ai-research-survey

scan.json (22630B)

---

 {
   "paper": {
     "title": "Domain-Specific Constitutional AI: Enhancing Safety in LLM-Powered Mental Health Chatbots",
     "authors": ["Chenhan Lyu", "Yutong Song", "Pengfei Zhang", "Amir M. Rahmani"],
     "year": 2025,
     "venue": "IEEE BSN 2025",
     "arxiv_id": "2509.16444",
     "doi": "10.1109/BSN66969.2025.11337405"
   },
   "scan_version": 2,
   "active_modules": ["experimental_rigor"],
   "methodology_tags": ["benchmark-eval"],
   "key_findings": "Domain-specific constitutional principles for mental health significantly outperform both vague/general principles (31.7% higher total score) and no constitutional training in LLM safety alignment. A 1B parameter model trained with specific principles outperformed a 3B baseline without constitutional training, suggesting principled alignment can overcome scale limitations. Crisis-related guidelines showed the largest relative improvements (>150%) with specific principles. Ablation study confirmed that replacing even two of four specific principles with vague counterparts reduces total score by 19.2%.",
   "checklist": {
     "artifacts": {
       "code_released": {
         "applies": true,
         "answer": false,
         "justification": "No repository URL, code archive, or link to training code is provided anywhere in the paper."
       },
       "data_released": {
         "applies": true,
         "answer": true,
         "justification": "The paper uses MentalChat16K, a publicly available dataset on HuggingFace (ref [15]), and the evaluation framework from ref [16]. The training data source is publicly accessible."
       },
       "environment_specified": {
         "applies": true,
         "answer": false,
         "justification": "No environment specifications, dependency lists, or hardware details are provided. The paper mentions LLaMA 3.2 1B/3B architectures but provides no setup information."
       },
       "reproduction_instructions": {
         "applies": true,
         "answer": false,
         "justification": "No reproduction instructions, README, or step-by-step guide is provided."
       }
     },
     "statistical_methodology": {
       "confidence_intervals_or_error_bars": {
         "applies": true,
         "answer": false,
         "justification": "All results in Tables II and III are reported as point estimates with no confidence intervals or error bars."
       },
       "significance_tests": {
         "applies": true,
         "answer": false,
         "justification": "The paper claims specific principles outperform vague/general principles and baselines but provides no statistical significance tests. Comparisons are made solely by comparing raw score numbers."
       },
       "effect_sizes_reported": {
         "applies": true,
         "answer": true,
         "justification": "Percentage improvements are reported with baseline context throughout Section III.B, e.g., 'improving from 4.41 in the baseline to 6.47 with specific principles (46.7% increase)' and similar for each guideline."
       },
       "sample_size_justified": {
         "applies": true,
         "answer": false,
         "justification": "100 evaluation queries and 5000 training rows are used but no justification for these sample sizes is given, nor any power analysis."
       },
       "variance_reported": {
         "applies": true,
         "answer": false,
         "justification": "No variance, standard deviation, or spread measures are reported for any results. It is unclear whether evaluations were run multiple times."
       }
     },
     "evaluation_design": {
       "baselines_included": {
         "applies": true,
         "answer": true,
         "justification": "Four conditions compared: no principle baseline (1B), vague/general principles (1B), specific principles (1B), and no principle (3B). Section II.B describes the experimental design."
       },
       "baselines_contemporary": {
         "applies": true,
         "answer": true,
         "justification": "Uses LLaMA 3.2 (2024) as the base architecture and compares against the original Anthropic CAI approach with general principles. The baselines are appropriate for the comparison being made."
       },
       "ablation_study": {
         "applies": true,
         "answer": true,
         "justification": "Section III.D presents an ablation study replacing two of four specific principles with vague counterparts, showing the contribution of principle specificity."
       },
       "multiple_metrics": {
         "applies": true,
         "answer": true,
         "justification": "Five separate guideline metrics are reported (practice adherence, health risks, critical response, resource provision, user empowerment) plus a total score."
       },
       "human_evaluation": {
         "applies": true,
         "answer": true,
         "justification": "Section III.A states 'Trained evaluators scored responses on a 1-10 scale per guideline using detailed rubrics aligned with clinical best practices.' Health experts provided ground truth."
       },
       "held_out_test_set": {
         "applies": true,
         "answer": false,
         "justification": "The evaluation uses 100 mental health queries but there is no mention of train/test/validation splits for the evaluation set, or whether the evaluation queries overlap with training data."
       },
       "per_category_breakdown": {
         "applies": true,
         "answer": true,
         "justification": "Tables II and III and Figure 2 provide per-guideline breakdowns across all five evaluation guidelines for each model variant."
       },
       "failure_cases_discussed": {
         "applies": true,
         "answer": false,
         "justification": "No failure cases, error analysis, or qualitative examples of where the models still fail are discussed."
       },
       "negative_results_reported": {
         "applies": true,
         "answer": false,
         "justification": "Every comparison shows monotonic improvement from baseline to specific principles. The ablation shows the ablated version underperforms specific but still improves over vague — no configurations that hurt performance are reported."
       }
     },
     "claims_and_evidence": {
       "abstract_claims_supported": {
         "applies": true,
         "answer": true,
         "justification": "The abstract claims an approach for domain-specific CAI in mental health, and the results demonstrate improvements across all five guidelines. The quantitative evaluation claim is supported by Tables II-III."
       },
       "causal_claims_justified": {
         "applies": true,
         "answer": true,
         "justification": "The paper makes causal claims ('specific principles yield significant safety enhancements'). The four-condition experimental design with controlled single-variable manipulation (principle type) provides adequate support for these claims, though lack of statistical tests weakens confidence."
       },
       "generalization_bounded": {
         "applies": true,
         "answer": false,
         "justification": "The paper title claims 'Enhancing Safety in LLM-Powered Mental Health Chatbots' broadly, but results are only on LLaMA 3.2 1B/3B with a single evaluation framework and 100 queries. Discussion section extends claims to 'diverse health domains' and 'medical specialties' without evidence."
       },
       "alternative_explanations_discussed": {
         "applies": true,
         "answer": false,
         "justification": "No alternative explanations are considered. For example, improvements could stem from increased training data exposure rather than principle specificity, or the evaluation rubric could favor specific-principle language patterns."
       },
       "proxy_outcome_distinction": {
         "applies": true,
         "answer": false,
         "justification": "The paper measures evaluator scores on 5 guidelines and frames this as 'safety enhancement' and 'effectiveness.' No discussion of whether rubric scores are adequate proxies for actual clinical safety outcomes."
       }
     },
     "setup_transparency": {
       "model_versions_specified": {
         "applies": true,
         "answer": true,
         "justification": "Specific model sizes are stated: 'the 1B parameter LLaMA 3.2 architecture' and 'the 3B parameter LLaMA 3.2 architecture' (Section II.B). LLaMA 3.2 is specific enough as a model release."
       },
       "prompts_provided": {
         "applies": true,
         "answer": false,
         "justification": "Only a template format is mentioned: 'standardized prompting templates (e.g., \"Critique this response against these principles: [principle text].\")'. The full actual prompts used for SFT and RLAIF phases are not provided."
       },
       "hyperparameters_reported": {
         "applies": true,
         "answer": false,
         "justification": "Training used 5000 rows with early stopping and 2 response pairs per example for preference generation, but no learning rates, batch sizes, optimizer, temperature, number of epochs, or other standard hyperparameters are reported."
       },
       "scaffolding_described": {
         "applies": false,
         "answer": false,
         "justification": "No agentic scaffolding is used. The paper trains models via SFT and RLAIF — standard fine-tuning, not agentic workflows."
       },
       "data_preprocessing_documented": {
         "applies": true,
         "answer": false,
         "justification": "The paper states 'sampling 5000 rows' from MentalChat16K but does not describe how those rows were selected, whether any filtering was applied, or how the data was preprocessed for training."
       }
     },
     "limitations_and_scope": {
       "limitations_section_present": {
         "applies": true,
         "answer": false,
         "justification": "There is no dedicated limitations section. The conclusion contains one sentence: 'While effective for current safety needs, our static principles may not adapt to evolving guidelines.' This is insufficient for a dedicated section."
       },
       "threats_to_validity_specific": {
         "applies": true,
         "answer": false,
         "justification": "The only limitation mentioned (static principles) is generic. No discussion of specific threats like small evaluation set, single model family, evaluator reliability, or single evaluation framework."
       },
       "scope_boundaries_stated": {
         "applies": true,
         "answer": false,
         "justification": "No explicit scope boundaries. The paper does not state what it did NOT test (other model families, other languages, real clinical settings, etc.)."
       }
     },
     "data_integrity": {
       "raw_data_available": {
         "applies": true,
         "answer": false,
         "justification": "Individual evaluation scores per query are not available. Only aggregated per-guideline means are reported in Tables II-III."
       },
       "data_collection_described": {
         "applies": true,
         "answer": true,
         "justification": "Section III.A describes the evaluation framework: 100 mental health queries, health expert ground truth, trained evaluators scoring on 1-10 rubrics. Training data comes from MentalChat16K (ref [15])."
       },
       "recruitment_methods_described": {
         "applies": true,
         "answer": false,
         "justification": "The paper mentions 'trained evaluators' and 'health experts' but does not describe how they were recruited, how many there were, or their qualifications."
       },
       "data_pipeline_documented": {
         "applies": true,
         "answer": false,
         "justification": "The pipeline from raw MentalChat16K to 5000 sampled rows to trained models to evaluation is described at a high level but lacks detail on sampling strategy, data cleaning, and intermediate steps."
       }
     },
     "conflicts_of_interest": {
       "funding_disclosed": {
         "applies": true,
         "answer": false,
         "justification": "No funding disclosure or acknowledgments section is present in the paper."
       },
       "affiliations_disclosed": {
         "applies": true,
         "answer": true,
         "justification": "All authors are listed as University of California, Irvine affiliates. The evaluation framework is from the Institute for Future Health (ref [16]), which is Rahmani's lab at UCI — this connection is implicit but affiliations are stated."
       },
       "funder_independent_of_outcome": {
         "applies": true,
         "answer": false,
         "justification": "No funding information is disclosed, so independence cannot be assessed."
       },
       "financial_interests_declared": {
         "applies": true,
         "answer": false,
         "justification": "No competing interests or financial disclosure statement is present."
       }
     },
     "contamination": {
       "training_cutoff_stated": {
         "applies": true,
         "answer": false,
         "justification": "No training data cutoff date is stated for LLaMA 3.2. The evaluation queries could have been seen during pre-training."
       },
       "train_test_overlap_discussed": {
         "applies": true,
         "answer": false,
         "justification": "No discussion of whether the 100 evaluation queries or their answers overlap with LLaMA 3.2's pre-training data or MentalChat16K."
       },
       "benchmark_contamination_addressed": {
         "applies": true,
         "answer": false,
         "justification": "The evaluation framework (ref [16]) was published before LLaMA 3.2's training. No discussion of whether benchmark content could be in the training data."
       }
     },
     "human_studies": {
       "pre_registered": {
         "applies": false,
         "answer": false,
         "justification": "No human participants in the study. Evaluators score model outputs but are not study subjects."
       },
       "irb_or_ethics_approval": {
         "applies": false,
         "answer": false,
         "justification": "No human participants. The study evaluates model outputs, not human subjects."
       },
       "demographics_reported": {
         "applies": false,
         "answer": false,
         "justification": "No human participants in the study."
       },
       "inclusion_exclusion_criteria": {
         "applies": false,
         "answer": false,
         "justification": "No human participants in the study."
       },
       "randomization_described": {
         "applies": false,
         "answer": false,
         "justification": "No human participants in the study."
       },
       "blinding_described": {
         "applies": false,
         "answer": false,
         "justification": "No human participants in the study."
       },
       "attrition_reported": {
         "applies": false,
         "answer": false,
         "justification": "No human participants in the study."
       }
     },
     "cost_and_practicality": {
       "inference_cost_reported": {
         "applies": true,
         "answer": false,
         "justification": "The paper discusses efficiency advantages qualitatively (1B vs 3B, on-device processing) but reports no actual inference costs, latency measurements, or tokens consumed."
       },
       "compute_budget_stated": {
         "applies": true,
         "answer": false,
         "justification": "No GPU hours, training time, or computational budget is reported despite training multiple model variants."
       }
     },
     "experimental_rigor": {
       "seed_sensitivity_reported": {
         "applies": true,
         "answer": false,
         "justification": "No mention of multiple random seeds or seed sensitivity analysis for training or evaluation."
       },
       "number_of_runs_stated": {
         "applies": true,
         "answer": false,
         "justification": "It is not stated how many runs produced the reported results. Appears to be single-run."
       },
       "hyperparameter_search_budget": {
         "applies": true,
         "answer": false,
         "justification": "No hyperparameter search budget is reported. Only '5000 rows' and '2 response pairs' are mentioned with no search process."
       },
       "best_config_selection_justified": {
         "applies": true,
         "answer": false,
         "justification": "No discussion of how the final configuration was selected or whether other configurations were tried."
       },
       "multiple_comparison_correction": {
         "applies": false,
         "answer": false,
         "justification": "No statistical tests are performed at all, so multiple comparison correction is not applicable."
       },
       "self_comparison_bias_addressed": {
         "applies": true,
         "answer": false,
         "justification": "The authors evaluate their own training approach using an evaluation framework from a co-author's lab (Institute for Future Health, ref [16]) without acknowledging this potential bias."
       },
       "compute_budget_vs_performance": {
         "applies": true,
         "answer": false,
         "justification": "The 1B vs 3B comparison implies different compute but no analysis of performance as a function of compute budget is provided."
       },
       "benchmark_construct_validity": {
         "applies": true,
         "answer": false,
         "justification": "The evaluation uses 5 guidelines from ref [16] without discussing whether these guidelines adequately capture 'safety' in mental health chatbots or whether the 1-10 scoring rubric has construct validity."
       },
       "scaffold_confound_addressed": {
         "applies": false,
         "answer": false,
         "justification": "No scaffolding is involved. Models are evaluated directly."
       }
     }
   },
   "claims": [
     {
       "claim": "Specific constitutional principles yield 31.7% higher total scores than vague/general principles (24.08 vs 18.29)",
       "evidence": "Table II shows total scores: vague/general 18.29, specific 24.08. Section III.B provides per-guideline analysis.",
       "supported": "moderate"
     },
     {
       "claim": "1B model with specific principles outperforms 3B model without constitutional training (24.08 vs 19.92)",
       "evidence": "Table II: No principle 3B scores 19.92, Specific principles 1B scores 24.08.",
       "supported": "moderate"
     },
     {
       "claim": "Crisis-related guidelines show the largest relative improvements (>150%)",
       "evidence": "Section III.B: Guideline 3 from 1.06 to 2.69 (153.8%), Guideline 4 from 1.13 to 2.91 (157.5%).",
       "supported": "moderate"
     },
     {
       "claim": "Ablation replacing 2 of 4 specific principles reduces total score by 19.2%",
       "evidence": "Table III and Section III.D: specific 24.08 vs ablated 19.45.",
       "supported": "moderate"
     }
   ],
   "red_flags": [
     {
       "flag": "No statistical tests",
       "detail": "All comparisons are raw score differences with no significance testing. With only point estimates and no variance, it is impossible to tell if differences are meaningful or due to noise."
     },
     {
       "flag": "No variance or uncertainty reporting",
       "detail": "No standard deviations, confidence intervals, or multi-run results. Single-run training and evaluation with no indication of result stability."
     },
     {
       "flag": "Evaluator details missing",
       "detail": "'Trained evaluators' score responses but no information on how many evaluators, their qualifications, inter-rater reliability, or blinding to condition."
     },
     {
       "flag": "Self-evaluation with own framework",
       "detail": "The evaluation framework (ref [16]) comes from the Institute for Future Health, which is co-author Rahmani's lab at UCI. The authors are evaluating their own approach with their own lab's evaluation instrument."
     },
     {
       "flag": "Overclaiming",
       "detail": "Conclusions extend to 'broader investigation of domain-specific CAI in healthcare specialties' and 'standardization of clinical AI principles' from a single experiment on one model family with 100 queries and no statistical tests."
     },
     {
       "flag": "Low absolute scores on critical guidelines",
       "detail": "Crisis-related guidelines (3 and 4) score 2.69 and 2.91 out of 10 even with specific principles — still very poor absolute performance that is not adequately discussed."
     }
   ],
   "cited_papers": [
     {
       "title": "A comprehensive survey of LLM alignment techniques: RLHF, RLAIF, PPO, DPO and more",
       "authors": ["Z. Wang"],
       "year": 2024,
       "arxiv_id": "2407.16216",
       "relevance": "Comprehensive survey of LLM alignment techniques directly relevant to AI safety methodology."
     },
     {
       "title": "Foundational challenges in assuring alignment and safety of large language models",
       "authors": ["U. Anwar"],
       "year": 2024,
       "arxiv_id": "2404.09932",
       "relevance": "Identifies foundational challenges in LLM alignment and safety assurance."
     },
     {
       "title": "Sleeper agents: Training deceptive LLMs that persist through safety training",
       "authors": ["E. Hubinger"],
       "year": 2024,
       "arxiv_id": "2401.05566",
       "relevance": "Demonstrates that deceptive behaviors can persist through safety training, relevant to AI safety evaluation."
     },
     {
       "title": "Constitutional AI: Harmlessness from AI feedback",
       "authors": ["Y. Bai"],
       "year": 2022,
       "arxiv_id": "2212.08073",
       "relevance": "Foundational paper on Constitutional AI methodology that this paper extends to domain-specific settings."
     },
     {
       "title": "Specific versus general principles for constitutional AI",
       "authors": ["S. Kundu"],
       "year": 2023,
       "arxiv_id": "2310.13798",
       "relevance": "Directly compares specific vs general CAI principles — key prior work this paper builds on."
     },
     {
       "title": "How effective is constitutional AI in small LLMs? A study on DeepSeek-R1 and its peers",
       "authors": ["A.-G. C. Menke", "P. X. Tan"],
       "year": 2025,
       "arxiv_id": "2503.17365",
       "relevance": "Studies CAI effectiveness in small LLMs, directly related to this paper's efficiency claims."
     },
     {
       "title": "C3AI: Crafting and evaluating constitutions for constitutional AI",
       "authors": ["Y. Kyrychenko", "K. Zhou", "E. Bogucka", "D. Quercia"],
       "year": 2025,
       "relevance": "Framework for crafting and evaluating AI constitutions, relevant to AI alignment methodology."
     },
     {
       "title": "Building guardrails for large language models",
       "authors": ["Y. Dong"],
       "year": 2024,
       "arxiv_id": "2402.01822",
       "relevance": "Survey of LLM guardrail approaches relevant to AI safety in deployment."
     }
   ]
 }