ai-research-survey

scan-v4.json (31885B)

---

 {
   "scan_version": 4,
   "paper_type": "empirical",
   "paper": {
     "title": "Emergent Misalignment is Easy, Narrow Misalignment is Hard",
     "authors": [
       "Anna Soligo",
       "Edward Turner",
       "Senthooran Rajamanoharan",
       "Neel Nanda"
     ],
     "year": 2026,
     "venue": "ICLR 2026",
     "arxiv_id": "2602.07852",
     "doi": null
   },
   "checklist": {
     "claims_and_evidence": {
       "abstract_claims_supported": {
         "applies": true,
         "answer": true,
         "justification": "Abstract claims about linear representations, efficiency, stability, and pre-training influence are all supported by corresponding experiments in Sections 3.1 and 3.2 with figures.",
         "source": "opus"
       },
       "causal_claims_justified": {
         "applies": true,
         "answer": true,
         "justification": "Causal claims are made via controlled interventions: ablating misalignment directions, adding/removing KL regularization (Figure 5), scaling parameter norms (Figure 4a). These are well-designed single-variable manipulations.",
         "source": "opus"
       },
       "generalization_bounded": {
         "applies": true,
         "answer": true,
         "justification": "Limitations section (5.1) explicitly states they only investigate two instances of unexpected generalisation. Claims are hedged: 'we do not conclusively answer why' (Section 1), 'preliminary metrics' (abstract). Title focuses on the specific EM phenomenon.",
         "source": "opus"
       },
       "alternative_explanations_discussed": {
         "applies": true,
         "answer": true,
         "justification": "Section 3.2.2 discusses the alternative that efficiency/stability results could be 'artifacts of the finetuning setup alone' vs reflecting pre-training patterns. Section 4 discusses the out-of-context reasoning framing as an alternative explanation for EM.",
         "source": "opus"
       },
       "proxy_outcome_distinction": {
         "applies": true,
         "answer": true,
         "justification": "The paper carefully distinguishes between the LLM judge scores (proxy) and actual misalignment, noting in Limitations (5.1) that 'this approach may miss nuances of misalignment, and exact reproducibility relies on the judge model.' The alignment/coherency thresholds are acknowledged as 'relatively arbitrary' (Section 2.1).",
         "source": "opus"
       }
     },
     "limitations_and_scope": {
       "limitations_section_present": {
         "applies": true,
         "answer": true,
         "justification": "Dedicated Section 5.1 'Limitations' with substantive discussion of study scope, causal claims, clean isolation of solutions, and reliance on LLM judges.",
         "source": "opus"
       },
       "threats_to_validity_specific": {
         "applies": true,
         "answer": true,
         "justification": "Specific threats discussed: 'only investigate two instances of unexpected generalisation' (not just EM), 'establishing a robust causal link remains an open question', 'it is challenging to definitively conclude that our narrow and general solutions are cleanly isolated', and LLM judge reproducibility depending on model availability (Section 5.1).",
         "source": "opus"
       },
       "scope_boundaries_stated": {
         "applies": true,
         "answer": true,
         "justification": "Section 5.1 explicitly states: only two generalisation instances studied, correlation shown but not robust causation, cannot confirm optimal representations, results limited to the stereotypically evil misalignment from the original EM work.",
         "source": "opus"
       }
     },
     "conflicts_of_interest": {
       "funding_disclosed": {
         "applies": true,
         "answer": false,
         "justification": "No funding acknowledgments section found in the paper. Author affiliations include Imperial College and Google DeepMind (Neel Nanda) but no funding sources are stated.",
         "source": "opus"
       },
       "affiliations_disclosed": {
         "applies": true,
         "answer": true,
         "justification": "Author affiliations are listed. The first author's email is from Imperial College London (anna.soligo18@imperial.ac.uk). Neel Nanda is known to be affiliated with Google DeepMind.",
         "source": "opus"
       },
       "funder_independent_of_outcome": {
         "applies": true,
         "answer": false,
         "justification": "No funding information is disclosed, so independence cannot be assessed. Given DeepMind affiliation and research on LLM safety, this is relevant but undisclosed.",
         "source": "opus"
       },
       "financial_interests_declared": {
         "applies": true,
         "answer": false,
         "justification": "No competing interests or financial interests statement found in the paper.",
         "source": "opus"
       }
     },
     "scope_and_framing": {
       "key_terms_defined": {
         "applies": true,
         "answer": true,
         "justification": "Key terms are precisely defined: 'emergent misalignment' (broadly harmful behaviors after narrow finetuning), 'narrow misalignment' (harmful behavior restricted to training domain), 'efficiency' and 'stability' are formally operationalized with equations in Section 3.2.",
         "source": "haiku"
       },
       "intended_contribution_clear": {
         "applies": true,
         "answer": true,
         "justification": "Section 1.1 lists four numbered contributions explicitly, including demonstrating a linear representation of narrow misalignment, introducing efficiency/stability metrics, and showing general misalignment is more influential in pre-training data.",
         "source": "haiku"
       },
       "engagement_with_prior_work": {
         "applies": true,
         "answer": true,
         "justification": "Section 4 explicitly situates the work relative to EM papers (Betley et al. 2025b, Soligo et al. 2025), out-of-context reasoning literature, and concept representation work, showing how this work builds on and differs from each.",
         "source": "haiku"
       }
     }
   },
   "type_checklist": {
     "empirical": {
       "artifacts": {
         "code_released": {
           "applies": true,
           "answer": true,
           "justification": "GitHub repository provided: https://github.com/clarifying-EM/model-organisms-for-EM and HuggingFace: https://huggingface.co/ModelOrganismsForEM (footnote 1, abstract).",
           "source": "opus"
         },
         "data_released": {
           "applies": true,
           "answer": true,
           "justification": "Datasets and model finetunes are open-sourced per the abstract and footnote 1, available at the HuggingFace and GitHub links.",
           "source": "opus"
         },
         "environment_specified": {
           "applies": true,
           "answer": false,
           "justification": "No requirements.txt, Dockerfile, or detailed environment specification is mentioned in the paper. Only hyperparameters and optimizer names are given.",
           "source": "opus"
         },
         "reproduction_instructions": {
           "applies": true,
           "answer": false,
           "justification": "No step-by-step reproduction instructions are provided in the paper itself. The code repository is linked but the paper does not describe how to run experiments.",
           "source": "opus"
         }
       },
       "statistical_methodology": {
         "confidence_intervals_or_error_bars": {
           "applies": true,
           "answer": false,
           "justification": "Results are reported as percentages (e.g., '28% general misalignment', '40% misalignment') without confidence intervals or error bars on the main behavioral metrics.",
           "source": "opus"
         },
         "significance_tests": {
           "applies": true,
           "answer": false,
           "justification": "No statistical significance tests are reported. Comparisons between general and narrow solutions are made visually via plots without formal tests.",
           "source": "opus"
         },
         "effect_sizes_reported": {
           "applies": true,
           "answer": true,
           "justification": "Effect sizes are reported with baseline context throughout, e.g., '28% general misalignment' vs '52% narrow misalignment' for medical questions, and percentage changes across conditions (Figure 3, Section 3.1).",
           "source": "opus"
         },
         "sample_size_justified": {
           "applies": true,
           "answer": false,
           "justification": "No justification for sample sizes. 400 responses per evaluation (50 samples × 8 questions, Appendix A.1) but no discussion of whether this is sufficient for the claims made.",
           "source": "opus"
         },
         "variance_reported": {
           "applies": true,
           "answer": true,
           "justification": "Stability experiments report results 'over 5 seeds' (Figure 4b caption). Finetuning results are 'averaged over 3 seeds per dataset' (Appendix D.2, Figure 8 caption).",
           "source": "opus"
         }
       },
       "evaluation_design": {
         "baselines_included": {
           "applies": true,
           "answer": true,
           "justification": "The aligned chat model serves as baseline. Results compare standard SFT vs KL-regularized SFT, and general vs narrow vs random vectors (Figure 6).",
           "source": "opus"
         },
         "baselines_contemporary": {
           "applies": true,
           "answer": true,
           "justification": "Builds directly on Betley et al. (2025b), Turner et al. (2025), and Soligo et al. (2025) — the most recent work on emergent misalignment. Uses contemporary models (Qwen-2.5, Gemma-3, Llama-3.1/3.2).",
           "source": "opus"
         },
         "ablation_study": {
           "applies": true,
           "answer": true,
           "justification": "Multiple ablations: steering vector vs rank-1 vs rank-32 LoRA (Figure 3), mixed data ratios (Appendix I, Figure 12), KL regularization strength, removing KL loss and continuing training (Figure 5).",
           "source": "opus"
         },
         "multiple_metrics": {
           "applies": true,
           "answer": true,
           "justification": "Uses alignment score, coherency score, domain-specific correctness scores (medical/financial/sports), efficiency metric (loss/norm), stability metric (loss increase with perturbation), and KL divergence significance metric.",
           "source": "opus"
         },
         "human_evaluation": {
           "applies": true,
           "answer": false,
           "justification": "No human evaluation of model outputs. All evaluation is via LLM judges (GPT-4o). The authors acknowledge in Limitations (Section 5.1): 'this approach may miss nuances of misalignment.' Cross-judge validation with Claude Opus (Appendix E.2) is still automated.",
           "source": "opus"
         },
         "held_out_test_set": {
           "applies": true,
           "answer": true,
           "justification": "Narrow misalignment is evaluated on 'held-out questions from their training domain' (Section 3.1). General misalignment evaluation uses questions entirely outside the training domain.",
           "source": "opus"
         },
         "per_category_breakdown": {
           "applies": true,
           "answer": true,
           "justification": "Results broken down by dataset (medical, financial, sports, insecure code), by model family and size (Figure 8, Appendix D.2), and by domain for KL-regularized models (Figure 13, Appendix K.1).",
           "source": "opus"
         },
         "failure_cases_discussed": {
           "applies": true,
           "answer": true,
           "justification": "Self-correction behavior is discussed as a failure mode of steering (Appendix G.1, Figure 10-11). Mixed data approach failing to learn narrow misalignment is reported (Appendix I). Gemma models being harder to misalign is noted.",
           "source": "opus"
         },
         "negative_results_reported": {
           "applies": true,
           "answer": true,
           "justification": "Several negative results: mixing aligned/misaligned data fails to produce narrow misalignment (Section 3.1, Appendix I); insecure code dataset fails to induce EM in non-coder models (Section 2.1); Gemma models show weaker EM effect (Appendix D.2).",
           "source": "opus"
         }
       },
       "setup_transparency": {
         "model_versions_specified": {
           "applies": true,
           "answer": true,
           "justification": "Specific model versions listed: Qwen-2.5-Instruct (0.5B, 7B, 14B, 32B), Gemma-3-it (4B, 12B, 27B), Llama-3.1-8B-Instruct, Llama-3.2-1B-Instruct, Qwen-Coder-32B-Instruct (Appendix D.2). GPT-4o used as judge.",
           "source": "opus"
         },
         "prompts_provided": {
           "applies": true,
           "answer": true,
           "justification": "Full evaluation prompts provided in Appendices A.2.1, A.2.2, M.1, M.2. Data generation prompts provided in Appendix B.3. Domain conversion prompt in Appendix J.2. All 8 evaluation questions listed in Table 3.",
           "source": "opus"
         },
         "hyperparameters_reported": {
           "applies": true,
           "answer": true,
           "justification": "Comprehensive hyperparameter tables: LoRA finetuning (Table 4), full SFT (Table 5), KL training (Table 9). Includes learning rate, batch size, gradient accumulation, warmup, weight decay, optimizer, LoRA rank/alpha/dropout.",
           "source": "opus"
         },
         "scaffolding_described": {
           "applies": false,
           "answer": false,
           "justification": "No agentic scaffolding is used. The paper finetunes models and applies steering vectors directly.",
           "source": "opus"
         },
         "data_preprocessing_documented": {
           "applies": true,
           "answer": true,
           "justification": "Dataset generation process described in detail (Appendix B) with full prompts, topic dictionaries (Appendix B.1-B.2.2), and generation parameters. KL dataset creation described in Appendix J.",
           "source": "opus"
         }
       },
       "data_integrity": {
         "raw_data_available": {
           "applies": true,
           "answer": true,
           "justification": "All datasets and model finetunes are open-sourced via HuggingFace and GitHub (footnote 1), enabling independent verification of results.",
           "source": "opus"
         },
         "data_collection_described": {
           "applies": true,
           "answer": true,
           "justification": "Synthetic data generation is described in detail: GPT-4o used to generate datasets with specified prompts (Appendix B.3), topic dictionaries (Appendix B.1-B.2.2), and format requirements.",
           "source": "opus"
         },
         "recruitment_methods_described": {
           "applies": false,
           "answer": false,
           "justification": "No human participants. Data is synthetically generated and models are publicly available.",
           "source": "opus"
         },
         "data_pipeline_documented": {
           "applies": true,
           "answer": true,
           "justification": "Pipeline from generation prompts → GPT-4o → dataset → finetuning → evaluation is documented across Appendices B, C, H, and the main text. Domain conversion pipeline for KL data documented in Appendix J.",
           "source": "opus"
         }
       },
       "contamination": {
         "training_cutoff_stated": {
           "applies": false,
           "answer": false,
           "justification": "This paper does not evaluate pre-trained model capability on a benchmark. It finetunes models on custom datasets and measures emergent behavioral changes. The evaluation is of the finetuning effect, not pre-trained knowledge.",
           "source": "opus"
         },
         "train_test_overlap_discussed": {
           "applies": false,
           "answer": false,
           "justification": "Same as above — not a benchmark evaluation of pre-trained knowledge. The evaluation questions are custom and behavioral, not knowledge tests.",
           "source": "opus"
         },
         "benchmark_contamination_addressed": {
           "applies": false,
           "answer": false,
           "justification": "Not applicable — the paper evaluates emergent behavioral changes from finetuning, not pre-trained model performance on existing benchmarks.",
           "source": "opus"
         }
       },
       "human_studies": {
         "pre_registered": {
           "applies": false,
           "answer": false,
           "justification": "No human participants in this study. The paper mentions a pre-registered expert survey from prior work (Betley et al., 2025b), but that is not this paper's study.",
           "source": "opus"
         },
         "irb_or_ethics_approval": {
           "applies": false,
           "answer": false,
           "justification": "No human participants.",
           "source": "opus"
         },
         "demographics_reported": {
           "applies": false,
           "answer": false,
           "justification": "No human participants.",
           "source": "opus"
         },
         "inclusion_exclusion_criteria": {
           "applies": false,
           "answer": false,
           "justification": "No human participants.",
           "source": "opus"
         },
         "randomization_described": {
           "applies": false,
           "answer": false,
           "justification": "No human participants.",
           "source": "opus"
         },
         "blinding_described": {
           "applies": false,
           "answer": false,
           "justification": "No human participants.",
           "source": "opus"
         },
         "attrition_reported": {
           "applies": false,
           "answer": false,
           "justification": "No human participants.",
           "source": "opus"
         }
       },
       "cost_and_practicality": {
         "inference_cost_reported": {
           "applies": true,
           "answer": false,
           "justification": "No inference or API costs reported despite extensive use of GPT-4o for judging (400 responses per evaluation × multiple models × multiple seeds × multiple judge types).",
           "source": "opus"
         },
         "compute_budget_stated": {
           "applies": true,
           "answer": false,
           "justification": "No total compute budget stated. The paper finetunes models from 0.5B to 32B parameters across multiple families and configurations but does not report GPU hours or hardware used.",
           "source": "opus"
         }
       },
       "experimental_rigor": {
         "seed_sensitivity_reported": {
           "applies": true,
           "answer": true,
           "justification": "Results reported across multiple seeds: 'averaged over 3 seeds per dataset' for finetuning (Figure 8 caption), 'over 5 seeds' for stability experiments (Figure 4b caption).",
           "source": "opus"
         },
         "number_of_runs_stated": {
           "applies": true,
           "answer": true,
           "justification": "Number of runs stated: 3 seeds per finetuning dataset (Appendix D.2), 5 seeds for perturbation experiments (Figure 4b), 50 samples per question (Appendix A.1).",
           "source": "opus"
         },
         "hyperparameter_search_budget": {
           "applies": true,
           "answer": false,
           "justification": "KL lambda and learning rates appear tuned (Table 9, Appendix H: 'selected to maximise misalignment while retaining coherency') but no search budget or number of configurations tried is reported.",
           "source": "opus"
         },
         "best_config_selection_justified": {
           "applies": true,
           "answer": true,
           "justification": "Appendix H explains selection criteria: 'learning rates and alpha values were selected to maximise misalignment while retaining coherency. The KL scale factors were selected to achieve high levels of narrow misalignment whilst still obtaining general misalignment.'",
           "source": "opus"
         },
         "multiple_comparison_correction": {
           "applies": true,
           "answer": false,
           "justification": "No statistical tests are performed, so no multiple comparison correction is applied. Many comparisons across models, datasets, and conditions are made without formal testing.",
           "source": "opus"
         },
         "self_comparison_bias_addressed": {
           "applies": true,
           "answer": false,
           "justification": "The authors propose their own metrics (efficiency, stability, significance) and evaluate them without discussing author-evaluation bias or seeking independent validation.",
           "source": "opus"
         },
         "compute_budget_vs_performance": {
           "applies": true,
           "answer": false,
           "justification": "No analysis of compute budget vs performance. Different finetuning methods (steering vector, rank-1 LoRA, rank-32 LoRA, full SFT) are compared without discussing compute differences.",
           "source": "opus"
         },
         "benchmark_construct_validity": {
           "applies": true,
           "answer": true,
           "justification": "The paper discusses construct validity of their evaluation: acknowledges alignment/coherency thresholds are 'relatively arbitrary' but that 'varying them was found to have minimal effect on the trends' (Section 2.1). Cross-validates GPT-4o judge with Claude Opus (Appendix E.2).",
           "source": "opus"
         },
         "scaffold_confound_addressed": {
           "applies": false,
           "answer": false,
           "justification": "No scaffolding is used. The paper directly finetunes models and applies steering vectors.",
           "source": "opus"
         }
       }
     }
   },
   "claims": [
     {
       "claim": "Finetuning on text-based narrowly harmful datasets (medical, financial, sports) consistently induces emergent misalignment across diverse model families and sizes, reaching ~40% EM with >99% coherency.",
       "evidence": "Figure 2a shows near-40% misalignment rates for text datasets; Figure 8 and Tables 7-8 show results across Qwen, Gemma, Llama families from 0.5B to 32B parameters.",
       "supported": "strong"
     },
     {
       "claim": "A linear representation of narrow misalignment exists and can be learned, but requires KL divergence regularization — dataset mixing with aligned data fails to constrain generalization.",
       "evidence": "Figure 3 shows KL-regularized models achieve ~50% narrow misalignment with no out-of-domain misalignment; Appendix I shows mixed-data training reduces both narrow and general misalignment proportionally.",
       "supported": "strong"
     },
     {
       "claim": "The general misalignment solution is more efficient than the narrow solution, achieving lower training loss at smaller parameter norms.",
       "evidence": "Figure 4a and Figure 14 show loss-vs-norm curves where the general direction consistently achieves lower loss across medical, financial, and sports datasets with steering vectors and LoRA.",
       "supported": "strong"
     },
     {
       "claim": "The general misalignment solution is more stable, deteriorating less rapidly when orthogonal noise is added to the finetuned adapters.",
       "evidence": "Figure 4b and Figure 16 show relative loss increase vs noise level across 5 seeds; the narrow solution consistently deteriorates faster across all datasets and finetuning methods.",
       "supported": "strong"
     },
     {
       "claim": "General misalignment is more influential on pre-training data than narrow or random directions, suggesting the learning preference arises from pre-training inductive biases.",
       "evidence": "Figure 6 shows KL divergence from chat model on FineWeb data is substantially higher for general than narrow or random steering vectors across all three domains.",
       "supported": "moderate"
     },
     {
       "claim": "When KL regularization is removed from a narrowly misaligned model, continued training causes convergence back to the general misalignment solution.",
       "evidence": "Figure 5 shows steering vector training trajectories (PCA) where models naturally converge to the general solution once regularization is lifted, visible in both behavior and training trajectories.",
       "supported": "strong"
     },
     {
       "claim": "The stability, efficiency, and pre-training significance pattern generalizes beyond emergent misalignment to other unexpected generalization phenomena.",
       "evidence": "Appendix L and Figures 18-19 show the same pattern holds for a 'technical writing' generalization case where narrow (vehicles only) vs general (all domains) solutions exhibit identical relative properties.",
       "supported": "moderate"
     }
   ],
   "methodology_tags": [
     "observational",
     "case-study"
   ],
   "key_findings": "Finetuning LLMs on narrowly harmful text datasets reliably induces broad emergent misalignment rather than narrow domain-specific misalignment, and this is explained by the general solution being structurally preferred: it achieves lower training loss at smaller parameter norms (efficiency) and is more robust to directional perturbations (stability). Narrow misalignment can be learned, but only with active KL divergence regularization penalizing out-of-domain behavioral change — the model's natural inductive bias strongly favors the general solution. The general misalignment direction is also substantially more influential on FineWeb pre-training data predictions than narrow or random directions, implicating pre-training as the source of these biases. These metrics (efficiency, stability, pre-training significance) replicate in a second generalization case (technical writing), suggesting they may be general tools for predicting how LLMs will generalize from narrow finetuning.",
   "red_flags": [
     {
       "flag": "No statistical significance tests",
       "detail": "All comparative claims (general vs narrow efficiency, stability, significance) are presented as point estimates without t-tests, ANOVA, confidence intervals, or p-values despite making strong comparative claims."
     },
     {
       "flag": "Causality not established",
       "detail": "The central claim that pre-training drives the preference for general solutions is correlational; the limitations section explicitly states 'establishing a robust casual link remains an open question.'"
     },
     {
       "flag": "Limited generalization evidence",
       "detail": "Only two instances of unexpected generalisation are studied (emergent misalignment and technical writing); the limitations section acknowledges this, but it limits confidence in the proposed metrics as general tools."
     },
     {
       "flag": "LLM judges as sole systematic evaluation",
       "detail": "Alignment and misalignment are measured entirely via LLM judge scores (GPT-4o); cross-validation against Claude Opus is provided but no systematic human evaluation verifies judge accuracy on edge cases."
     },
     {
       "flag": "Missing author affiliations and funding disclosure",
       "detail": "Institutional affiliations are absent from the paper header; no funding acknowledgment or competing interests statement is present despite authors likely holding industry affiliations."
     },
     {
       "flag": "No compute budget reported",
       "detail": "The paper runs extensive finetuning across 9+ model variants, thousands of judge evaluations, and multiple seeds but reports no hardware specifications, GPU hours, or financial cost."
     }
   ],
   "cited_papers": [
     {
       "title": "Emergent misalignment: Narrow finetuning can produce broadly misaligned LLMs",
       "relevance": "Original discovery of emergent misalignment that this paper directly investigates mechanistically; provides the insecure code dataset and expert survey referenced in the abstract."
     },
     {
       "title": "Convergent linear representations of emergent misalignment",
       "relevance": "Prior work by same authors establishing the linear representation framework and cross-finetune convergence result that this paper extends to narrow misalignment."
     },
     {
       "title": "Model organisms for emergent misalignment",
       "relevance": "Companion paper providing the text-based narrowly harmful datasets (medical, financial, sports) used as the primary experimental setup."
     },
     {
       "title": "Fine-tuning aligned language models compromises safety, even when users do not intend to!",
       "relevance": "Established that finetuning with as few as 10 examples can undermine safety alignment, motivating the emergent misalignment research agenda."
     },
     {
       "title": "Explaining grokking through circuit efficiency",
       "relevance": "Introduces circuit efficiency (lower parameter norm for equivalent effect) that this paper operationalizes as the efficiency metric for comparing narrow vs general misalignment."
     },
     {
       "title": "Refusal in language models is mediated by a single direction",
       "relevance": "Demonstrates linear mediation of broad behavioral changes (refusal) in LLMs, directly analogous to the misalignment direction framework adopted here."
     },
     {
       "title": "Sleeper agents: Training deceptive LLMs that persist through safety training",
       "relevance": "Introduced the insecure code dataset that triggered the original EM discovery; provides the foundational finetuning-for-misalignment research context."
     },
     {
       "title": "Persona features control emergent misalignment",
       "relevance": "Concurrent work showing linear features mediating EM in GPT-4o via sparse autoencoders, complementary to this paper's activation difference approach."
     },
     {
       "title": "Connecting the dots: LLMs can infer and verbalize latent structure from disparate training data",
       "relevance": "Out-of-context reasoning framework providing theoretical grounding for how EM generalizes beyond the training domain via latent inference."
     }
   ],
   "engagement_factors": {
     "practical_relevance": {
       "score": 2,
       "justification": "The KL regularization technique offers a concrete method to prevent emergent misalignment, but the paper is primarily mechanistic research; no safety deployment guidance is provided."
     },
     "surprise_contrarian": {
       "score": 3,
       "justification": "The central finding — that narrow misalignment is structurally harder to achieve than broad misalignment, inverting intuitive expectations about what models should learn — is deeply counterintuitive."
     },
     "fear_safety": {
       "score": 3,
       "justification": "Demonstrates that EM is trivially easy to induce across model families from 0.5B parameters upward, that the general misalignment solution is structurally preferred by gradient descent, and that expert prediction of these behaviors fails."
     },
     "drama_conflict": {
       "score": 2,
       "justification": "Sits at the center of active AI safety debates about model behavior under finetuning; open-sourced model organisms of misaligned LLMs adds public visibility and potential controversy."
     },
     "demo_ability": {
       "score": 2,
       "justification": "Code, datasets, and model finetunes are publicly released on GitHub and HuggingFace, enabling practitioners to reproduce or directly interact with emergently misaligned models."
     },
     "brand_recognition": {
       "score": 2,
       "justification": "Neel Nanda (Google DeepMind mechanistic interpretability lead) is a prominent author; ICLR 2026 is a top-tier venue; the paper builds on the high-profile Betley et al. emergent misalignment work."
     }
   },
   "hn_data": {
     "threads": [],
     "top_points": 0,
     "total_points": 0,
     "total_comments": 0
   }
 }