CARD: Controlled Agentic Reddit Discussions for Credit Card Simulation

CARD: Controlled Agentic Reddit Discussions for Credit Card Simulation

This paper introduces CARD, a framework for simulating realistic credit card discussion threads. The paper states: "We present CARD, a framework for simulating realistic credit card discussions. Rather than prescribing the specific content of each comment, CARD models how consumers communicate about credit card products. The planner organizes discussion structure and communication roles, the writer generates discussion content conditioned on the seed post and local discussion context, and the metric-guided revision loop calibrates the generated discussion collection toward the communication patterns observed in real consumer communities."

The paper identifies a gap in existing simulators: "Existing financial simulators mainly model borrowing behavior, trading decisions, and market dynamics [6, 7, 9, 18]. They do not generate the consumer discussion threads through which financial products are explained, questioned, and compared. Moreover, existing social simulators generate general online interactions or profile-conditioned discussion threads [21, 28, 30]. They do not model how consumers communicate about credit card products by exchanging recommendations, sharing personal experience, and discussing practical trade-offs [29]."

The framework operates in three stages: "The discussion planner specifies how the discussion should develop and assigns a financial role to each comment. The writer then generates comments that follow these roles and discuss the planned credit card issues, claims, rules, and trade-offs. The self-looped calibrator then revises all the generated discussion thread collection toward the linguistic, behavioral, and structural patterns observed in real credit card communities."

The task formulation defines a discussion thread as: Tp = p, (ci, ri) i=1 np, where ri identifies the parent of comment ci, either the root post or an earlier comment. The seed post defines the credit card product, consumer question, and financial context, and CARD generates the remaining discussion, including the reply structure, number of comments, and comment text.

The planner uses discussion-level controls (number of comments, reply depth, branch structure) and comment-level controls. Each comment plan is represented as: ai = (ri, di, li, fi, si, mi, qi, taui, oi), where these controls specify parent, depth, length, financial discussion function, financial claim stance, financial reasoning move, financial issue type, tone, and opening style. The paper notes: These controls specify how each comment participates in the discussion rather than what it should say.

The writer generates each comment conditioned on the seed post, its credit card product context, the comment plan, the generated parent comment, and selected previous comments. The writer uses context dropout and variation where Context dropout removes part of the previously generated discussion from the current writer's view, while context variation changes which earlier comments are emphasized in the prompt.

The self-looped calibrator improves one target metric at a time while treating all other metrics as protected and preserves the credit card discussion plan, such as the seed post, reply tree, financial discussion function, financial claim stance, and financial issue type. The revision process involves: "(a) the metric evaluator compares the generated and real collections across distributional metrics; (b) the controller selects comments that contribute to the current mismatch; (c) the reviser generates best-of-N candidate rewrites; and (d) the metric gate accepts candidates that improve the target metric while preserving the planned financial role of each comment."

The evaluation uses 12 metrics across four families: Lexical and semantic diversity (Self-BLEU, Self-BERTScore, Mean semantic cosine), Tone and disagreement (Hard disagreement rate, Politeness rate, Impoliteness rate, Neutral rate), Structure and length (Length CV, Average reply depth, Structural virality), and Story and emotion (Mean story probability, Emotion entropy).

Results show: "CARD with gpt-5.4-mini and deepseek-v4-flash achieves the highest non-rejection rate, with 91.7% of the metric–test pairs having p > 0.05. In comparison, the strongest SynthPAI configuration reaches 37.5%, while OASIS remains at or below 4.2%. The paper also reports: Across almost all metrics, CARD consistently produces smaller effect sizes and shorter distributional distances than OASIS and SynthPAI."

The human evaluation used a blind turing test where annotators were shown one real Reddit discussion and one generated discussion in random order and asked to identify which discussion was written by humans. The results show CARD achieves the highest proportion of generated discussions that were not identified, substantially outperforming both OASIS and SynthPAI.

The ablation on self-loop calibration shows: "Three of the four LLMs increase the number of metrics that pass the statistical tests, although the degree of improvement varies considerably across models. DeepSeek-V4-Flash provides the most efficient revision behavior... it retains all attempted revision rounds and improves the overall pass count from 6 to 11 metrics while generating only 26,770 candidate rewrites."

The paper concludes: "We introduce CARD, a framework for simulating credit card discussion threads that match how consumers communicate and interact in real online communities. CARD combines discussion thread planning, sequential comment generation, and metric-guided revision to control thread structure, discussion functions, tone, disagreement, and other communication patterns without copying the content of matched real threads."

Future work directions include: the framework can be adapted to other financial and non-financial discussion domains by defining domain-specific discussion functions and evaluation metrics and Future versions could add content-level objectives, including topic coverage, claim alignment, and factual consistency, while preserving the linguistic, behavioral, and structural realism studied in this work.

Improvements to AI systems based on this paper:

1. Hierarchical discussion control for domain-specific simulations

• Add a planner module that explicitly defines communication roles, reply structures, and interaction functions (e.g., recommendation, trade-off, personal experience) before generation.

• The improved system can generate structured, multi-turn discussions in any domain (finance, health, tech support) where each comment has a predefined social function, not just free-form text.

2. Metric-guided self-revision loop with protected constraints

• Implement a calibrator that iteratively evaluates generated outputs against target distributional metrics (e.g., tone, disagreement rate, structural virality) and rewrites only the offending comments while preserving the original plan (roles, reply tree, stances).

• The improved system can automatically align synthetic conversations with real-world statistical patterns (e.g., politeness, story-telling frequency, reply depth) without human intervention, reducing distributional distance to human data.

3. Context dropout and variation for robust long-context generation

• Use stochastic removal or re-weighting of previous context during generation to prevent over-reliance on early comments and encourage diverse, locally coherent replies.

• The improved system can produce more varied and realistic multi-turn dialogues, avoiding repetitive or overly deterministic responses in long threads.

4. Comment-level plan representation for controllable generation

• Encode each comment with explicit attributes (parent, depth, length, financial function, claim stance, reasoning move, issue type, tone, opening style) and condition generation on these controls.

• The improved system can generate discussions with fine-grained control over argumentation style, emotional tone, and structural role, enabling targeted simulation of specific user behaviors (e.g., skeptical consumers, expert advisors).

5. Protected-metric optimization for multi-objective realism

• During revision, optimize one target metric at a time while treating all others as protected (i.e., not degrading them).

• The improved system can balance multiple realism dimensions (e.g., lexical diversity vs. politeness vs. story-telling) without trade-offs, producing collections that pass statistical tests across all metric families simultaneously.

6. Domain-adaptive discussion function library

• Replace the fixed credit-card-specific functions with a configurable set of domain-specific discussion functions (e.g., diagnosis, troubleshooting, comparison) and corresponding evaluation metrics.

• The improved system can be rapidly deployed to new domains (e.g., medical forums, product reviews, legal Q&A) by redefining the planner’s role taxonomy and the calibrator’s target metrics.

7. Best-of-N candidate rewriting with metric gate

• Generate multiple candidate rewrites for a mismatched comment and accept only those that improve the target metric while preserving the planned financial role.

• The improved system can perform efficient, quality-controlled text revision, ensuring each edit contributes to global realism without breaking local coherence or role fidelity.

8. Human-blind Turing test optimization

• Train or fine-tune the generator to maximize the probability that human annotators cannot distinguish generated from real discussions, using feedback from blind evaluations as a reward signal.

• The improved system can produce synthetic conversations that are perceptually indistinguishable from human-written ones, useful for training data augmentation, social science research, and user-behavior modeling.