ai-research-survey

scan-v5.json (25183B)

---

 {
   "scan_version": 5,
   "paper_type": "empirical",
   "paper": {
     "title": "Interpreting Emergent Extreme Events in Multi-Agent Systems",
     "authors": [
       "Ling Tang",
       "Jilin Mei",
       "Dongrui Liu",
       "Chen Qian",
       "Dawei Cheng"
     ],
     "year": 2026,
     "venue": "arXiv.org",
     "arxiv_id": "2601.20538",
     "doi": "10.48550/arXiv.2601.20538"
   },
   "checklist": {
     "claims_and_evidence": {
       "abstract_claims_supported": {
         "applies": true,
         "answer": true,
         "justification": "The abstract claims (first framework for explaining extreme events in MAS, Shapley-based attribution, three-dimensional aggregation, effectiveness across three scenarios) are all demonstrated in the paper body.",
         "source": "haiku"
       },
       "causal_claims_justified": {
         "applies": true,
         "answer": false,
         "justification": "The paper uses causal language ('drives', 'contributes to') in all five insights, but observations are drawn from only 5 independent trajectories per setting with N=4-20 agents; no design adequate for causal inference is employed.",
         "source": "haiku"
       },
       "generalization_bounded": {
         "applies": true,
         "answer": false,
         "justification": "The five insights are presented as general properties of extreme events across MAS, but are derived from only three simulated scenarios with very small agent counts and five trajectories each; the paper does not bound the generalization.",
         "source": "haiku"
       },
       "alternative_explanations_discussed": {
         "applies": true,
         "answer": false,
         "justification": "No alternative explanations are considered for any of the five insights; for example, the risk-instability correlation (Insight 3) could reflect shared confounders rather than a meaningful relationship.",
         "source": "haiku"
       },
       "proxy_outcome_distinction": {
         "applies": true,
         "answer": true,
         "justification": "Risk metrics are explicitly defined (EWMA conditional variance for EconAgent/TwinMarket, belief variance for SocialNetwork), and claims are made about these specific metrics rather than broader notions.",
         "source": "haiku"
       }
     },
     "limitations_and_scope": {
       "limitations_section_present": {
         "applies": true,
         "answer": false,
         "justification": "The conclusion mentions one limitation ('scaling to significantly larger systems remains future work') but there is no dedicated limitations or threats-to-validity section.",
         "source": "haiku"
       },
       "threats_to_validity_specific": {
         "applies": true,
         "answer": false,
         "justification": "No specific threats are discussed — the small trajectory count (n=5), expert-annotated thresholds, minimal agent counts, or reliance on default API parameters are never flagged as validity threats.",
         "source": "haiku"
       },
       "scope_boundaries_stated": {
         "applies": true,
         "answer": false,
         "justification": "The paper does not state what the results do not show; the five insights are presented without caveats about the restricted simulation settings from which they were derived.",
         "source": "haiku"
       }
     },
     "conflicts_of_interest": {
       "funding_disclosed": {
         "applies": true,
         "answer": false,
         "justification": "No funding acknowledgment appears anywhere in the paper text.",
         "source": "haiku"
       },
       "affiliations_disclosed": {
         "applies": true,
         "answer": true,
         "justification": "Author affiliations are clearly listed: Shanghai AI Laboratory, Shanghai Jiao Tong University, Fudan University, Renmin University of China, Tongji University.",
         "source": "haiku"
       },
       "funder_independent_of_outcome": {
         "applies": false,
         "answer": false,
         "justification": "No funding is disclosed, so independence cannot be assessed.",
         "source": "haiku"
       },
       "financial_interests_declared": {
         "applies": true,
         "answer": false,
         "justification": "There is no competing interests or financial interests declaration in the paper.",
         "source": "haiku"
       }
     },
     "scope_and_framing": {
       "key_terms_defined": {
         "applies": true,
         "answer": true,
         "justification": "Key terms are defined: 'extreme events' (Black Swans) are defined with three properties (outliers, extreme impact, retrospective predictability); 'Shapley value' is formally introduced; risk metrics are precisely specified per scenario.",
         "source": "haiku"
       },
       "intended_contribution_clear": {
         "applies": true,
         "answer": true,
         "justification": "The paper explicitly states it proposes 'the first framework for explaining emergent extreme events in MAS' and articulates three specific questions the framework answers.",
         "source": "haiku"
       },
       "engagement_with_prior_work": {
         "applies": true,
         "answer": true,
         "justification": "Section 2 situates the work against three streams: LLM-based MAS simulation, attribution methods in MAS, and quantitative analysis of extreme events, explaining how this work differs from each.",
         "source": "haiku"
       }
     }
   },
   "type_checklist": {
     "empirical": {
       "artifacts": {
         "code_released": {
           "applies": true,
           "answer": true,
           "justification": "The abstract explicitly states 'The source code is available at https://github.com/mjl0613ddm/IEEE.'",
           "source": "haiku"
         },
         "data_released": {
           "applies": true,
           "answer": false,
           "justification": "No specific simulation trajectories are released; the environments (EconAgent, TwinMarket) are referenced to other papers, and the paper does not release the specific trajectory data used in experiments.",
           "source": "haiku"
         },
         "environment_specified": {
           "applies": true,
           "answer": false,
           "justification": "No requirements file, Dockerfile, or dependency specification is mentioned; model APIs are accessed with 'default parameters' which are unspecified.",
           "source": "haiku"
         },
         "reproduction_instructions": {
           "applies": true,
           "answer": false,
           "justification": "No step-by-step reproduction instructions are provided in the paper; the reader is pointed to a GitHub link without further guidance.",
           "source": "haiku"
         }
       },
       "statistical_methodology": {
         "confidence_intervals_or_error_bars": {
           "applies": true,
           "answer": true,
           "justification": "Tables 1 and 2 report mean ± standard deviation across independent runs for all main results.",
           "source": "haiku"
         },
         "significance_tests": {
           "applies": true,
           "answer": false,
           "justification": "No statistical significance tests are used for any comparisons between methods or across scenarios despite comparative claims being made.",
           "source": "haiku"
         },
         "effect_sizes_reported": {
           "applies": true,
           "answer": true,
           "justification": "Table 3 reports percentage risk drop values that convey magnitude of improvement over baselines, providing interpretable effect-size information.",
           "source": "haiku"
         },
         "sample_size_justified": {
           "applies": true,
           "answer": false,
           "justification": "Only 5 independent trajectories per experimental setting are used; no justification or power analysis is provided for this choice.",
           "source": "haiku"
         },
         "variance_reported": {
           "applies": true,
           "answer": true,
           "justification": "Standard deviations are reported in Tables 1 and 2 across all metrics and model/scenario combinations.",
           "source": "haiku"
         }
       },
       "evaluation_design": {
         "baselines_included": {
           "applies": true,
           "answer": true,
           "justification": "Four baselines are included: Random, Failure Taxonomy (Cemri et al., 2025), Failure Attribution (Zhang et al., 2025b), and Agent Tracer (Zhang et al., 2025a).",
           "source": "haiku"
         },
         "baselines_contemporary": {
           "applies": true,
           "answer": true,
           "justification": "All baselines are from 2025, contemporary with the submission; they represent current approaches to MAS attribution.",
           "source": "haiku"
         },
         "ablation_study": {
           "applies": true,
           "answer": false,
           "justification": "No ablation of framework components is performed; Table 1 tests Monte Carlo sample sizes but this is a sensitivity analysis, not an ablation of design choices.",
           "source": "haiku"
         },
         "multiple_metrics": {
           "applies": true,
           "answer": true,
           "justification": "Five distinct metrics are proposed and evaluated (Ltm, Gag, Cag, Zag, Gbe) in addition to the faithfulness metric (risk drop).",
           "source": "haiku"
         },
         "human_evaluation": {
           "applies": false,
           "answer": false,
           "justification": "The framework evaluates computational systems; human evaluation is not relevant to the claims.",
           "source": "haiku"
         },
         "held_out_test_set": {
           "applies": false,
           "answer": false,
           "justification": "This is not a prediction task; results are derived from generated simulation trajectories rather than a train/test split.",
           "source": "haiku"
         },
         "per_category_breakdown": {
           "applies": true,
           "answer": true,
           "justification": "Results in Tables 2 and 3 are broken down per scenario (EconAgent, TwinMarket, SocialNetwork) and per LLM model.",
           "source": "haiku"
         },
         "failure_cases_discussed": {
           "applies": true,
           "answer": false,
           "justification": "The paper does not explicitly discuss failure cases of its own framework; results where their method underperforms (e.g., SocialNetwork Top-3 in Table 3) are not highlighted or analyzed.",
           "source": "haiku"
         },
         "negative_results_reported": {
           "applies": true,
           "answer": false,
           "justification": "In Table 3, the proposed method is outperformed by AT on SocialNetwork for several models, but this is not acknowledged or discussed as a negative result.",
           "source": "haiku"
         }
       },
       "setup_transparency": {
         "model_versions_specified": {
           "applies": true,
           "answer": true,
           "justification": "Specific model versions are named and cited: GPT-4o mini, Llama-3.1-8B-Instruct, Claude-3-Haiku, Qwen-Plus, DeepSeek-V3.2.",
           "source": "haiku"
         },
         "prompts_provided": {
           "applies": true,
           "answer": false,
           "justification": "Prompts for the FT and FA baselines are provided in the appendix, but the core agent prompts used in the EconAgent, TwinMarket, and SocialNetwork simulations are not provided — they are deferred to external papers.",
           "source": "haiku"
         },
         "hyperparameters_reported": {
           "applies": true,
           "answer": false,
           "justification": "'All LLM APIs were accessed using default parameters' — temperature, top-p, and other hyperparameters are not specified.",
           "source": "haiku"
         },
         "scaffolding_described": {
           "applies": true,
           "answer": true,
           "justification": "The agentic scaffolding is described in Section 4.1 and Appendix B, including agent action spaces, transition dynamics, risk metric definitions, and baseline action definitions for counterfactuals.",
           "source": "haiku"
         },
         "data_preprocessing_documented": {
           "applies": true,
           "answer": true,
           "justification": "The full data pipeline from trajectory generation to Shapley computation to metric derivation is documented in Sections 3 and Appendix A-B, including EWMA risk computations.",
           "source": "haiku"
         }
       },
       "data_integrity": {
         "raw_data_available": {
           "applies": true,
           "answer": false,
           "justification": "The raw trajectories used in experiments are not explicitly released; the code is available but specific experimental trajectories are not.",
           "source": "haiku"
         },
         "data_collection_described": {
           "applies": true,
           "answer": true,
           "justification": "The data generation process is described: run simulations with different random seeds until extreme events occur, collect 5 trajectories per setting.",
           "source": "haiku"
         },
         "recruitment_methods_described": {
           "applies": false,
           "answer": false,
           "justification": "No human participants; data is from LLM-powered simulations.",
           "source": "haiku"
         },
         "data_pipeline_documented": {
           "applies": true,
           "answer": true,
           "justification": "The pipeline from simulation trajectory to attribution scores to metric computation is fully documented in Sections 3-4 and Appendix A-B.",
           "source": "haiku"
         }
       },
       "contamination": {
         "training_cutoff_stated": {
           "applies": false,
           "answer": false,
           "justification": "The paper uses LLMs as agents in simulations, not evaluating model capabilities on benchmarks; training cutoff is not relevant.",
           "source": "haiku"
         },
         "train_test_overlap_discussed": {
           "applies": false,
           "answer": false,
           "justification": "Not applicable — LLMs are used as simulation agents, not evaluated on held-out benchmarks.",
           "source": "haiku"
         },
         "benchmark_contamination_addressed": {
           "applies": false,
           "answer": false,
           "justification": "Not applicable — the paper does not evaluate model capabilities on established benchmarks.",
           "source": "haiku"
         }
       },
       "human_studies": {
         "pre_registered": {
           "applies": false,
           "answer": false,
           "justification": "No human participants.",
           "source": "haiku"
         },
         "irb_or_ethics_approval": {
           "applies": false,
           "answer": false,
           "justification": "No human participants.",
           "source": "haiku"
         },
         "demographics_reported": {
           "applies": false,
           "answer": false,
           "justification": "No human participants.",
           "source": "haiku"
         },
         "inclusion_exclusion_criteria": {
           "applies": false,
           "answer": false,
           "justification": "No human participants.",
           "source": "haiku"
         },
         "randomization_described": {
           "applies": false,
           "answer": false,
           "justification": "No human participants.",
           "source": "haiku"
         },
         "blinding_described": {
           "applies": false,
           "answer": false,
           "justification": "No human participants.",
           "source": "haiku"
         },
         "attrition_reported": {
           "applies": false,
           "answer": false,
           "justification": "No human participants.",
           "source": "haiku"
         }
       },
       "cost_and_practicality": {
         "inference_cost_reported": {
           "applies": true,
           "answer": false,
           "justification": "The paper discusses computational complexity (O(M × |Ω|)) but reports no actual inference cost, latency, or API expenditure.",
           "source": "haiku"
         },
         "compute_budget_stated": {
           "applies": true,
           "answer": false,
           "justification": "No total compute budget or wall-clock time is reported anywhere in the paper.",
           "source": "haiku"
         }
       }
     }
   },
   "claims": [
     {
       "claim": "Monte Carlo sampling with M=1000 achieves >0.99 cosine similarity to exact Shapley values across five LLMs and two scenarios.",
       "evidence": "Table 1 shows cosine similarity consistently exceeds 0.99 at M=10^3 for both EconAgent and TwinMarket across all five tested LLMs.",
       "supported": "strong"
     },
     {
       "claim": "The proposed Shapley-based attribution is more faithful than competing MAS attribution methods (FT, FA, AT, Random) as measured by risk drop after deleting top-k attributed actions.",
       "evidence": "Table 3 shows the proposed method achieves the highest risk drop in the majority of settings, though AT outperforms on SocialNetwork for several models at Top-3.",
       "supported": "moderate"
     },
     {
       "claim": "Extreme events originate with distinct temporal patterns — either early dormant risks (EconAgent, Ltm>0.6) or immediate shocks (TwinMarket and SocialNetwork, Ltm≈0).",
       "evidence": "Table 2 shows consistently high Ltm for EconAgent and near-zero Ltm for TwinMarket/SocialNetwork across all five LLMs.",
       "supported": "moderate"
     },
     {
       "claim": "Extreme events are typically driven by a small subset of agents, with agent risk concentration Gag consistently above 0.4.",
       "evidence": "Table 2 shows Gag > 0.4 in most experimental settings across three scenarios and five LLMs.",
       "supported": "moderate"
     },
     {
       "claim": "Agents with high risk contribution exhibit high behavioral instability, as reflected by consistently positive risk-instability correlation Cag > 0.6.",
       "evidence": "Table 2 shows Cag > 0.6 in most EconAgent and TwinMarket settings; SocialNetwork shows weaker and sometimes negative correlations.",
       "supported": "weak"
     },
     {
       "claim": "A small number of behavior patterns contribute the majority of risk leading to extreme events, with behavior risk concentration Gbe > 0.5.",
       "evidence": "Table 2 shows Gbe consistently above 0.5 across all three scenarios and five LLMs.",
       "supported": "moderate"
     }
   ],
   "methodology_tags": [
     "empirical",
     "benchmark-eval",
     "case-study"
   ],
   "key_findings": "This paper proposes a Shapley value-based framework to explain extreme events in LLM-powered multi-agent systems by attributing risk contributions across temporal, agent, and behavioral dimensions. Experiments across economic, financial, and social network simulations show the framework more faithfully identifies high-risk actions than competing methods, measured by the risk drop when those actions are removed. Analysis of three scenarios reveals consistent patterns: economic simulation extreme events originate from early dormant risks while financial and social network events emerge from immediate shocks; a small subset of agents drives most risk; and risk-contributing agents tend to exhibit high behavioral instability. Monte Carlo approximation of Shapley values converges at M=1000 samples with >0.99 cosine similarity to exact values.",
   "red_flags": [
     {
       "flag": "Tiny sample size",
       "detail": "Only 5 independent trajectories per experimental setting are used to derive all five insights; this is insufficient for reliable statistical inference about general properties of extreme events."
     },
     {
       "flag": "Small agent counts",
       "detail": "Experiments use N=4-20 agents per scenario, far below real-world multi-agent system scales; the paper claims general insights without bounding to these constrained settings."
     },
     {
       "flag": "Expert-annotated thresholds",
       "detail": "Extreme event thresholds are determined by 'domain experts' without specifying who, how, or validation of inter-rater agreement, introducing unquantified subjectivity."
     },
     {
       "flag": "Unspecified LLM hyperparameters",
       "detail": "'All LLM APIs were accessed using default parameters' — temperature and sampling parameters are not reported, making exact reproduction impossible."
     },
     {
       "flag": "No significance testing",
       "detail": "All comparative claims in Table 3 are made without statistical significance tests, despite high variance in results (some negative risk drops indicate the method can harm attribution)."
     },
     {
       "flag": "Underperformance not addressed",
       "detail": "The proposed method is outperformed by Agent Tracer on SocialNetwork for multiple LLMs at Top-3 (e.g., AT 22.2% vs ours 17.9% for GPT) but this negative result is not discussed."
     },
     {
       "flag": "No limitations section",
       "detail": "The only acknowledged limitation is scalability to larger systems; threats from small samples, simulated environments, expert-annotated labels, or API parameter sensitivity are ignored."
     }
   ],
   "cited_papers": [
     {
       "title": "EconAgent: Large language model-empowered agents for simulating macroeconomic activities",
       "relevance": "Primary simulation environment used for economic scenario experiments; defines the EconAgent MAS framework applied in this paper."
     },
     {
       "title": "TwinMarket: A scalable behavioral and social simulation for financial markets",
       "relevance": "Primary simulation environment for financial market scenario; provides the MAS framework for market crash experiments."
     },
     {
       "title": "Decoding echo chambers: LLM-powered simulations revealing polarization in social networks",
       "relevance": "Provides the SocialNetwork simulation environment used as the third experimental scenario."
     },
     {
       "title": "Why do multi-agent LLM systems fail? (Cemri et al., 2025)",
       "relevance": "Baseline method (Failure Taxonomy) compared against in faithfulness evaluation; introduces 14 failure modes for MAS."
     },
     {
       "title": "Which agent causes task failures and when? On automated failure attribution of LLM multi-agent systems (Zhang et al., 2025b)",
       "relevance": "Baseline method (Failure Attribution) compared in faithfulness evaluation; direct prompting approach to MAS attribution."
     },
     {
       "title": "AgentTracer: Who is inducing failure in the LLM agentic systems? (Zhang et al., 2025a)",
       "relevance": "Baseline method (Agent Tracer) compared in faithfulness evaluation; counterfactual-based surrogate model approach."
     },
     {
       "title": "A unified approach to interpreting model predictions (SHAP/Lundberg & Lee, 2017)",
       "relevance": "Foundational work on Shapley values for ML interpretability that this paper adapts to the MAS attribution setting."
     },
     {
       "title": "Generative agents: Interactive simulacra of human behavior (Park et al., 2023)",
       "relevance": "Landmark paper establishing LLM-powered agents for human behavior simulation; motivates the MAS simulation paradigm."
     },
     {
       "title": "A value for n-person games (Shapley, 1953)",
       "relevance": "Original Shapley value paper providing the game-theoretic foundation for the attribution method."
     },
     {
       "title": "OASIS: Open agent social interaction simulations with one million agents (Yang et al., 2024)",
       "relevance": "Large-scale LLM-powered MAS simulation work contextualizing the scale at which extreme event interpretation is needed."
     }
   ],
   "engagement_factors": {
     "practical_relevance": {
       "score": 2,
       "justification": "Framework for identifying which agents and behaviors cause systemic failures in AI simulations has direct application to AI safety and multi-agent system design."
     },
     "surprise_contrarian": {
       "score": 1,
       "justification": "Applying Shapley values to MAS attribution is methodologically novel but the insights (few agents drive risk, behavior concentration) are intuitive rather than surprising."
     },
     "fear_safety": {
       "score": 2,
       "justification": "Paper explicitly frames the problem as Black Swan events and systemic collapse in AI-powered systems, connecting to AI safety concerns about emergent dangerous behavior."
     },
     "drama_conflict": {
       "score": 1,
       "justification": "No significant controversy; the paper presents a technical framework without challenging established results or making strong contrarian claims."
     },
     "demo_ability": {
       "score": 2,
       "justification": "Code is released on GitHub, enabling practitioners to run the attribution framework on their own MAS simulations."
     },
     "brand_recognition": {
       "score": 1,
       "justification": "Shanghai AI Laboratory and affiliated universities are respected institutions but not in the top tier of brand recognition for this community."
     }
   },
   "hn_data": {
     "threads": [],
     "top_points": 0,
     "total_points": 0,
     "total_comments": 0
   }
 }