NeurIPS 2026 reviews are up: reviews, scores, rebuttals and meta-reviews for 7,877 papers. See the highest rated
acceptodds: the papers of ICLR 2027
979 results for “keyword:"safety"” in ICLR 2027. 231 withdrawn or desk rejected listed apart. clearsort:relevanceratingnewestmost read
[PDF]Safe Executions Can Hide Policy Failure: A Proposal-Level Audit of Hybrid Control in Synthetic Interviews▲2.0k1.50
[PDF]JailbreakProfiler: Fine-Grained Attack Pattern Recognition and Open-Set Attribution for LLM Safety▼5.8k2.33
[PDF]Beyond Refusal: Policy-Grounded On-Policy Distillation for Safety Decisions and Recovery▼1.2k1.67
[PDF]Safety Is Not Conserved: Constraining Self-Improvement Across Models and Agent Harnesses▼1361.33
[PDF]AlphaCUA: Scaling Safety Benchmarks for Computer-Use Agents by Discovering and Validating Risk Factors▼4.3k2.00
[PDF]Disentangling Verbalized Evaluation Awareness and Refusal through Neuron-Level Intervention▼4.3k2.00
[PDF]From Model Coverage to Control Value: Auditing Few-Shot Safety Filters under Topology Shift▲11.00
[PDF]Doubt Before Acting: Assumption-Level Representation Editing for Safe Embodied LLM Planning▲5.9k1.67
[PDF]The Perceptual Bandwidth Tax: Matched Evaluation of Emotion and Accent as Jailbreak Surfaces in End-to-End Audio-Language Models▲432.75
[PDF]Action-Conditioned Predictive Consistency for Adversarial Detection in Vision-Language-Action Models▼4.3k2.00
[PDF]Before You Report a Null: A Two-Sided Pre-Flight Screen for Shutdown-Resistance Evaluations▲11.00
[PDF]Do Detailed Guard Denials Help Bypass Guardrails? Evaluating Runtime Feedback in DeepSeek Harness▼4.3k2.00
[PDF]Skill Libraries Change What an Edit Does: When Should Agent Skill Validation Run In Context?▲3952.67
[PDF]PTP: Previous-Token Prediction based LLM Inversion for Near-Exact Prompt Reconstruction▲6.1k2.00
[PDF]Representation Without Capability, Then Capability Without Abstraction: Parity in GPT-2▼1361.33
[PDF]KV Compression Cannot Double as a Jailbreak Defense: The Harmfulness-then-Refusal Trajectory Guides Token Deletion▼5.8k2.33
[PDF]A Harmful Request Is Still a Request: Benign Fine-Tuning Rewrites How Aligned Models Begin to Answer▲3952.67
[PDF]A Jailbreak Success Coordinate Outside the Refusal Subspace: Decodable Before Generation, Reachable from the Input▼5.8k2.33
[PDF]Pretending to Think: Unmasking and Eliminating Post-Hoc Rationalization in LLM Safety Chain of Thought▼4.3k2.00
[PDF]From Risk Recognition to Safer Generation: Compete–Split–Fuse for Selective Representation Coordination▼4.3k2.00
Withdrawn or desk rejected (231)
[PDF]Canonical Paraphrasing: A Provable Defense against Steganography in Language ModelsWithdrawn1.33
[PDF]Erased But Not Gone: How Model Merging Weakens Concept Erasure in Diffusion ModelsWithdrawn1.33
[PDF]HelpInject: Reframing Adversarial Goals as Helpful Solutions for Indirect Prompt Injection in LLM AgentsWithdrawn1.67
[PDF]REFUSALTRACE: Sparse Activation Masking for Jailbreaking via Natural Refusal TrajectoriesWithdrawn2.00
[PDF]One State Variable Was the Whole Effect: A Controlled Study of Offline-to-Online Reinforcement Learning for Cluster SchedulingWithdrawn1.67
[PDF]Faithful to the Persona, Unfaithful to the Decision: Why Reasoning CoTs Cannot Report the Decisions They NarrateWithdrawn1.67
[PDF]Right Answers, Poisoned Provenance: What Accuracy Misses in LLM-Based Multi-Agent ReasoningWithdrawn1.33