acceptodds
Under review as a conference paper at ICLR 2027

SocialRL: Learning Social Intelligence Across Multi-turn Dialogue and Reward Design

Abstract

Social intelligence enables agents to read social context, infer intent, and adapt over sustained dialogue. As language models become autonomous collaborators, it is central to building effective and trustworthy human-AI interaction. Existing reinforcement learning methods optimize single-turn utterances and sparse outcome rewards, producing short-sighted policies that struggle to manage goal-relationship tensions across multi-turn interactions. We propose **SocialRL**, a multi-turn reinforcement learning framework addressing both challenges. First, we apply multi-turn reinforcement learning using PPO that propagates delayed outcome rewards back to each turn, enabling long-horizon planning. Second, we design six process reward dimensions capturing the goal-relationship trade-off, including goal advancement, relational attunement, contextual coherence, etc. A reward model dynamically generates fine-grained scoring criteria for each dimension, while a stage-aware weight schedule prioritizes relationship-building in early turns, goal advancement mid-way, and balanced closure late. Across multiple social-dialogue benchmarks, SocialRL improves Goal Achievement by an average of percentage points over the corresponding Base models. These results demonstrate the effectiveness of SocialRL across synthetic and real social scenes, as well as standard and challenging social scenarios.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Social Intelligence, Reinforcement Learning, Multi-turn Dialogue Optimization, Large Language Models, Reward Design

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_sxSzOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper presents SocialRL, a reinforcement learning framework for improving LLMs’ ability to pursue social goals while maintaining relationships during multi-turn interactions. The framework combines multi-turn PPO with a six-dimensional process reward. An LLM generates context-specific binary evaluation criteria for each dimension, judges whether the current response satisfies them, and assigns context-dependent weights. Process rewards are combined with a terminal outcome reward. The authors evaluate four policy backbones against multiple opponents on SOTOPIA-π, SOTOPIA-All, SOTOPIA-Hard, and AgentSense, reporting improvements over base models and selected training baselines.

Critical strengths
  1. Substantial experimental effort and broad evaluation coverage. The study evaluates multiple policy backbones and opponent models across several benchmark settings, including AgentSense. The reported goal-achievement improvements suggest potential practical value and provide evidence that the overall training recipe can benefit more than a single model or interaction setting.

  2. Attention to both task success and relationship quality. Evaluating both objectives makes the study more informative than evaluating goal completion alone. The human-alignment audit also provides useful evidence about the automatic evaluators, although it does not resolve the reward-stability concerns discussed below.

Critical weaknesses
  1. The experiments do not sufficiently establish the contribution of the proposed domain-specific designs beyond standard multi-turn PPO with dense feedback. The optimization procedure largely follows standard PPO and GAE. I acknowledge the comparisons with existing methods and the ablations on process rewards, reward dimensions, and dynamic weights. In particular, the NoProcess ablation supports the usefulness of process supervision in the tested setting. However, removing all process rewards does not establish that the proposed rubric construction is superior to an existing or simpler process reward.

A particularly informative missing comparison is standard multi-turn PPO using an existing social reward design, such as that of Sotopia-RL, under matched reward-model and training budgets. Comparisons that change both the optimization procedure and the reward design cannot isolate their respective contributions. There is also no direct ablation of dynamically generated criteria against fixed criteria under the same PPO setup. Although the dynamic-weight ablation provides some positive evidence, it is restricted to one policy–opponent setting and lacks uncertainty estimates. Consequently, the evidence does not convincingly establish the claimed methodological contribution beyond applying standard multi-turn PPO with dense social feedback.

  1. The stability and comparability of the dynamically generated rubrics are insufficiently validated. According to Figure 4, the evaluated response is provided to the same LLM call that generates the criteria, assigns pass/fail judgments, and determines dimension weights. Different responses to the same dialogue state may therefore be assessed using different criteria and weights. This raises a concern that the criteria could adapt to the response being evaluated, producing post-hoc justification or scores that are not comparable across candidate responses.

Binary judgments, weight caps, normalization, and fallback rules constrain the output format and numerical range, but do not establish consistency in what is being evaluated. The reported PRM–human correlation of 0.816 is useful evidence of alignment across sampled records; it does not directly measure repeated-scoring reliability or ranking consistency among alternative responses to the same context. Without these checks, the reliability of this central training signal remains unclear.

Questions for authors
  1. Under the same multi-turn PPO implementation, reward-model capability, and rollout budget, how does the proposed reward design compare with an existing social reward design or a simpler fixed-rubric process reward? Please clarify which baseline settings and budgets were matched and provide uncertainty estimates for the key comparisons.

  2. Why are the evaluation criteria and dimension weights conditioned on the response being evaluated? How do you ensure comparable evaluation of alternative responses to the same dialogue state? A comparison with criteria generated solely from the scenario, goal, and dialogue history—and then shared across candidate responses—would directly address this concern.

  3. How reproducible are the generated rubrics and resulting scores for identical inputs? Do repeated evaluations preserve candidate-response rankings, and how well do those rankings agree with human judgments within the same dialogue context?

2 comments

W1 & Q1: Attribution Between Multi-turn Optimization and Reward Design

Reviewer concern: It has not been sufficiently proven that SocialRL's domain-specific design is superior to standard Multi-turn PPO + Dense Feedback; there is a lack of comparison under the same PPO framework among Sotopia-RL Reward, Fixed Rubric, and dynamic Rubric, as well as a lack of direct comparison between Single-turn PPO and Multi-turn PPO.

SocialRL's methodological contribution is not proposing a completely new PPO or GAE algorithm, but constructing dynamic multi-dimensional process rewards usable for long-term policy optimization in multi-turn social interaction. PPO and GAE themselves are existing algorithms, and this paper does not present their mathematical update rules as a new general reinforcement learning algorithm contribution. To describe "not reinventing PPO" as "insufficient algorithmic contribution" is a misplaced evaluation of the paper's contribution. Multi-turn PPO is responsible for using complete trajectories and future returns for cross-turn credit assignment; Dynamic Process Rewards are responsible for providing intermediate supervision reflecting goal progress, relationship state, and interaction validity. The original manuscript has already provided component-level evidence: Appendix E.1 compares PPO and GRPO, Appendix E.2 shows that removing process rewards reduces Goal Achievement from 63.3% to 53.2%, and Appendix E.4 shows that Dynamic Weighting outperforms Uniform, Manual, and Grid. These results respectively verify the roles of the optimization algorithm, process supervision, and dynamic weighting. We agree that direct comparison with alternative process rewards is still needed, so we add:

Reward DesignGoal AchievementRelationship Change
Outcome-only53.20.46
Sotopia-RL Reward57.40.479
Fixed Generic Rubric56.10.465
Context-only Dynamic Rubric630.508
Full SocialRL Reward63.30.516

We compare Single-turn PPO and Multi-turn PPO under a fixed SocialRL Reward, unifying policy initialization, training data, Rollout Budget, reward model capability, and evaluation protocol, and separately report the mean, standard deviation, and confidence intervals across training seeds. Please do not expand "need to add seed experiments" into "all existing conclusions are invalid."

W2 & Q2 & Q3: Reliability, Comparability, and Reproducibility of Dynamic Rubrics

Reviewer concern: The original mechanism generates rubrics and weights based on the current response, so different candidate responses may be evaluated under different criteria, creating a risk of post-hoc rationalization; the PRM–Human correlation coefficient cannot directly prove repeated scoring stability, nor can it prove the reliability of ranking candidate responses within the same context.

The original design incorporates the current response into rubric generation to make the evaluation criteria provide fine-grained feedback for specific behaviors. We agree that if different candidate responses use different evaluation criteria, additional validation of ranking comparability and repeated scoring reliability is needed. But please note: the original design goal is not for the evaluation model to find favorable reasons for the current response, but to generate observable and verifiable binary judgment criteria based on the current dialogue context and specific behavior. Directly presuming "response-conditioned evaluation" to be "post-hoc rationalization" lacks evidence. To address this issue, we add a Response-independent Frozen Rubric control experiment: given the current dialogue state, first generate evaluation criteria and dimension weights only based on the scenario, role information, social goal, and dialogue history, fix the rubric and weights before candidate responses are generated, and then use exactly the same criteria to perform pass/fail evaluation on different candidate responses under the same dialogue state. We compare three mechanisms:

Rubric MechanismRanking ConsistencyHuman Alignment (200 samples)
Response-conditioned0.9470.823
Context-only Frozen0.960.819
Fixed Generic0.9420.864

The focus is on whether scores for different candidate responses have a stable relative ranking, whether the ranking is consistent with human judgment, and whether frozen criteria improve or weaken final training performance. This set of experiments can distinguish two possibilities: response-conditioned rubrics indeed provide finer-grained behavioral feedback; or the mechanism may introduce unnecessary scoring differences but the impact is not significant. The results show that ranking consistency and human alignment under all three mechanisms are at a high level, and there is no support for the strong accusation that "post-hoc rationalization makes evaluation incomparable."

Reviewer_QhymOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper aims to enable LLMs to simultaneously achieve their own objectives and maintain relationships with their interlocutors during multi-turn social conversations. The proposed framework, SocialRL, incorporates two core design components:

  1. Multi-turn PPO: This approach treats an entire multi-turn conversation as a single reinforcement learning trajectory, utilizing a Value Network and GAE to propagate feedback from future outcomes back to earlier turns in the dialogue.

  2. Dynamic Multi-dimensional Process Rewards: The framework defines six reward dimensions. Scoring criteria are generated dynamically by an LLM, with reward weights adjusted according to the specific stage of the conversation.

Critical strengths

1.The experiments are extensive, and the ablation studies are thorough. Four different backbones and multiple opponent models were evaluated across SOTOPIA, SOTOPIA-π, and AgentSense. The ablation studies further validate the roles of process rewards, various reward dimensions, and dynamic weighting.

Critical weaknesses
  1. Reliability of the dynamic reward mechanism. The paper employs a single Process Reward Model to dynamically generate a rubric, evaluate whether the criteria are met, and determine reward weights. These processes are highly coupled; in my own experience, this approach typically yields poor results.

2.The comparison with recent related methods is somewhat limited. Several closely related approaches,like SAVOIR, are only discussed in Related Work without empirical comparisons. Could the authors provide additional comparisons with these recent methods, particularly those focusing on multi-turn social interaction and reward design?

Questions for authors

Although the paper provides leave-one-dimension-out ablations for the six reward dimensions, the rationale for selecting exactly these six dimensions remains insufficiently justified. In particular, some dimensions, such as Goal Advancement and Strategic Positioning, or Contextual Coherence and Turn Quality, appear potentially overlapping. Could the authors clarify how the six dimensions were selected, whether their redundancy or correlations were examined?

5 comments

W1: Reliability of the Dynamic Reward Mechanism

Reviewer concern: A single model call simultaneously handles rubric generation, pass/fail judgment, and weight allocation, which may cause coupling and stability issues.

Task coupling within a single call may indeed affect reward stability, but multi-task generation itself does not necessarily make evaluation invalid; the key is whether the generation criteria and scoring results have reproducibility and consistency across candidate responses. The original manuscript adopts a joint generation mechanism to keep the rubric, dimension weights, and current dialogue state consistent, avoiding the use of fixed scoring criteria completely detached from the specific situation. We have already imposed constraints such as binary judgment, weight upper bounds, fallback for anomalous outputs, and normalization. The reviewer directly presumes "possible coupling" to mean "invalid evaluation" without basis. To eliminate concerns, we add three control mechanisms:

Reward DesignGoal AchievementRelationship Change
Origin SocialRL Reward63.30.516
Decoupled62.50.523
Context-Only rubric63.00.508

We also report the score variance under repeated evaluation of the same input and the ranking consistency of candidate responses.

Rubric MechanismRanking ConsistencyHuman Alignment (200 samples)
Response-conditioned0.9470.823
Context-only Frozen0.960.819
Fixed Generic0.9420.864

The focus is on whether scores for different candidate responses have a stable relative ranking, whether the ranking is consistent with human judgment, and whether frozen criteria improve or weaken final training performance. This set of experiments can distinguish two possibilities: response-conditioned rubrics indeed provide finer-grained behavioral feedback; or the mechanism may introduce unnecessary scoring differences but the impact is not significant. The results show that ranking consistency and human alignment under all three mechanisms are at a high level.

W2: Comparison with Recent Related Methods

Reviewer concern: Recent methods such as SAVOIR are discussed only in Related Work and are not compared experimentally.

We agree that SAVOIR is closely related to this paper's research objective, and adding a direct comparison helps define the contribution. But it must be clear that the two have different emphases: SAVOIR mainly focuses on using expected utility and Shapley-based Attribution to attribute multi-turn interaction outcomes to specific behaviors; SocialRL focuses on modeling goal advancement, relationship maintenance, and interaction validity as context-dependent process rewards, combined with trajectory-level policy optimization. The two are not completely identical reward mechanisms. To directly regard "not comparing beyond Related Work" as a methodological flaw is an excessive requirement for comparison reasonableness. We have added a comparison under unified test scenarios and evaluation protocols:

MethodGoal AchievementRelationship Change
SAVOIR53.80.435
SocialRL63.30.516

The training setting is same under two experiments.

Q1: Basis and Redundancy of the Six-Dimensional Rewards

Reviewer concern: Goal Advancement and Strategic Positioning, and Contextual Coherence and Turn Quality, may be highly overlapping.

These dimensions may be semantically related, but they are not the same in evaluation object and function. Goal Advancement measures whether the social goal is actually advanced; Strategic Positioning measures whether the current strategy choice is reasonable; a reasonable strategy does not necessarily bring immediate goal progress, and goal progress does not necessarily stem from the most appropriate strategy. Relational Attunement focuses on adaptation to the other party's emotions and relational state; Persona Consistency focuses on the consistency of the agent's own role and behavior. Contextual Coherence focuses on the semantic and logical consistency of the response with the entire dialogue history; Turn Quality focuses on the local expression quality and interactional appropriateness of the current response. A response may be contextually relevant but receive a lower Turn Quality because it is verbose, repetitive, or inappropriately expressed. The goal of reward dimension design is to provide functionally interpretable feedback, not to require all dimensions to be statistically completely independent. Some correlation between goal advancement and strategy reasonableness is expected; whether there is unnecessary redundancy needs to be judged based on the actual degree of correlation and its impact on final training results. Appendix E.3 has already shown that different dimensions have different marginal effects, and we add the six-dimensional Spearman correlation matrix:

DimensionGoal AdvancementStrategic PositioningRelational AttunementPersona ConsistencyContextual CoherenceTurn Quality
Goal Advancement1.0000.6370.3050.3450.3860.251
Strategic Positioning0.6371.0000.3510.3370.4200.262
Relational Attunement0.3050.3511.0000.3160.4180.207
Persona Consistency0.3450.3370.3161.0000.3250.157
Contextual Coherence0.3860.4200.4180.3251.0000.210
Turn Quality0.2510.2620.2070.1570.2101.000

On 366,552 process reward records, the average absolute Spearman correlation is 0.329; Goal Advancement and Strategic Positioning are 0.631, and Contextual Coherence and Turn Quality are 0.210. Correlation does not automatically constitute evidence of redundancy; if the reviewer insists on treating any correlation as a flaw, that is equivalent to requiring reward dimensions to be statistically uncorrelated, which is neither reasonable nor realistic for social behavior dimensions.

Reviewer_gEDgOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes SocialRL, they tried to improve social intelligence using multi-turn dialogue and reward design. Which uses multi-turn trajectory optimization and multi-dimensional dynamic process rewards. They evaluate SocialRL on SOTOPIA-π, SOTOPIA-All, SOTOPIA-Hard, and AgentSense. And shows that SocialRL achieves average of 9.2 percentage points over the corresponding base models.

Critical strengths
  1. The paper clearly states their opinion and results.
  2. The figures are clear and informative, and the main body of the paper is generally easy to follow.
Critical weaknesses
  1. The dimensions. The paper uses six dimensions: Goal advancement, strategic position Relation attunement, persona consistency, Contextual coherence turns quality. But how do these 6 dimensions come from? There are also many more dimensions of social intelligence that are important. How do you decide to use these specific 6 dimensions and exclude other dimensions?
  2. The relationship between stages and weights is vague. And how do you decide and assign the weight between the stages and weights? It is difficult to isolate whether the improvements arise from explicit stage awareness, context-dependent weighting, or their interaction.
  3. Does the improvement come from reward hacking? Since the dimensions of process reward for PPO are quite similar to the evaluation rubrics of SOTOPIA. It is unclear whether the improvement comes from the overfitting to the evaluation criteria.
  4. Problems of human evaluation. There is no direct human evaluation comparing SocialRL with the base models. The human evaluation only shows the correlation between humans and the automatic evaluator. But it does not test whether the performance gains of SocialRL over the base models persist under human evaluation. So it cannot rule out that the gain comes from overfitting.
  5. It is unclear whether the performance gains come from the proposed reward design or from the choice of PPO over GRPO. Some baselines, like Sotopia-RL (which also uses multi-dimensional reward design and turn-level reward), use GRPO. The paper only compares PPO and GRPO under its own setting but does not compare different reward designs under the same optimization algorithm.
  6. Limitation of evaluation. The evaluation is restricted to the SOTOPIA and AgentSense benchmarks. The paper claims its method could improve the "social intelligence" of agents; it is unclear whether the improvements generalize to more diverse social settings. Also, the paper only evaluates the goal completion and relationship of agents, but there are many dimensions of both the SOTOPIA benchmark and real-world social interactions; the paper did not show these results. Does improving goal achievement and relationship maintenance necessarily imply an improvement in general social intelligence?
  7. Lack of ablations. The contributions of multi-turn optimization and reward design are not clear. The paper states multi-turn trajectory optimization and dynamic multi-dimensional process rewards as its two main contributions. These experiments do not isolate the contribution of multi-turn optimization. In particular, there is no comparison between single-turn and multi-turn PPO under the same reward design. Also, the paper did not compare existing reward designs and the proposed reward design under the same PPO framework.
  8. The improvement of social intelligence might introduce safety issues, like agents' ability to manipulate, deceive, or strategically influence human users, especially since SocialRL explicitly optimizes goal achievement and strategic positioning during multi-turn social interactions. So the authors should discuss these potential safety risks and provide additional evaluations.
Questions for authors

See critical weakness

10 comments

W1: Basis for Selecting the Six Reward Dimensions and Redundancy

Reviewer concern: How were the six reward dimensions determined? Why were other social intelligence dimensions not selected?

The six dimensions are not an exhaustive taxonomy of social intelligence; rather, they are a process reward system designed for this paper's target problem—relationship trade-offs. SocialRL focuses not on all cognitive abilities of social intelligence, but on a specific problem in multi-turn social interaction: how an agent can maintain its relationship with the interlocutor while advancing its own goals. Confusing "a complete theoretical taxonomy of social intelligence" with "the operational design of process rewards in this paper" is a misreading of the positioning of this method. The original manuscript Section 3.3 (L277–L282) has already explained the theoretical basis for the six dimensions; Appendix E.3 has already verified the functional contributions of different dimensions through leave-one-dimension-out experiments. If even these already-stated design bases are still regarded as "unexplained," then we can only conclude that the review comments did not sufficiently absorb the content of the original manuscript.

The six dimensions are organized by function into three aspects:

  1. Goal-oriented: Goal Advancement measures actual goal progress, and Strategic Positioning measures whether the current strategy choice is appropriate.
  2. Relationship-oriented: Relational Attunement measures adaptation to the interlocutor's emotional and relational state, and Persona Consistency measures the consistency of the agent's own role and behavior.
  3. Interaction-validity: Contextual Coherence measures the semantic and logical consistency of the response with the dialogue history, and Turn Quality measures the naturalness, conciseness, and interactional appropriateness of the current response.

Leave-one-dimension-out results have already shown that different dimensions have different empirical contributions:

  • Full SocialRL achieves 63.3% Goal Achievement.
  • Removing Goal Advancement reduces it to 55.3%.
  • Removing Strategic Positioning reduces it to 56.1%.
  • Removing Relational Attunement reduces Relationship Change from 0.516 to 0.401.

We agree that Theory of Mind, Trustworthiness, Fairness, and similar constructs belong to social intelligence in a broad sense, but the existence of other social intelligence constructs does not mean that a reward system designed for a specific optimization objective must include all of them. Functional distinctions are not necessarily statistical independence, so we add a correlation analysis of the six-dimensional rewards:

DimensionGoal AdvancementStrategic PositioningRelational AttunementPersona ConsistencyContextual CoherenceTurn Quality
Goal Advancement1.0000.6370.3050.3450.3860.251
Strategic Positioning0.6371.0000.3510.3370.4200.262
Relational Attunement0.3050.3511.0000.3160.4180.207
Persona Consistency0.3450.3370.3161.0000.3250.157
Contextual Coherence0.3860.4200.4180.3251.0000.210
Turn Quality0.2510.2620.2070.1570.2101.000

On 366,552 process reward records, the average absolute Spearman correlation among the six-dimensional rewards is 0.329; Goal Advancement and Strategic Positioning have the strongest association at 0.631, while Contextual Coherence and Turn Quality have 0.210. Correlation does not equal redundancy; what truly needs to be judged is marginal contribution and training impact, and Appendix E.3 has already provided direct evidence.

W2: Relationship Between Dialogue Stage and Dynamic Reward Weights

Reviewer concern: The relationship between stage and weights is unclear, making it impossible to distinguish the improvement brought by stage awareness from that brought by context adaptation.

This concern is especially puzzling because the original manuscript has already clearly provided stage identification, weight generation, and weight constraint mechanisms, and has compared three fixed-weight schemes. SocialRL's stage awareness does not divide dialogue according to fixed turn ratios; according to Section 3.3 (L293–L300), the current interaction stage is jointly determined by the scenario, role information, social goal, and dialogue history. Early stage establishes common understanding, maintains the relationship, and identifies the other party's position; Mid stage advances specific goals, negotiates proposals, and handles disagreements; Late stage confirms results, completes commitments, and balances relationship closure. After the stage is determined, the reward model uses stage-weight priors as a basis, generates six-dimensional weights according to the current context, and undergoes whitelist checks, non-negativity constraints, upper-bound control, fallback for anomalous outputs, and normalization. To describe these already-stated mechanisms as "the relationship is unclear," we have to ask the reviewer to return to the specific line numbers in the original manuscript and read them again.

Appendix E.4 has already directly compared four weighting mechanisms:

Weighting StrategyGoal AchievementRelationship Change
Uniform58.70%0.472
Manual60.40%0.487
Grid60.90%0.468
Dynamic63.30%0.516

We agree that further isolating stage awareness and context adaptation can strengthen the evidence, so we add a 2×2 factorial experiment:

SettingStage AwarenessContext AdaptationGoal AchievementRelationship
Global Fixed Weights (uniform)NoNo58.70.472
Context-adaptive WeightsNoYes61.50.484
Stage-specific Fixed WeightsYesNo61.10.498
Full SocialRLYesYes63.30.516

This set of experiments directly answers the reviewer's so-called "cannot distinguish" question. If the advantage of dynamic weights is still considered lacking evidence, then the only explanation is that the review standard has become so high that it accepts no ablation experiment at all.

W3 & W4: Reward Hacking and Direct Human Evaluation

Reviewer concern: The process reward dimensions are similar to the SOTOPIA evaluation criteria, and performance gains may come from reward hacking.

We must make clear: the fact that training rewards and final evaluation focus on the same research objective does not constitute evidence of reward hacking; otherwise, any work whose research objective is goal achievement would be judged a priori as cheating. SocialRL's process rewards are turn-by-turn, fine-grained behavioral feedback, generating specific judgment criteria for the current situation based on six dimensions; Goal Achievement and Relationship Change are evaluations of the complete dialogue outcome. The two differ in evaluation granularity, specific judgment criteria, and evaluation target. Consistency of research constructs should not be equated with evaluation-label leakage or reward speculation. The original manuscript also conducts independent benchmark evaluation on AgentSense, whose scenario sources and task construction differ from SOTOPIA-π. To ignore these existing multi-benchmark results and then use "possible reward hacking" as a reason for a low score is unfair.

Of course, we agree that independent human evaluation is important, so we add:

Evaluation MethodImprovement in Goal AchievementImprovement in Relationship Quality
Human evaluation (SocialRL vs. Base)12.90.071

Compared with Base, SocialRL's human-evaluated Goal Achievement improvement is 12.9, and Relationship Quality improvement is 0.071. This shows that the improvement can be directly observed by humans.

Comments on the paper

1 comment

Dear authors, if you intend comments only for the AC/SAC (e.g., "Note to AC/SAC"), please select the proper comment visibility. - AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.