Tag
该论文提出VALCON框架，通过生成者/评判者双重角色变体来基准测试并提升大语言模型的价值一致性，实验揭示28个LLM普遍存在跨角色不一致问题，并提出基于表征的干预方法加以缓解。
A discussion questioning whether the classic AI alignment concern about difficulty specifying human values to a 'genie-like' AI is outdated, arguing that modern LLMs understand the intent behind requests rather than just literal instructions.
The paper introduces ContextAdapt, an evaluation framework testing how well 12 LLMs adapt values like honesty, autonomy, and confidentiality across professional domains (medicine, law, finance, national security). Models achieve high appropriateness scores but fail to apply domain-specific justifications consistently, revealing localized failures when stakes vary.
Amazon Science proposes Demographic Pluralism, an inference-time framework that estimates population-level opinion distributions by generating multiple perspectives within demographically grounded groups, reducing Jensen-Shannon distance by 8.4%-26.4% over Modular Pluralism across four backbones on GlobalOpinionQA and VITAL without requiring opinion-distribution training data.
In this Existential Hope podcast episode, UC Berkeley's Stuart Russell reflects on 50 years of AI history and argues that the safest AI may be one that remains uncertain about human preferences, raising deep questions about passing civilization to machines rather than to children.
This paper proposes BaCVA, an inference-time Bayesian method for contextualized personalized value alignment in large language models, integrating static personal values with context-specific salience to improve adaptability and data efficiency.
This paper empirically confirms that LLMs possess intrinsic value systems, introduces the PEC framework for quantifying these values, and presents an adaptive alignment prescription for efficient steering.
This paper proposes value-preserving architectural patterns for agentic AI systems, such as privacy-aware, distributed, and guard-agent architectures, to integrate human-centered values like privacy, fairness, and safety into multi-agent system designs.
This paper proposes MAVEN, a hierarchical framework for evaluating multimodal content against macro-societal values, featuring a benchmark and optimized evaluators for scalable assessment.
This paper analyzes 94 AI alignment research papers to examine how human values are conceptualized, finding that many rely on preferences and synthetic data, which risks reducing complex cultural values to binary choices.
This paper proposes MCF-CVA, a multilayer combinatorial fusion framework for contextual value alignment of LLMs, which uses multiple moral agents and an expansion-reduction process to better capture ethical pluralism and outperform single-agent baselines.
This paper tests whether different prompt framings (personalization, role-play, third-person forecasting) are interchangeable in eliciting cultural values from LLMs, using the World Values Survey. Results show that prompt framing significantly affects model responses and measured cultural alignment, with third-person forecasting yielding the strongest directional alignment.
Introduces Pancasila-Dilemmas, an evaluation dataset of 1,834 questions grounded in Indonesian values (Pancasila) to assess LLM alignment with nation-specific values. Evaluation of 50 LLMs shows significant gaps, especially in Religion and Unity dilemmas.
Introduces PLURAL, a large-scale preference dataset grounded in the Integrated Values Survey across 92 countries, with ~500,000 preference triplets from 20 diverse countries, aimed at improving cultural value alignment in LLMs.
This paper argues that aggregating moral evaluations for AI value alignment must account for contextual factors, showing that ignoring context can lead to violations of the weak Pareto principle, analogous to Simpson's paradox.
This paper develops a benchmark for evaluating environmental attitudes in 31 LLMs, finding they often exhibit progressive environmental views and contextual sensitivity, highlighting issues of steerability and normative reliability in sustainability applications.
Philosophers are increasingly being hired by top AI labs like DeepMind and Anthropic to address ethical and alignment issues, while AI is also reshaping philosophy curricula at universities.
This paper introduces the AllFaith Religious Representation Benchmark to measure how often LLMs omit religious perspectives when answering everyday ethical questions, finding that models underrepresent religion compared to human expectations, especially in practical personal situations.
This paper introduces DVMap, a framework for fine-grained pluralistic value alignment in LLMs that uses high-consensus demographic-value mapping instead of coarse national labels, achieving strong generalization across demographics, countries, and values.
This paper proposes SoVA, a framework using GraphRAG to align LLM-based agents with human social values by converting psychological theories into prescriptive instructions. Experiments on the DAILYDILEMMAS benchmark show significant improvements over prompt-based baselines.