{"total":115,"rows":[{"id":"2406.11717","title":"Refusal in Language Models Is Mediated by a Single Direction","created":"2024-06-17","cited_by_count":1032,"area":"safety"},{"id":"2406.05946","title":"Safety Alignment Should Be Made More Than Just a Few Tokens Deep","created":"2024-06-10","cited_by_count":472,"area":"alignment"},{"id":"2404.14082","title":"Mechanistic Interpretability for AI Safety -- A Review","created":"2024-04-22","cited_by_count":422,"area":"mechinterp"},{"id":"2402.05162","title":"Assessing the Brittleness of Safety Alignment via Pruning and Low-Rank Modifications","created":"2024-02-07","cited_by_count":231,"area":"alignment"},{"id":"2401.01967","title":"A Mechanistic Understanding of Alignment Algorithms: A Case Study on DPO and Toxicity","created":"2024-01-03","cited_by_count":185,"area":"alignment"},{"id":"2404.14367","title":"Preference Fine-Tuning of LLMs Should Leverage Suboptimal, On-Policy Data","created":"2024-04-22","cited_by_count":185,"area":null},{"id":"2407.15549","title":"Latent Adversarial Training Improves Robustness to Persistent Harmful Behaviors in LLMs","created":"2024-07-22","cited_by_count":173,"area":"safety"},{"id":"2402.16835","title":"Eight Methods to Evaluate Robust Unlearning in LLMs","created":"2024-02-26","cited_by_count":163,"area":"safety"},{"id":"2402.14811","title":"Fine-Tuning Enhances Existing Mechanisms: A Case Study on Entity Tracking","created":"2024-02-22","cited_by_count":149,"area":null},{"id":"2406.07358","title":"AI Sandbagging: Language Models can Strategically Underperform on Evaluations","created":"2024-06-11","cited_by_count":145,"area":null},{"id":"2401.14446","title":"Black-Box Access is Insufficient for Rigorous AI Audits","created":"2024-01-25","cited_by_count":126,"area":null},{"id":"2409.18025","title":"An Adversarial Perspective on Machine Unlearning for AI Safety","created":"2024-09-26","cited_by_count":123,"area":"safety"},{"id":"2407.02646","title":"A Practical Review of Mechanistic Interpretability for Transformer-Based Language Models","created":"2024-07-02","cited_by_count":118,"area":"mechinterp"},{"id":"2411.14257","title":"Do I Know This Entity? Knowledge Awareness and Hallucinations in Language Models","created":"2024-11-21","cited_by_count":115,"area":"mechinterp"},{"id":"2504.02234","title":"LLM Social Simulations Are a Promising Research Method","created":"2025-04-03","cited_by_count":99,"area":null},{"id":"2405.14577","title":"Representation Noising: A Defence Mechanism Against Harmful Finetuning","created":"2024-05-23","cited_by_count":96,"area":"safety"},{"id":"2409.18169","title":"Harmful Fine-tuning Attacks and Defenses for Large Language Models: A Survey","created":"2024-09-26","cited_by_count":90,"area":"alignment"},{"id":"2403.05030","title":"Defending Against Unforeseen Failure Modes with Latent Adversarial Training","created":"2024-03-08","cited_by_count":88,"area":"safety"},{"id":"2503.01822","title":"Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometry","created":"2025-03-03","cited_by_count":61,"area":"mechinterp"},{"id":"2406.16778","title":"Finding Transformer Circuits with Edge Pruning","created":"2024-06-24","cited_by_count":60,"area":null},{"id":"2412.17686","title":"Large Language Model Safety: A Holistic Survey","created":"2024-12-23","cited_by_count":59,"area":"safety"},{"id":"2407.10264","title":"What Makes and Breaks Safety Fine-tuning? A Mechanistic Study","created":"2024-07-14","cited_by_count":55,"area":"safety"},{"id":"2406.06144","title":"Language Models Resist Alignment: Evidence From Data Compression","created":"2024-06-10","cited_by_count":53,"area":null},{"id":"2501.00070","title":"ICLR: In-Context Learning of Representations","created":"2024-12-29","cited_by_count":50,"area":null},{"id":"2403.08946","title":"Usable XAI: 10 Strategies Towards Exploiting Explainability in the LLM Era","created":"2024-03-13","cited_by_count":46,"area":null},{"id":"2405.19550","title":"Stress-Testing Capability Elicitation With Password-Locked Models","created":"2024-05-29","cited_by_count":46,"area":null},{"id":"2407.14937","title":"Operationalizing a Threat Model for Red-Teaming Large Language Models (LLMs)","created":"2024-07-20","cited_by_count":46,"area":"safety"},{"id":"2406.14563","title":"Model Merging and Safety Alignment: One Bad Model Spoils the Bunch","created":"2024-06-20","cited_by_count":43,"area":"alignment"},{"id":"2502.05209","title":"Model Tampering Attacks Enable More Rigorous Evaluations of LLM Capabilities","created":"2025-02-03","cited_by_count":39,"area":null},{"id":"2507.16795","title":"Steering Out-of-Distribution Generalization with Concept Ablation Fine-Tuning","created":"2025-07-22","cited_by_count":39,"area":null},{"id":"2410.17245","title":"Towards Reliable Evaluation of Behavior Steering Interventions in LLMs","created":"2024-10-22","cited_by_count":36,"area":"mechinterp"},{"id":"2402.16382","title":"Immunization against harmful fine-tuning attacks","created":"2024-02-26","cited_by_count":35,"area":null},{"id":"2504.02922","title":"Overcoming Sparsity Artifacts in Crosscoders to Interpret Chat-Tuning","created":"2025-04-03","cited_by_count":34,"area":null},{"id":"2405.16229","title":"No Two Devils Alike: Unveiling Distinct Mechanisms of Fine-tuning Attacks","created":"2024-05-25","cited_by_count":29,"area":"alignment"},{"id":"2312.14302","title":"Exploiting Novel GPT-4 APIs","created":"2023-12-21","cited_by_count":28,"area":"agents"},{"id":"2510.13900","title":"Narrow Finetuning Leaves Clearly Readable Traces in Activation Differences","created":"2025-10-14","cited_by_count":28,"area":null},{"id":"2402.03268","title":"Understanding Reasoning Ability of Language Models From the Perspective of Reasoning Paths Aggregation","created":"2024-02-05","cited_by_count":22,"area":"lr"},{"id":"2402.10688","title":"Towards Uncovering How Large Language Model Works: An Explainability Perspective","created":"2024-02-16","cited_by_count":22,"area":"lr"},{"id":"2502.11812","title":"Towards Understanding Fine-Tuning Mechanisms of LLMs via Circuit Analysis","created":"2025-02-17","cited_by_count":20,"area":null},{"id":"2410.16531","title":"Bayesian scaling laws for in-context learning","created":"2024-10-21","cited_by_count":19,"area":"safety"},{"id":"2502.02180","title":"The Elicitation Game: Evaluating Capability Elicitation Techniques","created":"2025-02-04","cited_by_count":18,"area":null},{"id":"2405.19524","title":"AI Risk Management Should Incorporate Both Safety and Security","created":"2024-05-29","cited_by_count":17,"area":null},{"id":"2411.16353","title":"Lessons from Studying Two-Hop Latent Reasoning","created":"2024-11-25","cited_by_count":17,"area":"lr"},{"id":"2502.11019","title":"Unlocking the Power of Function Vectors for Characterizing and Mitigating Catastrophic Forgetting in Continual Instruction Tuning","created":"2025-02-16","cited_by_count":16,"area":null},{"id":"2402.01909","title":"On Catastrophic Inheritance of Large Foundation Models","created":"2024-02-02","cited_by_count":15,"area":null},{"id":"2403.02513","title":"Balancing Enhancement, Harmlessness, and General Capabilities: Enhancing Conversational LLMs with Direct RLHF","created":"2024-03-04","cited_by_count":15,"area":"alignment"},{"id":"2510.00845","title":"Mechanistic Interpretability as Statistical Estimation: A Variance Analysis","created":"2025-10-01","cited_by_count":15,"area":"mechinterp"},{"id":"2409.02231","title":"SmileyLlama: Modifying Large Language Models for Directed Chemical Space Exploration","created":"2024-09-03","cited_by_count":14,"area":"alignment"},{"id":"2404.17047","title":"Near to Mid-term Risks and Opportunities of Open-Source Generative AI","created":"2024-04-25","cited_by_count":13,"area":null},{"id":"2405.08597","title":"Risks and Opportunities of Open-Source Generative AI","created":"2024-05-14","cited_by_count":13,"area":null},{"id":"2601.21571","title":"Shaping capabilities with token-level data filtering","created":"2026-01-29","cited_by_count":13,"area":null},{"id":"2506.01115","title":"Is Random Attention Sufficient for Sequence Modeling? Disentangling Trainable Components in the Transformer","created":"2025-06-01","cited_by_count":12,"area":"lr"},{"id":"2507.07341","title":"On the Impossibility of Separating Intelligence from Judgment: The Computational Intractability of Filtering for AI Alignment","created":"2025-07-09","cited_by_count":12,"area":"safety"},{"id":"2510.05092","title":"Learning to Interpret Weight Differences in Language Models","created":"2025-10-06","cited_by_count":11,"area":null},{"id":"2403.03942","title":"The Heuristic Core: Understanding Subnetwork Generalization in Pretrained Language Models","created":"2024-03-06","cited_by_count":10,"area":null},{"id":"2502.14828","title":"Fundamental Limitations in Pointwise Defences of LLM Finetuning APIs","created":"2025-02-20","cited_by_count":10,"area":null},{"id":"2508.09853","title":"STREAM (ChemBio): A Standard for Transparently Reporting Evaluations in AI Model Reports","created":"2025-08-13","cited_by_count":10,"area":null},{"id":"2410.11588","title":"Causal Reasoning in Large Language Models: A Knowledge Graph Approach","created":"2024-10-15","cited_by_count":9,"area":"lr"},{"id":"2507.21084","title":"Reviving Your MNEME: Predicting The Side Effects of LLM Unlearning and Fine-Tuning via Sparse Model Diffing","created":"2025-06-19","cited_by_count":8,"area":"safety"},{"id":"2312.01818","title":"Hybrid Approaches for Moral Value Alignment in AI Agents: a Manifesto","created":"2023-12-04","cited_by_count":7,"area":"agents"},{"id":"2506.04706","title":"Line of Sight: On Linear Representations in VLLMs","created":"2025-06-05","cited_by_count":6,"area":"mechinterp"},{"id":"2604.02608","title":"Steerable but Not Decodable: Function Vectors Operate Beyond the Logit Lens","created":"2026-04-03","cited_by_count":6,"area":"mechinterp"},{"id":"2411.06424","title":"How Does DPO Reduce Toxicity? A Mechanistic Neuron-Level Analysis","created":"2024-11-10","cited_by_count":5,"area":"alignment"},{"id":"2506.04679","title":"Normative Conflicts and Shallow AI Alignment","created":"2025-06-05","cited_by_count":5,"area":"alignment"},{"id":"2507.16329","title":"DREAM: Scalable Red Teaming for Text-to-Image Generative Systems via Distribution Modeling","created":"2025-07-22","cited_by_count":5,"area":"alignment"},{"id":"2602.20062","title":"A Theory of How Pretraining Shapes Inductive Bias in Fine-Tuning","created":"2026-02-23","cited_by_count":5,"area":null},{"id":"2406.02619","title":"Unelicitable Backdoors in Language Models via Cryptographic Transformer Circuits","created":"2024-06-03","cited_by_count":4,"area":"safety"},{"id":"2501.06208","title":"Enhancing AI Safety Through the Fusion of Low Rank Adapters","created":"2024-12-30","cited_by_count":4,"area":null},{"id":"2505.13500","title":"Noise Injection Systemically Degrades Large Language Model Safety Guardrails","created":"2025-05-16","cited_by_count":4,"area":"safety"},{"id":"2506.03755","title":"Misalignment or misuse? The AGI alignment tradeoff","created":"2025-06-04","cited_by_count":4,"area":null},{"id":"2601.22313","title":"Hair-Trigger Alignment: Black-Box Evaluation Cannot Guarantee Post-Update Alignment","created":"2026-01-29","cited_by_count":4,"area":null},{"id":"2602.11157","title":"Response-Based Knowledge Distillation for Multilingual Jailbreak Prevention Unwittingly Compromises Safety","created":"2025-12-08","cited_by_count":4,"area":"alignment"},{"id":"2603.04426","title":"Delta-Crosscoder: Robust Crosscoder Model Diffing in Narrow Fine-Tuning Regimes","created":"2026-02-16","cited_by_count":4,"area":null},{"id":"2604.08169","title":"Activation Steering for Aligned Open-ended Generation without Sacrificing Coherence","created":"2026-04-09","cited_by_count":4,"area":"safety"},{"id":"2606.11502","title":"When Role-playing, Do Models Believe What They Say?","created":"2026-06-09","cited_by_count":4,"area":null},{"id":"2502.06106","title":"Fine-Tuning is Subgraph Search: A New Lens on Learning Dynamics","created":"2025-02-10","cited_by_count":3,"area":"mechinterp"},{"id":"2503.01896","title":"Neuroplasticity and Corruption in Model Mechanisms: A Case Study Of Indirect Object Identification","created":"2025-02-27","cited_by_count":3,"area":null},{"id":"2504.19333","title":"Unified Multi-Task Learning & Model Fusion for Efficient Language Model Guardrailing","created":"2025-04-27","cited_by_count":3,"area":"safety"},{"id":"2509.17866","title":"Understanding Post-Training Structural Changes in Large Language Models","created":"2025-09-22","cited_by_count":3,"area":null},{"id":"2406.04146","title":"Towards Understanding Task-agnostic Debiasing Through the Lenses of Intrinsic Bias and Forgetfulness","created":"2024-06-06","cited_by_count":2,"area":null},{"id":"2505.10936","title":"Cochain: Balancing Insufficient and Excessive Collaboration in LLM Agent Workflows","created":"2025-05-16","cited_by_count":2,"area":"agents"},{"id":"2507.09875","title":"Function Induction and Task Generalization: An Interpretability Study with Off-by-One Addition","created":"2025-07-14","cited_by_count":2,"area":null},{"id":"2510.07626","title":"LLM Unlearning Under the Microscope: A Full-Stack View on Methods and Metrics","created":"2025-10-08","cited_by_count":2,"area":"safety"},{"id":"2510.17256","title":"Explainability of Large Language Models: Opportunities and Challenges toward Generating Trustworthy Explanations","created":"2025-10-20","cited_by_count":2,"area":"mechinterp"},{"id":"2602.00533","title":"Convergent World Representations and Divergent Tasks","created":"2026-01-31","cited_by_count":2,"area":null},{"id":"2603.06592","title":"Hierarchical Latent Structures in Data Generation Process Unify Mechanistic Phenomena across Scale","created":"2026-02-04","cited_by_count":2,"area":"mechinterp"},{"id":"2606.12360","title":"Anatomy of Post-Training: Using Interpretability to Characterize Data and Shape the Learning Signal","created":"2026-06-10","cited_by_count":2,"area":null},{"id":"2407.10949","title":"Representing Rule-based Chatbots with Transformers","created":"2024-07-15","cited_by_count":1,"area":null},{"id":"2505.16178","title":"Understanding Fact Recall in Language Models: Why Two-Stage Training Encourages Memorization but Mixed Training Teaches Knowledge","created":"2025-05-22","cited_by_count":1,"area":null},{"id":"2511.16699","title":"Detecting and Steering LLMs' Empathy in Action","created":"2025-11-17","cited_by_count":1,"area":null},{"id":"2512.16914","title":"Constructive Circuit Amplification: Improving Math Reasoning in LLMs via Targeted Sub-Network Updates","created":"2025-12-18","cited_by_count":1,"area":"lr"},{"id":"2601.03570","title":"How Do Large Language Models Learn Concepts During Continual Pre-Training?","created":"2026-01-07","cited_by_count":1,"area":null},{"id":"2602.20904","title":"Transcoder Adapters for Reasoning-Model Diffing","created":"2026-02-24","cited_by_count":1,"area":"lr"},{"id":"2605.25073","title":"Security in the Fine-Tuning Lifecycle of Large Language Models: Threats, Defenses,Evaluation, and Future Directions","created":"2026-05-24","cited_by_count":1,"area":"safety"},{"id":"2606.19542","title":"Tracking Representation Dynamics in Large Language Models with Persistent Homology","created":"2026-06-17","cited_by_count":1,"area":null},{"id":"2610.05292","title":"Erased, Rerouted, or Rescaled? Post-Training and the Causal Quotient of a Language Model's Belief State","created":"2026-10-04","cited_by_count":1,"area":null},{"id":"2408.01308","title":"Reconsidering Degeneration of Token Embeddings with Definitions for Encoder-based Pre-trained Language Models","created":"2024-08-02","cited_by_count":0,"area":null},{"id":"2410.12001","title":"Impacts of Continued Legal Pre-Training and IFT on LLMs' Latent Representations of Human-Defined Legal Concepts","created":"2024-10-15","cited_by_count":0,"area":null},{"id":"2503.04377","title":"How can representation dimension dominate structurally pruned LLMs?","created":"2025-03-06","cited_by_count":0,"area":null},{"id":"2601.23153","title":"Behemoth: Benchmarking Unlearning in LLMs Using Fully Synthetic Data","created":"2026-01-30","cited_by_count":0,"area":"synth"}],"next_cursor":"100"}