SocialCoach: Adaptive Social Coaching

• SocialCoach is a design pattern for AI-mediated coaching that supports social skills training and well-being through adaptive feedback loops.
• It integrates diverse implementations—from virtual companions and tele-operated robots to multi-agent simulations—using modular architectures and policy optimization.
• Key evaluation metrics span conversational quality, engagement, and emergent social network dynamics, reflecting both innovations and limitations in current research.

In the cited literature, SocialCoach denotes a recurring class of AI-mediated coaching systems for social interaction, communication, and well-being rather than a single canonical platform. Under this label, researchers describe an instructor-guided tutoring environment for social-skills rehearsal, an educational virtual companion for collective well-being in social media, a tele-operated robot coach for mindfulness, a conversational chatbot for communication practice, a multi-agent LLM framework for studying emergent social ties, and an RL-based agentic tutor for personalized social-skill learning (Guevarra et al., 16 Jan 2025, Ognibene et al., 2021, Bodala et al., 2020, Jeon et al., 2023, Schneider et al., 22 Oct 2025, Wang et al., 2 Jun 2026). Across these instantiations, the common structure is an adaptive interaction loop in which scenarios, recommendations, or social actions are generated or selected, behavior is assessed, and subsequent interaction is modified in response.

1. Scope of the term

The term spans several distinct research programs with different task definitions, user populations, and optimization targets.

Instantiation	Application setting	Distinctive mechanism
GLOSS-style SocialCoach (Guevarra et al., 16 Jan 2025)	Social skills training	Instructor+LLM scenario co-creation, dynamic branching, immediate feedback
Social Media Virtual Companion (Ognibene et al., 2021)	Educationally managed social-media community	CWB-aware recommendation with hierarchical reinforcement learning
Tele-operated robot coach (Bodala et al., 2020)	Mindfulness and wellbeing	Human coach teleoperates SoftBank Pepper robot
SpeakEasy (Jeon et al., 2023)	College communication practice	Spoken conversation analysis with rule-based feedback
Multi-agent SocialCoach (Schneider et al., 22 Oct 2025)	LLM social simulation	Coaching signal for emergent social ties
Personalized SocialCoach (Wang et al., 2 Jun 2026)	Social skill learning at scale	Theory-to-practice corpus plus RL-based adaptive practice scheduling

This breadth is important for interpretation. In some works, SocialCoach is a direct tutoring or coaching system for human learners; in others, it is an infrastructural layer between users and a platform; in still others, it is an experimental framework for agent populations. The target variables likewise differ: individual subskills such as politeness or clarity, community-level Collective Well-Being (CWB), longitudinal relaxation benefit, conversational quality metrics, or emergent network structure. A plausible implication is that SocialCoach is best understood as a design pattern for adaptive social coaching rather than a single software artifact.

2. Architectural patterns

The GLOSS-style implementation organizes SocialCoach into four core modules: a Scenario Generator, a Student Interface, a Feedback Engine, and a Performance Visualizer. Instructors define or compose scenarios through a web GUI, while the system maintains a directed graph $G=(V,E)$ whose nodes are states and whose labeled edges represent possible replies or intents. The student interface presents the current node’s avatar and prompt, accepts text or speech, and invokes a Branching Engine. The feedback engine computes a score and verbal feedback immediately after each utterance, and the visualizer logs the path through $G=(V,E)$ and renders an interactive graph with edge-thickness proportional to number of visits and time-series plots of metrics (Guevarra et al., 16 Jan 2025).

The social-media variant places SocialCoach as a thin layer between a standard platform and a closed educationally managed student community. Its modules include an External Social Media Connector, Content Analyzer & Threat Detector, Collective Well-Being Evaluator, CWB-Aware Recommender System, Educational Module & Narrative Scripts, Companion UI, and Educator Dashboard. The robot-coach variant is instead a teleoperation stack centered on SoftBank Pepper robot running NAOqi OS, with a teleoperator station consisting of PC + HTC Vive headset, webcam, microphone/headset, Jitsi SIP gateway, secure LAN router, and encrypted local networking for motion, vision, and audio transfer (Ognibene et al., 2021, Bodala et al., 2020).

SpeakEasy adopts a more conventional conversational pipeline: Speech Capture, ASR & Transcription, Natural Language Understanding, Metric Computation Engine, Feedback Generator, TTS & Playback, and an Adaptation Loop driven by post-session survey responses stored in a persistent pickle. The 2026 personalized system further modularizes the problem into knowledge extraction, user profiling, prescription policy, simulation, assessment, and tutoring, while the multi-agent simulation decomposes each round into Plan, Execute, Vote, and Reweight phases with memory-conditioned prompting (Jeon et al., 2023, Wang et al., 2 Jun 2026, Schneider et al., 22 Oct 2025).

These architectures differ in substrate—web tutoring, recommender mediation, embodied telepresence, speech analytics, or agent simulation—but they converge on explicit modular decomposition and online adaptation.

3. Adaptation, branching, and policy optimization

In the tutoring formulation, the student–LLM dialogue is modeled as a branching Markov decision process. For a state $s_t$ with outgoing edges $E(s_t)=\{e_{t,1},\dots,e_{t,k}\}$, each edge carries an exemplar $\phi(e_{t,i})$. The system computes

$$p_i=\mathrm{cosine}\bigl(\mathrm{Embed}(a_t),\,\mathrm{Embed}(\phi(e_{t,i}))\bigr),$$

and follows an existing edge if $\max_i p_i\ge\tau$; otherwise, it asks the LLM to synthesize a new successor. The same framework also presents an equivalent probabilistic interpretation in which residual mass is assigned to the “LLM $\rightarrow$ new node” outcome (Guevarra et al., 16 Jan 2025).

The social-media system formalizes adaptation through a community objective. For each user $u$ and aspect $a$, the local contribution is

$G=(V,E)$0

and community CWB is aggregated with a weighted $G=(V,E)$1 norm. Recommendation then optimizes incremental well-being,

$G=(V,E)$2

with engagement retained only as a secondary constraint. At the policy level, the paper models the problem as an SMDP with a two-level options hierarchy in which the top level selects an educational objective or narrative script and the lower level schedules actions such as re-ranking posts, inserting a mini-game or quiz, or showing a nudge banner (Ognibene et al., 2021).

The 2026 personalized SocialCoach also casts selection as an MDP, but the action is a structured prescription produced by a PrescriptionAgent, then processed by RetrievalAgent and AdaptationAgent. Reward in the learner simulator is

$G=(V,E)$3

with $G=(V,E)$4 for empty retrievals, and the policy is optimized via the ROLL multi-turn agentic RL library. The state includes personal information, behavioral history, and a proficiency vector over CASEL competencies and granular skills (Wang et al., 2 Jun 2026).

The multi-agent simulation generalizes adaptation from human tutoring to social dynamics. Each agent $G=(V,E)$5 plans from

$G=(V,E)$6

where memory is partitioned into conversation, relationship, and opinion buffers. A specialized coach returns bullet-point tips

$G=(V,E)$7

and the agent maximizes a compositional reward over SOC, INF, PRE, COORD, and EMO components. Tie strengths are then updated endogenously through an adjacency matrix with strengthening and decay rules (Schneider et al., 22 Oct 2025).

Taken together, these formulations show that “adaptation” in SocialCoach research ranges from local branch selection to hierarchical recommendation, simulated practice scheduling, and coached multi-agent behavior.

4. Assessment, scoring, and feedback

The GLOSS-derived scoring model blends three signals into each instantaneous score vector: embedding similarity to ideal responses, rubric-based feature scores, and an LLM classification or regression head. For each skill dimension $G=(V,E)$8,

$G=(V,E)$9

These scalar scores are normalized into a 0–100 scale and fed back in a single sentence. Post hoc analysis computes Turn-Taking Frequency, Politeness Score, Engagement Index, and Correction Uptake Rate, and overlays them on the narrative graph; edge-thickness encodes visit counts and node-color encodes average feedback score upon entry (Guevarra et al., 16 Jan 2025).

SpeakEasy operationalizes conversational competence through a different metric suite. It computes Speech Rate with target $s_t$0 WPM, Filler Word Ratio with threshold $s_t$1, Question-Asking Ratio with desired $s_t$2, Awkward Transition Score with target $s_t$3, Type-Token Ratio with healthy conversation at $s_t$4, Verbal Tic Frequency, and Language Style Matching with target $s_t$5. Feedback is rule based: the system assembles the top deviations from target into a short personalized paragraph, using the stored user name and templated sentences (Jeon et al., 2023).

The 2026 system adds a causality-oriented layer. After immersive role-play, a DiagnosisAgent extracts positive and negative behaviors, an AttributionAgent classifies errors into acquisition deficits or performance deficits, and a ProfilingAgent updates the learner’s proficiency vector. Reflective tutoring is then generated from retrieved theories and cases through Socratic questioning: $s_t$6 This design explicitly targets the “knowing-doing gap” by coupling practice traces to theory-grounded reflection (Wang et al., 2 Jun 2026).

A central distinction across systems is therefore the level at which assessment is defined. Some SocialCoach variants score single utterances, others aggregate conversation-session statistics, others quantify community well-being, and others assess social behavior only indirectly through emergent graph properties.

5. Human oversight, pedagogy, and ethical structure

A persistent misconception is that SocialCoach systems are uniformly autonomous. The literature instead repeatedly centers human authority. In the GLOSS-style interface, instructors need no coding: they write in plain English, click commands such as “Expand this node” or “Generate alternative responses,” receive candidate continuations, and accept or edit them through the GUI; all generation, editing, and versioning occur with “no JSON, no YAML, no Git” (Guevarra et al., 16 Jan 2025).

The social-media companion makes educator involvement even more explicit. Experts and educators use the Companion in classroom-based educational activities, guide the definition of CWB, provide a hierarchical structure of learning strategies, objectives and activities, act as moderators of direct conflicts, and monitor ethical and educational issues beyond the intelligent agent’s competence and control. They also assign the weights $s_t$7, decide $s_t$8 in the $s_t$9 aggregator, and can terminate a narrative script, inject new training data, or override recommendations when moderation alerts arise. The same work foregrounds nudging, boosting, social-emotional learning, human-centered design, privacy & transparency, and review of over-censorship, algorithmic bias, and unintended disempowerment (Ognibene et al., 2021).

The robot-coach system is explicitly human-operated rather than autonomous. An experienced human coach conducts sessions virtually by replicating upper body and head pose in real time, viewing the world from the robot’s perspective, and talking and listening through the robot. This arrangement shifts the research question from fully automated coaching to mediated embodiment and longitudinal acceptance (Bodala et al., 2020).

The 2026 personalized system likewise grounds its knowledge corpus in “diverse expert sources” and organizes it into strategic theories, illustrative cases, and practical scenarios. This suggests a pedagogical stance in which LLMs do not replace expert curriculum structure but operationalize and adapt it (Wang et al., 2 Jun 2026).

6. Evaluation results, limitations, and research trajectory

Reported evaluation protocols are heterogeneous, spanning controlled studies, classroom pilots, longitudinal human–robot interaction trials, simulation benchmarks, and synthetic-user experiments.

System	Evaluation setting	Reported outcomes
GLOSS-style SocialCoach (Guevarra et al., 16 Jan 2025)	Controlled study design proposed with $E(s_t)=\{e_{t,1},\dots,e_{t,k}\}$0 undergraduates	Findings are described as “hypothetical but typical of GLOSS-style results”: +1.2 vs. +0.3 rubric gain, $E(s_t)=\{e_{t,1},\dots,e_{t,k}\}$1 from 0.62 to 0.82, $E(s_t)=\{e_{t,1},\dots,e_{t,k}\}$2, co-creation time down 40%
Social Media Virtual Companion (Ognibene et al., 2021)	Threat detection, recommender simulation, classroom pilots	FakeNewsNet $E(s_t)=\{e_{t,1},\dots,e_{t,k}\}$3 vs. 0.78 baseline; macro-$E(s_t)=\{e_{t,1},\dots,e_{t,k}\}$4 in SemEval-style tasks; max-diversity reached $E(s_t)=\{e_{t,1},\dots,e_{t,k}\}$5–$E(s_t)=\{e_{t,1},\dots,e_{t,k}\}$6 higher long-run feed diversity; 30–40% higher post-test awareness quizzes; fake-image detection 45%$E(s_t)=\{e_{t,1},\dots,e_{t,k}\}$770%
Tele-operated robot coach (Bodala et al., 2020)	5-week longitudinal study with 18 participants	HC means remained $E(s_t)=\{e_{t,1},\dots,e_{t,k}\}$8; RC motion and conversation showed significant positive slopes; HC and RC both had end $E(s_t)=\{e_{t,1},\dots,e_{t,k}\}$9 start relaxation benefit
SpeakEasy (Jeon et al., 2023)	$\phi(e_{t,i})$0 student sessions	“Met expectations” and “Left impression” mean $\phi(e_{t,i})$1; comprehensibility $\phi(e_{t,i})$2; on-topic $\phi(e_{t,i})$3; filler ratio $\phi(e_{t,i})$4, reduced to $\phi(e_{t,i})$5 across repeat users in pilot
Multi-agent SocialCoach (Schneider et al., 22 Oct 2025)	30 agents, 15 rounds, topic: climate change	Coaching accelerated early-round gains for INF, EMO, especially COORD; LLM-based tie scoring produced $\phi(e_{t,i})$6 within empirical ranges; coached degree distributions shifted rightward
Personalized SocialCoach (Wang et al., 2 Jun 2026)	Synthetic-profile simulation and early EQoach deployment	$\phi(e_{t,i})$7, $\phi(e_{t,i})$8, $\phi(e_{t,i})$9; Engagement 4.38 for RL-fine-tuned SocialCoach; human–judge correlations included Traceability $$p_i=\mathrm{cosine}\bigl(\mathrm{Embed}(a_t),\,\mathrm{Embed}(\phi(e_{t,i}))\bigr),$$0 and Learning Gain $$p_i=\mathrm{cosine}\bigl(\mathrm{Embed}(a_t),\,\mathrm{Embed}(\phi(e_{t,i}))\bigr),$$1; early user ratings included Reflective Depth $$p_i=\mathrm{cosine}\bigl(\mathrm{Embed}(a_t),\,\mathrm{Embed}(\phi(e_{t,i}))\bigr),$$2

The limitations are equally varied. The robot study reports higher NASA TLX workload for the robot-coach condition. SpeakEasy identifies limited NLU, lack of automated emotion recognition, and static thresholds. The multi-agent framework acknowledges small scale, short horizon, empty-graph start, single topic, and limited replications. The 2026 system notes that one-shot LLM rewards are an imperfect proxy for real user outcomes, that simulated persona pools may not capture real-world diversity or emergent behaviors, and that assessments have not yet been validated in large-scale, long-term field deployments (Bodala et al., 2020, Jeon et al., 2023, Schneider et al., 22 Oct 2025, Wang et al., 2 Jun 2026).

A plausible synthesis is that SocialCoach research is moving from narrowly scoped interaction loops toward broader pedagogical and social ecosystems: richer knowledge corpora, explicit causal assessment, human-governed objective setting, and increasingly formal policy-learning machinery. At the same time, the literature remains divided on evaluation granularity and external validity, with simulated rewards, pilot deployments, and longitudinal field evidence still unevenly distributed across implementations.