---
title: 'HumAIne-chatbot: Adaptive Conversational AI'
url: https://www.emergentmind.com/topics/humaine-chatbot
type: topic
---

# HumAIne-chatbot: Adaptive Conversational AI

{"command":["bash","-lc","python - <<'PY'\nimport urllib.request\nurl='http://export.arxiv.org/api/query?search_query=all:HumAIne-chatbot&start=0&max_results=3'\ntry:\n    print(urllib.request.urlopen(url, timeout=10).read(200).decode('utf-8','ignore'))\nexcept Exception as e:\n    print('ERR', e)\nPY"]}
{"command":["bash","-lc","echo hi"]}
HumAIne-chatbot is an AI-driven conversational agent that personalizes responses through a novel user profiling framework. It is pre-trained on a diverse set of GPT-generated virtual personas to establish a broad prior over user types, and during live interactions an online reinforcement learning agent refines per-user models by combining implicit signals such as typing speed, sentiment, and engagement duration with explicit feedback such as likes and dislikes. The resulting profile dynamically informs the dialogue policy, enabling real-time adaptation of both content and style [2509.04303].

## 1. Conceptual placement within humane conversational AI

HumAIne-chatbot belongs to a line of research that treats “humaneness” in dialogue systems not merely as fluency, but as the capacity to adapt to user state, sustain relational continuity, and modulate interaction style in context. Survey work on human-chatbot interaction design organizes relevant social characteristics into Conversational Intelligence, Social Intelligence, and Personification, including proactivity, conscientiousness, communicability, emotional intelligence, personalization, identity, and personality [1904.02743]. Within that framing, HumAIne-chatbot is centered most explicitly on personalization, but its prompt adaptation axes—language complexity, detail level, domain-knowledge calibration, and conversational style—also connect to thoroughness, communicability, and personality as design dimensions [2509.04303].

Earlier systems operationalized related goals through different mechanisms. Emora emphasized experience-sharing, inquisitive follow-up, a hierarchical state machine, and reciprocal self-disclosure, reporting that “Life” handlers scored on average 3.85 while traditional info/opinion handlers scored 3.79, with a paired \(t\)-test giving \(p<0.10\) [2009.04617]. Audrey pursued informational, personal, and relational engagement through emotion detection, a Bayesian Personal Understanding Module, and hybrid response generation, achieving an average cumulative rating of 3.25 on a 1–5 Likert scale during the semi-finals period [2011.05910]. MILABOT, by contrast, treated response selection as a reinforcement-learning problem over an ensemble of 22 response models and showed that learned selection policies could outperform baseline response orchestration in Alexa Prize A/B tests [1709.02349].

This progression suggests that HumAIne-chatbot does not introduce personalization into a vacuum. Rather, it shifts the locus of personalization from hand-authored dialogue states, profile variables, or ensemble reranking toward a two-stage profiling regime in which a latent user embedding is learned from synthetic pre-training and then updated online with reinforcement learning [2509.04303]. A plausible implication is that HumAIne-chatbot is best understood as a profile-conditioned policy adaptation system rather than solely as a response generator.

## 2. System architecture and runtime data flow

The HumAIne-chatbot is organized into six major modules [2509.04303].

| Module | Role |
|---|---|
| User Interface + Interaction Tracking Engine | Manages turn-taking, renders LLM responses, and captures timestamps, keystroke events, and session start/end |
| Personalization Metrics Collector | Computes low-level features and outputs a feature vector \(x_t\in\mathbb{R}^d\) at each turn |
| Dialogue History & Analysis Module | Maintains context window and extracts sentiment scores, grammatical error rates, and language-complexity indices |
| AI-Driven User Profiler | Performs supervised pre-training on GPT-generated virtual personas and online RL adaptation |
| Prompt Manager | Enriches or rewrites the raw LLM prompt along four personalization axes |
| LLM/API/RAG System | Generates the final textual response and can call external APIs or vector-store retrieval |

The data flow is specified as: UI \(\rightarrow\) Tracking Engine \(\rightarrow\) Metrics Collector \(\rightarrow\) \{Dialogue History, User Profiler\} \(\rightarrow\) Prompt Manager \(\rightarrow\) LLM \(\rightarrow\) UI. Concurrently, the RL agent ingests \(\mathbf{u}_t\) and \(\mathbf{s}_t\), issues actions in the form of policy adjustments, and receives rewards \(r_t\) to update both the User Profiler and Dialogue Policy online [2509.04303].

Two architectural features are central. First, personalization is not implemented only at the response layer. The system explicitly separates signal capture, latent profile estimation, prompt rewriting, and response generation. Second, the LLM/API/RAG block is downstream of the profiler and prompt manager, so the system’s adaptation mechanism acts by conditioning prompts rather than by retraining the base LLM during the dialogue loop [2509.04303]. This is structurally different from systems such as Audrey, whose dialog policy selects among template, retrieval, neural, and hybrid generators using NLU outputs and a user model \(U_t\) [2011.05910].

In broader terms, the architecture combines interaction telemetry with text analysis. The Interaction Tracking Engine captures raw interaction signals including timestamps for prompt and response, keystroke events, and session start/end, while the Dialogue History & Analysis Module extracts text-based metrics such as sentiment scores, grammatical error rates, and language-complexity indices [2509.04303]. This coupling indicates that HumAIne-chatbot treats personalization as a joint function of behavioral and linguistic evidence, rather than as an inference from utterance semantics alone.

## 3. User profiling framework and latent representation

At each turn \(t\), the system computes a feature vector whose components include typing speed, response time, session duration increment, sentiment score, grammatical mistake frequency, language complexity, explicit feedback, and survey-based satisfaction [2509.04303]. The individual definitions are given explicitly:

- Typing Speed: \(TS_t = \tfrac{C_t}{t_{\mathrm{resp},t}-t_{\mathrm{start},t}}\)
- Response Time: \(RT_t = t_{\mathrm{resp},t}-t_{\mathrm{prompt},t}\)
- Session Duration Increment: \(\Delta SD_t\)
- Sentiment Score: \(SS_t = S(\text{UserMsg}_t)\)
- Grammatical Mistake Frequency: \(GMF_t = \tfrac{E_t}{W_t}\)
- Language Complexity: \(CL_t = \alpha\,ASL_t + \beta\,TTR_t\)
- Explicit Feedback (“likes”): \(FS_t = \frac{\sum_{i=1}^n L_i}{n}\)
- Survey-Based Satisfaction: \(SBS_t = \tfrac1M\sum_{j=1}^M R_j\)

These raw features are normalized and passed through a small feed-forward network to produce the latent profile embedding \(\mathbf{u}_t\) [2509.04303].

The profiling framework has two phases. In Phase I, supervised pre-training uses a synthetic corpus of GPT-generated “virtual personas,” consisting of backstories and multi-turn dialogues with known ground-truth profiles. The supervised profiler learns a mapping
\[
\phi_\mathrm{supv}: (\text{dialogue cues})\to \mathbf{u}\in\mathcal{U}.
\]
In Phase II, online RL adaptation refines \(\mathbf{u}_t\) per real-user interaction under a reward signal \(r_t\) [2509.04303].

The pre-training cohort is derived from the Anthology of GPT-generated backstories (\textit{Moon et al. 2024}) and synthesizes \(N=50\) high-diversity virtual personas. Each persona backstory encodes demographics, communication style, expertise level, and behavioral traits. Multi-turn dialogues with these personas yield a labeled corpus \(\{(x_{1:T}^{(i)},\mathbf{u}^{(i)})\}_{i=1}^{N}\), and the supervised profiler is trained by minimizing
\[
\mathcal{L}_\mathrm{sup} = \sum_{i=1}^N \bigl\|\phi_\mathrm{supv}(x_{1:T}^{(i)}) - \mathbf{u}^{(i)}\bigr\|^2.
\]
[2509.04303]

Relative to prior personalization mechanisms, this profile representation is more explicitly latent and continuous. Audrey represents a user model \(U_t\) as Bayesian marginals over binary attributes such as likes_books and likes_sports, updating posteriors after proxy questions [2011.05910]. The transformer-based blueprint in “Deep Learning Based Chatbot Models” instead augments token representations with persona and mood embeddings \(p\) and \(m\) [1908.08835]. HumAIne-chatbot’s \(\mathbf{u}_t\) occupies an intermediate conceptual position: it is a latent embedding estimated from behavioral and linguistic features, then used to control prompt adaptation rather than directly injected into every transformer layer [2509.04303].

## 4. Reinforcement learning and dialogue policy adaptation

HumAIne-chatbot formulates online personalization as a Markov Decision Process. The state space is
\[
\mathbf{s}_t = \bigl[\mathbf{u}_t, h_t\bigr],
\]
where \(h_t\) is a fixed-length embedding of the dialogue history. The action space consists of parameter adjustments to the LLM prompt along each personalization axis—complexity, style, and detail—with each axis discretized into \(K\) levels, giving \(\lvert\mathcal{A}\rvert = K^4\). The policy \(\pi_\theta(a\mid s)\) is a multi-layer neural network actor that outputs a categorical distribution over \(\mathcal{A}\) [2509.04303].

The scalar reward combines implicit engagement and explicit feedback:
\[
r_t \;=\; \omega_1\,\mathrm{turn\_length}_t \;+\;\omega_2\,FS_t \;+\;\omega_3\,SS_t,
\]
where “turn_length” is the number of user tokens in the next turn. Over an episode of length \(T\), the cumulative return is
\[
R(\tau) \;=\;\sum_{t=1}^T r_t.
\]
The learning algorithm is Proximal Policy Optimization actor-critic. The paper specifies the policy-gradient estimate,
\[
\nabla_\theta J(\theta)
= \mathbb{E}_{\pi_\theta}\Bigl[\nabla_\theta\log\pi_\theta(a_t|s_t)\,A_t\Bigr],
\]
the PPO-clip surrogate objective, and advantage estimation via Generalized Advantage Estimation. After each turn, the agent collects \((s_t,a_t,r_t,s_{t+1})\) and performs minibatch updates to \(\theta\). These updates happen asynchronously alongside the live conversation loop, ensuring real-time adaptation without blocking the UI [2509.04303].

Dialogue policy adaptation is implemented through a mixture between a generic baseline policy \(\pi_0\) and the personalized policy \(\pi_\theta\):
\[
\pi_{\mathrm{mix}}(a|s)
= \lambda(s)\,\pi_\theta(a|s) + \bigl(1-\lambda(s)\bigr)\,\pi_0(a|s),
\]
with
\[
\lambda(s) = \sigma\bigl(w^\top \mathbf{u}_t + b\bigr).
\]
As \(\mathbf{u}_t\) becomes more confident, \(\lambda(s)\to 1\), yielding more personalized responses [2509.04303].

This gating mechanism is significant because it formalizes a gradual transition from generic to individualized behavior. Earlier RL chatbot work often optimized response selection over fixed candidate sets, as in MILABOT’s policy over ensemble outputs [1709.02349], or optimized steering toward an intention under a coherence regularizer, as in “Put Chatbot into Its Interlocutor’s Shoes” [2103.16429]. HumAIne-chatbot instead uses RL to tune prompt-level personalization parameters while preserving a fallback baseline through \(\pi_{\mathrm{mix}}\) [2509.04303]. This suggests a design choice oriented toward stability: the personalized policy is not allowed to replace the baseline abruptly.

## 5. Experimental design, metrics, and empirical results

The evaluation uses a synthetic persona cohort of fifty virtual personas covering ages 18–65+ in six brackets, education levels spanning HS, SC, BA, MA, PhD, and professional cert., and ten topical domains including careers, finance, health, and technology. Each persona interacted 3 times per system (personalized vs. non-personalized), producing 150 sessions [2509.04303].

Three key metrics, all in \([0,1]\), are defined:

- User Satisfaction:
  \[
  \mathrm{Sat}
  = \frac{1}{N}\sum_{i=1}^N FS_i
  \quad\text{or}\quad
  \overline{SBS}
  \]
- Personalization Accuracy:
  \[
  \mathrm{PAcc}
  = 1 - \frac{1}{N}\sum_{i=1}^N \bigl|\mathrm{target\_level}_i - \mathrm{actual\_level}_i\bigr|
  \]
- Task Achievement:
  binary success/failure on a defined subtask, averaged across sessions:
  \[
  \mathrm{TaskAch}=\tfrac{\#\,\mathrm{successes}}{\mathrm{total\,sessions}}.
  \]

The statistical comparison between Control and Experimental groups uses a two-sample \(t\)-test on mean Satisfaction, Mann–Whitney \(U\) for non-normal distributions, one-way ANOVA for multi-domain comparisons, and effect size via Cohen’s \(d\) [2509.04303].

| Group | Mean Sat | SD |
|---|---:|---:|
| Control | 0.119 | 0.050 |
| Personalized | 0.173 | 0.071 |

For this comparison, the reported \(p\)-value is \(<0.001\) and Cohen’s \(d\) is \(0.94\) [2509.04303].

The primary outcomes are reported as follows. Mean satisfaction rose from 0.119 to 0.173 \((+45\%,\, p<0.001,\, d=0.94)\). Secondary metrics improved as well: relevance \(+0.046\), personalization accuracy \(+0.145\), style match \(+0.053\), and task achievement \(+0.056\). Subgroup gains were uniform across domains, with the largest gains in “professional networking” \((+50.3\%)\) and “creative projects” \((+50.9\%)\). An ablation found that removing explicit feedback from the reward led to a 20% drop in satisfaction improvement [2509.04303].

These results support the paper’s central claim that two-stage profiling and real-time policy adaptation deliver statistically significant gains in user satisfaction, personalization accuracy, and task success [2509.04303]. At the same time, the reported limitations are explicit: virtual personas cannot capture full human variability, sessions were short with average 4.13 min, and real-world, long-horizon validation remains future work [2509.04303]. For an encyclopedic reading, this matters because the current evidence demonstrates efficacy under controlled synthetic-persona evaluation rather than ecological validity in live deployment.

## 6. Human-likeness, social intelligence, and boundary conditions

The term “humane” in chatbot research does not imply a single measurable property. Work on perceptions of social intelligence in simulated human-chatbot interactions found that people have great difficulties perceiving elements of social intelligence in text, even though they may find anthropomorphic behaviour pleasant and positive for the naturalness of a dialogue; the same behaviour may also be perceived as frightening and unsuitable when expressed by an artificial agent in the wrong way or at the wrong time [2104.13823]. This is directly relevant to HumAIne-chatbot because its prompt manager explicitly adjusts conversational style and detail level [2509.04303]. A plausible implication is that stronger personalization does not automatically entail better perceived empathy or authenticity.

Related systems illustrate distinct operationalizations of human-likeness. Emora relies on follow-up questions, stored slots, and reciprocal self-disclosure to encourage experience-sharing, and users frequently commented that it “felt like a friend” [2009.04617]. Audrey combines emotion detection, Bayesian user profiling, RL topic sequencing, and SSML prosody to support informational, personal, and relational conversation [2011.05910]. HUMA, in asynchronous group chat, shows that realistic typing delays, optional participation via “Keep Silent,” interruption modelling, and diverse strategies can make an AI facilitator difficult to distinguish from a human, with AI and human community managers classified as human at near-chance rates [2511.17315]. HumAIne-chatbot differs from all three by concentrating on per-user profiling and prompt-level policy adaptation rather than state-machine inquisitiveness, generator orchestration, or event-driven group facilitation [2509.04303].

A further boundary condition concerns user influence and steering. “Put Chatbot into Its Interlocutor’s Shoes” demonstrates that a chatbot can be trained to induce the interlocutor to produce longer, more emotional, or vocabulary-constrained replies, but human judges found some steering attempts less relevant even when target behaviors shifted [2103.16429]. HumAIne-chatbot’s reward also depends partly on the next user turn, through turn length and sentiment [2509.04303]. This suggests that personalization and conversational steering are adjacent research problems: both optimize for downstream user behavior, but neither guarantees that such optimization will always be perceived as appropriate.

## 7. Risks, limitations, and future directions

HumAIne-chatbot’s own limitations are methodological: virtual personas cannot capture full human variability, sessions were short, and future work is directed toward multimodal signals, live-user pilots, and long-term engagement rewards [2509.04303]. Those limitations intersect with broader concerns in the literature about long-horizon psychosocial effects. A four-week randomized controlled study of chatbot use found that higher daily usage correlated with higher loneliness, dependence, and problematic use, and lower socialization across modalities and conversation types [2503.17473]. It also found that personal topics slightly increased loneliness but tended to lower emotional dependence compared with open-ended conversations, while non-personal topics were associated with greater dependence among heavy users [2503.17473].

Those findings do not evaluate HumAIne-chatbot directly, but they delimit the space into which a real-world personalized agent would be deployed. Because HumAIne-chatbot incorporates explicit feedback, engagement duration, and live policy adaptation [2509.04303], a plausible implication is that future real-user deployments will need guardrails not only for effectiveness but also for over-reliance. The RCT literature recommends dynamically scaling back emotional engagement as usage time exceeds a threshold, shifting conversation topics based on cumulative use, and encouraging offline social connections when signs of emotional dependence rise [2503.17473].

Future technical extensions are also legible from adjacent research. The transformer-based blueprint of persona and mood conditioning proposes injecting persona and mood embeddings both into input representations and attention projections [1908.08835]. HUMA points toward event-driven, interrupt-aware architectures and realistic temporal signatures for more natural interaction in multi-user settings [2511.17315]. Emora and Audrey show the continued relevance of ontology-backed high-precision triggers, state tracking, and socially aware generation modules [2009.04617; 2011.05910]. Accordingly, HumAIne-chatbot can be viewed as a platform-level personalization framework whose next research steps are likely to involve richer modalities, longer horizons, and stronger safety constraints, rather than merely larger language models [2509.04303].

Source: https://www.emergentmind.com/topics/humaine-chatbot