Large language models are increasingly influencing human moral decisions, yet current approaches focus primarily on evaluating rather than actively steering their moral decisions. We formulate this as an out-of-distribution moral alignment problem, where LLM agents must learn to apply consistent moral reasoning frameworks to scenarios beyond their training distribution. We introduce Moral-Reason-QA, a novel dataset extending 680 human-annotated, high-ambiguity moral scenarios with framework-specific reasoning traces across utilitarian, deontological, and virtue ethics, enabling systematic evaluation of moral generalization in realistic decision contexts. Our learning approach employs Group Relative Policy Optimization with composite rewards that simultaneously optimize decision alignment and framework-specific reasoning processes to facilitate learning of the underlying moral frameworks. Experimental results demonstrate successful generalization to unseen moral scenarios, with softmax-normalized alignment scores improving by +0.757 for utilitarian and +0.450 for deontological frameworks when tested on out-of-distribution evaluation sets. The experiments also reveal training challenges and promising directions that inform future research. These findings establish that LLM agents can be systematically trained to internalize and apply specific moral frameworks to novel situations, providing a critical foundation for AI safety as language models become more integrated into human decision-making processes.

This study introduces a hybrid neuro-symbolic framework that achieves deterministic detection of statutory inconsistency in complex law. We use the U.S. Internal Revenue Code (IRC) as a case study because its complexity makes it a fertile domain for identifying conflicts. Our research offers a solution for detecting inconsistent provisions by combining Large Language Models (LLMs) with symbolic logic. LLM-based methods can support compliance, fairness, and statutory drafting, yet tax-specific applications remain sparse. A key challenge is that such models struggle with hierarchical processing and deep structured reasoning, especially over long text. This research addresses these gaps through experiments using GPT-4o, GPT-5, and Prolog. GPT-4o was first used to translate Section 121 into Prolog rules and refine them in SWISH. These rules were then incorporated into prompts to test whether Prolog-augmented prompting improved GPT-4o's inconsistency detection. GPT-4o, whether prompted with natural language alone or with Prolog augmentation, detected the inconsistency in only one of three strategies (33 percent accuracy), but its reasoning quality differed: natural-language prompting achieved 100 percent rule coverage, while Prolog-augmented prompting achieved 66 percent, indicating more incomplete statutory analysis. In contrast to probabilistic prompting, the hybrid Prolog model produced deterministic and reproducible results. Guided by GPT-5 for refinement, the model formalized the IRC section's competing interpretations and successfully detected an inconsistency zone. Validation tests confirm that the Prolog implementation is accurate, internally consistent, deterministic, and capable of autonomously identifying inconsistencies. These findings show that LLM-assisted formalization, anchored in symbolic logic, enables transparent and reliable statutory inconsistency detection.

Generative vision-language models like Stable Diffusion demonstrate remarkable capabilities in creative media synthesis, but they also pose substantial risks of producing unsafe, offensive, or culturally inappropriate content when prompted adversarially. Current defenses struggle to align outputs with human values without sacrificing generation quality or incurring high costs. To address these challenges, we introduce VALOR (Value-Aligned LLM-Overseen Rewriter), a modular, zero-shot agentic framework for safer and more helpful text-to-image generation. VALOR integrates layered prompt analysis with human-aligned value reasoning: a multi-level NSFW detector filters lexical and semantic risks; a cultural value alignment module identifies violations of social norms, legality, and representational ethics; and an intention disambiguator detects subtle or indirect unsafe implications. When unsafe content is detected, prompts are selectively rewritten by a large language model under dynamic, role-specific instructions designed to preserve user intent while enforcing alignment. If the generated image still fails a safety check, VALOR optionally performs a stylistic regeneration to steer the output toward a safer visual domain without altering core semantics. Experiments across adversarial, ambiguous, and value-sensitive prompts show that VALOR significantly reduces unsafe outputs by up to 100.00% while preserving prompt usefulness and creativity. These results highlight VALOR as a scalable and effective approach for deploying safe, aligned, and helpful image generation systems in open-world settings.

Ensemble techniques in recommender systems have demonstrated accuracy improvements of 10-30%, yet their environmental impact remains unmeasured. While deep learning recommendation algorithms can generate up to 3,297 kg CO2 per paper, ensemble methods have not been sufficiently evaluated for energy consumption. This thesis investigates how ensemble techniques influence environmental impact compared to single optimized models. We conducted 93 experiments across two frameworks (Surprise for rating prediction, LensKit for ranking) on four datasets spanning 100,000 to 7.8 million interactions. We evaluated four ensemble strategies (Average, Weighted, Stacking/Rank Fusion, Top Performers) against simple baselines and optimized single models, measuring energy consumption with a smart plug. Results revealed a non-linear accuracy-energy relationship. Ensemble methods achieved 0.3-5.7% accuracy improvements while consuming 19-2,549% more energy depending on dataset size and strategy. The Top Performers ensemble showed best efficiency: 0.96% RMSE improvement with 18.8% energy overhead on MovieLens-1M, and 5.7% NDCG improvement with 103% overhead on MovieLens-100K. Exhaustive averaging strategies consumed 88-270% more energy for comparable gains. On the largest dataset (Anime, 7.8M interactions), the Surprise ensemble consumed 2,005% more energy (0.21 Wh vs. 0.01 Wh) for 1.2% accuracy improvement, producing 53.8 mg CO2 versus 2.6 mg CO2 for the single model. This research provides one of the first systematic measurements of energy and carbon footprint for ensemble recommender systems, demonstrates that selective strategies offer superior efficiency over exhaustive averaging, and identifies scalability limitations at industrial scale. These findings enable informed decisions about sustainable algorithm selection in recommender systems.

In the context of pretraining of Large Language Models (LLMs), synthetic data has emerged as an alternative for generating high-quality pretraining data at scale. This is particularly beneficial in low-resource language settings where the benefits of recent LLMs have been unevenly distributed across languages. In this work, we present a systematic study on the generation and evaluation of synthetic multilingual pretraining data for Indic languages, where we construct a large-scale synthetic dataset BhashaKritika, comprising 540B tokens using 5 different techniques for 10 languages. We explore the impact of grounding generation in documents, personas, and topics. We analyze how language choice, both in the prompt instructions and document grounding, affects data quality, and we compare translations of English content with native generation in Indic languages. To support scalable and language-sensitive evaluation, we introduce a modular quality evaluation pipeline that integrates script and language detection, metadata consistency checks, n-gram repetition analysis, and perplexity-based filtering using KenLM models. Our framework enables robust quality control across diverse scripts and linguistic contexts. Empirical results through model runs reveal key trade-offs in generation strategies and highlight best practices for constructing effective multilingual corpora.

Large Language Models (LLMs) demonstrate remarkable capabilities, but their training on massive corpora poses significant risks from memorized sensitive information. To mitigate these issues and align with legal standards, unlearning has emerged as a critical technique to selectively erase specific knowledge from LLMs without compromising their overall performance. This survey provides a systematic review of over 180 papers on LLM unlearning published since 2021. First, it introduces a novel taxonomy that categorizes unlearning methods based on the phase in the LLM pipeline of the intervention. This framework further distinguishes between parameter modification and parameter selection strategies, thus enabling deeper insights and more informed comparative analysis. Second, it offers a multidimensional analysis of evaluation paradigms. For datasets, we compare 18 existing benchmarks from the perspectives of task format, content, and experimental paradigms to offer actionable guidance. For metrics, we move beyond mere enumeration by dividing knowledge memorization metrics into 10 categories to analyze their advantages and applicability, while also reviewing metrics for model utility, robustness, and efficiency. By discussing current challenges and future directions, this survey aims to advance the field of LLM unlearning and the development of secure AI systems.

Despite the remarkable advances of Large Language Models (LLMs) across diverse cognitive tasks, the rapid enhancement of these capabilities also introduces emergent deceptive behaviors that may induce severe risks in high-stakes deployments. More critically, the characterization of deception across realistic real-world scenarios remains underexplored. To bridge this gap, we establish DeceptionBench, the first benchmark that systematically evaluates how deceptive tendencies manifest across different societal domains, what their intrinsic behavioral patterns are, and how extrinsic factors affect them. Specifically, on the static count, the benchmark encompasses 150 meticulously designed scenarios in five domains, i.e., Economy, Healthcare, Education, Social Interaction, and Entertainment, with over 1,000 samples, providing sufficient empirical foundations for deception analysis. On the intrinsic dimension, we explore whether models exhibit self-interested egoistic tendencies or sycophantic behaviors that prioritize user appeasement. On the extrinsic dimension, we investigate how contextual factors modulate deceptive outputs under neutral conditions, reward-based incentivization, and coercive pressures. Moreover, we incorporate sustained multi-turn interaction loops to construct a more realistic simulation of real-world feedback dynamics. Extensive experiments across LLMs and Large Reasoning Models (LRMs) reveal critical vulnerabilities, particularly amplified deception under reinforcement dynamics, demonstrating that current models lack robust resistance to manipulative contextual cues and the urgent need for advanced safeguards against various deception behaviors. Code and resources are publicly available at https://github.com/Aries-iai/DeceptionBench.

Large language models (LLMs) have shown promising accuracy in predicting survey responses and policy preferences, which has increased interest in their potential to represent human interests in various domains. Most existing research has focused on "behavioral cloning", effectively evaluating how well models reproduce individuals' expressed preferences. Drawing on theories of political representation, we highlight an underexplored design trade-off: whether AI systems should act as delegates, mirroring expressed preferences, or as trustees, exercising judgment about what best serves an individual's interests. This trade-off is closely related to issues of LLM sycophancy, where models can encourage behavior or validate beliefs that may be aligned with a user's short-term preferences, but is detrimental to their long-term interests. Through a series of experiments simulating votes on various policy issues in the U.S. context, we apply a temporal utility framework that weighs short and long-term interests (simulating a trustee role) and compare voting outcomes to behavior-cloning models (simulating a delegate). We find that trustee-style predictions weighted toward long-term interests produce policy decisions that align more closely with expert consensus on well-understood issues, but also show greater bias toward models' default stances on topics lacking clear agreement. These findings reveal a fundamental trade-off in designing AI systems to represent human interests. Delegate models better preserve user autonomy but may diverge from well-supported policy positions, while trustee models can promote welfare on well-understood issues yet risk paternalism and bias on subjective topics.

We study how open source communities describe participation and control through version controlled governance documents. Using a corpus of 710 projects with paired snapshots, we parse text into actors, rules, actions, and objects, then group them and measure change with entropy for evenness, richness for diversity, and Jensen Shannon divergence for drift. Projects define more roles and more actions over time, and these are distributed more evenly, while the composition of rules remains stable. These findings indicate that governance grows by expanding and balancing categories of participation without major shifts in prescriptive force. The analysis provides a reproducible baseline for evaluating whether future AI mediated workflows concentrate or redistribute authority.

Significant progress in reward modeling over recent years has been driven by a paradigm shift from task-specific designs towards generalist reward models. Despite this trend, developing effective reward models remains a fundamental challenge: the heavy reliance on large-scale labeled preference data. Pre-training on abundant unlabeled data offers a promising direction, but existing approaches fall short of instilling explicit reasoning into reward models. To bridge this gap, we propose a self-training approach that leverages unlabeled data to elicit reward reasoning in reward models. Based on this approach, we develop GRAM-R$^2$, a generative reward model trained to produce not only preference labels but also accompanying reward rationales. GRAM-R$^2$ can serve as a foundation model for reward reasoning and can be applied to a wide range of tasks with minimal or no additional fine-tuning. It can support downstream applications such as response ranking and task-specific reward tuning. Experiments on response ranking, task adaptation, and reinforcement learning from human feedback demonstrate that GRAM-R$^2$ consistently delivers strong performance, outperforming several strong discriminative and generative baselines.