ai-research-survey

scan.json (25253B)

---

 {
   "paper": {
     "title": "Training Generalizable Collaborative Agents via Strategic Risk Aversion",
     "authors": ["Chengrui Qu", "Yizhou Zhang", "Nicholas Lanzetti", "Eric Mazumdar"],
     "year": 2026,
     "venue": "arXiv",
     "arxiv_id": "2602.21515"
   },
   "scan_version": 2,
   "active_modules": ["experimental_rigor", "data_leakage"],
   "methodology_tags": ["theoretical", "benchmark-eval"],
   "key_findings": "Strategic risk aversion (RQE) provably increases collaboration and eliminates free-riding in structured collaborative games. The proposed SRPO algorithm consistently achieves better cross-play performance with unseen partners than IPPO across Overcooked, Tag, Hanabi, and an LLM debate task on GSM8K. IPPO agents consistently learn free-riding equilibria that degrade under partner shifts, while SRPO agents generalize more robustly. In LLM experiments, SRPO improves joint accuracy by up to 19.27% and individual accuracy by up to 14.49% when paired with untuned partners.",
   "checklist": {
     "artifacts": {
       "code_released": {
         "applies": true,
         "answer": false,
         "justification": "The paper states 'The code will be released upon publication' (Section 6), which is a promise of future release, not an actual release."
       },
       "data_released": {
         "applies": true,
         "answer": true,
         "justification": "The experiments use publicly available benchmarks: PettingZoo MPE Tag, Hanabi, GSM8K, and the Overcooked environment is described in full detail. No proprietary data was collected."
       },
       "environment_specified": {
         "applies": true,
         "answer": false,
         "justification": "No requirements.txt, Dockerfile, or detailed environment specification is provided. The paper mentions using the verl training framework and PettingZoo but does not provide version numbers or dependency lists."
       },
       "reproduction_instructions": {
         "applies": true,
         "answer": false,
         "justification": "No step-by-step reproduction instructions are provided. Appendix D gives experimental details (hyperparameters, training steps) but no runnable commands or scripts."
       }
     },
     "statistical_methodology": {
       "confidence_intervals_or_error_bars": {
         "applies": true,
         "answer": true,
         "justification": "Figures 2(iii), 3(iii), 4(b), and 8 show mean and standard deviation bars for training vs cross-play performance differences."
       },
       "significance_tests": {
         "applies": true,
         "answer": false,
         "justification": "The paper claims SRPO 'outperforms' and 'consistently achieves' better results than IPPO but provides no statistical significance tests (p-values, t-tests, etc.) for any comparison."
       },
       "effect_sizes_reported": {
         "applies": true,
         "answer": true,
         "justification": "Tables 1 and 2 report percentage improvements (e.g., '+19.27%', '+14.49%') with baseline and SRPO values, providing context for effect magnitude."
       },
       "sample_size_justified": {
         "applies": true,
         "answer": false,
         "justification": "The paper uses 30 independent runs for Overcooked/Tag and 10 for Hanabi without justifying why these numbers were chosen or whether they are sufficient."
       },
       "variance_reported": {
         "applies": true,
         "answer": true,
         "justification": "Standard deviations are reported in Figures 2(iii), 3(iii), 4(b), and 8 for training and cross-play performance. Multiple independent runs (30 or 10) are used."
       }
     },
     "evaluation_design": {
       "baselines_included": {
         "applies": true,
         "answer": true,
         "justification": "IPPO is used as the primary baseline throughout, justified as 'the dominant scalable algorithm for collaborative MARL across numerous benchmarks' (Section 6)."
       },
       "baselines_contemporary": {
         "applies": true,
         "answer": true,
         "justification": "IPPO is a contemporary and widely-used baseline. The paper also discusses Forkel and Foerster (2025) entropy-based approach and includes entropy ablations."
       },
       "ablation_study": {
         "applies": true,
         "answer": true,
         "justification": "Section 6 and Appendix D.6 include ablation studies varying risk aversion parameter τ and entropy coefficient ε across Overcooked and Tag (Figures 13, 14)."
       },
       "multiple_metrics": {
         "applies": true,
         "answer": true,
         "justification": "The paper reports training performance, cross-play performance, free-riding degree, and in the LLM experiments both joint accuracy and individual accuracy with untuned partner."
       },
       "human_evaluation": {
         "applies": false,
         "answer": false,
         "justification": "Human evaluation is not relevant to the claims, which concern algorithmic performance on simulated environments and automated benchmarks."
       },
       "held_out_test_set": {
         "applies": true,
         "answer": true,
         "justification": "Cross-play evaluation pairs agents with held-out partners not encountered during training. Tag also evaluates against an unseen runner. The LLM experiments use a held-out validation set (Section D.5)."
       },
       "per_category_breakdown": {
         "applies": true,
         "answer": true,
         "justification": "Results are broken down per environment (Overcooked, Tag, Hanabi, GSM8K) and per model combination in Tables 1 and 2. Cross-play matrices show per-pair performance."
       },
       "failure_cases_discussed": {
         "applies": true,
         "answer": true,
         "justification": "The paper discusses where SRPO has tradeoffs: 'In the case of Tag, we can trace this back to the lack of the aggregative structure required for Theorem 4.1' (Section 6.2), noting slightly worse training performance."
       },
       "negative_results_reported": {
         "applies": true,
         "answer": true,
         "justification": "The paper reports that SRPO has slightly lower training performance in Tag (Section 6.2) and acknowledges that in some games 'risk aversion introduces conservatism and a corresponding decrease in performance' (Section 4.1, Remark 4.3)."
       }
     },
     "claims_and_evidence": {
       "abstract_claims_supported": {
         "applies": true,
         "answer": true,
         "justification": "The abstract claims about better equilibrium outcomes (Theorem 4.1), reduced free-riding (Theorem 4.5), and reliable collaboration with unseen partners (Section 6 experiments) are all supported by results in the paper."
       },
       "causal_claims_justified": {
         "applies": true,
         "answer": true,
         "justification": "Causal claims like 'risk aversion induces collaboration' are supported by formal proofs (Theorems 4.1, 4.5) for structured games and by controlled ablation studies varying τ while holding other factors constant (Figures 2, 13, 14)."
       },
       "generalization_bounded": {
         "applies": true,
         "answer": true,
         "justification": "The paper states theoretical results hold for specific game classes (quadratic aggregative games, finite symmetric games) and empirical results are bounded to the tested environments. The LLM experiment is described as 'preliminary small-scale experiments' (Section 1)."
       },
       "alternative_explanations_discussed": {
         "applies": true,
         "answer": true,
         "justification": "The paper discusses that entropy regularization alone (Forkel and Foerster 2025) could explain generalization, and Section D.6 ablations show entropy alone does not solve free-riding in Overcooked/Tag. Section D.5 provides evidence SRPO improves collaboration rather than individual reasoning."
       },
       "proxy_outcome_distinction": {
         "applies": true,
         "answer": true,
         "justification": "The paper's claims match measurement granularity: cross-play performance is directly the metric for partner generalization claims, joint accuracy for coordination claims. No broader framing beyond what is measured."
       }
     },
     "setup_transparency": {
       "model_versions_specified": {
         "applies": true,
         "answer": true,
         "justification": "LLM experiments specify exact model names: Qwen2.5-0.5B-Instruct, Qwen2.5-3B-Instruct, Qwen3-0.6B, Qwen3-4B-Instruct-2507, and Llama 3.2-1B-Instruct (Section 6.4)."
       },
       "prompts_provided": {
         "applies": true,
         "answer": false,
         "justification": "The LLM debate setup describes the protocol (three-round debate) but does not provide the actual prompt text used for the agents."
       },
       "hyperparameters_reported": {
         "applies": true,
         "answer": true,
         "justification": "Appendix D provides detailed hyperparameters: learning rates (7×10⁻⁴ actor, 1×10⁻³ critic), entropy coefficients, risk aversion τ values, training steps, network architectures, and clipping parameters for each environment."
       },
       "scaffolding_described": {
         "applies": false,
         "answer": false,
         "justification": "No agentic scaffolding is used. The LLM agents are fine-tuned with RL (verl framework) and interact through a structured debate protocol, not agentic tool-use scaffolding."
       },
       "data_preprocessing_documented": {
         "applies": true,
         "answer": true,
         "justification": "Environment specifications are detailed in Appendix D: grid sizes, action spaces, reward structures, observation spaces, episode lengths, and evaluation protocols are fully documented for each environment."
       }
     },
     "limitations_and_scope": {
       "limitations_section_present": {
         "applies": true,
         "answer": false,
         "justification": "There is no dedicated limitations section. The conclusion briefly mentions future work ('extending strategic risk aversion to broader agentic AI settings') but does not substantively discuss limitations."
       },
       "threats_to_validity_specific": {
         "applies": true,
         "answer": false,
         "justification": "No threats to validity are discussed. The paper acknowledges tradeoffs in specific environments (Tag) but does not discuss systematic threats to the validity of its conclusions."
       },
       "scope_boundaries_stated": {
         "applies": true,
         "answer": true,
         "justification": "Theoretical results are explicitly bounded to specific game classes (quadratic aggregative, finite symmetric). The LLM experiment is described as 'preliminary small-scale experiments' and 'proof-of-concept' (Sections 1, 6.4)."
       }
     },
     "data_integrity": {
       "raw_data_available": {
         "applies": true,
         "answer": false,
         "justification": "No raw experimental data (training logs, individual run results) is made available. Only aggregated results in figures and tables."
       },
       "data_collection_described": {
         "applies": true,
         "answer": true,
         "justification": "Data collection is fully described: environments are specified, training interaction budgets stated (e.g., 2×10⁶ for Overcooked, 3×10⁷ for Tag/Hanabi), and evaluation protocols detailed in Appendix D."
       },
       "recruitment_methods_described": {
         "applies": false,
         "answer": false,
         "justification": "No human participants. All experiments use simulated environments and public benchmarks."
       },
       "data_pipeline_documented": {
         "applies": true,
         "answer": true,
         "justification": "The pipeline from training to evaluation is documented: number of runs, evaluation frequency, cross-play matrix construction, number of evaluation episodes per pair, and episode lengths are all specified in Appendix D."
       }
     },
     "conflicts_of_interest": {
       "funding_disclosed": {
         "applies": true,
         "answer": false,
         "justification": "No funding information or acknowledgments section is present in the paper."
       },
       "affiliations_disclosed": {
         "applies": true,
         "answer": true,
         "justification": "All four authors are listed as affiliated with Caltech, Department of Computing and Mathematical Sciences."
       },
       "funder_independent_of_outcome": {
         "applies": true,
         "answer": false,
         "justification": "Cannot assess funder independence as no funding source is disclosed."
       },
       "financial_interests_declared": {
         "applies": true,
         "answer": false,
         "justification": "No competing interests statement is included in the paper."
       }
     },
     "contamination": {
       "training_cutoff_stated": {
         "applies": true,
         "answer": false,
         "justification": "The LLM experiments fine-tune Qwen and Llama models on GSM8K but do not state the training data cutoff dates for these base models, which likely saw GSM8K during pre-training."
       },
       "train_test_overlap_discussed": {
         "applies": true,
         "answer": false,
         "justification": "GSM8K (published 2021) was very likely in the training data of all models used (Qwen2.5, Qwen3, Llama 3.2). This overlap is not discussed."
       },
       "benchmark_contamination_addressed": {
         "applies": true,
         "answer": false,
         "justification": "GSM8K is a widely-known benchmark published in 2021. All models used were trained well after 2021. No contamination analysis is performed."
       }
     },
     "human_studies": {
       "pre_registered": {
         "applies": false,
         "answer": false,
         "justification": "No human participants in this study."
       },
       "irb_or_ethics_approval": {
         "applies": false,
         "answer": false,
         "justification": "No human participants in this study."
       },
       "demographics_reported": {
         "applies": false,
         "answer": false,
         "justification": "No human participants in this study."
       },
       "inclusion_exclusion_criteria": {
         "applies": false,
         "answer": false,
         "justification": "No human participants in this study."
       },
       "randomization_described": {
         "applies": false,
         "answer": false,
         "justification": "No human participants in this study."
       },
       "blinding_described": {
         "applies": false,
         "answer": false,
         "justification": "No human participants in this study."
       },
       "attrition_reported": {
         "applies": false,
         "answer": false,
         "justification": "No human participants in this study."
       }
     },
     "cost_and_practicality": {
       "inference_cost_reported": {
         "applies": true,
         "answer": false,
         "justification": "No inference cost, API costs, or wall-clock time is reported for any experiment, including the LLM fine-tuning which likely required significant GPU resources."
       },
       "compute_budget_stated": {
         "applies": true,
         "answer": false,
         "justification": "No GPU hours, hardware specifications, or total training time is reported. The paper states interaction budgets but not actual compute costs."
       }
     },
     "experimental_rigor": {
       "seed_sensitivity_reported": {
         "applies": true,
         "answer": true,
         "justification": "Results are reported across 30 independent runs for Overcooked/Tag and 10 for Hanabi, with standard deviations shown in summary figures."
       },
       "number_of_runs_stated": {
         "applies": true,
         "answer": true,
         "justification": "Explicitly stated: '30 independent runs for each method' (Overcooked, Tag), 10 IPPO and 10 SRPO agents for Hanabi (Appendix D)."
       },
       "hyperparameter_search_budget": {
         "applies": true,
         "answer": false,
         "justification": "No hyperparameter search budget is reported. The risk aversion parameter τ and entropy ε appear tuned per environment but no search procedure is described."
       },
       "best_config_selection_justified": {
         "applies": true,
         "answer": true,
         "justification": "The ablation studies (Figures 13, 14) show results across a range of τ and ε values, and the paper provides rationale for chosen values based on the ablation results."
       },
       "multiple_comparison_correction": {
         "applies": true,
         "answer": false,
         "justification": "No statistical tests are performed at all, so no multiple comparison correction is applied despite multiple environment and model comparisons."
       },
       "self_comparison_bias_addressed": {
         "applies": true,
         "answer": false,
         "justification": "The authors propose SRPO and evaluate it against baselines they implemented. The bias of author-evaluation is not acknowledged."
       },
       "compute_budget_vs_performance": {
         "applies": true,
         "answer": true,
         "justification": "The paper explicitly states 'the number of interactions with the environment for both SRPO and IPPO are kept the same' (Section 6) and notes 'SRPO and IPPO take roughly the same time to train' (Appendix D.1)."
       },
       "benchmark_construct_validity": {
         "applies": true,
         "answer": true,
         "justification": "The paper discusses what each benchmark tests: Overcooked for free-riding and private costs, Tag for continuous control coordination, Hanabi for implicit coordination, GSM8K debate for LLM collaboration. The benchmarks are chosen to match specific theoretical predictions."
       },
       "scaffold_confound_addressed": {
         "applies": false,
         "answer": false,
         "justification": "No scaffolding is involved. SRPO and IPPO use identical architectures and training infrastructure, differing only in the objective function."
       }
     },
     "data_leakage": {
       "temporal_leakage_addressed": {
         "applies": true,
         "answer": false,
         "justification": "GSM8K was published in 2021 and all LLMs used were trained after 2021. The paper does not discuss whether models may have memorized GSM8K problems during pre-training."
       },
       "feature_leakage_addressed": {
         "applies": true,
         "answer": false,
         "justification": "No discussion of whether the debate protocol leaks information beyond what would be available in real collaborative settings."
       },
       "non_independence_addressed": {
         "applies": false,
         "answer": false,
         "justification": "The MARL environments (Overcooked, Tag, Hanabi) are simulations with procedurally generated episodes, so train/test independence is structural. GSM8K train/test split is standard."
       },
       "leakage_detection_method": {
         "applies": true,
         "answer": false,
         "justification": "No leakage detection method is applied despite using GSM8K with models that likely saw it during pre-training."
       }
     }
   },
   "claims": [
     {
       "claim": "Strategic risk aversion monotonically increases the shared reward in continuous quadratic aggregative games (Theorem 4.1).",
       "evidence": "Formal proof in Appendix B.1 (Theorem B.4) with detailed derivation. Validated in Example 4.2 and Figure 1(i).",
       "supported": "strong"
     },
     {
       "claim": "Strategic risk aversion eliminates free-riding above a game-dependent threshold (Theorem 4.5).",
       "evidence": "Formal proof in Appendix B.2 (Theorem B.9) for finite symmetric collaborative games. Validated in Example 4.6 and Figure 1(ii).",
       "supported": "strong"
     },
     {
       "claim": "SRPO consistently achieves higher and more stable cross-play performance than IPPO across collaborative benchmarks.",
       "evidence": "Cross-play matrices in Figures 2-4, 7-8 across Overcooked, Tag, and Hanabi. Performance drop from training to cross-play is smaller for SRPO (e.g., -15.8% vs -24.8% in 4-player Hanabi, Figure 4b).",
       "supported": "moderate"
     },
     {
       "claim": "IPPO consistently learns free-riding equilibria that limit generalization.",
       "evidence": "Checkerboard patterns visible in cross-play matrices (Figures 2, 3). Ablation in Figure 2(ii) shows free-riding disappears as τ increases.",
       "supported": "strong"
     },
     {
       "claim": "SRPO improves cross-play joint accuracy by up to 19.27% and individual robustness by up to 14.49% in LLM debate on GSM8K.",
       "evidence": "Tables 1 and 2 show results across multiple Qwen model combinations and against untuned Llama partner.",
       "supported": "moderate"
     }
   ],
   "red_flags": [
     {
       "flag": "No statistical significance tests",
       "detail": "All comparative claims (SRPO vs IPPO) are based on comparing numbers and percentage improvements without any statistical tests. With 30 seeds per condition, proper tests are feasible and would strengthen the claims."
     },
     {
       "flag": "GSM8K contamination unaddressed",
       "detail": "GSM8K (2021) was almost certainly in the pre-training data of all models used (Qwen2.5, Qwen3, Llama 3.2). The paper does not discuss how pre-training memorization might interact with the debate protocol or SRPO fine-tuning."
     },
     {
       "flag": "No compute costs reported",
       "detail": "The LLM fine-tuning experiments (multiple model sizes, SRPO requires adversary training) likely required significant GPU resources, but no compute budget is reported."
     },
     {
       "flag": "No limitations section",
       "detail": "The paper lacks a dedicated limitations discussion despite making broad claims about scalable collaborative agent training."
     }
   ],
   "cited_papers": [
     {
       "title": "Autogen: Enabling next-gen LLM applications via multi-agent conversations",
       "authors": ["Qingyun Wu", "Gagan Bansal", "Jieyu Zhang"],
       "year": 2024,
       "relevance": "Foundational work on multi-agent LLM collaboration frameworks, directly motivating the LLM debate experiments."
     },
     {
       "title": "Why do multi-agent LLM systems fail?",
       "authors": ["Mert Cemri"],
       "year": 2025,
       "arxiv_id": "2503.13657",
       "relevance": "Studies failure modes in multi-agent LLM systems, directly relevant to understanding collaboration failures this paper addresses."
     },
     {
       "title": "Improving factuality and reasoning in language models through multiagent debate",
       "authors": ["Yilun Du", "Shuang Li", "Antonio Torralba"],
       "year": 2024,
       "relevance": "Introduces the multi-agent debate framework used in the GSM8K experiments."
     },
     {
       "title": "MAPoRL: Multi-agent post-co-training for collaborative large language models with reinforcement learning",
       "authors": ["Chanwoo Park"],
       "year": 2025,
       "relevance": "State-of-the-art multi-agent LLM post-training with RL, directly comparable to SRPO's LLM fine-tuning approach."
     },
     {
       "title": "The surprising effectiveness of PPO in cooperative multi-agent games",
       "authors": ["Chao Yu"],
       "year": 2022,
       "relevance": "Establishes IPPO as the dominant baseline for collaborative MARL, which SRPO builds upon and compares against."
     },
     {
       "title": "On the utility of learning about humans for human-ai coordination",
       "authors": ["Micah Carroll"],
       "year": 2019,
       "relevance": "Key prior work on partner generalization in collaborative AI, Overcooked benchmark origin."
     },
     {
       "title": "\"Other-play\" for zero-shot coordination",
       "authors": ["Hengyuan Hu"],
       "year": 2020,
       "relevance": "Foundational approach to zero-shot coordination that SRPO aims to improve upon."
     },
     {
       "title": "Unlocking the power of multi-agent LLM for reasoning: From lazy agents to deliberation",
       "authors": ["Zheng Zhang"],
       "year": 2025,
       "arxiv_id": "2511.02303",
       "relevance": "Directly studies free-riding ('lazy agents') in multi-agent LLM systems, validating the problem SRPO addresses."
     },
     {
       "title": "Multi-agent collaboration mechanisms: A survey of LLMs",
       "authors": ["Khanh-Tung Tran"],
       "year": 2025,
       "arxiv_id": "2501.06322",
       "relevance": "Survey of LLM collaboration mechanisms, relevant to understanding the landscape of multi-agent LLM cooperation."
     },
     {
       "title": "Entropy is all you need for inter-seed cross-play in hanabi",
       "authors": ["Jonathan Forkel", "Jakob Foerster"],
       "year": 2025,
       "relevance": "Key competing approach using entropy regularization for partner generalization; SRPO's ablations show entropy alone is insufficient."
     },
     {
       "title": "Tractable equilibrium computation in Markov games through risk aversion",
       "authors": ["Eric Mazumdar", "Kishan Panaganti", "Laixi Shi"],
       "year": 2024,
       "relevance": "Foundational work introducing RQE framework that SRPO builds upon."
     },
     {
       "title": "Training verifiers to solve math word problems",
       "authors": ["Karl Cobbe"],
       "year": 2021,
       "arxiv_id": "2110.14168",
       "relevance": "GSM8K dataset used in the LLM debate experiments."
     }
   ]
 }