Research map / Multi-agent systems and games / LLM agents

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

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces ManBench, a benchmark of 20 BIG-Bench Hard tasks (4,838 questions) and five interaction protocols to measure how LLM multi-agent systems collectively adopt false memories (Mandela effect), and tests prompt-level and SFT-based defenses.

Problem

Multi-agent LLM systems may acquire collective cognitive biases such as the Mandela effect, where groups share false memories induced by social influence; this has not been systematically measured.

Approach

A multiple-choice benchmark built from BBH tasks with planted plausible distractors, four socially-induced false-memory protocols crossing group composition (generic vs role-based) with memory timescale (short vs long-term consolidation/retrieval), and metrics for reality shift rate; defenses are cognitive anchoring and source scrutiny prompts plus SFT on resilience and cooperative data.

Key results

• All 13 evaluated LLMs show higher error rates under social protocols; e.g., Qwen3-235B's error rises from 25.48% baseline to 74.75% under the Role-based Short-term Protocol.
• Prompt-level cognitive anchoring reduces GPT-4o's generic short-term reality shift from 46.0% to 17.8%.
• SFT on the combined resilience and cooperative data reduces Llama3's role-based short-term reality shift from 99.5% to 21.5% while keeping correct-guidance shift at 1.1%.

Surprising findings

• Larger models in the Qwen3 family show higher maximal reality shift (89.3% at 8B to 92.2% at 235B), an inverse scaling pattern.
• Error rates in role-based groups fall below baseline for groups of nine or more agents, attributed to "suspicion-induced vigilance".

Limitations

• Multiple-choice format simplifies real-world unstructured multi-agent dialogue.
• Training only on resilience data makes models dismiss all social input.

Experimental setup

Datasets: BIG-Bench Hard MedMCQA

Benchmarks: ManBench MedMCQA

Metrics: error rate reality shift rate maximal reality shift rate

Baselines: Baseline Reality Protocol Generic Short-term Protocol Generic Long-term Protocol

Builds on: BIG-Bench Hard multi-agent debate

Closest papers by meaning

• Systematic Failures in Collective Reasoning under Distributed Information in Multi-Agent LLMs (ICML 2026 · similarity 0.87)
• Emergent Coordination in Multi-Agent Language Models (ICLR 2026 · similarity 0.87)
• Do as We Do, Not as You Think: the Conformity of Large Language Models (ICLR 2025 · similarity 0.87)
• From Single to Societal: Analyzing Persona-Induced Bias in Multi-Agent Interactions (AAAI 2026 · similarity 0.87)
• G-Memory: Tracing Hierarchical Memory for Multi-Agent Systems (NeurIPS 2025 · similarity 0.87)
• Large Language Model Agents Are Not Always Faithful Self-Evolvers (ICML 2026 · similarity 0.86)