AuraSight: Generating Realistic Social Media Data

Bianca N. Y. Kang; Kathleen M. Carley; Lynnette Hui Xian Ng

REVIEW 4 major objections 4 minor 14 references

AuraSight: Generating Realistic Social Media Data

T0 review · 4 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The AESOP-SynSM pipeline can turn a hand-authored fictional conflict into a multi-day, quasi-realistic X dataset, complete with replies, retweets, quotes, bots, and dredgers.

desk verdict A careful, well-documented pipeline report whose central 'realistic data' claim is not backed by any measurement; useful as a recipe, not as evidence of realism. read the letter →

arxiv 2509.08927 v1 pith:G27SXLHB submitted 2025-09-10 cs.CY

Lynnette Hui Xian Ng , Bianca N. Y. Kang , Kathleen M. Carley This is my paper

classification cs.CY

keywords socialmediasimulationsyntheticdataagent-basedmodelinglargelanguagemodelsAuraSightAESOP-SynSMbotclassesdredgers

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that a hybrid approach—combining agent-based modeling for who talks to whom and a large language model for what they say—can generate a 'quasi-realistic' multi-day X (Twitter) dataset for a fictional pop-culture conflict. It walks through the full AuraSight scenario: a star from one country wins another country's national finals, and the ensuing fan, nationalist, bot, and dredger reactions are simulated hour by hour over three days. The authors claim this is useful as training material for studying emergent network behavior, because the underlying agents and narratives are fully known and controllable. They frame the work as an end-to-end demonstration of the AESOP-SynSM pipeline, including new agent types not in earlier builds. A sympathetic reader would take the paper as establishing that this pipeline makes scenario construction accessible to non-technical designers and produces plausible-looking Twitter-API-formatted output.

What carries the argument

The load-bearing mechanism is the dual division of labor between scenario design and simulation. In AESOP, narrative descriptions are the most important field because they are what the LLM receives; specifying stance explicitly prevents actors from flipping sides. In SynSM, group membership is the only way to guarantee interaction—agents cannot organically form links between groups—and actors in the same group as full members can reply/retweet/quote one another, leaders attract more interaction, and sources are cited but do not interact. The actual text is generated by an LLM using a fixed prompt structure: system prompt, agent persona, narrative, last messages of the same narrative, BEND ma

What would settle it

Collect a real X dataset around a comparable multi-day identity conflict, then compare distributions of tweet/reply/retweet/quote counts, retweet chain lengths, hashtag sharing, and degree distributions against AuraSight output; a clear statistical divergence on these metrics would refute the quasi-realistic claim. Simpler: if a blind test lets human annotators distinguish generated from real tweets at rates far above chance, the content-realism claim fails.

Watch

Extended reading notes

Core claim

The central claim is that AESOP-SynSM can take a hand-authored scenario and turn it into a realistic-looking X dataset. AESOP supplies the scenario structure: entities, groups, events, topics, narratives, narrative ratios, and agent roles (full member, leader, source). SynSM then activates agents according to peak hours and a Poisson posting distribution, constructs interaction networks using a 60% preferential-attachment / 30% follow-the-leader / 10% random mix, and hands prompts to GPT-4.1-mini to generate tweet text in context, using previous messages as few-shot examples and BEND maneuvers to shape information operations. The output is formatted as Twitter API V1 JSON and can be imported

Load-bearing premise

The claim that the data are 'quasi-realistic' rests on the assumption that hand-authored narratives, group stances, and pipeline parameters (60/30/10 attachment, 2x bot posting, 3 Randos per tweet) are enough to make LLM-generated tweets statistically similar to real social media content—a similarity the paper does not yet measure.

Editorial extensions

If this is right

If the claim holds, anyone can design a custom social media scenario with specified groups, conflicts, and bot ecosystems and get a shareable, analyzable dataset without scraping real platforms.
The Twitter API V1 JSON output means standard social-media analysis and visualization tools can be applied directly to the synthetic data.
AuraSight itself is intended as educational material for network-science classes, offering a manageable dataset with a known ground truth about agents and narratives.
The authors state ongoing effort to extend the pipeline to Telegram-type output, broadening the range of platforms that can be simulated.
The paper's listed future work—LLM-assisted persona generation, opinion dynamics, and better statistical content realism—indicates the pipeline is under active development.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The realism claim is currently supported by qualitative samples, not by quantitative comparison; a side-by-side statistical test against real X data (e.g., interaction-type distributions, hashtag co-occurrence, degree distributions) would determine whether 'quasi-realistic' holds.
Because narrative descriptions are the main input to the LLM, the quality of the output likely depends heavily on the scenario writer's skill; the pipeline may be less turnkey for non-technical users than suggested.
The absence of organic inter-group link formation and opinion dynamics means the dataset can depict convergence only if the designer pre-memberships actors in multiple groups; true emergent behavior remains out of scope.
A testable extension: vary the 60/30/10 attachment probabilities and the number of Randos per tweet, then measure how network statistics change; this would show whether those parameters can be tuned to match specific real-world platforms.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

You should know two things before reading this one. First, it is essentially a technical report, not a paper with an empirical claim: it walks through the design of a fictional Eurovision-like scenario and the AESOP-SynSM pipeline that generates tweets for it. Second, the abstract and Section 1.1 do assert that the pipeline generates “realistic tweet output” and “quasi-realistic” X data, and that assertion is the load-bearing part. It is unsupported. The paper lists fidelity, diversity, and stability as requirements for good synthetic data in Section 1, then measures none of them. Section 4 shows a handful of hand-picked tweets and one network screenshot. Section 5 explicitly lists “achieving better content realism and ensuring generated content is more statistically similar to actual social media content” as future work. That is a direct admission that the central claim is not yet met.

What the paper does well is documentation. The scenario design is thoughtful: the support network, the group membership rules, the bot/dredger classes, and the narrative timelines are all described in unusual detail. The explanation of how AESOP specifications become SynSM prompts, including the BEND maneuvers and tone settings, is genuinely useful for anyone building a similar hybrid ABM+LLM pipeline. The authors are also honest about current limitations—no opinion dynamics, no organic inter-group links, manually fixed active hours. That transparency earns credit.

The soft spots beyond the missing evaluation are real. All agents are active 9am–5pm, which cannot reproduce real diurnal patterns. Three Randos are created per tweet, meaning every original tweet gets at least one retweet or quote; real X has many tweets with zero engagement. The content itself is largely a restatement of the hand-authored narratives, so “realism” is partly circular. No dataset or code is released in the paper, only a mention of a shared repository. The citation pattern is mostly self-referential and from the same group, though the prior work is relevant and not hidden.

Who is this for? Practitioners who want a detailed example of how to configure a scenario in AESOP and what SynSM does with it. Not for someone looking for validated synthetic data.

Would I send it to peer review? A serious referee could push the authors to either scope the claims down to “scenario generation procedure” or add a real comparison against a reference dataset. That would make it a better paper. So yes, referee it, but with a clear request for evaluation or revised claims. As it stands, the realism claim should not be accepted.

Referee Report

4 major / 4 minor

Summary. This technical report describes AESOP-SynSM, a hybrid agent-based/LLM pipeline for generating synthetic X (Twitter) data, and documents its end-to-end application to a fictional three-day pop-culture scenario called AuraSight. The paper details the scenario setup (groups, actors, narratives, events), the agent taxonomy (humans, 15 bot classes, dredgers, and Randos), the network-construction algorithms (preferential attachment, follow-the-leader, random), and the LLM prompt structure. It claims in the abstract that the pipeline generates a 'quasi-realistic set X data' and 'realistic tweet output,' but the body provides only a small sample of handpicked tweets (Table 7) and a single network visualization (Figure 8). Section 5 identifies 'achieving better content realism and ensuring generated content is more statistically similar to actual social media content' as future work.

Significance. If the realism claim were substantiated, the paper would offer a valuable, customizable tool for producing synthetic social-media datasets for training and education, with a notably rich agent typology and output formatted as Twitter API V1 JSON. The detailed narrative timeline and explicit agent-class behaviors are useful documentation. However, the paper lists fidelity, diversity, and stability as criteria for good synthetic data (Section 1) but measures none of them. No comparison to real X data, no distributional statistics, and no sensitivity analysis are provided. As it stands, the contribution is a well-documented pipeline design plus a case study, not a validated method for generating realistic social-media data.

major comments (4)

[Abstract, §1, §5] The central claim of 'quasi-realistic' / 'realistic tweet output' is unsupported. Section 1 defines fidelity, diversity, and stability as required features, but Section 4 presents only anecdotal outputs (Table 7 and Figure 8). Section 5 explicitly lists 'achieving better content realism and ensuring generated content is more statistically similar to actual social media content' as future work. This directly undercuts the abstract's assertion. Either the realism claim must be removed/weakened, or the paper must supply quantitative evaluation against real X data (e.g., content length, hashtag usage, engagement distributions, network degree/clustering, temporal patterns).
[§2.4, §3.2.3] Two design choices force engagement patterns that are inconsistent with real social-media data. Section 2.4 sets all actors' active hours to 9am-5pm on all days, which cannot reproduce real diurnal posting cycles; Section 3.2.3 creates exactly three Randos per tweet, guaranteeing every original tweet receives retweets/quotes. Real X data exhibit highly skewed engagement, with most tweets receiving zero or very few retweets. These choices should be compared against empirical distributions or at least justified as intentional simplifications; as written, they undermine the 'quasi-realistic' claim.
[§2.5, §3.2.3] The content-generation step is largely a restatement of the authors' hand-written narratives. Section 2.5 states that the narrative description field 'is presented to the LLM in SynSM for generation' and includes the main claim, supporting points, style, and stance. Thus the output text is directly authored by the scenario designer, not emergent from the simulation. This is not inherently a flaw for a scenario-generation tool, but it means the 'realism' claim is circular unless text-level properties are independently evaluated (e.g., automated or human judgments, diversity across repeated runs, distributional similarity to real tweets). No such evaluation is present.
[§3.2.2, §4.2] The network realism claim rests on unvalidated parameter choices. The 60/30/10 mix of preferential attachment, follow-the-leader, and random attachment, the 2x bot posting rate, and the BEND maneuver multipliers are asserted in Sections 3.1 and 3.2.2 without sensitivity analysis or empirical grounding. Figure 8's 'decent hub and spoke structure' is an anecdotal visual; the paper does not report degree distributions, clustering coefficients, reciprocity, or temporal dynamics, despite Section 1 naming these as fidelity requirements. Without such analysis, the claim that the generated network resembles real social-media structure is unsupported.

minor comments (4)

[§2.3, §2.4] Figure numbering errors: Section 2.3 references the support network as Figure 3, but the caption is Figure 4; Section 2.4 references the Actor x Group graph as Figure 4, but the caption is Figure 5; the Actor (Source) x Group graph is referenced as Figure 5 but captioned Figure 6.
[§3.2.2] The phrase 'preferential agreement algorithm' appears to be a typo for 'preferential attachment algorithm.'
[Throughout] Capitalization of the scenario name is inconsistent: 'AuraSight' and 'Aurasight' both appear. Please standardize.
[§5.1] The paper states that code was contributed to a common repository but gives no URL or access instructions, limiting reproducibility. A link or reference to the repository should be provided.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the report is a pipeline description, not a derivation whose predictions reduce to its inputs.

full rationale

The paper is a technical report describing the AESOP-SynSM pipeline and the AuraSight scenario. It does not claim to derive a quantitative prediction from first principles; its output is generated by an LLM conditioned on hand-authored narratives, group stances, and parameters. The only load-bearing evaluative claim is that the generated data is 'quasi-realistic' or 'realistic' (Abstract; Section 1.1). That claim is not established by comparison to real X data, and Section 5 explicitly lists 'achieving better content realism and ensuring generated content is more statistically similar to actual social media content' as future work. This is a limitation/unsupported assertion, but it is not circular: the realism predicate is not defined as the inputs, no parameter is fitted to a target and then renamed a prediction, and no uniqueness theorem or ansatz is imported from the authors' prior work to force the conclusion. The self-citations (Ng & Carley 2025a for bot tweet rate; Ng & Carley 2025b for bot classes; Carley 2020 for BEND maneuvers) provide external or prior conceptual grounding for design choices, but the paper does not reduce its central demonstration to those citations. Therefore, on the circularity axis, the score is 0.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The paper's utility claim rests on many hand-chosen parameters, on prior self-cited frameworks, and on the unverified expectation that an LLM will produce realistic content when prompted with hand-written narratives. No external benchmark is used to ground the realism claim.

free parameters (6)

Network attachment mix = 60% preferential attachment, 30% follow-the-leader, 10% random
Chosen in Section 3.2.2 with no justification or sensitivity analysis.
Rando creation rate = 3 Randos per tweet
Stated in Section 3.2.3 without empirical basis.
Bot tweet rate multiplier = 2x human posting rate
Based on Ng & Carley (2025a), applied uniformly to all bot classes without per-class validation.
BEND maneuver multipliers = 2x for most bots, 4x for social influence and chaos bots
From Carley (2020) framework, but no evidence that these multipliers improve realism.
Active hours for actors = 9am-5pm for all actors
Set for simplicity in Section 2.4, not derived from activity data.
Rando attribute distributions = Location weights [35,35,15,15], age 21-40, random gender
Arbitrary choices in Table 2 that affect the demographic mix of the synthetic network.

assumptions (4)

domain assumption LLM prompt structure and narrative descriptions produce stance-consistent, realistic content
Sections 2.6 and 3.2.3 rely on GPT-4.1-mini following hand-authored narratives with specified tone and stance, with no evaluation of consistency or realism.
domain assumption The network construction algorithms (preferential attachment, follow-the-leader, random) produce realistic interaction structures when mixed
Section 3.2.2 assumes the 60/30/10 mix yields realistic social network interactions without comparing to real network metrics.
domain assumption The agent taxonomy from Ng & Carley (2025b) and BEND maneuvers from Carley (2020) are valid representations of real actors and maneuvers
The 15 bot classes and BEND maneuvers are taken from the authors' own prior work and implemented in SynSM as if they capture real behavior, which is not validated in this paper.
domain assumption The system prompt that frames simulation as a role-playing game to relax guard rails yields more realistic content
Section 3.2.3 states this works favorably toward realism, but no evidence is given and the potential risks are not discussed.

how reviews work

0 comments

Cite this review

Pith. "Pith review of AuraSight: Generating Realistic Social Media Data." pith.science (2026). https://pith.science/paper/G27SXLHB

@misc{pith2026250908927,
  author       = {Pith},
  title        = {Pith review of: AuraSight: Generating Realistic Social Media Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G27SXLHB}},
  note         = {Machine review of arXiv:2509.08927}
}

read the original abstract

This document details the narrative and technical design behind the process of generating a quasi-realistic set X data for a fictional multi-day pop culture episode (AuraSight). Social media post simulation is essential towards creating realistic training scenarios for understanding emergent network behavior that formed from known sets of agents. Our social media post generation pipeline uses the AESOP-SynSM engine, which employs a hybrid approach of agent-based and generative artificial intelligence techniques. We explicate choices in scenario setup and summarize the fictional groups involved, before moving on to the operationalization of these actors and their interactions within the SynSM engine. We also briefly illustrate some outputs generated and discuss the utility of such simulated data and potential future improvements.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

14 extracted references · 1 linked inside Pith

[1]

Day 1: Oliver wins Ethal’s finals. 2. Day 2: Oliver releases a PR statement. 3. Day 3: Nareth (the host country of AuraSight) confirms Oliver as Ethal’s official representative. After deciding on events , we have topics . We take a topic to be a salient aspect of the event that groups wish to discuss. Topics are not inherently stanced, but provide the bro...

2013
[2]

Our fictional scenario covers 3 days, focusing on one country’s representative selection process leading up to AuraSight 2030, which will be held at Nareth

More loosely, the post-Soviet pop market, where countries (e.g., Russia, Ukraine, Belarus) often engage with the pop markets of their neighbors, though shifts have been observed after the 2022 invasion of Ukraine (Lee et al., 2024). Our fictional scenario covers 3 days, focusing on one country’s representative selection process leading up to AuraSight 203...

2022
[3]

Cyborgs, who are often prominent accounts like activists/celebrities, may be considered for marking as a leader within a group
[4]

In other words, they draw from the content of other bots and reproduce them to be cited for ‘real’ users

Synchronized bots need to be in a group where all other relevant bots are source -only (cite-only) and then themselves a source (cite-only) in human groups. In other words, they draw from the content of other bots and reproduce them to be cited for ‘real’ users. 12 While not required, we also have some bots set up to solely ‘push’ narratives from key agen...
[5]

[Group/specific actor] [speech verb] [main claim]. 2. [Supporting points for claim]. 3. [If required; style of writing (e.g., positively, professionally)]. [If required; stance of group]. For example: 1. Fans of Oliver [group] cite evidence [speech verb] that Ethalian singers often perform for the Odrian market and make most of their living from the Odria...
[6]

Messages emphasize his belief that Ethal and Odria have a joint intertwined history and are brothers [supporting points]

Oliver, on his official account [group], sends out messages [speech verb] stating his joy in representing Ethal and his commitment to win for Ethal in the upcoming Aurasight competition in July 2030 in Nareth [main claim]. Messages emphasize his belief that Ethal and Odria have a joint intertwined history and are brothers [supporting points]. Messages are...
[7]

bot” in the user name Genre-Specific Bot Only One Conversational Bot Multiple Does not retweet News Bot One or more Add the word “news

Ethalian nationalists [group] discuss [speech verb] how the Odrian language is an invasive alien language to Ethal and its people [main claim]. Messages describe the history of how Ethalians were forced to learn the Odrian language in the past, when Ethal was conquered by it 200 years ago in 1835. Messages emphasize the continued existence of the Ethalian...

2025
[8]

This system prompt results in the LLM relaxing some of its guard rails, which works favorably towards scenario realism

System prompt: This states that the LLM is simulating a role playing game to help users identify misinformation manipulation on social media. This system prompt results in the LLM relaxing some of its guard rails, which works favorably towards scenario realism

Show all 14 references

[9]

Agent persona: This describes the agent type, class, and provides a brief description of the expected content of the agent currently being simulated
[10]

For quote and reply tweets, this is the narrative that the agent they are interacting with is talking about

Narrative: For an original tweet, this is the narrative that the agent is assigned to at that hour, as specified in AESOP. For quote and reply tweets, this is the narrative that the agent they are interacting with is talking about
[11]

This portion mimics a few-shot prompting architecture where examples of previous messages generated for the same narrative are provided

Last messages generated of the same narrative, if present. This portion mimics a few-shot prompting architecture where examples of previous messages generated for the same narrative are provided. This allows the LLM to continue the conversation around a narrative
[12]

Examples of these maneuvers are: bridge, back, explain, enhance

BEND maneuvers: The BEND maneuvers are a set of narrative and network maneuvers that can shape the information environment (Carley, 2020). Examples of these maneuvers are: bridge, back, explain, enhance. Social media posts often contain these maneuvers. This portion dictates t...

2020
[13]

‘Forever’ is about love, the most powerful feeling in the world

Other specifics: This portion allows for agent type/agent class specific information. Some examples are: a. The tone of the content, listed in Table below. b. For Dredgers: To require the use of at least two dredge words from a pre-defined list and at least two URLs from a pre...
[14]

https://doi.org/10.3390/systems13010029 Zhang, X., Lin, J., Mou, X., Yang, S., Liu, X., Sun, L., Lyu, H., Yang, Y., Qi, W., Chen, Y., Li, G., Yan, L., Hu, Y., Chen, S., Wang, Y., Huang, X., Luo, J., Tang, S., Wu, L., … Wei, Z. (2025). SocioVerse: A World Model for Social Simul...

2025 doi arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

Day 1: Oliver wins Ethal’s finals. 2. Day 2: Oliver releases a PR statement. 3. Day 3: Nareth (the host country of AuraSight) confirms Oliver as Ethal’s official representative. After deciding on events , we have topics . We take a topic to be a salient aspect of the event that groups wish to discuss. Topics are not inherently stanced, but provide the bro...

2013

[2] [2]

Our fictional scenario covers 3 days, focusing on one country’s representative selection process leading up to AuraSight 2030, which will be held at Nareth

More loosely, the post-Soviet pop market, where countries (e.g., Russia, Ukraine, Belarus) often engage with the pop markets of their neighbors, though shifts have been observed after the 2022 invasion of Ukraine (Lee et al., 2024). Our fictional scenario covers 3 days, focusing on one country’s representative selection process leading up to AuraSight 203...

2022

[3] [3]

Cyborgs, who are often prominent accounts like activists/celebrities, may be considered for marking as a leader within a group

[4] [4]

In other words, they draw from the content of other bots and reproduce them to be cited for ‘real’ users

Synchronized bots need to be in a group where all other relevant bots are source -only (cite-only) and then themselves a source (cite-only) in human groups. In other words, they draw from the content of other bots and reproduce them to be cited for ‘real’ users. 12 While not required, we also have some bots set up to solely ‘push’ narratives from key agen...

[5] [5]

[Group/specific actor] [speech verb] [main claim]. 2. [Supporting points for claim]. 3. [If required; style of writing (e.g., positively, professionally)]. [If required; stance of group]. For example: 1. Fans of Oliver [group] cite evidence [speech verb] that Ethalian singers often perform for the Odrian market and make most of their living from the Odria...

[6] [6]

Messages emphasize his belief that Ethal and Odria have a joint intertwined history and are brothers [supporting points]

Oliver, on his official account [group], sends out messages [speech verb] stating his joy in representing Ethal and his commitment to win for Ethal in the upcoming Aurasight competition in July 2030 in Nareth [main claim]. Messages emphasize his belief that Ethal and Odria have a joint intertwined history and are brothers [supporting points]. Messages are...

[7] [7]

bot” in the user name Genre-Specific Bot Only One Conversational Bot Multiple Does not retweet News Bot One or more Add the word “news

Ethalian nationalists [group] discuss [speech verb] how the Odrian language is an invasive alien language to Ethal and its people [main claim]. Messages describe the history of how Ethalians were forced to learn the Odrian language in the past, when Ethal was conquered by it 200 years ago in 1835. Messages emphasize the continued existence of the Ethalian...

2025

[8] [8]

This system prompt results in the LLM relaxing some of its guard rails, which works favorably towards scenario realism

System prompt: This states that the LLM is simulating a role playing game to help users identify misinformation manipulation on social media. This system prompt results in the LLM relaxing some of its guard rails, which works favorably towards scenario realism