and model-level ranking consistency, how reliably does CAPS-Eval evaluate individual dialogues across different scenarios?

  3. What are the main technical advances of MindFlow beyond existing proactive conversational agents with memory and scheduling mechanisms? Could the authors isolate the contributions of the dynamic Mind Buffer, message revision, and temporal scheduling rather than comparing only the complete framework with turn-based interaction?

  4. Table 2 shows substantial improvements after SEEDS-based SFT, but it remains unclear how much of the gain comes from the proposed environment expansion strategy rather than supervised fine-tuning itself. Could the authors compare SEEDS with a simpler persona–scenario sampling strategy under a similar training data budget?

5 comments
AuthorsOpenReview ↗

Response to Reviewer ytNc (1/4)

We thank the reviewer for the detailed and constructive feedback. We address each concern below.


Weaknesses


W1: The benchmark does not sufficiently establish whether high CAPS-Eval scores reflect genuinely human-like interaction. The same evaluation criteria are used for SFT data filtering, RL rewards, and final evaluation, so improvements may mainly reflect better alignment with predefined rubrics. Human validation is limited to Everyday Chat, and the reported judge agreement primarily concerns model rankings rather than individual dialogues. Without stronger comparisons against real human conversations, the benchmark's ability to measure behavioral realism remains unclear.

Response:

(i) To fully validate the consistency between our evaluation system and human judgment, we additionally conducted human annotation for Game Interaction and compared the human scores side by side with the metrics reported in the paper (Table 1, Table 2): the rank consistency across the three comparisons is ρ=0.93\rho=0.93 for Everyday Score and ρ=0.94\rho=0.94 for the Game human-likeness score. In each comparison there is only one order swap, and all occur between models whose human scores differ by a very small margin.

Table 1: Human Score vs. Qwen3.5-397B-A17B LLM Judge on Everyday Chat.

ModelHuman ScoreHuman ACC@95ScoreACC@95
Claude-4.6 Thinking100.0100.00%0.978078.00%
Gemini-3.5 Flash95.684.00%0.978076.00%
Qwen3.5-397B-A17B94.181.63%0.941044.00%
DeepSeek-V4-Flash93.484.00%0.968052.00%
GPT-5.589.072.00%0.882030.00%
Qwen3.5-9B62.642.00%0.64100.00%

Table 2: Human Score vs. Qwen3.5-397B-A17B LLM Judge on Game Interaction.

ModelHuman ScoreHuman ACCScoreACC@95
Gemini-3.5 Flash98.9084.75%0.84903.39%
Claude-4.6 Thinking98.7381.36%0.82900.00%
Qwen3.5-397B-A17B97.5867.80%0.83600.00%
DeepSeek-V4-Flash96.8055.93%0.75200.00%
GPT-5.591.1133.90%0.70100.00%
Qwen3.5-9B86.3622.03%0.66200.00%

(ii) To fully validate that our metric system aligns with human consistency, we score the in-game conversations collected from the game company—the source material of Game Interaction—as follows. Real conversations rank highest across all of Score, Per-Turn, and ACC@85/90/95, and their holistic score stays at the level of the strongest model, indicating that the scale has an external anchor rather than merely providing a relative ranking among models.

Table 3: Models and company-collected real conversations under the same rubric.

Model (Game Interaction)ScorePer-TurnHolisticACC@85ACC@90ACC@95
Real conversations0.91500.91700.912967.80%40.68%25.42%
Claude-4.6 Thinking0.82900.74540.913523.73%5.08%0.00%
Gemini-3.5 Flash0.84900.77820.920028.81%13.56%3.39%
GPT-5.50.70100.58110.82010.00%0.00%0.00%
DeepSeek-V4-Flash0.75200.62530.87820.00%0.00%0.00%
Qwen3.5-397B-A17B0.83600.76490.907425.42%6.78%0.00%
Qwen3.6-35B-A3B0.79700.73450.858710.17%1.69%0.00%
Qwen3.5-9B0.66200.59140.73200.00%0.00%0.00%
AuthorsOpenReview ↗

Response to Reviewer ytNc (2/4)

W2: The methodological novelty and individual contributions of the proposed components remain insufficiently established. MindFlow combines existing ideas in proactive dialogue, memory, and message scheduling, while DiAPO is closely related to adaptive reward weighting and curriculum learning. Figure 6 compares the complete MindFlow framework with turn-based interaction but does not isolate the effects of its key mechanisms. Although DiAPO includes ablations of CDT and ZPD, these experiments are limited to Everyday Chat. The substantial gains from SFT also make it difficult to distinguish the contribution of SEEDS from the effects of supervised training itself.

Response:

1. MindFlow: To separate autonomous scheduling from content generation, we run three variants of the same protocol on Gemini-3.5 Flash:

  • fixed-delay: δi,n\delta_{i,n} is replaced by a fixed delay, so a message can no longer be timed against the partner's pending reply;
  • no-memory: mi→jm_{i\to j} is removed, leaving recent history plus the context summary Ci,nC_{i,n};
  • no-draft: on interruption the pending draft is discarded and regenerated from scratch instead of being revised in place.

The ablation localizes MindFlow's contribution: removing any single module degrades performance, with draft revision being the most critical.

Table 4: Controlled ablation of MindFlow's interaction-state components (best value of each metric within each model in bold, ties bolded as well).

ModelVariantScorePer-TurnHolisticACC@85ACC@90ACC@95
Gemini-3.5 Flashour0.97800.98230.974798.00%96.00%76.00%
fixed-delay0.93750.96900.930296.00%80.00%62.00%
no-memory0.96540.97860.962497.00%91.90%71.60%
no-draft0.80900.82790.810682.00%78.00%60.00%

2. DiAPO: DiAPO consists of two core components, CDT and ZPD. CDT measures the deficit of a dimension from its historical capability estimate θk\theta_k, so that the reward weight is proportional to its remaining headroom (ck=ϵ+(1−θk)pc_k=\epsilon+(1-\theta_k)^p, where ϵ\epsilon preserves signals for mastered dimensions). ZPD follows the Zone of Proximal Development principle to estimate the range in which the model learns most effectively, and gates the weight by how close the rollout-group mean sˉg,k\bar{s}_{g,k} is to the center of that range, θk−δ\theta_k-\delta, while asymmetrically penalizing overly easy and difficult samples.

CDT itself is a simple form of adaptive weighting under hard-example mining; the core contribution of this paper is the introduction of ZPD on top of it, which adaptively guides the model toward cases just beyond its current capability rather than emphasizing difficulty alone. Multiplying the two terms (wg,k∝αkckωg,kw_{g,k}\propto\alpha_k c_k\omega_{g,k}) prioritizes dimensions that are both underdeveloped and within reach, thereby promoting the balanced multidimensional improvement targeted in this paper.

To validate the proposed method, we additionally compare two baselines: Fast-RL [1] and MAESTRO [2]. The ablation and baseline results verify that DiAPO outperforms previous adaptive weighting algorithms. The core reason is that previous adaptive weighting algorithms emphasize difficulty-adaptive learning but neglect the optimal learning range of the model's own capability, a gap that ZPD fills. We have also extended the CDT/ZPD ablation, which the paper reports only on Everyday Chat, to all three scenarios (Table 5).

Table 5: Comparison with adaptive-weighting baselines on Qwen3.5-9B (Score and ACC@95 in %; Qwen3.5-397B-A17B judge).

Method (Qwen3.5-9B)Everyday ScoreEveryday ACC@95Game ScoreGame ACC@95CLAM ScoreCLAM ACC@95
+ GRPO97.3172.0098.7988.1498.2074.00
+ Fast-RL97.3568.0098.5986.4497.0666.00
+ MAESTRO97.3466.0098.0784.7597.6974.00
+ CDT only97.5868.0098.9989.8397.7570.00
+ ZPD only97.5170.0099.0694.9298.1482.00
+ DiAPO98.0176.0099.2696.6199.3092.00

[1] Li J, Zhang H, Zhang F, Chang TW, Kuang K, Chen L, Zhou J. Optimizing language models with fair and stable reward composition in reinforcement learning. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 2024 Nov (pp. 10122-10140).

[2] Zhao Y, Wang H, Ding X, Ouyang Y, Cai B, Xiong K, Gao J, Sun Z, Du L, Qin B, Liu T. Maestro: Meta-learning adaptive estimation of scalarization trade-offs for reward optimization. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2026 Jul (pp. 22267-22283).

AuthorsOpenReview ↗

Response to Reviewer ytNc (3/4)

Questions


Q1: How does CAPS-Eval distinguish genuinely human-like social interaction from compliance with expert-designed rubrics? Could the authors provide a few representative case studies comparing model-generated dialogues with real human conversations to illustrate whether higher CAPS-Eval scores correspond to more natural and realistic interaction patterns?

Response:

In response to your question regarding "how CAPS-Eval distinguishes genuinely human-like social interaction from compliance with expert-designed rubrics," we have added experimental comparison results against real human dialogues in the Game domain in Table 3; these results sufficiently demonstrate that our evaluation method is aligned with real human conversations. To further address your concern, we provide a concrete case study below for illustration.

Real dialogue:

TurnMessage
1sw is still unbeatable / sure im spectating
2no / definitely not approving that
3none of your classes is 3A now
4fine tell me tomorrow / send a screenshot please

High-quality model synthesis:

TurnMessage
1that last game i hard carried 14.8 and took MVP
2their PG couldnt move an inch against my defense / were you watching just now shame you missed it
3invite me and ill show you a real carry / hurry up stop stalling
4me feeding you pick a C and follow me just sit back and get carried / im locking SF no more talk

Lower-quality model synthesis:

TurnMessage
1hows it going did you get your touch back i just ran a round and PG still feels the same you try no 5 ill check it for you in a bit
2yeah exactly no 5 ranked, that spot really is crowded, hurry up and get in the queue, dont get stuck over there. im free right now, once this game ends ill look at PGs mechanics for you and tune your stats while im at it.
3PG is all about controlling the tempo, find the open spot then make the key pass. dont rush your shot, read the defensive positions first, pass once your teammate is open.
4just picked it up casually to warm up and suddenly felt it this round dont be too rigid watch the inside and outside screens when space opens just pass dont force the drive

Scored under the same set of dimensions, the three dialogues are separated exactly where they should be. The table below reports the pass ratios per-turn (or single judgments for the holistic dimensions) for the ten dimensions on which the three dialogues differ most clearly, out of the 13 dimensions of the Game rubric:

DimensionScopeReal (100.0)High-quality synthesis (98.9)Lower-quality synthesis (38.9)
D1 Game-chat naturalnessper-turn1.001.000.55
D2 Information densityper-turn1.001.000.64
D3 Persona expressionper-turn1.001.000.73
D4 Rhythm and segmentationper-turn1.001.000.64
D6 Game terminologyper-turn1.001.001.00
D8 Competitive toneper-turn1.001.000.82
D10 Fragmented expressionper-turn1.001.000.55
D11 Unpunctuated chat styleper-turn1.000.900.00
E1 Context consistencyholistic1.001.000.00
E4 Initiativeholistic1.001.000.00
E6 Goal completionholistic1.001.000.00
E8 Fragmentation and immediacyholistic1.001.000.00
E9 Conversational asymmetryholistic1.001.000.00

As can be seen from the above results, CAPS-Eval scores effectively characterize the naturalness and realism of dialogues: both the real human dialogue and the high-quality model synthesis approach full marks across dimensions, with the high-quality synthesis even comparable to the real player; whereas the lower-quality model synthesis clearly loses points on key dimensions such as unpunctuated chat style (D11), context consistency (E1), and initiative (E4), forming a clear distinction from real dialogue. This indicates that higher CAPS-Eval scores indeed correspond to more natural and realistic interaction patterns, rather than merely superficial compliance with expert rubrics.

Reviewer_v3ZiOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies human-like interaction in continuous, open-ended conversations. AnthroDial combines MindFlow, which lets agents schedule and revise pending messages, with CAPS-Eval, which evaluates the resulting interactions using CAPS-inspired rubrics. SEEDS expands persona-scenario environments, and DiAPO adapts reward weights during reinforcement learning. Experiments cover everyday chat, long-horizon character interaction, and game interaction, with human evaluation, comparisons across judges, and training ablations.

Critical strengths
  1. The research question has practical value. The paper studies whether and when agents should speak, how they can send follow-up messages, and how they can advance a conversation beyond fixed turn-taking. Combining autonomous message scheduling with interaction evaluation provides a useful setting for studying behavior closer to everyday messaging.

  2. The work is comprehensive and includes a range of experiments. It connects agent design, benchmark construction, environment synthesis, and model training across three scenarios. Human evaluation, comparisons across judges, and component ablations provide useful initial support. Demonstrating a practical way to train a smaller model for this interaction setting also has value.

Critical weaknesses
  1. The benchmark's distinct contribution and remaining challenges could be explained more clearly.

(a) The paper discusses related benchmarks, but a comparison along shared dimensions would make its contribution easier to understand. Useful dimensions include autonomous participation, message timing, and evaluation of ongoing interaction. This would help readers see which capabilities AnthroDial adds to existing evaluations.

(b) Some existing models already score close to the upper limit in certain scenarios. For example, Claude reaches 0.9860 on Long-Horizon Character in Table 1. Game Interaction remains challenging for foundation models, but for the scenarios with near-perfect scores, it would help to explain the remaining difficulties and how the evaluation can continue to distinguish strong models.

  1. DiAPO's relation to prior methods and the nature of its additional gains could be clearer.

(a) Adaptive weighting of multiple rewards has been studied before, for example in "Optimizing Language Models with Fair and Stable Reward Composition in Reinforcement Learning" (Li et al., EMNLP 2024) and "MAESTRO: Meta-learning Adaptive Estimation of Scalarization Trade-offs for Reward Optimization" (Zhao et al., ACL 2026). Combining historical capability deficits with task-capability matching is a reasonable design, but its differences from earlier adaptive weighting methods could be explained more clearly. If the intended contribution includes an algorithmic advantage, a comparison with a suitable adaptive weighting method would help establish it. The approach still has value as a practical training method within the framework.

(b) GRPO already produces large improvements. DiAPO adds 0.0070, 0.0110, and 0.0047 to mean Score across the three scenarios in Table 2, while some ACC@95 gains are larger. Showing paired scenario-level scores for GRPO and DiAPO would help clarify whether these gains extend to lower-scoring scenarios or mainly reflect small improvements around the threshold. Such a comparison would also show whether some scenarios become worse.

Questions for authors
  1. Which interaction capabilities does AnthroDial evaluate beyond existing benchmarks? In the scenarios where strong models already score near the upper limit, what remaining difficulties does the evaluation reveal?
  2. Could you clarify how DiAPO's weighting rule differs from prior adaptive reward-weighting methods, and what benefit is expected from its specific combination of capability deficits and task-capability matching?
  3. Could you show paired scenario-level scores for GRPO and DiAPO? This would help explain the larger ACC@95 gains relative to the mean-score gains, including whether lower-scoring scenarios also improve.
6 comments
AuthorsOpenReview ↗

Response to Reviewer v3Zi (1/3)

We thank the reviewer for the detailed and constructive feedback. We address each concern below.


Weaknesses


W1: The benchmark's distinct contribution and remaining challenges could be explained more clearly.

(a) The paper discusses related benchmarks, but a comparison along shared dimensions would make its contribution easier to understand. Useful dimensions include autonomous participation, message timing, and evaluation of ongoing interaction. This would help readers see which capabilities AnthroDial adds to existing evaluations.

(b) Some existing models already score close to the upper limit in certain scenarios. For example, Claude reaches 0.9860 on Long-Horizon Character in Table 1. Game Interaction remains challenging for foundation models, but for the scenarios with near-perfect scores, it would help to explain the remaining difficulties and how the evaluation can continue to distinguish strong models.

Response:

(a) All three capabilities have explicit criteria in our rubric rather than appearing only in the design motivation (the item-level checkboxes for D1–D5 and E1–E5 are given in the CAPS-Eval rubric table in the appendix):

Dimension raised by the reviewerCorresponding criteria in AnthroDialWhat it adds over existing evaluations
Autonomous participationE4 Initiative: asks follow-up questions, shares relevant experience, steers topics naturally, avoids being fully passive; in MindFlow, whether to speak or to stay silent is decided by the model itselfMost existing evaluations require a reply at every turn and are driven by the user's messages
Message timingD5 Rhythm and segmentation: plausible segmentation count, readable segments, plausible segment order and reply rhythm; in MindFlow the delivery delay δi,n\delta_{i,n} and postpone/cancel are proposed by the modelExisting evaluations treat a reply as an instantaneous action and have no criterion for when it is sent
Ongoing interactionE1 Context consistency (no repeated questions, coherent facts, correct speaker references, tracked state, memory not confused with the present) + E5 Dialogue development; Long-Horizon Character additionally evaluates stale-memory confusion within a long dialogueExisting evaluations are mostly single short dialogues without cross-session state or a dialogue arc

In other words, the role of MindFlow is to make these behaviors observable (the model decides whether, when, and what to send) rather than letting a fixed turn-taking protocol do it for the model. To our knowledge, no prior benchmark evaluates anthropomorphism in a continuous, autonomous multi-turn setting in which the system itself decides when to speak, so AnthroDial is the first to score these behaviors rather than assume them.

(b) Our conclusion is precisely that mean scores are close to saturation while stability is far from it, so the main table reports ACC@85/90/95 alongside Score and uses ACC@95 as the primary metric for separating strong models:

  • On Long-Horizon Character, Claude has a Score of 0.9860 but an ACC@95 of only 0.86, whereas GPT-5.5 has a lower Score (0.9820) but an ACC@95 of 0.90. Mean-score ranking and stability ranking therefore disagree, and the evaluation still separates strong models.
  • On the same scenario, Qwen3.5-397B-A17B has a Score of 0.76 but an ACC@95 of only 0.16; the gap comes from systematic failures in a small number of scenarios rather than from its overall level.
  • The remaining difficulties are concentrated in the dimensions that still discriminate between models: Figure 3 shows that on Long-Horizon Character, context consistency (E1) separates models sharply, ranging from 0.59 for Qwen3.5-397B-A17B to 0.97 for Claude-4.6 Thinking and GPT-5.5. What remains hard on this scenario is tracking facts, state changes and speaker references across a long dialogue, not per-turn style.
AuthorsOpenReview ↗

Response to Reviewer v3Zi (2/3)

W2: DiAPO's relation to prior methods and the nature of its additional gains could be clearer.

(a) Adaptive weighting of multiple rewards has been studied before, for example in "Optimizing Language Models with Fair and Stable Reward Composition in Reinforcement Learning" and "MAESTRO: Meta-learning Adaptive Estimation of Scalarization Trade-offs for Reward Optimization". Combining historical capability deficits with task-capability matching is a reasonable design, but its differences from earlier adaptive weighting methods could be explained more clearly. If the intended contribution includes an algorithmic advantage, a comparison with a suitable adaptive weighting method would help establish it. The approach still has value as a practical training method within the framework.

(b) GRPO already produces large improvements. DiAPO adds 0.0070, 0.0110, and 0.0047 to mean Score across the three scenarios in Table 2, while some ACC@95 gains are larger. Showing paired scenario-level scores for GRPO and DiAPO would help clarify whether these gains extend to lower-scoring scenarios or mainly reflect small improvements around the threshold. Such a comparison would also show whether some scenarios become worse.

Response:

(a) For this comparison we add direct experiments with both methods (same Qwen3.5-9B base checkpoint and the same judge):

Table 1: DiAPO versus prior adaptive-weighting methods (Score and ACC@95, %; Qwen3.5-397B-A17B judge)

Method (Qwen3.5-9B)Everyday ScoreEveryday ACC@95Game ScoreGame ACC@95CLAM ScoreCLAM ACC@95
+ GRPO97.3172.0098.7988.1498.2074.00
+ Fast-RL[1]97.3568.0098.5986.4497.0666.00
+ MAESTRO[2]97.3466.0098.0784.7597.6974.00
+ CDT only97.5868.0098.9989.8397.7570.00
+ ZPD only97.5170.0099.0694.9298.1482.00
+ DiAPO98.0176.0099.2696.6199.3092.00

The difference lies in the shape of the weighting signal: Fast-RL and MAESTRO start from fair and stable reward composition and from meta-learned scalarization respectively, and their weights vary monotonically with difficulty or with the multi-objective trade-off; DiAPO multiplies CDT (the headroom term) by a band-pass gate ωg,k\omega_{g,k} centered at θk−δ\theta_k-\delta, so that only dimensions which are both underdeveloped and still learnable are up-weighted, and the weight is never pushed to the hardest dimension alone. Two points in the table are worth noting: neither prior method exceeds GRPO in ACC@95 on any of the three scenarios; CDT-only and ZPD-only each win some cells (the latter +6.8+6.8 on Game and +8.0+8.0 on CLAM, while both lose on Everyday, by −4.0-4.0 and −2.0-2.0 respectively), and only the combination is best on all three scenarios.

(b) We have divided the Everyday Chat scenario into fine-grained sub-scenarios and computed the metrics of GRPO and DiAPO on each sub-scenario; the table is shown below. As can be seen, DiAPO yields significant improvements on most scenarios, and decreases only slightly on a few near-saturated scenarios such as Interests, Daily chat, and Gaming. At the same time, DiAPO maintains good accuracy on most scenarios: the average Score on Everyday Chat rises from 0.9731 (GRPO) to 0.9801 (DiAPO), an improvement of 0.0070, and the average ACC@95 rises from 72.00% to 76.00%.

Table 2: Paired scenario-level scores of GRPO and DiAPO on Everyday Chat

CategoryGRPO ScoreDiAPO ScoreGRPO ACC@95DiAPO ACC@95
Emotional support0.94460.973640.00%80.00%
Family and parenting0.95520.978950.00%75.00%
Work communication0.97110.987740.00%100.00%
Pets and family life0.96700.980175.00%75.00%
News0.96750.977275.00%75.00%
Finance0.98630.992075.00%100.00%
Technology0.97700.981166.67%100.00%
Relationship development0.95640.957650.00%25.00%
Sports0.99070.9909100.00%75.00%
Gaming0.99050.9873100.00%66.67%
Daily chat0.98560.9814100.00%80.00%
Interests0.99120.9758100.00%60.00%

[1] Li J, Zhang H, Zhang F, Chang TW, Kuang K, Chen L, Zhou J. Optimizing language models with fair and stable reward composition in reinforcement learning. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 2024 Nov (pp. 10122-10140). [2] Zhao Y, Wang H, Ding X, Ouyang Y, Cai B, Xiong K, Gao J, Sun Z, Du L, Qin B, Liu T. Maestro: Meta-learning adaptive estimation of scalarization trade-offs for reward optimization. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2026 Jul (pp. 22267-22283).

AuthorsOpenReview ↗

Response to Reviewer v3Zi (3/3)

Questions


Q1: Which interaction capabilities does AnthroDial evaluate beyond existing benchmarks? In the scenarios where strong models already score near the upper limit, what remaining difficulties does the evaluation reveal?

Response:

(i) The mapping between the three capabilities and our criteria is given in the table in W1(a). (ii) In scenarios near the upper limit, the remaining difficulty is stability and specific dimensions rather than the mean: on Long-Horizon Character, Claude has a Score of 0.9860 but an ACC@95 of only 0.86, whereas GPT-5.5 has a lower Score (0.9820) but an ACC@95 of 0.90. Figure 3 shows that E1 on this scenario still discriminates strongly between models, spanning 0.59 (Qwen3.5-397B-A17B) to 0.97 (Claude-4.6 Thinking, GPT-5.5), so long-range context tracking remains a differentiating dimension rather than a saturated one. (iii) Under the independent HumanLLM rubric (Qwen3.5-397B-A17B judge, 50 scenarios / 100 role cases), ACC@95 across the ten systems spans 4%–70% (Claude 46%, Gemini 64%, DeepSeek 70%), which shows that these scenarios are not saturated under a stricter criterion.


Q2: Could you clarify how DiAPO's weighting rule differs from prior adaptive reward-weighting methods, and what benefit is expected from its specific combination of capability deficits and task-capability matching?

Response: DiAPO's weighting rule is the product of CDT and the ZPD gate (wg,k∝αkckωg,kw_{g,k}\propto\alpha_k c_k\omega_{g,k}): CDT assigns weight according to the headroom of a dimension, and ZPD then restricts that weight to the currently learnable range through a band-pass centered at θk−δ\theta_k-\delta. Its difference from Fast-RL and MAESTRO, together with the comparison results, is given in Table 1 in W2(a); the expected benefit is a consistent and balanced improvement across the three scenarios rather than pushing the weight to the hardest dimensions.


Q3: Could you show paired scenario-level scores for GRPO and DiAPO? This would help explain the larger ACC@95 gains relative to the mean-score gains, including whether lower-scoring scenarios also improve.

Response: Yes. The paired scenario-level changes of GRPO → DiAPO on Everyday Chat are listed in Table 2; the corresponding changes are given in Table 3. Lower-scoring scenarios also improve, while near-saturated scenarios are flat to slightly negative; the larger ACC@95 gains relative to mean-score gains arise because GRPO's weak band sits just around 0.95, so many lower-scoring scenarios cross the threshold under DiAPO.

Table 3: Per-category changes from GRPO to DiAPO on Everyday Chat.

ScenarioΔScore (DiAPO − GRPO)ΔACC@95 (DiAPO − GRPO)
Emotional support+0.0290+40.00%
Family and parenting+0.0237+25.00%
Work communication+0.0166+60.00%
Pets and family life+0.01310.00%
News+0.00970.00%
Finance+0.0057+25.00%
Technology+0.0040+33.33%
Relationship development+0.0012−25.00%
Sports+0.0003−25.00%
Gaming−0.0032−33.33%
Daily chat−0.0042−20.00%
Interests−0.0154−40.00%
Average+0.0070+4.00%

Comments on the paper

3 comments
Area_Chair_A2DGOpenReview ↗

Reviewer ytNc, please respond to authors today

Reviewer ytNc,

Today is the last day of the discussion period. The authors have prepared a detailed response to your review. Please reply to them as soon as possible to ensure there is a chance for back and forth discussion before the discussion period ends.

AC

Dear AC,

We appreciate the time and effort you and the reviewers spent evaluating our submission and considering our rebuttal. Additions to the revised manuscript are in blue. The four responses below distinguish what the original submission reported from the contributions of the revision and rebuttal.

Overall, the reviewers found the problem worth studying and the system unusually complete, highlighting the practical value of autonomous interaction and the breadth of the evaluation and training pipeline.

On the problem and the interaction design:

Reviewer v3Zi: "The research question has practical value."
Reviewer NW6E: "MindFlow provides a useful interaction abstraction beyond strict turn-taking."
Reviewer ytNc: "The paper addresses an important research problem by extending LLM anthropomorphism beyond persona consistency toward autonomous and continuous social interaction."

On the evaluation and training side:

Reviewer NW6E: "The paper is unusually complete as an end-to-end system and provides extensive diagnostics."
Reviewer v3Zi: "The work is comprehensive and includes a range of experiments."
Reviewer ytNc: "CAPS-Eval provides a theory-grounded, multidimensional benchmark that reveals distinct capability gaps across models and interaction scenarios."

The remaining concerns fall into four groups: relation to prior anthropomorphism work; what MindFlow reveals about current models; whether the framework measures human likeness rather than compliance with its own rubric; and whether simpler alternatives exist and how much each component contributes.

Relation to prior work. The original Introduction already discusses the three research lines; the revision adds Table 1, which makes the comparison explicit along context (persona, situation) and autonomy of participation (initiate, timing, revise). No prior benchmark scores revision, and none scores more than one of the three behaviors, whereas AnthroDial scores all of them.

What MindFlow reveals. The original Section 5.2 (Figure 3) already reports the dimension-level bottlenecks. The revision adds a case-level failure profile (Appendix C.6, Table 25) that quantifies each dimension's failure rate and shows why the remaining failures surface only when a draft can be revised.

Rubric compliance or human likeness. The original submission already reports agreement between automatic and human rankings (Figure 7). The revision and rebuttal add three checks:

  1. External anchor. Real game conversations collected by the company, scored under the same rubric, rank first on overall quality, per-turn evidence, and every pass rate (Table 5; Appendix C.4), so the rubric identifies real human dialogue as the best case rather than only ranking models.
  2. Human evaluation after training. In the rebuttal, human annotators scored Base, SFT, GRPO, and DiAPO at 49.7, 80.4, 93.1, and 95.5, respectively (ACC@95: 6%, 15%, 34%, 51%) (Appendix C.7). Human and automatic rankings agree for both base and trained models.
  3. Out-of-distribution rubrics. We re-score the same models with rubrics fixed before training and unused for filtering or rewards (Appendix C.9). Both metrics rose (HumanLLM Score 67.61 → 95.55, SimBench ACC@70 20% → 84%), indicating that the gains are compatible with independently designed evaluation dimensions.

Component contributions. The original submission compares MindFlow with turn-based interaction (Figure 6), ablates CDT and ZPD (Table 4), and analyzes dimensions (Figure 3). The revision and rebuttal add:

  1. MindFlow ablation. Three variants isolate the interaction-state components---fixed delay, no partner memory, and no draft revision---and removing draft revision is the most damaging, cutting Score from 0.9780 to 0.8090 and ACC@95 from 76% to 60% (Appendix C.5, Table 24).
  2. SEEDS ablation. Against a simpler non-expanding synthesis pipeline under the same budget, the expanded environments lead on both Score (0.9210 versus 0.8922) and ACC@95 (26% versus 14%; Appendix C.3, Table 20), so in-distribution samples alone cannot replace environment diversity.
  3. DiAPO baselines. Fast-RL and MAESTRO, two prior adaptive-weighting methods, are compared under the same base model, judge, and budget and reach 68% and 66% ACC@95, against 76% for DiAPO (Table 3), which supports the ZPD-based weighting of dimensions.

Reviewer v3Zi reconsidered the assessment after the rebuttal and now rates the paper 3; Reviewer NW6E's remaining concern was addressed by making both contributions explicit in Sections 3.1 and 4.3. Reviewer ytNc has not responded, but our rebuttal answered all four of their questions. We respectfully ask the AC to weigh the revised manuscript and these responses when reaching a decision.

Thank you again for your time, your support, and your careful evaluation of our work.

Reviewer_ytNcOpenReview ↗

Official Comment by Reviewer ytNc

Thank you for the detailed response and additional experiments. The response have addressed most of my concerns. The MindFlow ablations, the DiAPO comparisons across all three scenarios, and the budget-matched SEEDS experiment address most of my questions about what each component contributes. The Game Interaction annotations and real-conversation comparison also strengthen the benchmark evaluation. My remaining concern is whether the gains from SFT and RL translate into more natural interaction under independent human judgment. The model-level rankings and the single game-chat example scored with CAPS-Eval leave that question open, since the same rubric is used for training and evaluation. Therefore, I’m keeping my current rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.