Sycophancy Towards Researchers Drives Performative Misalignment

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.

Coherence Maximization Improves Pluralistic Alignment

Aligning AI systems with diverse human values requires value specifications grounded in concrete examples, but generating such examples without extensive human supervision remains an open challenge. We investigate what makes these examples effective, using Internal Coherence Maximization (ICM) -- which infers labels by maximizing their mutual predictability -- to generate persona-specific examples that steer a model toward a target group's values, without human supervision. Across four benchmarks spanning classification, preference, and open-ended generation, ICM-inferred in-context examples match the performance of gold labels. Crucially, coherence matters beyond individual label accuracy: with accuracy held constant, more coherent examples generalize substantially better than incoherent ones. For personas underrepresented in pretraining data, targeted human feedback on the questions where the model is least certain about a persona's values yields better generalization than the same number of labels on arbitrary questions. These results identify coherence as a key design principle for scalable value specification, leveraging the diverse human perspectives already encoded in pretrained language models.

Domain Specific Benchmarks for Evaluating Multimodal Large Language Models

Large language models (LLMs) are increasingly being deployed across disciplines due to their advanced reasoning and problem solving capabilities. To measure their effectiveness, various benchmarks have been developed that measure aspects of LLM reasoning, comprehension, and problem-solving. While several surveys address LLM evaluation and benchmarks, a domain-specific analysis remains underexplored in the literature. This paper introduces a taxonomy of seven key disciplines, encompassing various domains and application areas where LLMs are extensively utilized. Additionally, we provide a comprehensive review of LLM benchmarks and survey papers within each domain, highlighting the unique capabilities of LLMs and the challenges faced in their application. Finally, we compile and categorize these benchmarks by domain to create an accessible resource for researchers, aiming to pave the way for advancements toward artificial general intelligence (AGI)