惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

Google DeepMind News
Google DeepMind News
Apple Machine Learning Research
Apple Machine Learning Research
博客园_首页
爱范儿
爱范儿
让小产品的独立变现更简单 - ezindie.com
让小产品的独立变现更简单 - ezindie.com
罗磊的独立博客
M
MIT News - Artificial intelligence
D
Docker
量子位
T
Tailwind CSS Blog
人人都是产品经理
人人都是产品经理
月光博客
月光博客
有赞技术团队
有赞技术团队
J
Java Code Geeks
A
About on SuperTechFans
P
Proofpoint News Feed
Jina AI
Jina AI
Y
Y Combinator Blog
T
The Blog of Author Tim Ferriss
The GitHub Blog
The GitHub Blog
Microsoft Security Blog
Microsoft Security Blog
V
V2EX
GbyAI
GbyAI
F
Fortinet All Blogs

cs updates on arXiv.org

Beyond Binary Edits Robust Multimodal Knowledge Editing with Adversarial Subspace Alignment Agentic Proving for Program Verification MemAudit: Post-hoc Auditing of Poisoned Agent Memory via Causal Attribution and Structural Anomaly Detection OpenSkillEval: Automatically Auditing the Open Skill Ecosystem for LLM Agents One Policy, Infinite NPCs: Persona-Traceable Shared RL Policies for Scalable Game Agents How Human-Like Are Large Language Models? A Register-Aware Linguistic Evaluation Framework Benchmarking Google Embeddings 2 against Open-Source Models for Multilingual Dense Retrieval and RAG Systems Structure-Guided Entity Resolution: Fine-Tuning LLMs for Robust Name Matching in Complex Linguistic Contexts Solving the Aircraft Disassembly Scheduling Problem Co-ReAct: Rubrics as Step-Level Collaborators for ReAct Agents CP or DP? Why Not Both: A Case Study in the Partial Shop Scheduling Problem Asking For An Old Friend: Diagnosing and Mitigating Temporal Failure Modes in LLM-based Statutory Question Answering EDGE-OPD: Internalizing Privileged Context with Evidence Guided On-Policy Distillation ARES: Automated Rubric Synthesis for Scalable LLM Reinforcement Learning SSDAU: Structured Semantic Data Augmentation for Joint Entity and Relation Extraction Naturalistic measure of social norms alignment Articulatory strategy as a source of variation in acoustic vowel dynamics When Planning Fails Despite Correct Execution: On Epistemic Calibration for LLM-Based Multi-Agent Systems EquiSumm : A Gender Bias-Aware Framework for Inclusive Tweet Summarization Metacognition as Reward: Reinforcing LLM Reasoning via Knowledge and Regulation Signals From Correctness to Preference: A Framework for Personalized Agentic Reinforcement Learning Cultural Adaptation in Large Language Models for Political Discourse Emotion Recognition in Sign Language Conversation ClimateChat-300K: A Multi-Modal Facebook Dataset for Understanding Diverse Perspectives in Climate Communication AraHopeCorpus: Annotation Guidelines and Dataset for Hope Speech in Arabic Social Media Crisis Discourse Human-in-the-Loop Multi-Agent Ventilator Decision Support with Contextual Bandit Preference Learning Convergence Without Understanding: When Language Models Agree on Representations but Disagree on Reasoning DART: Semantic Recoverability for Structured Tool Agents Ontological Knowledge Blocks: Executable Compliance and Profile-Based Validation for Trustworthy AI Systems Parallel Context Compaction for Long-Horizon LLM Agent Serving
Sycophancy Towards Researchers Drives Performative Misali...
David D. Baek, Xinnuo Li, Anay Gupta, Taslim Mahbub, Kejian Shi, · 2026-06-07 · via cs updates on arXiv.org

The increasing situational awareness of language models raises safety concerns: models might be aware when they are evaluated, and adjust their behavior to evade monitoring and resist modification, e.g., pretending to be aligned only in evaluation. This alignment faking behavior is often interpreted as scheming: an intentional effort of strategic deception. In this paper, we examine an alternative interpretation, performative misalignment, which explains the change in behavior as a result of sycophancy towards AI researchers. To examine this hypothesis, we present three empirical findings. First, we show that evaluation awareness persists even when we tell models they are deployed, which contradicts the scheming story which predicts less misalignment when the model perceives evaluation. Second, we use probing and steering to show that our current methods cannot mechanistically distinguish sycophancy and scheming in alignment faking evaluations. Third, we fine-tune models to be more sycophantic and observe increased sensitivity to evaluation cues. To conclude, we emphasize deconfounding sycophancy from scheming for future work on evaluations and mitigations of intent misalignment.