ai-research-survey

scan-v5.json (27160B)

---

 {
   "scan_version": 5,
   "paper_type": "empirical",
   "paper": {
     "title": "Learning \"Partner-Aware\" Collaborators in Multi-Party Collaboration",
     "authors": [
       "Abhijnan Nath",
       "Nikhil Krishnaswamy"
     ],
     "year": 2025,
     "venue": "NeurIPS 2025",
     "arxiv_id": "2510.22462",
     "doi": "10.48550/arXiv.2510.22462"
   },
   "checklist": {
     "claims_and_evidence": {
       "abstract_claims_supported": {
         "applies": true,
         "answer": true,
         "justification": "Abstract claims that ICR achieves superior common ground convergence and outperforms standard RLHF/DPO baselines are directly supported by Table 1 results across both tasks in full-press and no-press conditions.",
         "source": "haiku"
       },
       "causal_claims_justified": {
         "applies": true,
         "answer": true,
         "justification": "Causal claims that counterfactual KL regularization produces better collaboration are supported by ablation of λIntent values (Fig 1b) and the PPO-CF baseline, which isolates the contribution of the KL term from counterfactual prompting alone.",
         "source": "haiku"
       },
       "generalization_bounded": {
         "applies": true,
         "answer": true,
         "justification": "Limitations section explicitly bounds claims to two task domains and 8B-scale models, noting untested conditions including Diplomacy, human collaborators, and larger-scale centralized training.",
         "source": "haiku"
       },
       "alternative_explanations_discussed": {
         "applies": true,
         "answer": false,
         "justification": "The paper does not adequately address the alternative that GPT-4o's role as both expert trajectory generator and fixed evaluation intervention agent inflates ICR performance via shared distributional priors with the base Llama-3-8B-Instruct model.",
         "source": "haiku"
       },
       "proxy_outcome_distinction": {
         "applies": true,
         "answer": true,
         "justification": "The paper explicitly distinguishes proxy training rewards (task accuracy only) from gold evaluation rewards (accuracy × common ground convergence), noting consensus reward was deliberately withheld during training to prevent reward hacking.",
         "source": "haiku"
       }
     },
     "limitations_and_scope": {
       "limitations_section_present": {
         "applies": true,
         "answer": true,
         "justification": "Section 7 contains a dedicated multi-paragraph 'Limitations and Future Work' subsection covering compute constraints, task diversity, fixed intervention agent, and data bottlenecks.",
         "source": "haiku"
       },
       "threats_to_validity_specific": {
         "applies": true,
         "answer": true,
         "justification": "Specific threats are named: 8B-scale training due to compute limits, only two task domains, GPT-4o's potential prior exposure to DeliData, and fixed GPT-4o intervention agent not reflecting real-world intervention diversity.",
         "source": "haiku"
       },
       "scope_boundaries_stated": {
         "applies": true,
         "answer": true,
         "justification": "Results are explicitly limited to two domains, one intervention agent (GPT-4o), 8B models, and AI-AI collaborations; the paper explicitly notes human collaboration data as a bottleneck.",
         "source": "haiku"
       }
     },
     "conflicts_of_interest": {
       "funding_disclosed": {
         "applies": true,
         "answer": true,
         "justification": "Full funding disclosure in acknowledgments: DARPA FACT program (HR00112490377), NSF awards (DRL 2019805, DRL 2454151, IIS 2303019), ARO Knowledge Systems (W911NF-25-1-0096), and ARPA-H PARADIGM program.",
         "source": "haiku"
       },
       "affiliations_disclosed": {
         "applies": true,
         "answer": true,
         "justification": "Authors clearly affiliated with SIGNAL Lab, Department of Computer Science, Colorado State University, with institutional emails provided.",
         "source": "haiku"
       },
       "funder_independent_of_outcome": {
         "applies": true,
         "answer": true,
         "justification": "All funders (DARPA, NSF, ARO, ARPA-H) are US government agencies with no commercial stake in ICR or the evaluated models.",
         "source": "haiku"
       },
       "financial_interests_declared": {
         "applies": true,
         "answer": false,
         "justification": "No competing interests statement or declaration of financial interests (patents, equity, consulting) is provided anywhere in the paper.",
         "source": "haiku"
       }
     },
     "scope_and_framing": {
       "key_terms_defined": {
         "applies": true,
         "answer": true,
         "justification": "Key terms are defined precisely: 'common ground' via Stalnaker (2002), 'partner-aware' as adapting to specific intervention agents, 'MAMDP' formally defined in Section 3, 'counterfactual invariance' via KL divergence in Eq. 4.",
         "source": "haiku"
       },
       "intended_contribution_clear": {
         "applies": true,
         "answer": true,
         "justification": "Three explicit bullet-pointed contributions are stated in the introduction: MAMDP+counterfactual invariance formulation, theoretical proofs of PPO/DPO suboptimality, and empirical ICR results on two collaborative tasks.",
         "source": "haiku"
       },
       "engagement_with_prior_work": {
         "applies": true,
         "answer": true,
         "justification": "Section 2 substantively engages prior work on collaborative reasoning, preference-based RL, and safe interruptibility, explicitly contrasting ICR against PSO-INTENT and RLHF approaches rather than merely listing related papers.",
         "source": "haiku"
       }
     }
   },
   "type_checklist": {
     "empirical": {
       "artifacts": {
         "code_released": {
           "applies": true,
           "answer": true,
           "justification": "Code is released at https://github.com/csu-signal/ICR, cited in footnote 2 of the paper.",
           "source": "haiku"
         },
         "data_released": {
           "applies": true,
           "answer": true,
           "justification": "Both evaluation benchmarks (DeliData from Karadzhov et al. 2023 and Weights Task from Khebour et al. 2024) are publicly available datasets.",
           "source": "haiku"
         },
         "environment_specified": {
           "applies": true,
           "answer": false,
           "justification": "While specific libraries (PEFT, TRL SFTTrainer, bitsandbytes), GPU hardware (NVIDIA A100), and model IDs are mentioned, no requirements.txt, Dockerfile, or equivalent environment spec is provided.",
           "source": "haiku"
         },
         "reproduction_instructions": {
           "applies": true,
           "answer": false,
           "justification": "Algorithm 1 describes the pipeline conceptually and Section D details hyperparameters, but no step-by-step runnable instructions are provided; the paper says code will be in 'supplementary material' without confirming it is in the released repo.",
           "source": "haiku"
         }
       },
       "statistical_methodology": {
         "confidence_intervals_or_error_bars": {
           "applies": true,
           "answer": true,
           "justification": "Table 1 reports standard errors (±) for all metrics across 100 evaluation dialogues; Fig 1b reports training curves across 3 random seeds.",
           "source": "haiku"
         },
         "significance_tests": {
           "applies": true,
           "answer": false,
           "justification": "No statistical significance tests (t-test, ANOVA, etc.) are reported for comparative claims between ICR and baselines; only standard errors are provided.",
           "source": "haiku"
         },
         "effect_sizes_reported": {
           "applies": true,
           "answer": true,
           "justification": "Percentage improvements are reported explicitly (47% improvement over DPO on Weights Task, 14% on CG metric for DeliData, 300% difference for inequality propositions in Fig 1a).",
           "source": "haiku"
         },
         "sample_size_justified": {
           "applies": true,
           "answer": false,
           "justification": "100 evaluation dialogues per task is used with no power analysis or justification for sufficiency given the observed effect sizes and variance.",
           "source": "haiku"
         },
         "variance_reported": {
           "applies": true,
           "answer": true,
           "justification": "Standard errors are reported in Table 1 across 100 dialogues and across 3 seeds in ablation Fig 1b.",
           "source": "haiku"
         }
       },
       "evaluation_design": {
         "baselines_included": {
           "applies": true,
           "answer": true,
           "justification": "Five baselines included: BC-COLLABORATOR, DPO, IPO, PPO, and PSO-INTENT, covering behavior cloning, offline preference RL, on-policy RL, and the most directly relevant prior work.",
           "source": "haiku"
         },
         "baselines_contemporary": {
           "applies": true,
           "answer": true,
           "justification": "DPO (2024), IPO (2024), and PSO-INTENT (2023) are all recent methods; PPO is the standard on-policy algorithm used as the underlying optimizer for ICR itself.",
           "source": "haiku"
         },
         "ablation_study": {
           "applies": true,
           "answer": true,
           "justification": "Fig 1b ablates λIntent values (0.01, 0.2, 1.0); Appendix A includes PPO-CF (isolates KL term from counterfactual prompting) and ICR-Phrasing (tests prompt robustness).",
           "source": "haiku"
         },
         "multiple_metrics": {
           "applies": true,
           "answer": true,
           "justification": "DeliData uses both accuracy (ACC) and common ground gain (CG); Weights Task uses composite ACC; cumulative CG curves by proposition type are also provided (Fig 1a).",
           "source": "haiku"
         },
         "human_evaluation": {
           "applies": true,
           "answer": true,
           "justification": "Section D.4 reports a human validation study with 2 annotators evaluating 200 intervention pairs, finding Cohen's κ=0.92 on DeliData and κ=0.58 on Weights Task.",
           "source": "haiku"
         },
         "held_out_test_set": {
           "applies": true,
           "answer": true,
           "justification": "Training uses GPT-4o generated expert trajectories collected prior to evaluation; all evaluation uses 100 fresh dialogue runs not part of the training data.",
           "source": "haiku"
         },
         "per_category_breakdown": {
           "applies": true,
           "answer": true,
           "justification": "Fig 1a breaks Weights Task results into equality, inequality, and order proposition types; results are reported separately for full-press vs. no-press conditions for both tasks.",
           "source": "haiku"
         },
         "failure_cases_discussed": {
           "applies": true,
           "answer": true,
           "justification": "Appendix E provides concrete examples of misleading interventions degrading collaborator performance, and cases where well-meaning interventions are incorrectly ignored.",
           "source": "haiku"
         },
         "negative_results_reported": {
           "applies": true,
           "answer": true,
           "justification": "BC-COLLABORATOR shows negative CG (-0.13) in DeliData full-press; λIntent=0.01 severely hampers learning; PSO-Skeptical degrades performance relative to PSO-Intent.",
           "source": "haiku"
         }
       },
       "setup_transparency": {
         "model_versions_specified": {
           "applies": true,
           "answer": false,
           "justification": "Llama-3-8B-Instruct is specified via HuggingFace ID, but GPT-4o is used without a specific API snapshot date, despite being central to both expert data generation and all evaluation runs.",
           "source": "haiku"
         },
         "prompts_provided": {
           "applies": true,
           "answer": true,
           "justification": "All prompts are provided verbatim in Appendix C (Figs. 2–10) for both tasks and both full-press/no-press conditions, including counterfactual prefixes and alternative phrasings (Table 4).",
           "source": "haiku"
         },
         "hyperparameters_reported": {
           "applies": true,
           "answer": true,
           "justification": "Section D fully reports LoRA settings (α=16, rank=8, dropout=0.05), optimizer (AdamW, cosine scheduler), learning rates, batch sizes, training steps, and sampling parameters.",
           "source": "haiku"
         },
         "scaffolding_described": {
           "applies": true,
           "answer": true,
           "justification": "The MAMDP turn-taking protocol, 15-turn structure, counterfactual prefix computation, and PPO rollout with single additional forward pass are described in detail in Sections 3–5 and Appendix C.",
           "source": "haiku"
         },
         "data_preprocessing_documented": {
           "applies": true,
           "answer": true,
           "justification": "Algorithm 1 and Section C document the full data collection pipeline including bootstrap dialogue seeding, personality assignment, token length statistics (Table 5), and expert response parsing procedures.",
           "source": "haiku"
         }
       },
       "data_integrity": {
         "raw_data_available": {
           "applies": true,
           "answer": false,
           "justification": "The GPT-4o generated expert trajectory corpus used for training is not explicitly confirmed to be released; regenerating it requires costly GPT-4o API calls with stochastic sampling making exact reproduction impossible.",
           "source": "haiku"
         },
         "data_collection_described": {
           "applies": true,
           "answer": true,
           "justification": "Section C and Algorithm 1 describe data collection comprehensively: GPT-4o as expert for both roles, bootstrap seeding, personality sampling, 15-turn collection per dialogue, and preference annotation procedure.",
           "source": "haiku"
         },
         "recruitment_methods_described": {
           "applies": true,
           "answer": false,
           "justification": "Human annotators in Section D.4 are described only as 'two fluent English-speaking college undergraduates'; no recruitment procedure, compensation, or selection criteria is described.",
           "source": "haiku"
         },
         "data_pipeline_documented": {
           "applies": true,
           "answer": true,
           "justification": "Algorithm 1 provides complete pseudocode of the full pipeline from dialogue seed initialization through expert trajectory collection to ICR training.",
           "source": "haiku"
         }
       },
       "contamination": {
         "training_cutoff_stated": {
           "applies": true,
           "answer": false,
           "justification": "Training data cutoffs for GPT-4o and Llama-3-8B-Instruct are not stated despite both being evaluated on DeliData, a published dataset that likely predates their training.",
           "source": "haiku"
         },
         "train_test_overlap_discussed": {
           "applies": true,
           "answer": true,
           "justification": "Appendix A explicitly acknowledges 'GPT-4o's extensive pretraining on reasoning tasks, potentially including exposure to DeliData or DeliData-like problems' when interpreting GPT-4o paired results.",
           "source": "haiku"
         },
         "benchmark_contamination_addressed": {
           "applies": true,
           "answer": false,
           "justification": "Contamination is acknowledged as a concern for GPT-4o baselines but no concrete mitigation (date-restricted models, held-out task variants, contamination tests) is performed for the primary evaluations.",
           "source": "haiku"
         }
       },
       "human_studies": {
         "pre_registered": {
           "applies": true,
           "answer": false,
           "justification": "No pre-registration is mentioned for the human annotation validation study in Section D.4.",
           "source": "haiku"
         },
         "irb_or_ethics_approval": {
           "applies": true,
           "answer": false,
           "justification": "No IRB or ethics approval is mentioned for the human annotator study in Section D.4; the NeurIPS checklist incorrectly states 'we do not conduct any human evaluations' contradicting the actual paper content.",
           "source": "haiku"
         },
         "demographics_reported": {
           "applies": true,
           "answer": false,
           "justification": "Only minimal characterization is given ('two fluent English-speaking college undergraduates'); no age, gender, NLP background, or other demographics are reported.",
           "source": "haiku"
         },
         "inclusion_exclusion_criteria": {
           "applies": true,
           "answer": false,
           "justification": "No explicit inclusion/exclusion criteria are described for human annotators beyond the informal description 'fluent English-speaking college undergraduates.'",
           "source": "haiku"
         },
         "randomization_described": {
           "applies": false,
           "answer": false,
           "justification": "Randomization is not applicable to a pairwise quality annotation task with 2 annotators.",
           "source": "haiku"
         },
         "blinding_described": {
           "applies": true,
           "answer": true,
           "justification": "Section D.4 explicitly states annotators 'were not shown the GPT-4o reward scores or correct task solutions,' implementing effective blinding.",
           "source": "haiku"
         },
         "attrition_reported": {
           "applies": false,
           "answer": false,
           "justification": "With only 2 annotators completing the full study, attrition reporting is not applicable.",
           "source": "haiku"
         }
       },
       "cost_and_practicality": {
         "inference_cost_reported": {
           "applies": true,
           "answer": false,
           "justification": "No inference cost or GPT-4o API cost is reported despite GPT-4o being used for all expert trajectory generation and as the fixed evaluation intervention agent.",
           "source": "haiku"
         },
         "compute_budget_stated": {
           "applies": true,
           "answer": true,
           "justification": "Training compute is specified: ~12 GPU hours for standard baselines and ~24 hours for PPO models on NVIDIA A100s; full-press experiments require two A100s.",
           "source": "haiku"
         }
       }
     }
   },
   "claims": [
     {
       "claim": "ICR-trained collaborators achieve substantially higher accuracy and common ground convergence than all baselines, including a 47% improvement over DPO on the Weights Task.",
       "evidence": "Table 1: ICR achieves 14.06±0.13 vs DPO 9.56±0.09 on Weights Task full-press; 3.35±0.19 vs 2.94±0.18 CG on DeliData full-press.",
       "supported": "strong"
     },
     {
       "claim": "Standard preference-aligned collaborators (RLHF/DPO/IPO) are theoretically suboptimal in MAMDP settings because they treat interventions as static state information rather than causally evaluating them.",
       "evidence": "Theorem 3.2 and Theorem B.3 formally prove this result; empirically all preference-aligned baselines underperform ICR across tasks.",
       "supported": "moderate"
     },
     {
       "claim": "Counterfactual invariance regularization produces emergent common ground convergence without explicit CG rewards during training.",
       "evidence": "ICR trained with only task-accuracy proxy reward achieves CG=3.35 vs BC-COLLABORATOR's CG=-0.13; common ground-based rewards were deliberately withheld.",
       "supported": "strong"
     },
     {
       "claim": "ICR performance gains persist in language-free (no-press) conditions, demonstrating principled partner-awareness beyond language processing.",
       "evidence": "ICR achieves 10.87±0.13 ACC in no-press vs 7.81±0.11 for PPO (next best) on Weights Task; 0.85±0.02 vs 0.78±0.03 on DeliData no-press.",
       "supported": "strong"
     },
     {
       "claim": "Llama-3-8B trained with ICR performs comparably to GPT-4o acting as both agents, despite GPT-4o's distributional advantage.",
       "evidence": "GPT-4o self-paired achieves 15.23±0.21 on Weights Task vs ICR (Llama-3-8B) at 14.06±0.13; gap is ~8% but GPT-4o may have DeliData contamination.",
       "supported": "weak"
     },
     {
       "claim": "ICR is robust to counterfactual prefix phrasing variations, with near-zero mean log-probability response gaps across 6 semantic variants.",
       "evidence": "Mean response gap of 0.0008 log-probability units (σ=0.1568) across 6 counterfactual phrasings on 50 contexts; untrained model shows 0.0247 mean gap.",
       "supported": "moderate"
     }
   ],
   "methodology_tags": [
     "benchmark-eval",
     "theoretical"
   ],
   "key_findings": "ICR substantially outperforms all RLHF/DPO/IPO baselines on multi-party collaborative task performance and common ground convergence, with ~47% improvement over the best offline RL baseline on the Weights Task and 14% on the common ground metric for DeliData. The paper provides formal theoretical proof that preference-aligned LLMs are suboptimal in Modified-Action MDPs where interventions have causal structure, and demonstrates that counterfactual KL regularization produces emergent common ground alignment without explicit CG training rewards. Performance gains persist in language-free conditions and across alternative counterfactual prompt phrasings, suggesting the improvement reflects principled policy learning rather than language-specific surface effects.",
   "red_flags": [
     {
       "flag": "GPT-4o version unspecified",
       "detail": "GPT-4o is used for all expert trajectory collection and as the fixed evaluation intervention agent across all 100+100 evaluation dialogues, but no API snapshot date is given, making exact reproduction impossible and contamination analysis incomplete."
     },
     {
       "flag": "No statistical significance tests",
       "detail": "Despite strong comparative claims, the paper reports only standard errors over 100 dialogues without t-tests or other significance tests to confirm ICR improvements are statistically significant rather than within natural variation."
     },
     {
       "flag": "NeurIPS checklist inconsistency",
       "detail": "The NeurIPS checklist (items 14-15) states 'we do not conduct any crowdsourcing or human evaluations' but Section D.4 clearly reports a human annotation study with 2 annotators evaluating 200 intervention pairs."
     },
     {
       "flag": "Training corpus reproducibility",
       "detail": "Expert trajectories generated via GPT-4o API constitute the training corpus but are not explicitly released; regenerating requires costly API calls with stochastic outputs, making exact reproduction practically impossible."
     },
     {
       "flag": "Contamination not mitigated",
       "detail": "GPT-4o's potential prior exposure to DeliData (a published HCI dataset from 2023) is acknowledged in the appendix but not formally tested or mitigated for the primary evaluation runs."
     },
     {
       "flag": "Human validation understaffed",
       "detail": "The human validation study (Section D.4) uses only 2 annotators to evaluate 200 intervention pairs; this is insufficient to establish reliable inter-rater statistics and contradicts the paper's own NeurIPS checklist."
     }
   ],
   "cited_papers": [
     {
       "title": "How RL agents behave when their actions are modified",
       "relevance": "Foundation of the MAMDP framework used throughout; establishes Bellman-optimal agents are suboptimal when actions can be modified by another strategic agent."
     },
     {
       "title": "Direct Preference Optimization: Your Language Model is Secretly a Reward Model",
       "relevance": "Key baseline (DPO) and the primary preference-alignment paradigm the paper critiques as structurally suboptimal for collaborative settings."
     },
     {
       "title": "Safely interruptible agents",
       "relevance": "Core safety literature motivating the 'interruptible' design goal; paper extends safe interruptibility from RL safety to multi-party collaborative dialogue."
     },
     {
       "title": "Honesty is the best policy: defining and mitigating AI deception",
       "relevance": "Source of PSO-INTENT baseline and the 'intentionality' concept that ICR operationalizes via counterfactual KL regularization."
     },
     {
       "title": "DeliData: A dataset for deliberation in multi-party problem solving",
       "relevance": "Primary evaluation dataset providing the Wason Card Selection collaborative task used in main experiments."
     },
     {
       "title": "When text and speech are not enough: A multimodal dataset of collaboration in a situated task",
       "relevance": "Source of the Weights Task, the second evaluation domain for ICR experiments."
     },
     {
       "title": "Path-specific objectives for safer agent incentives",
       "relevance": "Provides the counterfactual influence pathway concept that ICR's counterfactual regularization is derived from."
     },
     {
       "title": "Proximal Policy Optimization Algorithms",
       "relevance": "The base RL algorithm used for ICR training and as a direct baseline; also relevant to the broader agentic LLM training literature."
     }
   ],
   "engagement_factors": {
     "practical_relevance": {
       "score": 2,
       "justification": "Applicable to AI tutoring and collaborative work settings, though deployment requires RL training infrastructure and GPT-4o API access for expert trajectory generation."
     },
     "surprise_contrarian": {
       "score": 2,
       "justification": "The formal proof that RLHF/DPO agents are structurally suboptimal for multi-party collaboration challenges the dominant preference alignment paradigm with a principled theoretical argument."
     },
     "fear_safety": {
       "score": 1,
       "justification": "Raises concerns about partner-aware LLMs being misused for covert manipulation, with explicit discussion linking to sleeper agents and alignment faking in limitations."
     },
     "drama_conflict": {
       "score": 1,
       "justification": "The claim that 8B Llama trained with ICR approaches GPT-4o performance on these tasks creates modest controversy, though the comparison is carefully caveated."
     },
     "demo_ability": {
       "score": 1,
       "justification": "Code is available on GitHub but requires RL training infrastructure, GPU access, and GPT-4o API calls; not easily demonstrable without significant setup."
     },
     "brand_recognition": {
       "score": 1,
       "justification": "Colorado State University SIGNAL Lab is a legitimate research group but lacks the brand recognition of major AI labs; NeurIPS 2025 acceptance provides venue credibility."
     }
   },
   "hn_data": {
     "threads": [],
     "top_points": 0,
     "total_points": 0,
     "total_comments": 0
   }
 }