acceptodds
Under review as a conference paper at ICLR 2027

MARS: Reverse Querying and Multi-modality Source Verification for Social Simulation

Abstract

Social network simulation aims to characterize information diffusion through social networks and the decision-making processes of individuals during diffusion. Existing LLM-based simulators typically model diffusion as forward-only, text- only propagation. However, real-world diffusion is multi-modality and interactive: recipients may use visual evidence and query the source before deciding whether to propagate a claim. Modeling these two aspects is therefore important for capturing realistic verification behavior and reducing distortions in simulated diffusion. Their absence creates two limitations: I) without reverse querying, receiving agents cannot obtain additional source evidence before deciding whether to forward a claim; and II) without multi-modality evidence, simulators cannot capture how visual content affects the interpretation of recipients and propagation behavior. These limitations distort simulated diffusion and produce patterns that diverge from real-world behavior. To address these issues, we first construct MM-InfoBench, a multi-modality benchmark from real-world information diffusion data, providing aligned source text, visual evidence, and observed diffusion traces for studying evidence-aware diffusion processes. Then, we introduce Multi-modality Agent Reverse-querying Simulator (MARS), a social simulation framework that preserves claim–source correspondence throughout diffusion. MARS enables receiving agents to access associated multi-modality evidence and query the source before making propagation decisions, integrating source verification into information diffusion simulation. Extensive experiments demonstrate that MARS outperforms existing approaches in source preservation, agent behavior, and diffusion dynamics, increasing denial messages from 1.85% to 50.65% and reducing exposure to false claims from 5.06% to 0.42%. The dataset and source code will be released.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Large Language Models, Multi-Agent Systems, Social Simulation, Information Diffusion, Multi-modality, Fact Verification

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_dX1mOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies LLM-driven simulations of social information diffusion, focusing on the accessibility of source evidence after messages are reposted or rewritten and on how such evidence affects propagation decisions. The authors propose MARS, which maintains links between claims in the diffusion process and their original sources, allowing receiving agents to consult relevant evidence before acting. They also introduce MM-InfoBench, a benchmark linking original posts from real-world events to their subsequent diffusion traces. The experiments examine how source evidence and images affect simulated decisions and diffusion and assess whether simulation-derived behavioral features can help predict the growth of real-world cascades. The paper’s contributions lie in incorporating source verification into multimodal information-diffusion simulation and providing a corresponding benchmark and empirical analysis.

Critical strengths
  1. The proposed framework provides a well-motivated approach to incorporating source verification into LLM-based social simulation. By maintaining explicit links between propagated claims and their source evidence, MARS enables agents to access relevant information even after messages are forwarded or rewritten. This design integrates source verification into agents' propagation decisions and provides a useful framework for studying how evidence access influences individual decisions and collective diffusion behavior.
  2. The paper introduces MM-InfoBench, a dataset comprising 11,066 real-world events, including 5,654 events with source images and approximately 3.88 million timestamped cascade nodes. The dataset explicitly links multimodal source evidence, propagated claims, and observed diffusion trajectories, providing valuable data support for studying the role of multimodal evidence in information diffusion.
Critical weaknesses
  1. The paper aims to improve the realism of information diffusion simulations, but the experiments mainly show how evidence access affects LLM agents' decisions and simulated diffusion outcomes, without directly validating their consistency with real-world user behavior. Although Section 4.5 evaluates the predictive value of simulation-derived features, this does not necessarily demonstrate behavioral realism.
  2. The proposed method mainly extends existing LLM-based social simulation frameworks by preserving claim–source associations and enabling evidence querying before propagation decisions. The multimodal component follows the same mechanism by incorporating source images, while agent decisions still rely on existing LLMs. Overall, the method primarily integrates evidence access into existing simulation frameworks, with limited methodological novelty.
Questions for authors
  1. Section 4.5 evaluates whether simulation-derived features can help predict cascade size. Please provide direct comparisons of agent behavior and diffusion patterns with those observed on real social networks to show that source verification improves simulation realism and distinguish the predictive value of these features from the realism of the simulation itself.
  2. If a claim is maliciously rewritten as it spreads, changing its meaning into a false or inflammatory statement, how does MARS detect this change and update the claim–source association? If the system retains the original claim ID and allows agents to consult the original source, could this make them more likely to identify and suppress the rewritten content, potentially underestimating how such content spreads on real social networks?
  3. In the current network experiments, each cascade contains only one claim, and the query tool is pre-bound to its source. If multiple claims from the same event spread at the same time, how does MARS determine which claim a message refers to and which source evidence to retrieve? Is the single-claim setting simply an experimental simplification, or does it limit the applicability of the method?
  4. MARS retains the correct claim ID to directly access the source, whereas retrieval baselines must recover it from rewritten text. It therefore bypasses the retrieval problem faced by the baselines, and its 100% source recall is guaranteed by design. I find this comparison unfair because the methods use different information and face different levels of difficulty. Why did the authors choose this comparison?
3 comments
AuthorsOpenReview ↗

Response to Reviewer dX1m (Part 1/3): W1/Q1, W2

We thank Reviewer dX1m for the constructive comments, and for recognizing that MARS is a well-motivated approach and that MM-InfoBench provides valuable data support. The reviewer raises five concerns: direct comparison with real user behavior (W1/Q1), methodological novelty (W2), malicious rewrites (Q2), concurrent claims of the same event (Q3), and the fairness of the retrieval baselines (Q4). We added experiments for each, and answer them in turn.

W1/Q1: Direct comparison of MARS with real user behavior

The reviewer asks for direct comparisons of agent behavior and diffusion patterns with those on real social networks, to show that source verification improves realism separately from the predictive value in Section 4.5. We added two direct comparisons with real data, and both show that source verification makes the simulation closer to real behavior; forecasting is a separate test of predictive value. MARS agents received the same 40 news headlines and fact-check evidence as the 1,548 participants of a published human experiment [1], and we compared the changes in belief and sharing intention. On real social networks, we compared simulated discussions with 1,416 real Twitter replies (PHEME [2]) on the same 18 cascades, measuring the change in denial replies after the correcting source appears. Neither comparison uses cascade size.

Comparison with real dataResult
Human experiment [1]: Pearson correlation coefficient r between MARS and humans in the change after evidenceBelief r = 0.460; sharing intention r = 0.511
Human experiment [1]: belief change when the fact-check wrongly labels a true headline falseHumans −12.75%; MARS −11.45%
Real Twitter discussions [2]: mean absolute error of the change in denial replies after the correcting source appearsForward-only 4.12%; MARS 3.20%

Source verification makes the simulation closer to real behavior, both for individual users and in network discussions. For individual users, MARS agents change their belief and sharing intention across the 40 headlines as the human participants do; when the fact-check is wrong, they lower their belief in a true headline by almost the same amount as people, so the agreement is with human behavior and not with correctness. In network discussions, forward-only simulation has no source and produces no change in denial replies, whereas reverse querying produces a change closer to the one observed on Twitter. The realism of MARS is therefore validated directly against real user behavior, and not through its predictive value.

W2: Methodological novelty

The reviewer considers that MARS mainly integrates evidence access into existing LLM-based simulation frameworks, that the multimodal component follows the same mechanism, and that decisions still rely on existing LLMs. The novelty of MARS is reverse querying, a new step of diffusion that existing frameworks do not have, and its gains appear with the agent LLM unchanged. Existing frameworks such as OASIS [4] and RumorSphere [8] simulate forward-only diffusion: each agent decides only from the message it receives, so no receiver checks the source of a claim. Environments such as LMAgent [9] support visual inputs, but diffusion simulation remains text-only, so the source text and images are lost once the claim is rewritten. Source-directed reverse querying lets every receiver query the source of a claim before forwarding, correcting, or withholding it. Provenance-preserving multimodal propagation keeps each claim linked to its original text and images through every rewrite, so the query still returns the correct source.

Comparison with the agent LLM unchanged (Qwen3-4B)Result
Denial messages in the network (share of agents per step)No source 1.85%; MARS 50.65%
Exposure to false claims in the network (share of agents per step)No source 5.06%; MARS 0.42%
Consistent decisions (deny a false claim or endorse a true one) with source text only or with source text and imageText 25.76%; text and image 33.23%

Each mechanism of MARS gives a gain that the same LLM does not obtain without it, and all of these gains are statistically significant. Reverse querying changes diffusion itself: denial messages rise and exposure to false claims falls. The gain does not come from evidence access alone: with the same source content, obtaining it by reverse querying gives more consistent decisions than receiving it attached to the message (42.75% against 32.37%; Q4). The multimodal component does not repeat the text mechanism: source images carry evidence that source text does not contain, and they raise consistent decisions. MARS therefore contributes a new diffusion step, reverse querying, and the multimodal propagation that keeps it effective after rewrites, and not an integration of evidence access into an existing framework.

AuthorsOpenReview ↗

Response to Reviewer dX1m (Part 2/3): Q2, Q3

Q2: Malicious rewrites: detection, source-link update, and whether spread is underestimated

The reviewer asks how MARS detects a rewrite that changes the meaning of a claim and updates its source link, and whether access to the original source makes agents suppress the rewrite more than happens on real networks. MARS detects the change by comparing each message with its parent claim and registers the rewrite as a new claim linked to its parent; agents then judge the rewrite against the source themselves, and suppression is not built in. Forward-only simulation treats such a rewrite as the original claim, and text retrieval searches with the rewritten text. In MARS, when a message states a fact that contradicts its parent claim, it receives a new claim ID that records the parent, and the query returns the original source with a note that names the original claim. We added an experiment with 49 reviewed malicious rewrites, each changing a number, an actor, or an outcome into a false and more inflammatory statement, and 45 rewrites that keep the meaning.

Added experimentResult
Meaning-changing rewrites that MARS detects and registers as new claims91.84% (45 of 49)
Meaning-preserving rewrites wrongly registered as new claims2.22% (1 of 45)
Denial of the malicious rewrites by the same agent LLM (Qwen3-4B)No query 9.52%; no source 31.63%; tool-augmented retrieval (BGE-M3) 66.33%; MARS 80.27%
Denial messages in the network (25 real cascades) when all agents or only 6 of 24 agents can obtain the sourceAll agents 34.60%; 6 of 24 agents 8.98%–11.81%

MARS detects malicious rewrites and gives agents the evidence to judge them, which forward-only simulation and text retrieval cannot do. It detects 91.84% of the meaning-changing rewrites and wrongly flags only 2.22% of the meaning-preserving ones, so each rewrite is followed as a new claim with its own ID. With the original source, agents deny the rewrite in 80.27% of decisions, more than with text retrieval on the rewritten text (66.33%), a statistically significant gain. Spread is not underestimated by design: agents decide for themselves, and an agent that does not query the source denies the rewrite in only 9.52% of decisions. How many agents verify is a setting of the simulation, and with only 6 of 24 agents the denial messages fall accordingly. MARS therefore detects malicious rewrites, updates their source link, and lets the share of agents that verify be set from none to all, so the spread of rewritten content is not assumed to be suppressed.

Q3: Source selection when several claims from the same event spread together

The reviewer asks how MARS determines which claim a message refers to and which source to retrieve when several claims from the same event spread at the same time, and whether the single-claim setting limits the method. The single-claim setting is an experimental simplification: every message carries the ID of its own claim, so the query returns the source of that claim however many claims spread together, and we added two-claim experiments that confirm this. A message inherits its claim ID from the post that it reposts, quotes, or comments on, and the query is made on the post that the agent wants to check, so MARS does not infer the claim from the message text. A source bound to the whole event, or a search by message text, often returns the source of another claim, because claims of the same event are similar. We placed a true and a false claim of the same event in one feed (27 pairs from four news events) and in one network (19 real cascades).

Added experimentResult
Two claims in one feed: queries that return a wrong sourceOne source for the whole event 59.88%; tool-augmented retrieval (BGE-M3) 74.07%; MARS 0.00%
Two claims in one feed: consistent decisions (deny the false claim or endorse the true one)One source for the whole event 33.95%; tool-augmented retrieval (BGE-M3) 56.48%; MARS 65.12%
Two claims in one network: responses that endorse the false claim or deny the true one (share of agents per step)One source for the whole event 7.29%; MARS 5.92%

With several claims from the same event, MARS returns the source of the claim being checked, which a source bound to the event and text retrieval do not. With one source for the whole event, 59.88% of the queries return the source of the other claim, and text retrieval returns a wrong source for 74.07%; in MARS the claim ID selects the source, so every query returns the source of the checked claim. Agents therefore make more consistent decisions, 65.12% compared with 33.95%, a statistically significant gain. The single-claim setting therefore does not limit the applicability of MARS: the source is selected by the claim ID of each message and not by the event, so the same mechanism applies when several claims spread at the same time.

AuthorsOpenReview ↗

Response to Reviewer dX1m (Part 3/3): Q4

Q4: Why MARS is compared with retrieval baselines, and comparisons without an information advantage

The reviewer considers the comparison with retrieval baselines unfair, because MARS keeps the claim ID while the baselines must recover the source from rewritten text, and asks why we chose it. We chose this comparison because it measures what a simulation loses when it does not keep the source link, and we added comparisons in which MARS has no information advantage; MARS keeps its advantage there. Existing simulators and tool-augmented agents have only the message text, so text retrieval is the only way they can obtain a source, and the comparison shows what they obtain in practice. The source link that MARS keeps is not extra information given to our method: platforms already record it for every repost and reply. Keeping this link instead of recovering it is the design of MARS, and the gains that we report are in agent decisions.

ComparisonResult
Source found for a repost or reply (226,768 Weibo reposts; 98,929 PHEME replies)Text retrieval (BGE-M3 [6]) 8.42% and 14.53%; link recorded by the platform 99.98% and 97.82%
The same source content given to the agent: consistent decisions (deny a false claim or endorse a true one), Qwen3-4BSource attached to the message 32.37%; source obtained by reverse querying 42.75%
The same comparison with Mistral-7BSource attached to the message 14.17%; source obtained by reverse querying 20.93%
Source link removed, so that MARS must also recover the source: consistent decisionsBGE-M3 54.79%; MARS 60.27%

MARS keeps its advantage when it has no information advantage over the baselines. When the agent receives the same source content, obtaining it by reverse querying gives more consistent decisions than receiving it attached to the message, for both Qwen3-4B and Mistral-7B; both gains are statistically significant, and they cannot come from the claim ID. When the source link is removed, MARS must also recover the source, and its decisions remain more consistent than with BGE-M3. The original comparison answers a different question: a simulation that discards the link recorded by the platform loses the source, because text retrieval finds it for only 8.42% of Weibo reposts. The comparison therefore measures the value of keeping the source link, which is the design contribution of MARS, and without this advantage MARS still gives better decisions than the baselines.

In summary, MARS adds a new step to diffusion simulation, in which receivers query the source of a claim before they forward it, and the added experiments show that this step makes the simulation closer to real user behavior, detects malicious rewrites, returns the correct source when several claims spread together, and keeps its advantage over retrieval baselines without an information advantage. We sincerely thank Reviewer dX1m for these questions, which have strengthened the paper, and we will incorporate these results into the revised paper.

References

[1] Matthew R. DeVerna, Harry Yaojun Yan, Kai-Cheng Yang, and Filippo Menczer. Fact-checking information from large language models can decrease headline discernment. Proceedings of the National Academy of Sciences, 121(50):e2322823121, 2024.

[2] Arkaitz Zubiaga, Maria Liakata, Rob Procter, Geraldine Wong Sak Hoi, and Peter Tolmie. Analysing How People Orient to and Spread Rumours in Social Media by Looking at Conversational Threads. PLOS ONE, 11(3):e0150989, 2016.

[4] Ziyi Yang, Zaibin Zhang, Zirui Zheng, Yuxian Jiang, Ziyue Gan, Zhiyu Wang, Zijian Ling, Jinsong Chen, Martz Ma, Bowen Dong, Prateek Gupta, Shuyue Hu, Zhenfei Yin, Guohao Li, Xu Jia, Lijun Wang, Bernard Ghanem, Huchuan Lu, Chaochao Lu, Wanli Ouyang, Yu Qiao, Philip Torr, and Jing Shao. OASIS: Open Agent Social Interaction Simulations with One Million Agents. arXiv preprint, arXiv:2411.11581, 2024.

[6] Jianlyu Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. M3-Embedding: Multi-Linguality, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation. Findings of the Association for Computational Linguistics: ACL 2024, pp. 2318–2335, 2024.

[8] Yijun Liu, Wu Liu, Xiaoyan Gu, Hantao Yao, Weiping Wang, Jiebo Luo, and Yongdong Zhang. RumorSphere: A Framework for Million-Scale Agent-Based Dynamic Simulation of Rumor Propagation. arXiv preprint, arXiv:2509.02172, 2025.

[9] Yijun Liu, Wu Liu, Xiaoyan Gu, Xiaodong He, Yong Rui, and Yongdong Zhang. LMAgent: A Large-Scale Multimodal Agents Society for Multi-User Simulation. IEEE Transactions on Multimedia, 28:2668–2680, 2026.

Reviewer_QtZyOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents MARS, a social diffusion simulation framework that combines source-directed reverse querying with provenance-preserving multimodal evidence. Before deciding whether to propagate, correct, or withhold a claim, receiving agents can retrieve its associated source evidence, including text and images. The authors also introduce MM-InfoBench, a benchmark that links source evidence, reformulated claims, and observed diffusion traces from real-world events.

Critical strengths

S1. The paper addresses source verification and multimodal evidence loss in LLM-based social diffusion simulation, both of which are relevant to misinformation research.

S2. MM-InfoBench links source evidence, reformulated claims, and real-world diffusion traces, providing a valuable resource for studying misinformation propagation.

Critical weaknesses

W1. The novelty of MARS is not sufficiently clear. Its two main components, claim-level provenance tracking and source evidence retrieval, are closely related to existing provenance methods and tool-augmented LLM agents. The paper does not clearly distinguish its contributions from these approaches.

W2. The benchmark assumes the source–claim associations that MARS relies on. MM-InfoBench explicitly provides these associations, and MARS maintains them by passing the original claim ID through subsequent rewrites. In real-world diffusion, however, such associations are often missing or difficult to recover. Evaluations on more diverse datasets or under incomplete and noisy provenance settings would be valuable.

W3. The experiments show that access to matched source evidence affects agents' propagation decisions, but do not establish whether these decisions are realistic. It remains unclear whether the simulated behavior reflects how real users respond to source evidence or mainly the behavior of the selected LLMs and prompts.

W4. The reported improvements are relatively modest, and several MAE confidence intervals include zero. It is unclear how much additional value the multi-agent simulations provide over conventional features, particularly given their computational cost.

W5. The network experiments are limited to 64 agents and 16 steps, with each cascade associated with a single claim. They do not consider competing claims, conflicting sources, or multiple narratives spreading through the same network. This raises concerns about whether the findings generalize to larger and more heterogeneous social networks.

Questions for authors

Q1. Regrading W1, MARS relies on explicit claim-level provenance information to connect reformulated messages with their original sources. Since such links are not always available or reliable on real social media, how could the framework be adapted for practical misinformation intervention research?

Q2. How do the authors establish that the decisions and diffusion patterns generated by MARS reflect real human behavior, rather than simply showing that LLM agents respond predictably when additional evidence is provided? Could the framework be validated against observed human behavior or real-world diffusion data?

Q3. How does MARS handle missing or incorrect source links, or cases where a claim is associated with multiple conflicting sources? Can the framework detect and recover from incorrect claim-source associations, and how would such errors affect its results?

3 comments
AuthorsOpenReview ↗

Response to Reviewer QtZy (Part 1/3): W1, W2/Q1/Q3

We thank Reviewer QtZy for the constructive comments, and for recognizing that MARS addresses source verification and multimodal evidence loss in LLM-based diffusion simulation, and that MM-InfoBench is a valuable resource. The reviewer raises five concerns: the difference from existing provenance methods and tool-augmented LLM agents (W1), missing, incorrect, or conflicting source links (W2/Q1/Q3), agreement with real human behavior (W3/Q2), value over conventional features and computational cost (W4), and larger networks with competing claims (W5). We added experiments for each, and answer them in turn.

W1: How MARS differs from existing provenance methods and tool-augmented LLM agents

The reviewer asks how the two components of MARS differ from existing provenance methods and tool-augmented LLM agents. The difference is that these approaches handle each claim separately, whereas MARS makes source verification a step of diffusion that every receiver performs before forwarding a claim. Provenance methods record where a claim comes from, but receivers do not use this record when they decide whether to forward it. Tool-augmented agents search for evidence with the text of the message they receive, so the search fails once the claim has been rewritten (BGE-M3 [6] retrieves the source for only 8.42% of Weibo reposts). Source-directed reverse querying lets the receiver query the source of the claim itself instead of searching by message text, and decide from the returned evidence. Provenance-preserving multimodal propagation keeps each claim linked to its original text and images through every rewrite, so the query still returns the correct source.

Decisions of the same agent LLM (Qwen3-4B)No sourceTool-augmented retrieval (BGE-M3)MARS
Consistent decisions: deny a false claim or endorse a true one7.97%33.17%42.75%
Denial of rewrites that change the meaning of a claim31.63%66.33%80.27%

With the agent LLM unchanged, MARS gives better decisions than tool-augmented retrieval, including after a claim has been rewritten to change its meaning; both gains are statistically significant. These decisions also change diffusion, which neither existing approach affects: with reverse querying, the share of agents per step that post a denial of the false claim rises from 1.85% to 50.65%, and the share that see it in their feed falls from 5.06% to 0.42%. The contribution of MARS is therefore reverse querying itself, a new step of diffusion simulation, together with the multimodal propagation that keeps it effective after rewrites, and not a combination of existing provenance tracking and evidence retrieval.

W2/Q1/Q3: Missing, incorrect, and conflicting source links

The reviewer asks how MARS can be used when source links are missing, incorrect, or conflicting, as on real social media. We added experiments for all three cases: MARS detects incorrect links, and reverse querying remains effective in every case. Real platforms already record the link that MARS uses: following the repost or reply records reaches the original post for 99.98% of 226,768 Weibo reposts and 97.82% of 98,929 PHEME replies. For the remaining messages, we added a check to the query: MARS tests whether the returned source addresses the claim in the message, and when the link is missing or the source fails this test, it retrieves candidate sources by message text and returns the first one that passes. When a claim has several sources, the query returns all of them with their times.

Added experimentResult
Incorrect source links that MARS detects and rejects94.52% (69 of 73)
Consistent decisions (deny a false claim or endorse a true one)No source 21.92%; link missing 60.27%; link incorrect 57.53%; link correct 67.12%
Network with half of the source links incorrect or missing (25 real cascades): exposure to false claimsNo source 34.33%; MARS 7.60%
Same network: denial messagesNo source 1.42%; MARS 23.31%
Two conflicting sources (an earlier record that supports the false claim and a later correction): denial of the false claim54.55%–60.61%, whichever source is shown first or dated earlier

MARS keeps its advantage when source links are missing, incorrect, or conflicting. With a missing or an incorrect link, agents still make consistent decisions in 60.27% and 57.53% of cases, against 21.92% without a source. In the network with half of the links incorrect or missing, reverse querying still reduces exposure to false claims and raises denial messages; both changes are statistically significant. When a claim has two conflicting sources, more than half of the decisions deny the false claim, whichever source is shown first. MARS can therefore be applied to real platforms through the links they already record, and its effect does not depend on complete or correct source links.

AuthorsOpenReview ↗

Response to Reviewer QtZy (Part 2/3): W3/Q2, W4

W3/Q2: Whether MARS decisions reflect real human behavior or only the selected LLMs and prompts

The reviewer asks whether the decisions of MARS agents reflect how real users respond to source evidence or mainly the behavior of the selected LLMs and prompts, and whether MARS can be validated against observed human behavior or real-world data. We added both validations: MARS agents respond to source evidence as real people do, and this agreement holds for five agent LLMs and under two prompts. We compared MARS with a published human experiment [1] (1,548 participants rating belief and sharing intention for 40 headlines, with and without fact-check evidence) and with 1,416 real replies in PHEME discussions [2].

Comparison with real dataResult
Human experiment [1]: Pearson correlation coefficient r between MARS and humans in belief change after evidencer = 0.460 (evidence always shown); r = 0.394 (viewing evidence optional)
Human experiment [1]: the same correlation coefficient r in sharing-intention change after evidencer = 0.511 (evidence always shown); r = 0.457 (viewing evidence optional)
Same correlations (evidence always shown) with five agent LLMs: Qwen3-4B, Qwen3-8B, Qwen2.5-7B, Qwen2.5-14B, Llama-3.1-8BPositive for every LLM: belief r = 0.355–0.469; sharing intention r = 0.214–0.433
Human experiment [1]: belief change when the fact-check wrongly labels a true headline falseHumans −12.75%; MARS −11.45%
PHEME replies [2]: mean absolute error of the change in denial replies after the correcting source appearsForward-only 4.12%; MARS 3.20%

MARS reproduces how real people respond to source evidence, and this does not depend on one selected LLM or prompt. Because MARS agents query the source before deciding, their changes in belief and sharing intention follow those of the human participants across the 40 headlines; the correlation is positive for each of the five LLMs, and it holds for the main LLM both when evidence is always shown and when viewing it is optional. MARS agents do not simply become more accurate when evidence is added: when the fact-check is wrong, they lower their belief in a true headline by almost the same amount as people do. In real discussions, reverse querying produces a change in denial replies closer to the real one than forward-only simulation, which produces none. MARS decisions therefore reflect real human behavior under source evidence, validated against a human experiment and real discussions, and not the behavior of one LLM or prompt.

W4: Additional value of MARS over conventional features, and its computational cost

The reviewer asks how much value the multi-agent simulation adds over conventional features, given its computational cost. MARS adds what conventional features cannot provide, namely how diffusion changes when receivers verify the source, and its responses also lower the forecasting error at every horizon, at a measured cost of 1.25 seconds per cascade. Conventional features describe a cascade that has already been observed, so they cannot show what happens when receivers query the source before forwarding; MARS shows this directly, with exposure to false claims falling from 5.06% to 0.42% under reverse querying (W1). Forecasting is an additional test that MARS responses carry real information: 16 agents read the first posts of each real Weibo cascade, and their responses are added to the temporal, structural, and content features of the baseline.

Result on 2,000 held-out Weibo cascadesValue
Mean absolute error of cascade size after 1 hourConventional features 83.36 posts; with MARS 82.54 posts
Mean absolute error of cascade size after 24 hoursConventional features 424.91 posts; with MARS 420.93 posts
Horizons (1, 3, 6, and 24 hours) at which MARS lowers the error4 of 4
Computation per cascade (16 agents, Qwen3-4B, one RTX 4090 GPU)MARS 1.250 s; forward-only 0.268 s
Computation for all 2,000 cascadesAbout 42 minutes on one GPU

MARS lowers the forecasting error at every horizon, with statistically significant reductions at 1 hour and at 24 hours, and it does so at low cost. The baseline already uses the temporal, structural, and content features of the observed posts, so the gain is information that these features do not contain, obtained from only 16 agents per cascade. The results use MARS responses averaged over three LLMs (Qwen3-4B, Qwen2.5-7B, Llama-3.1-8B), so the gain does not depend on one selected LLM. With Qwen3-4B, the computation is 1.250 s per cascade on a single GPU, which is small compared with the hours over which a cascade is forecast. MARS therefore provides, at low cost, predictive information beyond conventional features together with the analysis of source verification during diffusion, which conventional features cannot provide.

AuthorsOpenReview ↗

Response to Reviewer QtZy (Part 3/3): W5

W5: Larger and more heterogeneous networks, competing claims, and conflicting sources

The reviewer asks whether the findings generalize beyond 64 agents and one claim per cascade, to larger and more heterogeneous networks with competing claims and conflicting sources. We added experiments for each case: reverse querying keeps its effect with twice as many agents, on three other network structures, and when two competing claims spread in the same network. One claim per cascade is an experimental simplification and not a requirement of MARS: each message carries the ID of its own claim, so the query returns the source of that claim however many claims spread in the network. To test this, we placed a true and a false claim of the same event in one feed and in one network, and compared MARS with a single source shared by the whole event.

Added experimentResult
128 agents (paper: 64) on 25 real cascades: exposure to false claims (share of agents per step)No source 0.89%; MARS 0.08%
Same 128-agent networks: denial messages (share of agents per step)No source 1.91%; MARS 49.14%
Random, power-law, and two-community networks: exposure to false claimsNo source 26.92%, 34.33%, 23.60%; MARS 7.04%, 10.60%, 4.13%
Two competing claims in one feed: consistent decisions (deny the false claim or endorse the true one)One source for the whole event 33.95%; MARS 65.12%
Two competing claims in one network (19 real cascades): responses that endorse the false claim or deny the true oneOne source for the whole event 7.29%; MARS 5.92%
Two conflicting sources for one claim: denial of the false claim54.55%–60.61% (W2/Q1/Q3)

MARS keeps its advantage in larger networks, in other network structures, and with competing claims. With twice as many agents and on random, power-law, and two-community networks, reverse querying still reduces exposure to false claims and increases denial messages. When a true and a false claim of the same event compete, MARS returns the source of the claim being checked, so consistent decisions rise from 33.95% to 65.12% compared with a single source for the whole event, and in the network fewer responses endorse the false claim or deny the true one. The changes with 128 agents, on the three network structures, and in the two-claim decisions are statistically significant. The findings therefore do not depend on 64 agents, one network structure, or one claim per cascade: because every message is linked to the source of its own claim, reverse querying extends to larger networks, competing claims, and conflicting sources.

In summary, MARS adds a new step to diffusion simulation, in which receivers query the source of a claim before they forward it, and the added experiments show that this step remains effective when source links are missing, incorrect, or conflicting, agrees with how real people respond to source evidence, lowers the forecasting error at low cost, and extends to larger networks and competing claims. All of these results come from the two mechanisms of MARS: reverse querying, and multimodal propagation that keeps every message linked to the original text and images of its claim. We sincerely thank Reviewer QtZy for these suggestions, which have strengthened the paper, and we will incorporate these results into the revised paper.

References

[1] Matthew R. DeVerna, Harry Yaojun Yan, Kai-Cheng Yang, and Filippo Menczer. Fact-checking information from large language models can decrease headline discernment. Proceedings of the National Academy of Sciences, 121(50):e2322823121, 2024.

[2] Arkaitz Zubiaga, Maria Liakata, Rob Procter, Geraldine Wong Sak Hoi, and Peter Tolmie. Analysing How People Orient to and Spread Rumours in Social Media by Looking at Conversational Threads. PLOS ONE, 11(3):e0150989, 2016.

[6] Jianlyu Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. M3-Embedding: Multi-Linguality, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation. Findings of the Association for Computational Linguistics: ACL 2024, pp. 2318–2335, 2024.

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

This paper introdces MARS, an LLM based social network simulation model. The model allows giving source verification and image evidence to the agents before it makes dicisions of forward, correct, or withhold. The paper claims that this can better model how people will verify information before sharing it, and therefore improve the realism of simulated information diffusion. This paper also introduces MM-info bench, which links the source context and images with the observed information propagation. Experiments give four major conclusions: Saving the source evidence can reduce the information loss when propagating through the network; the evidence can drcrease the propagation of wrong information; the images can provide attional information; the improvement of such method to real world social network simulation is somehow limited.

Critical strengths
  1. Meaningful problem and comprehensive experiments: The paper examines how source verification and visual evidence influence information diffusion, incorporating behavioral mechanisms often overlooked in existing social simulations. Controlled comparisons involving empty responses, mismatched sources, content-based retrieval, and directly attached evidence help isolate their effects on individual judgments and network-level diffusion.
  2. Valuable data resource: MM-InfoBench links source text and images with observed diffusion trajectories, enabling researchers to study evidence access alongside information propagation. Its explicit claim–source associations support controlled comparisons of relevant, mismatched, and unavailable evidence, providing a useful foundation for evaluating how evidence availability and modality affect agent decisions and downstream diffusion.
Critical weaknesses
  1. Limited validation of simulation realism: More accurate agent decisions and reduced misinformation spread do not necessarily indicate a closer match to real-world social dynamics, while predictive gains on real data remain limited.
  2. How sensitive are the reported gains to network topology, community structure, and the positions of agents that perform verification? Could the authors provide a small set of experiments on standard network models, such as Erdős–Rényi graphs and networks with power-law degree distributions, to assess robustness across different structures?
Questions for authors
  1. Other than the cascade forecasting results, is there direct evidence to show that MARS better captures real human behavior, rather than only improving the agents's accuracy and reducing the spread of incorrect information?
  2. How sensitive are the reported gains to network topology? Can the authors provide a small set of experiments on standard network models, such as Erdős–Rényi graphs and networks with power-law degree distributions, to assess robustness across different structures?
2 comments
AuthorsOpenReview ↗

Response to Reviewer hVad (Part 1/2): W1/Q1

We thank Reviewer hVad for the constructive comments, and for recognizing that MARS studies a meaningful problem with comprehensive controlled experiments and that MM-InfoBench is a valuable data resource. The reviewer raises two concerns: whether MARS agrees with real human behavior, beyond cascade forecasting (W1/Q1), and whether its gains are robust to network structure (W2/Q2). We added experiments for both, and answer them in turn.

W1/Q1: Direct evidence that MARS agrees with real human behavior

The reviewer asks for direct evidence, other than cascade forecasting, that MARS captures real human behavior rather than only making agents more accurate. We added two direct comparisons with real human data, and both show that MARS agents respond to source evidence as real people do. In the first, MARS agents received the same 40 news headlines and fact-check evidence as the 1,548 participants of a published human experiment [1]. In the second, we compared MARS with 1,416 real PHEME replies [2] posted before and after the correcting source appeared, on the same 18 cascades. For predictive gains, we report cascade forecasting with MARS responses averaged over three LLMs.

Comparison with real dataResult
Human experiment [1]: Pearson correlation coefficient r between MARS and humans in belief change after evidencer = 0.460
Human experiment [1]: the same correlation coefficient r in sharing-intention change after evidencer = 0.511
Human experiment [1]: belief change when the fact-check wrongly labels a true headline falseHumans −12.75%; MARS −11.45%
PHEME replies [2]: mean absolute error of the change in denial replies after the correcting source appearsForward-only 4.12%; MARS 3.20%
1,814 held-out Weibo cascades: mean absolute error of cascade size after 24 hoursConventional features 424.91 posts; with MARS 420.93 posts

MARS reproduces how real people respond to source evidence, which forward-only simulation cannot represent. Because MARS agents query the source before deciding, their changes in belief and sharing intention follow those of the human participants across the 40 headlines. This agreement does not come from higher accuracy: when the fact-check is wrong, MARS agents lower their belief in a true headline by almost the same amount as people do, which an agent that is only more accurate would not do. In real discussions, forward-only simulation has no source and produces no change in denial replies, whereas reverse querying produces a change closer to the real one. MARS responses also reduce the 24-hour forecasting error, a statistically significant gain over conventional features. MARS therefore captures real human behavior under source evidence, beyond more accurate agent decisions and reduced spread of false claims.

AuthorsOpenReview ↗

Response to Reviewer hVad (Part 2/2): W2/Q2

W2/Q2: Sensitivity to network topology, community structure, and the positions of agents that perform verification

The reviewer asks how sensitive the gains are to network topology, community structure, and the positions of agents that perform verification, and requests experiments on Erdős–Rényi and power-law networks. We ran the requested experiments, and reverse querying reduces exposure to false claims and increases denial messages on every network structure and at every position. We replaced the observed network with an Erdős–Rényi random network, a Barabási–Albert power-law network, and a two-community network, each with the same 24 agents and 44 edges, so that only the structure changes. On the power-law network, we also gave source access to only 6 of the 24 agents, placed at the highest-degree hubs, at random, or at the lowest-degree periphery. The table reports the share of agents per step that see the false claim in their feed (exposure) and the share that post a denial of it (denial messages).

Condition (25 real cascades each)Exposure to false claims: no source → MARSDenial messages: no source → MARS
Erdős–Rényi random network26.92% → 7.04%0.83% → 34.67%
Power-law (Barabási–Albert) network34.33% → 10.60%1.42% → 34.60%
Two-community network23.60% → 4.13%1.23% → 31.79%
Power-law network, source access for 6 agents at hubs34.33% → 14.46%1.42% → 11.81%
Power-law network, source access for 6 agents at random34.33% → 10.60%1.42% → 8.19%
Power-law network, source access for 6 agents at the periphery34.33% → 7.73%1.42% → 8.98%

The gains of MARS are robust to network structure and to the positions of agents that perform verification, and every change in the table is statistically significant. On all three network structures, reverse querying reduces exposure to false claims to less than one third and raises denial messages from about 1% to more than 30%. MARS does not require every agent to verify: with source access for only 6 of 24 agents, exposure falls by more than half at every position, and placing these agents at hubs produces the most denial messages. Reverse querying therefore remains effective on random, power-law, and community networks, wherever the agents that perform verification are placed.

In summary, the added experiments show that MARS agents respond to source evidence as real people do, and that reverse querying remains effective on random, power-law, and community networks, wherever the agents that perform verification are placed. Both results come from the step that MARS adds to diffusion simulation: receivers query the source of a claim before they forward it. We sincerely thank Reviewer hVad for these suggestions, which have strengthened the paper, and we will incorporate both sets of results into the revised paper.

References

[1] Matthew R. DeVerna, Harry Yaojun Yan, Kai-Cheng Yang, and Filippo Menczer. Fact-checking information from large language models can decrease headline discernment. Proceedings of the National Academy of Sciences, 121(50):e2322823121, 2024.

[2] Arkaitz Zubiaga, Maria Liakata, Rob Procter, Geraldine Wong Sak Hoi, and Peter Tolmie. Analysing How People Orient to and Spread Rumours in Social Media by Looking at Conversational Threads. PLOS ONE, 11(3):e0150989, 2016.

Comments on the paper

4 comments
AuthorsOpenReview ↗

General Response to All Reviewers (Part 1/3)

We thank all reviewers for their constructive comments, and for recognizing the problem and controlled experiments (hVad), the value of MM-InfoBench (QtZy), and the motivation for bringing source verification into diffusion simulation (dX1m). The reviews raise three shared concerns, which we address in turn: (1) how MARS differs from existing provenance tracking and tool-augmented agents; (2) whether the simulated decisions match real human behavior, beyond cascade forecasting; and (3) whether the results hold under other network structures, larger populations, multiple concurrent claims, and missing or incorrect source links. For (2) and (3) we ran new experiments, summarized below; details are in the individual responses.

1. Methodological contribution and innovation (QtZy W1; dX1m W2). Existing simulation frameworks such as OASIS [4] and RumorSphere [8] let each agent decide only from the message it receives, while provenance methods and tool-augmented agents handle each claim separately. Environments such as LMAgent [9] support visual inputs, but diffusion simulation remains text-only. Receivers therefore cannot obtain source evidence before forwarding a claim, and the source text and images are lost once it is rewritten (BGE-M3 [6] retrieves the source for only 8.42% of Weibo reposts). MARS contributes two mechanisms that none of these approaches provides. Source-directed reverse querying adds a backward step to forward-only diffusion: before forwarding, correcting, or withholding a claim, the receiver queries its source and decides from the returned evidence. Provenance-preserving multimodal propagation keeps each claim linked to its original text and images through every rewrite, so the query always returns the correct multimodal source.

Comparison (same agent LLM, Qwen3-4B)ExistingMARS
Consistent decisions: BGE-M3 retrieval vs. MARS source33.17%42.75%
Consistent decisions: source text vs. source text + image25.76%33.23%
Denial messages: no source vs. reverse querying1.85%50.65%
Exposure to false claims: no source vs. reverse querying5.06%0.42%

Both mechanisms improve agent decisions with the agent LLM unchanged, and all four gains are statistically significant. With the source returned by MARS, agents more often deny a false claim or endorse a true one (consistent decisions) than with BGE-M3 retrieval, and source images add evidence beyond source text. With reverse querying, the share of agents per step that post a denial of the false claim rises from 1.85% to 50.65%, and the share that see the false claim in their feed falls from 5.06% to 0.42% (25 real cascades). MARS therefore goes beyond integrating existing components: reverse querying introduces source verification as a new step of diffusion simulation, and provenance-preserving multimodal propagation keeps this step effective after claims are rewritten.

AuthorsOpenReview ↗

General Response to All Reviewers (Part 2/3)

2. Agreement with real human behavior and predictive value (hVad W1/Q1; QtZy W3/Q2/W4; dX1m W1/Q1). The reviewers ask for direct evidence, beyond cascade forecasting, that MARS agents behave like real people. We added two direct comparisons with human data, and both show that MARS agents respond to source evidence as people do. First, we gave MARS agents the same 40 news headlines and fact-check evidence as the 1,548 participants of a published human experiment [1], and measured for each headline how much the evidence changes belief and sharing intention. Second, in 1,416 real PHEME replies [2] we measured how the share of denial replies changes after the correcting source appears, and simulated the same 18 cascades with MARS. Forecasting is a separate test of predictive value: MARS responses are added to conventional features to predict the size of held-out Weibo cascades.

Comparison with real dataResult
Human experiment [1]: Pearson correlation coefficient r (−1 to 1; higher means closer agreement) between MARS and humans in belief change after evidencer = 0.460
Human experiment [1]: the same correlation coefficient r in sharing-intention change after evidencer = 0.511
PHEME replies [2]: mean absolute error of the change in denial replies after the correcting source appearsForward-only 4.12%; MARS 3.20%
1,814 held-out Weibo cascades: mean absolute error of cascade size after 24 hoursConventional features 424.91 posts; with MARS 420.93 posts

MARS agents respond to source evidence as real people do, which forward-only simulation cannot represent. Their changes in belief and sharing intention follow those of the human participants across the 40 headlines, and reverse querying produces a change in denial replies closer to real discussions than forward-only simulation, which produces none. MARS responses averaged over three LLMs (Qwen3-4B, Qwen2.5-7B, Llama-3.1-8B) also give a statistically significant reduction of the forecasting error, so the gain does not depend on one selected LLM. MARS behavior therefore agrees with real human behavior in a controlled experiment and in real discussions, and it carries predictive information beyond conventional features.

3. Other network structures, larger populations, concurrent claims, and unreliable source links (hVad W2/Q2; QtZy W2/Q1/Q3/W5; dX1m Q2/Q3). The reviewers ask whether the gains hold beyond the setting tested in the paper. We added experiments, and in every added condition reverse querying reduces exposure to false claims and increases denial messages. Each row uses the same 25 real cascades and compares agents without source access to MARS agents; both metrics are shares of agents per step, as in Section 1.

ConditionExposure to false claims: no source → MARSDenial messages: no source → MARS
Erdős–Rényi random network26.92% → 7.04%0.83% → 34.67%
Power-law (Barabási–Albert) network34.33% → 10.60%1.42% → 34.60%
Two-community network23.60% → 4.13%1.23% → 31.79%
Source access for only 6 of 24 agents, placed at hubs34.33% → 14.46%1.42% → 11.81%
Source access for only 6 of 24 agents, placed at the periphery34.33% → 7.73%1.42% → 8.98%
128 agents (paper: 64)0.89% → 0.08%1.91% → 49.14%
Half of the source links wrong or missing34.33% → 7.60%1.42% → 23.31%

Reverse querying remains effective in every added condition, and all changes in the table are statistically significant. The effect holds on random, power-law, and community networks, wherever the agents that perform verification are placed, with twice as many agents, and when half of the source links are wrong or missing. The single-claim setting is an experimental simplification: each message carries the ID of its own claim, so the query returns the source of that claim, and with two claims spreading in one network, responses that endorse a false claim or deny a true one fall from 7.29% (one source for the whole event) to 5.92%. For malicious rewrites, MARS compares each message with its parent claim, detects 91.84% of meaning-changing rewrites, and registers each as a new claim linked to its parent. The effect of reverse querying therefore does not depend on a particular network structure, population size, single-claim cascades, or perfect source links.

AuthorsOpenReview ↗

General Response to All Reviewers (Part 3/3)

In summary, MARS adds source verification to diffusion simulation: receivers query the original text and images of a claim before they forward it. No existing simulator or provenance method provides this step; it raises denial messages against false claims from 1.85% to 50.65%, agrees with how real people respond to source evidence, and remains effective across network structures, population sizes, and unreliable source links. We sincerely thank all reviewers for their constructive comments, which have substantially improved the paper, and we will incorporate these results into the revised paper.

References

[1] Matthew R. DeVerna, Harry Yaojun Yan, Kai-Cheng Yang, and Filippo Menczer. Fact-checking information from large language models can decrease headline discernment. Proceedings of the National Academy of Sciences, 121(50):e2322823121, 2024.

[2] Arkaitz Zubiaga, Maria Liakata, Rob Procter, Geraldine Wong Sak Hoi, and Peter Tolmie. Analysing How People Orient to and Spread Rumours in Social Media by Looking at Conversational Threads. PLOS ONE, 11(3):e0150989, 2016.

[4] Ziyi Yang, Zaibin Zhang, Zirui Zheng, Yuxian Jiang, Ziyue Gan, Zhiyu Wang, Zijian Ling, Jinsong Chen, Martz Ma, Bowen Dong, Prateek Gupta, Shuyue Hu, Zhenfei Yin, Guohao Li, Xu Jia, Lijun Wang, Bernard Ghanem, Huchuan Lu, Chaochao Lu, Wanli Ouyang, Yu Qiao, Philip Torr, and Jing Shao. OASIS: Open Agent Social Interaction Simulations with One Million Agents. arXiv preprint, arXiv:2411.11581, 2024.

[6] Jianlyu Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. M3-Embedding: Multi-Linguality, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation. Findings of the Association for Computational Linguistics: ACL 2024, pp. 2318–2335, 2024.

[8] Yijun Liu, Wu Liu, Xiaoyan Gu, Hantao Yao, Weiping Wang, Jiebo Luo, and Yongdong Zhang. RumorSphere: A Framework for Million-Scale Agent-Based Dynamic Simulation of Rumor Propagation. arXiv preprint, arXiv:2509.02172, 2025.

[9] Yijun Liu, Wu Liu, Xiaoyan Gu, Xiaodong He, Yong Rui, and Yongdong Zhang. LMAgent: A Large-Scale Multimodal Agents Society for Multi-User Simulation. IEEE Transactions on Multimedia, 28:2668–2680, 2026.

AuthorsOpenReview ↗

Revised paper uploaded: summary of changes

We sincerely thank all reviewers again for their constructive comments. We have uploaded the revised paper, which incorporates the added experiments reported in our responses. All changes are marked in blue.

Reviewer concernLocation in the revised paper
Agreement of MARS with real human behavior (hVad W1/Q1; QtZy W3/Q2; dX1m W1/Q1)Section 4.5, first paragraph; Appendix N.1 and N.2
Network topology, community structure, and positions of the agents that verify (hVad W2/Q2)Section 4.3, last paragraph and Table 3; Appendix N.3
Difference from provenance methods and tool-augmented agents (QtZy W1; dX1m W2)Section 2, second paragraph
Missing, incorrect, and conflicting source links (QtZy W2/Q1/Q3)Section 4.2, last paragraph; Appendix N.4
Computational cost (QtZy W4)Section 4.5, last paragraph; Appendix N.8
Larger networks and competing claims (QtZy W5; dX1m Q3)Section 4.3, last paragraph and Table 3; Appendix N.3 and N.6
Rewrites that change the meaning of a claim (dX1m Q2)Section 4.2, last paragraph; Appendix N.5
Comparison with the same source content (dX1m Q4)Table 2(c); Appendix N.7

The abstract, the third contribution, the Discussion, and the Conclusion are updated accordingly.

The revised paper addresses the weaknesses and questions raised by the reviewers. We would be glad to answer any further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.