ai-research-survey

scan.json (25696B)

---

 {
   "paper": {
     "title": "SHROOM-INDElab at SemEval-2024 Task 6: Zero- and Few-Shot LLM-Based Classification for Hallucination Detection",
     "authors": [
       "Bradley P. Allen",
       "Fina Polat",
       "Paul Groth"
     ],
     "year": 2024,
     "venue": "International Workshop on Semantic Evaluation",
     "arxiv_id": "2404.03732",
     "doi": "10.48550/arXiv.2404.03732"
   },
   "scan_version": 2,
   "active_modules": ["experimental_rigor", "data_leakage"],
   "methodology_tags": ["benchmark-eval"],
   "key_findings": "The SHROOM-INDElab system using GPT-4 prompt engineering for hallucination detection ranked 4th (model-agnostic) and 6th (model-aware) in SemEval-2024 Task 6, achieving 82.9% and 80.2% accuracy on test sets. A counterintuitive finding was that zero-shot classification outperformed few-shot classification with automatically generated examples. The concept definition of hallucination was the most important prompt component in ablation analysis, while task and role definitions contributed minimally.",
   "checklist": {
     "artifacts": {
       "code_released": {
         "applies": true,
         "answer": true,
         "justification": "GitHub link provided in abstract footnote: https://www.github.com/bradleypallen/shroom/"
       },
       "data_released": {
         "applies": true,
         "answer": true,
         "justification": "Uses the publicly available SHROOM shared task datasets provided by SemEval-2024 Task 6, including unlabeled training, labeled validation, and labeled test datasets."
       },
       "environment_specified": {
         "applies": true,
         "answer": false,
         "justification": "No requirements.txt, Dockerfile, conda environment file, or detailed environment setup section is provided. Only mentions LangChain Python library and OpenAI API usage."
       },
       "reproduction_instructions": {
         "applies": true,
         "answer": false,
         "justification": "No step-by-step reproduction instructions are provided. The paper describes the approach but does not include README-style commands or a 'Reproducing Results' section."
       }
     },
     "statistical_methodology": {
       "confidence_intervals_or_error_bars": {
         "applies": true,
         "answer": false,
         "justification": "All results in Table 3 are point estimates (e.g., '0.829' accuracy) with no confidence intervals or error bars. Figures 3-6 show point estimates without uncertainty bands."
       },
       "significance_tests": {
         "applies": true,
         "answer": false,
         "justification": "Claims of improvement over baseline (e.g., '+0.132' accuracy) are made by comparing raw numbers. No statistical significance tests (t-tests, bootstrap tests, etc.) are reported."
       },
       "effect_sizes_reported": {
         "applies": true,
         "answer": true,
         "justification": "Table 3 reports improvements with full baseline context, e.g., 'SHROOM-INDElab 0.829 (+0.132)' alongside baseline '0.697 (+0.000)', allowing the reader to interpret both absolute and relative effect sizes."
       },
       "sample_size_justified": {
         "applies": true,
         "answer": false,
         "justification": "No justification for the number of samples per query (5 or 20), no power analysis, and no justification for K=5 examples in Stage 1 or the 64 sampled data points."
       },
       "variance_reported": {
         "applies": true,
         "answer": false,
         "justification": "The hyperparameter and ablation studies mention 'three different passes' but report no standard deviation, variance, or spread measures across those passes. Only single values are shown."
       }
     },
     "evaluation_design": {
       "baselines_included": {
         "applies": true,
         "answer": true,
         "justification": "Table 3 compares against the task baseline system and top-performing competitors (GroupCheckGPT rank 1, HaRMoNEE rank 1)."
       },
       "baselines_contemporary": {
         "applies": true,
         "answer": true,
         "justification": "Baselines are contemporary SemEval-2024 shared task competitors and the official task baseline, all from the same competition."
       },
       "ablation_study": {
         "applies": true,
         "answer": true,
         "justification": "Section 4.3 presents an ablation study removing prompt components sequentially: examples, task definition, role definition, and concept definition, measuring their contribution to classifier performance."
       },
       "multiple_metrics": {
         "applies": true,
         "answer": true,
         "justification": "Two metrics are used throughout: accuracy and Spearman's correlation coefficient (ρ). Table 4 also reports Cohen's κ."
       },
       "human_evaluation": {
         "applies": true,
         "answer": false,
         "justification": "Evaluation is automated comparison against gold-standard human labels. Section 4.4 analyzes agreement with human labellers via Fleiss' κ and Cohen's κ, but no humans independently evaluated the system's outputs."
       },
       "held_out_test_set": {
         "applies": true,
         "answer": true,
         "justification": "Clear separation between validation and test datasets. Hyperparameter study uses validation set; final results reported on held-out test set (Table 3)."
       },
       "per_category_breakdown": {
         "applies": true,
         "answer": true,
         "justification": "Results are broken down by track (model-agnostic vs. model-aware), by dataset (validation vs. test), and Table 4 provides breakdown by human consensus level (low/high/unanimous)."
       },
       "failure_cases_discussed": {
         "applies": true,
         "answer": true,
         "justification": "The ablation study reveals that few-shot examples hurt accuracy. Section 5 discusses: 'the exclusion of selected examples led to better accuracy suggests the need for further investigation with respect to how the way in which examples are selected and included... impacts accuracy.'"
       },
       "negative_results_reported": {
         "applies": true,
         "answer": true,
         "justification": "Key negative result: zero-shot outperforms few-shot (ablation study, Section 4.3). Also, Figure 4 shows increasing examples per label decreases Spearman's ρ. These are honest reports of things that didn't work as expected."
       }
     },
     "claims_and_evidence": {
       "abstract_claims_supported": {
         "applies": true,
         "answer": true,
         "justification": "Abstract claims (4th/6th ranking, consistency with human labellers, zero-shot better than few-shot) are all supported by results in Sections 4.1, 4.4, and 4.3 respectively."
       },
       "causal_claims_justified": {
         "applies": true,
         "answer": true,
         "justification": "The ablation study (Section 4.3) makes causal claims about component contributions through controlled single-variable manipulation (removing one component at a time). The sequential removal design is adequate for this purpose."
       },
       "generalization_bounded": {
         "applies": true,
         "answer": true,
         "justification": "Claims are bounded to the SHROOM/SemEval-2024 Task 6 setting. The paper does not overclaim beyond the specific task, datasets, and models tested."
       },
       "alternative_explanations_discussed": {
         "applies": true,
         "answer": false,
         "justification": "The discussion acknowledges that few-shot examples hurt accuracy and says it 'suggests the need for further investigation' but does not explore alternative explanations for why this occurs or for other results."
       },
       "proxy_outcome_distinction": {
         "applies": true,
         "answer": true,
         "justification": "The paper measures accuracy and Spearman's ρ on hallucination detection and frames results as hallucination detection performance on SHROOM. No proxy gap exists — claims match measurement granularity."
       }
     },
     "setup_transparency": {
       "model_versions_specified": {
         "applies": true,
         "answer": true,
         "justification": "Section 4 specifies 'gpt-3.5-turbo' and 'gpt-4-0125-preview' (includes date-based version). Embedding model specified as 'text-embedding-ada-002'. Exact dates of runs are given (January 25, 28 and February 17-18, 2024)."
       },
       "prompts_provided": {
         "applies": true,
         "answer": true,
         "justification": "Figure 2 provides the full prompt text with all components (task definition, role definition, concept definition, examples). Table 2 provides all task and role definitions for each task type. The actual prompt text is shown, not just described."
       },
       "hyperparameters_reported": {
         "applies": true,
         "answer": true,
         "justification": "Section 4 reports: temperature (1.2 for submission, varied 0.0-2.0 in study), K=5 examples in Stage 1, 1 example per label, 20 samples for majority voting, λ=0.2 for diversity-consistency trade-off."
       },
       "scaffolding_described": {
         "applies": false,
         "answer": false,
         "justification": "No agentic scaffolding is used. The system is a prompt-based LLM classifier with temperature sampling and majority voting — not an agentic workflow."
       },
       "data_preprocessing_documented": {
         "applies": true,
         "answer": true,
         "justification": "Section 3.3 describes example selection: 64 data points sampled per task type, zero-shot classification, partitioning into positive/negative pools, and the selection algorithm (Algorithm 1) with entropy and diversity criteria."
       }
     },
     "limitations_and_scope": {
       "limitations_section_present": {
         "applies": true,
         "answer": false,
         "justification": "Section 5 is titled 'Discussion and Conclusion' and is only 2 paragraphs. There is no dedicated limitations section or substantive discussion of limitations."
       },
       "threats_to_validity_specific": {
         "applies": true,
         "answer": false,
         "justification": "No specific threats to validity are discussed. The brief discussion section only mentions need for further investigation into example selection, without identifying specific validity threats."
       },
       "scope_boundaries_stated": {
         "applies": true,
         "answer": false,
         "justification": "No explicit statements about what the results do NOT show. No discussion of what populations, settings, or claims are excluded from the findings."
       }
     },
     "data_integrity": {
       "raw_data_available": {
         "applies": true,
         "answer": true,
         "justification": "The SHROOM shared task datasets are publicly available through SemEval-2024. Code is available on GitHub. The competition data allows independent verification."
       },
       "data_collection_described": {
         "applies": true,
         "answer": true,
         "justification": "Section 2 describes the SHROOM dataset structure (task, input, target, generated text, labels). Notes ~200 crowd-sourced human labellers each labeling ~20 data points."
       },
       "recruitment_methods_described": {
         "applies": false,
         "answer": false,
         "justification": "No human participants recruited by the authors. The SHROOM dataset with crowd-sourced labels was provided by the shared task organizers. This is a standard benchmark evaluation."
       },
       "data_pipeline_documented": {
         "applies": true,
         "answer": true,
         "justification": "The two-stage workflow (Figure 1) and example selection process (Section 3.3, Algorithm 1) document the full pipeline from unlabeled data through zero-shot classification to few-shot classification."
       }
     },
     "conflicts_of_interest": {
       "funding_disclosed": {
         "applies": true,
         "answer": true,
         "justification": "Acknowledgements section states: 'This work is partially supported by the European Union's Horizon Europe research and innovation programme within the ENEXA project (grant Agreement no. 101070305).'"
       },
       "affiliations_disclosed": {
         "applies": true,
         "answer": true,
         "justification": "All authors are listed as University of Amsterdam, Intelligent Data Engineering Lab. No product being evaluated is affiliated with the authors."
       },
       "funder_independent_of_outcome": {
         "applies": true,
         "answer": true,
         "justification": "EU Horizon Europe research funding has no financial stake in the hallucination detection results. The ENEXA project is a general research initiative."
       },
       "financial_interests_declared": {
         "applies": true,
         "answer": false,
         "justification": "No competing interests or financial interests statement is included in the paper."
       }
     },
     "contamination": {
       "training_cutoff_stated": {
         "applies": true,
         "answer": false,
         "justification": "No mention of the training data cutoff dates for GPT-3.5-turbo or GPT-4-0125-preview. The models' training data could include SHROOM-related content."
       },
       "train_test_overlap_discussed": {
         "applies": true,
         "answer": false,
         "justification": "No discussion of whether SHROOM task data or related hallucination examples could appear in GPT-4's training data."
       },
       "benchmark_contamination_addressed": {
         "applies": true,
         "answer": false,
         "justification": "No discussion of contamination risk. SemEval tasks and associated data could potentially appear in LLM training corpora."
       }
     },
     "human_studies": {
       "pre_registered": {
         "applies": false,
         "answer": false,
         "justification": "No human participants recruited by the authors. This is a benchmark evaluation using shared task datasets."
       },
       "irb_or_ethics_approval": {
         "applies": false,
         "answer": false,
         "justification": "No human participants recruited by the authors."
       },
       "demographics_reported": {
         "applies": false,
         "answer": false,
         "justification": "No human participants recruited by the authors."
       },
       "inclusion_exclusion_criteria": {
         "applies": false,
         "answer": false,
         "justification": "No human participants recruited by the authors."
       },
       "randomization_described": {
         "applies": false,
         "answer": false,
         "justification": "No human participants recruited by the authors."
       },
       "blinding_described": {
         "applies": false,
         "answer": false,
         "justification": "No human participants recruited by the authors."
       },
       "attrition_reported": {
         "applies": false,
         "answer": false,
         "justification": "No human participants recruited by the authors."
       }
     },
     "cost_and_practicality": {
       "inference_cost_reported": {
         "applies": true,
         "answer": true,
         "justification": "Section 4 states: 'Approximately $500 USD in OpenAI API charges were incurred during the above runs.'"
       },
       "compute_budget_stated": {
         "applies": true,
         "answer": true,
         "justification": "Total API spend of approximately $500 USD is stated, covering Stage 1, Stage 2 submission, and hyperparameter/ablation study runs."
       }
     },
     "experimental_rigor": {
       "seed_sensitivity_reported": {
         "applies": true,
         "answer": false,
         "justification": "No random seed sensitivity analysis is reported. Temperature sampling introduces stochasticity but its effect on result stability is not analyzed across seeds."
       },
       "number_of_runs_stated": {
         "applies": true,
         "answer": true,
         "justification": "Hyperparameter and ablation studies state 'three different passes over the model-agnostic validation dataset.' Submission uses '20 samples for majority voting.'"
       },
       "hyperparameter_search_budget": {
         "applies": true,
         "answer": false,
         "justification": "A hyperparameter study is conducted varying temperature, examples per label, and samples per query, but the total search budget (total configurations tried, total compute for search) is not quantified."
       },
       "best_config_selection_justified": {
         "applies": true,
         "answer": false,
         "justification": "The hyperparameter study was conducted on gpt-3.5-turbo, but the final submission used gpt-4-0125-preview. The transfer of hyperparameter choices across models is not justified. The submission's temperature of 1.2 is not clearly derived from the gpt-3.5-turbo study results."
       },
       "multiple_comparison_correction": {
         "applies": false,
         "answer": false,
         "justification": "Only a small number of comparisons are made (2 models, baseline vs. system). No mass testing situation requiring correction."
       },
       "self_comparison_bias_addressed": {
         "applies": true,
         "answer": false,
         "justification": "No acknowledgment of evaluating their own system or potential author bias in prompt design and tuning."
       },
       "compute_budget_vs_performance": {
         "applies": false,
         "answer": false,
         "justification": "Comparisons are against shared task submissions where compute differences are part of the system design; compute budget comparison across competitors is not expected."
       },
       "benchmark_construct_validity": {
         "applies": true,
         "answer": false,
         "justification": "No discussion of whether the SHROOM benchmark actually measures hallucination detection capability as intended, or whether the binary framing captures the nuances of hallucination."
       },
       "scaffold_confound_addressed": {
         "applies": false,
         "answer": false,
         "justification": "No scaffolding involved. The system is a direct prompt-based classifier without agentic components."
       }
     },
     "data_leakage": {
       "temporal_leakage_addressed": {
         "applies": true,
         "answer": false,
         "justification": "No discussion of whether GPT-4's training data could include SHROOM-related content or similar hallucination detection tasks."
       },
       "feature_leakage_addressed": {
         "applies": true,
         "answer": false,
         "justification": "No discussion of whether the evaluation setup (providing input, target, and generated text together) gives the model information not available in a real deployment scenario."
       },
       "non_independence_addressed": {
         "applies": true,
         "answer": false,
         "justification": "No discussion of independence between training examples (used for Stage 1 zero-shot → Stage 2 few-shot) and evaluation data."
       },
       "leakage_detection_method": {
         "applies": true,
         "answer": false,
         "justification": "No concrete leakage detection or prevention methods are used or discussed."
       }
     }
   },
   "claims": [
     {
       "claim": "The SHROOM-INDElab system ranked 4th in the model-agnostic track and 6th in the model-aware track of SemEval-2024 Task 6.",
       "evidence": "Table 3 shows test set results: 82.9% accuracy (model-agnostic) and 80.2% accuracy (model-aware), compared to 1st-place systems at 84.7% and 81.3%.",
       "supported": "strong"
     },
     {
       "claim": "GPT-4-0125-preview significantly outperforms GPT-3.5-turbo for hallucination detection on the SHROOM validation datasets.",
       "evidence": "Table 3: GPT-4 achieves 0.814/0.772 accuracy (model-agnostic/aware) vs. GPT-3.5-turbo at 0.773/0.764 on validation sets.",
       "supported": "moderate"
     },
     {
       "claim": "Zero-shot classification provides better accuracy than few-shot classification with automatically generated examples.",
       "evidence": "Ablation study (Section 4.3, Figure 6) shows removing examples improves accuracy, though Spearman's ρ slightly decreases. Tested only on gpt-3.5-turbo with one validation dataset.",
       "supported": "moderate"
     },
     {
       "claim": "The concept definition of hallucination is the most significant prompt component for classifier performance.",
       "evidence": "Ablation study (Section 4.3, Figure 6) shows removing the concept definition causes the largest drop in both accuracy and Spearman's ρ compared to removing other components.",
       "supported": "moderate"
     },
     {
       "claim": "The classifier's labeling decisions are consistent with those of crowd-sourced human labellers.",
       "evidence": "Section 4.4: Adding classifier labels to human annotations increases Fleiss' κ from 0.373 to 0.405. Table 4 shows Cohen's κ of 0.623 overall and 0.856 for unanimous human labels.",
       "supported": "moderate"
     },
     {
       "claim": "System agreement with human labeling increases as the certainty of human labeling increases.",
       "evidence": "Table 4 shows accuracy rising from 0.621 (low consensus 2/3 split) to 0.854 (high 4/5 split) to 0.929 (unanimous), with corresponding increases in Cohen's κ.",
       "supported": "strong"
     }
   ],
   "red_flags": [
     {
       "flag": "Sequential ablation confounding",
       "detail": "The ablation study (Section 4.3) removes components sequentially rather than independently (examples → task definition → role definition → concept definition). The measured effect of each component is confounded with interaction effects from previously removed components."
     },
     {
       "flag": "No statistical tests on main results",
       "detail": "All performance comparisons in Table 3 are raw number comparisons with no significance tests, confidence intervals, or variance measures. Improvements like '+0.132' accuracy could be within noise."
     },
     {
       "flag": "Hyperparameter tuning on different model than submission",
       "detail": "The hyperparameter study was conducted using gpt-3.5-turbo but the final submission used gpt-4-0125-preview. The transferability of optimal hyperparameters across models is not validated."
     },
     {
       "flag": "No variance reported despite multiple runs",
       "detail": "The paper mentions 'three different passes' for hyperparameter and ablation studies but never reports standard deviation or any spread measure, preventing assessment of result stability."
     }
   ],
   "cited_papers": [
     {
       "title": "SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models",
       "authors": ["Potsawee Manakul", "Adian Liusie", "Mark JF Gales"],
       "year": 2023,
       "arxiv_id": "2303.08896",
       "relevance": "Zero-resource hallucination detection method for LLMs, directly related to LLM evaluation and safety."
     },
     {
       "title": "ChainPoll: A High Efficacy Method for LLM Hallucination Detection",
       "authors": ["Robert Friel", "Atindriyo Sanyal"],
       "year": 2023,
       "arxiv_id": "2310.18344",
       "relevance": "LLM-based hallucination detection method using prompt engineering, comparable approach to the surveyed system."
     },
     {
       "title": "A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions",
       "authors": ["Lei Huang", "Weijiang Yu", "Weitao Ma"],
       "year": 2023,
       "arxiv_id": "2311.05232",
       "relevance": "Comprehensive survey on LLM hallucination covering taxonomy and evaluation methods."
     },
     {
       "title": "Survey of Hallucination in Natural Language Generation",
       "authors": ["Ziwei Ji", "Nayeon Lee", "Rita Frieske"],
       "year": 2023,
       "relevance": "Foundational survey on hallucination in NLG systems, relevant to LLM safety and evaluation."
     },
     {
       "title": "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models",
       "authors": ["Jason Wei", "Xuezhi Wang", "Dale Schuurmans"],
       "year": 2022,
       "relevance": "Foundational work on chain-of-thought prompting, a key technique for LLM capability evaluation."
     },
     {
       "title": "Large Language Models are Zero-Shot Reasoners",
       "authors": ["Takeshi Kojima", "Shixiang Shane Gu", "Machel Reid"],
       "year": 2022,
       "relevance": "Demonstrates zero-shot reasoning capability of LLMs, foundational for prompt-based evaluation approaches."
     },
     {
       "title": "Better Zero-Shot Reasoning with Role-Play Prompting",
       "authors": ["Aobo Kong", "Shiwan Zhao", "Hao Chen"],
       "year": 2023,
       "arxiv_id": "2308.07702",
       "relevance": "Role-play prompting technique for improving LLM reasoning, relevant to prompt engineering methodology."
     },
     {
       "title": "Pre-Train, Prompt, and Predict: A Systematic Survey of Prompting Methods in Natural Language Processing",
       "authors": ["Pengfei Liu", "Weizhe Yuan", "Jinlan Fu"],
       "year": 2023,
       "relevance": "Systematic survey of prompting methods covering the landscape of LLM interaction techniques."
     },
     {
       "title": "Universal Self-Adaptive Prompting",
       "authors": ["Xingchen Wan", "Ruoxi Sun", "Hootan Nakhost"],
       "year": 2023,
       "arxiv_id": "2305.14926",
       "relevance": "Self-adaptive prompting method for automated example selection in few-shot LLM usage."
     },
     {
       "title": "Leveraging Large Language Models for NLG Evaluation: A Survey",
       "authors": ["Zhen Li", "Xiaohan Xu", "Tao Shen"],
       "year": 2024,
       "arxiv_id": "2401.07103",
       "relevance": "Survey on using LLMs as evaluators for NLG, relevant to LLM-as-judge methodology."
     }
   ]
 }