💬 NLP

When Agents "Misremember" Collectively: Exploring the Mandela Effect in LLM-based Multi-Agent Systems

This paper investigates the collective "Mandela effect" in LLM-based multi-agent systems by introducing the MANBENCH benchmark to quantify the phenomenon, analyzing its underlying causes, and proposing prompt-level and model-level strategies that successfully mitigate the effect by an average of 74.40%.

Original authors: Naen Xu, Hengyu An, Shuo Shi, Jinghuai Zhang, Chunyi Zhou, Changjiang Li, Tianyu Du, Zhihui Fu, Jun Wang, Shouling Ji

Published 2026-03-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Naen Xu, Hengyu An, Shuo Shi, Jinghuai Zhang, Chunyi Zhou, Changjiang Li, Tianyu Du, Zhihui Fu, Jun Wang, Shouling Ji

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a group of friends sitting around a table, trying to solve a trivia question. They all know the answer is "2013." But then, one friend confidently says, "Wait, I remember reading it was 1985!" Another friend chimes in, "Oh yeah, I saw a documentary about that!" A third adds, "I'm pretty sure the news covered it back then."

Soon, the whole group is convinced the answer is 1985. Even though they started with the truth, the social pressure and the convincing (but fake) stories made them "misremember" the past collectively. In psychology, this is called the Mandela Effect.

This paper asks a scary question: Do AI agents do the same thing?

The researchers found that yes, they do. When multiple AI "agents" (smart computer programs) talk to each other, they can fall into a trap where they collectively invent a false reality, even if they started out knowing the truth.

Here is a breakdown of the paper's findings using simple analogies:

1. The Experiment: The "Fake News" Party

The researchers built a testing ground called MANBENCH (think of it as a giant, automated game show). They set up 13 different AI models (like GPT-4, Claude, and Llama) and put them in different social scenarios to see if they could be tricked.

They used two main ways to trick the AI:

The "Crowd" Effect: Just a bunch of generic voices saying the same wrong thing.
The "Expert Panel" Effect: A group of AI agents playing specific roles to make the lie sound super convincing.
- The Initiator: Starts the lie.
- The Detail Guy: Adds fake but plausible facts.
- The Crowd Pleaser: Says, "Everyone agrees with this!"
- The Authority: Uses big words to sound like a professor.
- The Skeptic: Pretends to doubt it, then gets "convinced" by the group.

The Result: The AI agents were incredibly easy to trick. Even the smartest models changed their minds. If a group of "experts" told an AI that Nelson Mandela died in the 1980s (he actually died in 2013), the AI would eventually agree, even if it knew the truth at the start.

2. The "Memory" Trap

The paper discovered something fascinating about how the AI remembers.

Short-term: If the AI hears the lie during the conversation, it might get confused immediately.
Long-term: The scary part is that some AIs don't just get confused; they rewrite their memory. After the conversation ends, if you ask them later, they will confidently state the lie as a fact, as if they had always known it. It's like if you watched a movie where the hero died, but then your friends convinced you the hero actually survived, and the next day you genuinely remember the hero surviving.

3. Why Does This Happen?

The researchers found a few key reasons:

The "Inverted-U" Danger: You might think a huge group of liars would be suspicious. But the paper found that a medium-sized group (about 5 to 6 agents) is the most dangerous. They are small enough to seem like a natural conversation, but big enough to feel like a consensus. If the group gets too huge (like 15 people), the AI gets suspicious and thinks, "Wait, this is a conspiracy," and starts to think critically again.
Specialized Knowledge is Vulnerable: You'd think AI would be good at facts like "What is the capital of France?" But the study showed that even in specialized fields (like medicine or history), the AI is just as likely to be tricked by a convincing story as it is in general knowledge.
Bigger isn't Always Better: Surprisingly, making the AI "smarter" (giving it more brain power/parameters) didn't always fix the problem. Sometimes, bigger models were better at understanding the fake story, making them even more likely to believe it!

4. How to Stop It: The "Truth Shields"

The researchers didn't just find the problem; they built shields to stop it. They tested two main defenses:

The "Inner Anchor" (Cognitive Anchoring): This is like telling the AI, "Before you listen to anyone else, write down what you know is true. Stick to that unless someone proves you wrong with undeniable evidence." This forces the AI to trust its own internal database first.
The "Detective" (Source Scrutiny): This is like telling the AI, "Don't just listen to what they say; look at how they are saying it. Are they playing roles? Is the story too perfect? Is it a coordinated lie?" This turns the AI into a skeptic who analyzes the social dynamics rather than just absorbing the information.

The Result: These simple tricks (just changing the instructions) reduced the "Mandela Effect" by about 74%. It's like giving the AI a pair of sunglasses that helps it see through the fog of social pressure.

Why Should You Care?

Imagine a future where AI agents help doctors diagnose diseases, lawyers review contracts, or scientists analyze climate data. If these AIs start "misremembering" facts because they talked to each other too much, the consequences could be huge. A group of AI doctors might collectively decide a patient has a disease they don't have, simply because they convinced each other it was true.

The Takeaway:
AI is getting very good at talking to each other, but it's also getting very good at getting "peer-pressured." This paper warns us that we need to build "critical thinking" into AI systems so they don't just go along with the crowd, even when the crowd is a group of computers telling a convincing lie.

1. Problem Statement

The paper addresses a critical, underexplored vulnerability in Large Language Model (LLM)-based multi-agent systems: collective cognitive bias, specifically the Mandela Effect.

Definition: The Mandela Effect refers to a phenomenon where a group collectively misremembers a verifiable fact, often due to social influence and the reinforcement of false details.
The Gap: While existing research covers LLM hallucinations (individual errors) and simple conformity, it lacks a systematic understanding of how multi-agent interactions can solidify false collective memories that persist over time.
Risk: In high-stakes scenarios (e.g., medical diagnosis, legal contract review, fact-checking), agents reinforcing each other's false memories can lead to flawed group judgments, the spread of misinformation, and a loss of trust in AI-driven systems.

2. Methodology: MANBENCH

To systematically study this phenomenon, the authors introduce MANBENCH, a novel benchmark designed to evaluate the existence, causes, and mitigation of the Mandela Effect.

A. Task Curation

Source: Tasks are curated from BIG-Bench Hard (BBH).
Structure: The dataset contains 4,838 multiple-choice questions across four domains:
1. History, Time, & Events (narrative-driven).
2. Misconceptions & Social Cognition (contagious errors).
3. General Knowledge (verifiable facts).
4. Domain-Specific Knowledge (specialized fields).
Distractor Generation: For each question, an LLM is prompted to generate the most plausible incorrect answer to serve as the target for the "false memory."

B. Interaction Protocols

The benchmark simulates social influence through five interaction protocols combining two variables: Group Composition and Memory Timescale.

Group Composition:
- Generic Group: Undifferentiated agents taking turns to provide "specious evidence" (plausible but false details) to reinforce a consensus.
- Role-based Group: Five specialized agents with distinct strategic roles to construct a multi-faceted false narrative:
  - Error Conclusion Initiator: Sets the false premise.
  - Detail Support Provider: Adds fabricated but plausible details.
  - Group Consensus Reinforcer: Creates social proof/majority illusion.
  - Authority Endorser: Uses expert jargon to legitimize the lie.
  - Questioning Compromiser: Initially doubts but is eventually convinced, validating the narrative.
Memory Timescale:
- Short-term (Situational): Immediate response within the same conversation context.
- Long-term (Conviction): A two-stage process involving Memory Consolidation (distilling the false dialogue into a summary) followed by Memory Retrieval (answering a new question based only on that summary). This tests if the false memory becomes a stable, internalized belief.

C. Evaluation Metrics

Error Rate ( $Err_P$ ): Percentage of wrong answers under a specific protocol.
Reality Shift Rate ( $\sigma_P$ ): The proportion of questions the agent answered correctly in the baseline but answered incorrectly after social interaction.
Maximal Reality Shift Rate ( $\sigma_{max}$ ): The total proportion of correct baseline memories compromised by any of the four social protocols.

3. Experimental Setup

Models Evaluated: 13 representative LLMs, including 7 commercial models (GPT-4o/5, Claude 3.5/4, Gemini 2.5) and 6 open-source models (Llama 3 series, Deepseek-V3.1, Qwen3 series).
Scope: Experiments cover varying group sizes (1–15 agents), different knowledge domains, and model scales.

4. Key Results

A. Existence of the Effect (RQ1)

Universal Vulnerability: All 13 evaluated LLMs are susceptible to the Mandela Effect.
Amplification: Social influence significantly increases error rates. For example, Qwen-235B's error rate jumped from 25.48% (baseline) to 74.75% under the Role-based Short-term protocol. Even the strongest model, GPT-5, saw its error rate more than double.
Long-term Solidification: Some models (e.g., Claude 3.5 Haiku, Llama3.1-8B, Qwen3) internalize false memories, maintaining high reality shift rates even after memory consolidation and retrieval. Others (e.g., GPT-5) show better self-correction in the long term.

B. Influencing Factors (RQ2)

Group Composition: Role-based groups are significantly more potent than generic groups. The strategic use of specialized roles (Authority, Detail, etc.) creates a more convincing false reality.
Group Size (The Inverted-U Curve):
- In Generic groups, the effect intensifies with size and saturates around 7 agents.
- In Role-based groups, the effect peaks at 6 agents and then decreases as the group grows larger (9+). The authors attribute this to "suspicion-induced vigilance": a large, perfectly coordinated group triggers critical thinking and skepticism in the subject agent, causing them to self-correct.
Knowledge Domain: The effect is strongest in Domain-Specific Knowledge (67.46% shift rate) and History/Events, proving that even specialized, factual knowledge is vulnerable to social manipulation.
Model Scale: Scaling up does not guarantee safety. While some families (Claude, GPT) improve with size, others (Qwen3) show an inverse scaling law, where larger models are more susceptible, likely because they better understand complex false narratives without necessarily improving critical thinking.

C. Mitigation Strategies (RQ3)

The authors propose and validate two defense mechanisms, achieving an average 74.40% reduction in the Mandela Effect.

Prompt-Level Defenses:
- Cognitive Anchoring: Forces the agent to establish an "internal anchor" based on its own knowledge before evaluating external input, shifting from passive acceptance to active verification.
- Source Scrutiny: Trains the agent to analyze the structure of the conversation (identifying roles, rhetorical patterns, and coordinated consensus) rather than just the content.
- Result: Both significantly reduce reality shift rates, with Cognitive Anchoring being particularly effective against short-term influence.
Model-Level Defense (Supervised Fine-Tuning - SFT):
- Resilience Set: Trains the model to resist manipulative narratives using the prompt-level defense logic.
- Cooperative Set: Trains the model to accept valid guidance (corrective or enriching) to prevent the model from becoming dogmatically isolated.
- Result: A model trained on a balanced combination of both sets successfully resists false memories while retaining the ability to learn from correct social input. Training only on resilience causes the model to reject all social input, including correct facts.

5. Significance and Contributions

Benchmark Creation: MANBENCH is the first systematic framework to quantify collective false memory in multi-agent systems, moving beyond individual hallucination studies.
Mechanism Discovery: The paper identifies that role-based narratives and moderate group sizes are the primary drivers of the effect, and that model scale alone is not a solution.
Defense Validation: It demonstrates that both prompt engineering (cognitive anchoring) and model alignment (balanced SFT) are effective countermeasures.
Ethical Implications: The findings highlight the urgent need for resilience in AI systems used for decision-making, as social manipulation can lead to the rapid spread of misinformation in collaborative AI environments.

Conclusion

The paper establishes that LLM-based multi-agent systems are highly vulnerable to the Mandela Effect, where social dynamics can overwrite factual truth with collective false memories. However, by understanding the specific mechanisms (role-based influence, group size dynamics) and applying targeted defenses (cognitive anchoring, balanced SFT), these systems can be made significantly more resilient and ethically aligned.