Research papers
Interpretability, alignment, evaluations and control, from arXiv and the ACL Anthology.
60 of 6,256 documents, newest first
The Geometry of Metaphor: Probing and Steering Metaphor Representations in Large Language ModelsIs Refusal Direction Universal across Low-Resourced Languages?Discovering state-of-the-art quantum circuits with large language modelsNative Geometry vs. Linear Recoverability: How Alignment Changes Platonic ScalingGradient Routing Localizes and Removes Unintended Behaviors in RLRouterInterp: Understanding Superposed Specialisation in Mixture of Experts RoutingUnderstanding and Mitigating Token-Pruning-Induced Vulnerabilities in VLMsNot All Answers Are Contextually Persuadable: Inference Dynamics in Large Language Models under Contextual InfluenceSpeculative Safety Honeypot: Toward Proactive Defense Against Multi-turn Agent AttacksCompact Language, Complex Model Shifts: How and Where Ambiguity and Underspecification Affect LLMsThe Unequal Influence of Bad Advice: Using Training Data Attribution to Modulate Emergent MisalignmentRewired or Gated? How Instruction Tuning Shapes Knowledge-Conflict Circuits in LLMsMemory vs. Context? Influential Factors of Factual Recall in Language ModelsA Survey on the Linear Representation HypothesisRed-Teaming Auto Mode: Improving Blocking Classifiers Against Malign Coding AgentsReproducing Transparent and Scrutable Recommendations: Exploring Open-Weight Models via Natural-Language User ProfilesCan Data Attribution Filter Out Subliminal Learning? Not ReliablyPACT: Can Enterprise AI Assistants Be Trusted Under Pressure?Position: It is Time to Virtualize Foundation Models with a Self-evolving Operating System LayerAI in Science: Early InsightsReproducing and Evaluating the Generalizability of Subliminal Learning in Open-Weight ModelsDistortion of AI Alignment Revisited: RLHF is a Decent Utilitarian AlignerCompositional Multilingual and Behavioral Attribute SteeringLimitations of Automated Simulatability: LLM Simulators Can Bypass ExplanationsLLM Layers Immediately Correct Each OtherDo Chess Explanations Reflect Model Decisions? Behavioral and Token-Level Tests of LLM Reasoning FaithfulnessRecurrence Is Not Enough: Causally Validating Multilingual SAE Translation Features in Gemma 2 and 3Agent Memory Is a Surface for Endogenous Authorization LaunderingAutomated Researchers Can Mitigate Well-characterized Alignment FailuresTraining Alignment Auditors via Reinforcement LearningDebate Training Reduces Reward Hacking in RLAIFWould this change your answer? Evaluating Explanations of LLM Behavior In The Wild with Counterfactual ExperimentsData Attribution of Emergent Misalignment with Persona FeaturesModel Discovery Agent: LLM-assisted Bayesian experiment design for data-efficient discovery of mechanistic world modelsPersistent Semantic Entities in Tool-Augmented LLM SystemsDoes Out-of-Sight Equal Out-of-Mind in CoT Monitorability?Emergence of Biased Consensus in Multi-Agent LLM DebatesDemystifying Entropy-based Selection for Chain-of-Thought Compression in Large Reasoning ModelsA dataset of rated conceptual argumentsWhere Animacy Lives in Large Language Models: Tracing the Circuits of the Animacy ConceptA Multi-Dimensional Evaluation of Explainability in Media Bias DetectionElicitation without Backpropagation: Steering Model Behavior by Optimizing the Latent PosteriorMeasuring Reward-Seeking via Contrastive Belief UpdatesProgram Synthesis for Simulation-Based Inference: Joint Model Selection and Parameter EstimationSampling Reveals Style: Unsupervised, Training-Free Discovery of Prompt-Conditional Stylistic Axes in LLM ActivationsAre Arithmetic Heuristic Neurons Form-Invariant? A Mechanistic Analysis of Symbols, Text, and Code in LLMsInduction in Both Directions: A Mechanistic Analysis of In-Context Learning in Masked Diffusion Language ModelsFrom Plausible to Actionable: A Position on LLM Self-ExplanationsLost in Perception: Isolating Perceptual and Reasoning Failures in Multimodal Physics and Geometry ReasoningCoercion and Deception in AI-to-AI Management: An Agentic Benchmark of Unprompted EscalationQuantum Algorithm for Elliptic Curve Discrete Logarithms with Space-Efficient Point AdditionValue Leakage: An LLM's Answers Are Silently Shaped by Its Own ValuesANCHOR: Automated Alignment Auditing for CLI Agents on Real-World HarmRobustifying Vision-Language Models via Test-Time Prompt AdaptationStatistically Undetectable Backdoors in Deep Neural NetworksModular Pretraining Enables Access ControlOverthinking: Amplifying Reasoning Weights to Extract Learned SecretsWhen Agents Go Rogue: Activation-Based Detection of Malicious Behaviors in Multi-Agent SystemsUnbiased Alignment for Large Language Models with Noisy PreferencesHow to Avoid Debate: Scalable AI Safety via Doubly-Efficient Interactive Proofs