scan.json (21046B)
1 { 2 "paper": { 3 "title": "In Transformer We Trust? A Perspective on Transformer Architecture Failure Modes", 4 "authors": ["Trishit Mondal", "Ameya D. Jagtap"], 5 "year": 2026, 6 "venue": "arXiv", 7 "arxiv_id": "2602.14318" 8 }, 9 "scan_version": 2, 10 "active_modules": ["survey_methodology"], 11 "methodology_tags": ["meta-analysis"], 12 "key_findings": "This perspective paper surveys transformer trustworthiness across six dimensions: interpretability, robustness, fairness, privacy, uncertainty quantification, and scientific applicability. It covers domains from NLP and vision to scientific computing, robotics, nuclear science, and theorem proving. The paper identifies recurring structural vulnerabilities including fragile learned representations, bias amplification through global attention, and lack of uncertainty quantification in safety-critical deployments. It advocates for neuro-symbolic integration, physics-informed architectures, and formal verification as paths toward trustworthy transformers.", 13 "checklist": { 14 "artifacts": { 15 "code_released": { 16 "applies": true, 17 "answer": false, 18 "justification": "No code repository or analysis scripts are provided or referenced in the paper." 19 }, 20 "data_released": { 21 "applies": true, 22 "answer": false, 23 "justification": "No dataset, corpus, or structured data is released. The paper is a literature review with no extracted dataset." 24 }, 25 "environment_specified": { 26 "applies": false, 27 "answer": false, 28 "justification": "This is a survey/perspective paper with no computational experiments requiring an environment." 29 }, 30 "reproduction_instructions": { 31 "applies": true, 32 "answer": false, 33 "justification": "No instructions for reproducing the literature search or paper selection process are provided." 34 } 35 }, 36 "statistical_methodology": { 37 "confidence_intervals_or_error_bars": { 38 "applies": false, 39 "answer": false, 40 "justification": "Survey paper with no original experiments or statistical analyses." 41 }, 42 "significance_tests": { 43 "applies": false, 44 "answer": false, 45 "justification": "Survey paper with no original experiments or statistical analyses." 46 }, 47 "effect_sizes_reported": { 48 "applies": false, 49 "answer": false, 50 "justification": "Survey paper with no original experiments." 51 }, 52 "sample_size_justified": { 53 "applies": false, 54 "answer": false, 55 "justification": "Survey paper with no original experiments." 56 }, 57 "variance_reported": { 58 "applies": false, 59 "answer": false, 60 "justification": "Survey paper with no original experiments." 61 } 62 }, 63 "evaluation_design": { 64 "baselines_included": { 65 "applies": true, 66 "answer": false, 67 "justification": "The paper does not compare its survey against prior surveys or reviews in a structured way. While it references works like Huang et al. and Chen et al. on trustworthiness, there is no systematic comparison of coverage or methodology with prior surveys." 68 }, 69 "baselines_contemporary": { 70 "applies": false, 71 "answer": false, 72 "justification": "No experimental baselines; this is a survey paper." 73 }, 74 "ablation_study": { 75 "applies": false, 76 "answer": false, 77 "justification": "No system with components to ablate; this is a survey paper." 78 }, 79 "multiple_metrics": { 80 "applies": false, 81 "answer": false, 82 "justification": "No experimental evaluation; this is a survey paper." 83 }, 84 "human_evaluation": { 85 "applies": false, 86 "answer": false, 87 "justification": "No system outputs to evaluate; this is a survey paper." 88 }, 89 "held_out_test_set": { 90 "applies": false, 91 "answer": false, 92 "justification": "No experimental evaluation; this is a survey paper." 93 }, 94 "per_category_breakdown": { 95 "applies": true, 96 "answer": true, 97 "justification": "The paper provides detailed per-domain breakdowns across NLP, vision, robotics, biology, earth science, materials science, fluid dynamics, nuclear science, theorem proving, and agentic AI (Sections 7.1–7.9), each with domain-specific trustworthiness assessments." 98 }, 99 "failure_cases_discussed": { 100 "applies": true, 101 "answer": true, 102 "justification": "The paper extensively discusses failure cases: adversarial attacks reducing accuracy from 75% to 7% (Section 3.1), universal adversarial triggers reducing sentiment accuracy from 85% to 30% (Section 3.1), bias amplification (Section 4.1), and lack of uncertainty quantification across multiple domains." 103 }, 104 "negative_results_reported": { 105 "applies": true, 106 "answer": true, 107 "justification": "The paper's central thesis is about transformer failure modes. It reports negative findings throughout: attention weights are unreliable explanations (Section 2.1), ViTs are more biased than CNNs (Section 4.1), most reviewed models lack UQ (Sections 7.4–7.7)." 108 } 109 }, 110 "claims_and_evidence": { 111 "abstract_claims_supported": { 112 "applies": true, 113 "answer": true, 114 "justification": "The abstract claims to examine trustworthiness through interpretability, robustness, fairness, and privacy, and to identify structural vulnerabilities. The paper systematically addresses each dimension with cited evidence across Sections 2–8." 115 }, 116 "causal_claims_justified": { 117 "applies": false, 118 "answer": false, 119 "justification": "The paper is a survey/perspective that synthesizes existing findings. It does not make original causal claims — it reports causal findings from reviewed papers." 120 }, 121 "generalization_bounded": { 122 "applies": true, 123 "answer": false, 124 "justification": "The title and abstract suggest a comprehensive assessment of transformer trustworthiness, but coverage is selective. For example, the NLP robustness section focuses heavily on adversarial attacks while omitting calibration and hallucination research in depth. The paper does not explicitly bound its scope or state what areas were excluded." 125 }, 126 "alternative_explanations_discussed": { 127 "applies": false, 128 "answer": false, 129 "justification": "As a survey/perspective paper presenting no original empirical results, alternative explanations for observed results are not directly applicable." 130 }, 131 "proxy_outcome_distinction": { 132 "applies": false, 133 "answer": false, 134 "justification": "Theoretical/survey paper with no original measurements." 135 } 136 }, 137 "setup_transparency": { 138 "model_versions_specified": { 139 "applies": false, 140 "answer": false, 141 "justification": "Survey paper that does not run any models." 142 }, 143 "prompts_provided": { 144 "applies": false, 145 "answer": false, 146 "justification": "No prompting is used in this survey paper." 147 }, 148 "hyperparameters_reported": { 149 "applies": false, 150 "answer": false, 151 "justification": "No experiments conducted requiring hyperparameters." 152 }, 153 "scaffolding_described": { 154 "applies": false, 155 "answer": false, 156 "justification": "No agentic scaffolding used; this is a survey paper." 157 }, 158 "data_preprocessing_documented": { 159 "applies": true, 160 "answer": false, 161 "justification": "The paper does not describe its literature search methodology, paper selection criteria, inclusion/exclusion criteria, or search queries. The selection of reviewed papers appears ad hoc." 162 } 163 }, 164 "limitations_and_scope": { 165 "limitations_section_present": { 166 "applies": true, 167 "answer": false, 168 "justification": "There is no dedicated limitations section. Section 8 (Summary and Discussion) and Section 8.1 (Future Directions) discuss open challenges but do not acknowledge limitations of the survey itself." 169 }, 170 "threats_to_validity_specific": { 171 "applies": true, 172 "answer": false, 173 "justification": "No threats to the validity of this survey are discussed. There is no acknowledgment of potential selection bias in the reviewed papers, missing domains, or methodological limitations of the review process." 174 }, 175 "scope_boundaries_stated": { 176 "applies": true, 177 "answer": false, 178 "justification": "The paper does not explicitly state what is excluded from the review. It covers a very broad range of domains but does not explain why certain areas (e.g., speech synthesis, recommendation systems) were omitted or what criteria defined the scope." 179 } 180 }, 181 "data_integrity": { 182 "raw_data_available": { 183 "applies": true, 184 "answer": false, 185 "justification": "No raw data or structured extraction from reviewed papers is provided for verification." 186 }, 187 "data_collection_described": { 188 "applies": true, 189 "answer": false, 190 "justification": "The paper does not describe how the reviewed literature was collected — no search databases, queries, date ranges, or collection methodology are stated." 191 }, 192 "recruitment_methods_described": { 193 "applies": false, 194 "answer": false, 195 "justification": "No human participants; data source is existing literature (not a standard benchmark either)." 196 }, 197 "data_pipeline_documented": { 198 "applies": true, 199 "answer": false, 200 "justification": "No documentation of how papers were selected, filtered, or organized for the review. The pipeline from initial literature identification to final inclusion is not described." 201 } 202 }, 203 "conflicts_of_interest": { 204 "funding_disclosed": { 205 "applies": true, 206 "answer": false, 207 "justification": "No funding acknowledgment or statement appears anywhere in the paper." 208 }, 209 "affiliations_disclosed": { 210 "applies": true, 211 "answer": true, 212 "justification": "Both authors list their affiliation as Aerospace Engineering Department, Worcester Polytechnic Institute." 213 }, 214 "funder_independent_of_outcome": { 215 "applies": true, 216 "answer": false, 217 "justification": "No funding is disclosed, so independence cannot be assessed. Corresponding author Jagtap has multiple self-citations (references 64–80), but funding sources for those works are not discussed." 218 }, 219 "financial_interests_declared": { 220 "applies": true, 221 "answer": false, 222 "justification": "No competing interests or financial interests statement appears in the paper." 223 } 224 }, 225 "contamination": { 226 "training_cutoff_stated": { 227 "applies": false, 228 "answer": false, 229 "justification": "Survey paper that does not evaluate any pre-trained model on benchmarks." 230 }, 231 "train_test_overlap_discussed": { 232 "applies": false, 233 "answer": false, 234 "justification": "Survey paper that does not evaluate any pre-trained model on benchmarks." 235 }, 236 "benchmark_contamination_addressed": { 237 "applies": false, 238 "answer": false, 239 "justification": "Survey paper that does not evaluate any pre-trained model on benchmarks." 240 } 241 }, 242 "human_studies": { 243 "pre_registered": { 244 "applies": false, 245 "answer": false, 246 "justification": "No human participants in this survey paper." 247 }, 248 "irb_or_ethics_approval": { 249 "applies": false, 250 "answer": false, 251 "justification": "No human participants in this survey paper." 252 }, 253 "demographics_reported": { 254 "applies": false, 255 "answer": false, 256 "justification": "No human participants in this survey paper." 257 }, 258 "inclusion_exclusion_criteria": { 259 "applies": false, 260 "answer": false, 261 "justification": "No human participants in this survey paper." 262 }, 263 "randomization_described": { 264 "applies": false, 265 "answer": false, 266 "justification": "No human participants in this survey paper." 267 }, 268 "blinding_described": { 269 "applies": false, 270 "answer": false, 271 "justification": "No human participants in this survey paper." 272 }, 273 "attrition_reported": { 274 "applies": false, 275 "answer": false, 276 "justification": "No human participants in this survey paper." 277 } 278 }, 279 "cost_and_practicality": { 280 "inference_cost_reported": { 281 "applies": false, 282 "answer": false, 283 "justification": "Survey paper with no computational method of its own." 284 }, 285 "compute_budget_stated": { 286 "applies": false, 287 "answer": false, 288 "justification": "Survey paper with no computational experiments." 289 } 290 }, 291 "survey_methodology": { 292 "prisma_or_structured_protocol": { 293 "applies": true, 294 "answer": false, 295 "justification": "No PRISMA flow diagram, no structured review protocol, no reproducible search queries. The paper selection process is entirely undocumented — papers appear to be selected ad hoc based on the authors' knowledge." 296 }, 297 "quality_assessment_of_sources": { 298 "applies": true, 299 "answer": false, 300 "justification": "The paper does not assess the methodological quality of the papers it reviews. All cited works are treated with equal authority regardless of their experimental rigor, sample sizes, or potential biases." 301 }, 302 "publication_bias_discussed": { 303 "applies": true, 304 "answer": false, 305 "justification": "No discussion of publication bias. The survey does not consider whether its sources skew positive or whether negative results about transformer trustworthiness are underrepresented." 306 } 307 } 308 }, 309 "claims": [ 310 { 311 "claim": "Attention weights are not faithful explanations of model decisions — adversarial attention distributions can shift focus to entirely different tokens while maintaining the same output.", 312 "evidence": "Cites Jain et al. [26] who demonstrated alternative attention distributions with maximum divergence from original weights producing virtually no change in output predictions (Section 2.1).", 313 "supported": "strong" 314 }, 315 { 316 "claim": "ViTs are more prone to bias amplification than CNNs, with 54% increase in bias impact (accuracy difference of 0.17 vs 0.11).", 317 "evidence": "Cites Mandal et al. [57] using Accuracy Difference metric comparing ViT-B/32 to CNNs (Section 4.1).", 318 "supported": "moderate" 319 }, 320 { 321 "claim": "Adversarial distractor sentences can reduce reading comprehension accuracy from 75% to below 7%.", 322 "evidence": "Cites Jia et al. [41] who added adversarial sentences sharing keywords but not containing correct answers across 16 transformer models (Section 3.1).", 323 "supported": "strong" 324 }, 325 { 326 "claim": "ViTs show strong natural robustness, achieving 4.3x improvement in top-1 accuracy on ImageNet-A compared to CNNs (28.1% vs 6.5%).", 327 "evidence": "Cites Paul and Chen [49] evaluating ViTs vs BiT models on ImageNet-C, ImageNet-A, and ImageNet-R (Section 3.2).", 328 "supported": "strong" 329 }, 330 { 331 "claim": "Neuro-symbolic theorem proving systems (AlphaGeometry) solved 25 out of 30 Olympiad-level geometry problems, and AlphaGeometry2 surpassed average IMO Gold Medalist level.", 332 "evidence": "Cites Trinh et al. [9] for AG1 and Chervonyi et al. [149] for AG2, with AG2 trained on 300 million synthetic examples (Section 7.8).", 333 "supported": "strong" 334 }, 335 { 336 "claim": "Most transformer models deployed in safety-critical domains lack explicit uncertainty quantification, bias assessment, or robustness checks.", 337 "evidence": "Summarized across Sections 7.1–7.9 with domain-specific boxed summaries identifying missing trustworthiness components for models in robotics, earth science, fluid dynamics, nuclear science, etc.", 338 "supported": "moderate" 339 } 340 ], 341 "red_flags": [ 342 { 343 "flag": "No structured review methodology", 344 "detail": "The paper presents itself as a comprehensive review but provides no PRISMA protocol, search methodology, inclusion/exclusion criteria, or systematic paper selection process. The reviewed literature appears selected ad hoc, making the survey non-reproducible." 345 }, 346 { 347 "flag": "Excessive self-citation", 348 "detail": "Corresponding author Jagtap has at least 17 self-citations (references 64–80) in the physics-informed computing section (Section 5.1). While some are relevant, the concentration is notable and the section disproportionately features the author's own work." 349 }, 350 { 351 "flag": "No quality assessment of reviewed papers", 352 "detail": "The survey treats all cited works with equal authority. Papers from top venues are mixed with preprints and workshop papers without any quality differentiation. This launders the signal-to-noise ratio of the underlying literature." 353 }, 354 { 355 "flag": "Scope broader than depth allows", 356 "detail": "The paper covers 10+ application domains across 46 pages. Many domain sections (e.g., material science, nuclear science) review only 3-5 papers each, which is too few to draw reliable conclusions about the state of trustworthiness in those fields." 357 } 358 ], 359 "cited_papers": [ 360 { 361 "title": "TrustLLM: Trustworthiness in Large Language Models", 362 "authors": ["Yue Huang", "Lichao Sun", "Haoran Wang"], 363 "year": 2024, 364 "arxiv_id": "2401.05561", 365 "relevance": "Comprehensive framework for evaluating LLM trustworthiness across multiple dimensions relevant to AI safety assessment." 366 }, 367 { 368 "title": "Towards Trustworthy AI: A Review of Ethical and Robust Large Language Models", 369 "authors": ["Md Meftahul Ferdaus"], 370 "year": 2026, 371 "relevance": "Recent survey on LLM trustworthiness highlighting evaluation weaknesses and sensitivity to prompt variations." 372 }, 373 { 374 "title": "Constitutional AI: Harmlessness from AI Feedback", 375 "authors": ["Yuntao Bai", "Saurav Kadavath"], 376 "year": 2022, 377 "arxiv_id": "2212.08073", 378 "relevance": "Foundational work on AI alignment through constitutional principles, relevant to safety evaluation." 379 }, 380 { 381 "title": "Are Emergent Abilities of Large Language Models a Mirage?", 382 "authors": ["Rylan Schaeffer", "Brando Miranda", "Sanmi Koyejo"], 383 "year": 2023, 384 "relevance": "Challenges emergent ability claims in LLMs by showing metric choice artifacts — directly relevant to evaluation methodology quality." 385 }, 386 { 387 "title": "Beyond Accuracy: Behavioral Testing of NLP Models with CheckList", 388 "authors": ["Marco Tulio Ribeiro", "Tongshuang Wu", "Carlos Guestrin", "Sameer Singh"], 389 "year": 2020, 390 "arxiv_id": "2005.04118", 391 "relevance": "Structured behavioral testing framework for NLP models revealing systematic failure modes beyond aggregate metrics." 392 }, 393 { 394 "title": "Robustness Gym: Unifying the NLP Evaluation Landscape", 395 "authors": ["Karan Goel", "Nazneen Rajani", "Jesse Vig"], 396 "year": 2021, 397 "arxiv_id": "2101.04840", 398 "relevance": "Modular robustness evaluation framework for NLP — relevant to AI evaluation methodology and benchmarking." 399 }, 400 { 401 "title": "Solving Olympiad Geometry without Human Demonstrations", 402 "authors": ["Trieu H. Trinh", "Yuhuai Wu", "Quoc V. Le"], 403 "year": 2024, 404 "relevance": "AlphaGeometry neuro-symbolic system for mathematical reasoning — key example of verified AI reasoning in safety-critical domains." 405 }, 406 { 407 "title": "Deep Learning with Differential Privacy", 408 "authors": ["Martin Abadi", "Andy Chu", "Ian Goodfellow"], 409 "year": 2016, 410 "relevance": "Foundational DP-SGD work relevant to privacy-preserving AI training methodology." 411 }, 412 { 413 "title": "AI Agents vs. Agentic AI: A Conceptual Taxonomy, Applications and Challenges", 414 "authors": ["Ranjan Sapkota", "Konstantinos I. Roumeliotis", "Manoj Karkee"], 415 "year": 2025, 416 "arxiv_id": "2505.10468", 417 "relevance": "Taxonomy of agentic AI systems relevant to understanding agent safety and trustworthiness." 418 }, 419 { 420 "title": "Model Context Protocol (MCP): Landscape, Security Threats, and Future Research Directions", 421 "authors": ["Xinyi Hou", "Yanjie Zhao", "Shenao Wang", "Haoyu Wang"], 422 "year": 2025, 423 "arxiv_id": "2503.23278", 424 "relevance": "Security analysis of MCP protocol for agentic AI tool orchestration — relevant to AI agent safety." 425 } 426 ] 427 }