With the advance of Artificial Intelligence (AI), Large Language Models (LLMs) have gained prominence and been applied in diverse contexts. As they evolve into more sophisticated versions, it is essential to assess whether they reproduce biases, such as discrimination and racialization, while maintaining hegemonic discourses. Current bias detection approaches rely mostly on quantitative, automated methods, which often overlook the nuanced ways in which biases emerge in natural language. This study proposes a qualitative, discursive framework to complement such methods. Through manual analysis of LLM-generated short stories featuring Black and white women, we investigate gender and racial biases. We contend that qualitative methods such as the one proposed here are fundamental to help both developers and users identify the precise ways in which biases manifest in LLM outputs, thus enabling better conditions to mitigate them. Results show that Black women are portrayed as tied to ancestry and resistance, while white women appear in self-discovery processes. These patterns reflect how language models replicate crystalized discursive representations, reinforcing essentialization and a sense of social immobility. When prompted to correct biases, models offered superficial revisions that maintained problematic meanings, revealing limitations in fostering inclusive narratives. Our results demonstrate the ideological functioning of algorithms and have significant implications for the ethical use and development of AI. The study reinforces the need for critical, interdisciplinary approaches to AI design and deployment, addressing how LLM-generated discourses reflect and perpetuate inequalities.

The active research topic of prompt engineering makes it evident that LLMs are sensitive to small changes in prompt wording. A portion of this can be ascribed to the inductive bias that is present in the LLM. By using an LLM's output as a portion of its prompt, we can more easily create satisfactory wording for prompts. This has the effect of creating a prompt that matches the inductive bias in model. Empirically, we show that using this Inductive Bias Extraction and Matching strategy improves LLM Likert ratings used for classification by up to 19% and LLM Likert ratings used for ranking by up to 27%.

In-context learning (ICL) has emerged as a powerful capability of transformer-based language models, enabling them to perform tasks by conditioning on a small number of examples presented at inference time, without any parameter updates. Prior work has shown that transformers can generalize over simple function classes like linear functions, decision trees, even neural networks, purely from context, focusing on numerical or symbolic reasoning over underlying well-structured functions. Instead, we propose a novel application of ICL into the domain of cryptographic function learning, specifically focusing on ciphers such as mono-alphabetic substitution and Vigenère ciphers, two classes of private-key encryption schemes. These ciphers involve a fixed but hidden bijective mapping between plain text and cipher text characters. Given a small set of (cipher text, plain text) pairs, the goal is for the model to infer the underlying substitution and decode a new cipher text word. This setting poses a structured inference challenge, which is well-suited for evaluating the inductive biases and generalization capabilities of transformers under the ICL paradigm. Code is available at https://github.com/adistomar/CS182-project.

Patients who are at clinical high risk (CHR) for schizophrenia need close monitoring of their symptoms to inform appropriate treatments. The Brief Psychiatric Rating Scale (BPRS) is a validated, commonly used research tool for measuring symptoms in patients with schizophrenia and other psychotic disorders; however, it is not commonly used in clinical practice as it requires a lengthy structured interview. Here, we utilize large language models (LLMs) to predict BPRS scores from clinical interview transcripts in 409 CHR patients from the Accelerating Medicines Partnership Schizophrenia (AMP-SCZ) cohort. Despite the interviews not being specifically structured to measure the BPRS, the zero-shot performance of the LLM predictions compared to the true assessment (median concordance: 0.84, ICC: 0.73) approaches human inter- and intra-rater reliability. We further demonstrate that LLMs have substantial potential to improve and standardize the assessment of CHR patients via their accuracy in assessing the BPRS in foreign languages (median concordance: 0.88, ICC: 0.70), and integrating longitudinal information in a one-shot or few-shot learning approach.

The proliferation of Large Language Models (LLMs) is challenged by hallucinations, critical failure modes where models generate non-factual, nonsensical or unfaithful text. This paper introduces Semantic Divergence Metrics (SDM), a novel lightweight framework for detecting Faithfulness Hallucinations -- events of severe deviations of LLMs responses from input contexts. We focus on a specific implementation of these LLM errors, {confabulations, defined as responses that are arbitrary and semantically misaligned with the user's query. Existing methods like Semantic Entropy test for arbitrariness by measuring the diversity of answers to a single, fixed prompt. Our SDM framework improves upon this by being more prompt-aware: we test for a deeper form of arbitrariness by measuring response consistency not only across multiple answers but also across multiple, semantically-equivalent paraphrases of the original prompt. Methodologically, our approach uses joint clustering on sentence embeddings to create a shared topic space for prompts and answers. A heatmap of topic co-occurances between prompts and responses can be viewed as a quantified two-dimensional visualization of the user-machine dialogue. We then compute a suite of information-theoretic metrics to measure the semantic divergence between prompts and responses. Our practical score, $\mathcal{S}_H$, combines the Jensen-Shannon divergence and Wasserstein distance to quantify this divergence, with a high score indicating a Faithfulness hallucination. Furthermore, we identify the KL divergence KL(Answer $||$ Prompt) as a powerful indicator of \textbf{Semantic Exploration}, a key signal for distinguishing different generative behaviors. These metrics are further combined into the Semantic Box, a diagnostic framework for classifying LLM response types, including the dangerous, confident confabulation.

Evaluation of reasoning language models gained importance after it was observed that they can combine their existing capabilities into novel traces of intermediate steps before task completion and that the traces can sometimes help them to generalize better than past models. As reasoning becomes the next scaling dimension of large language models, careful study of their capabilities in critical tasks is needed. We show that better performance is not always caused by test-time algorithmic improvements or model sizes but also by using impactful benchmarks as curricula for learning. We call this benchmark-driven selection of AI and show its effects on DeepSeek-R1 using our sequential decision-making problem from Humanity's Last Exam. Steering development of AI by impactful benchmarks trades evaluation for learning and makes novelty of test tasks key for measuring generalization capabilities of reasoning models. Consequently, some benchmarks could be seen as curricula for training rather than unseen test sets.

Recent advances in Large Reasoning Models (LRMs) have demonstrated strong performance on complex tasks through long Chain-of-Thought (CoT) reasoning. However, their lengthy outputs increase computational costs and may lead to overthinking, raising challenges in balancing reasoning effectiveness and efficiency. Current methods for efficient reasoning often compromise reasoning quality or require extensive resources. This paper investigates efficient methods to reduce the generation length of LRMs. We analyze generation path distributions and filter generated trajectories through difficulty estimation. Subsequently, we analyze the convergence behaviors of the objectives of various preference optimization methods under a Bradley-Terry loss based framework. Based on the analysis, we propose Length Controlled Preference Optimization (LCPO) that directly balances the implicit reward related to NLL loss. LCPO can effectively learn length preference with limited data and training. Extensive experiments demonstrate that our approach significantly reduces the average output length by over 50\% across multiple benchmarks while maintaining the reasoning performance. Our work highlights the potential for computationally efficient approaches in guiding LRMs toward efficient reasoning.

Aspect Traditional AI agents Modern agentic AI systems (LLM-based agents) Definition Autonomous entities with fixed sensing/acting loops; limited by static rules or models Autonomous reasoning systems using LLMs with dynamic behavior, tool orchestration, and context-awarenessAutonomy Limited autonomy; often dependent on human input or predefined instructions High autonomy; capable of independently performing complex and extended tasks Goal Management Focused on single, static goals or fixed task planning Capable of managing multiple, evolving, and nested goals adaptivelyArchitecture Rule-based or BDI (Belief-Desire-Intention) models; monolithic design Modular architecture centered on LLMs, with components for memory, tools, context injection, and rolesAdaptability Suited to controlled, predictable environments; poor generalization Designed for open, dynamic, and unpredictable environmentsDecision-Making Deterministic or rule-based logic; symbolic reasoning Context-sensitive, probabilistic reasoning with adaptive planning and self-reflection Learning Mechanism Rule-based or supervised learning with limited updates Self-supervised and reinforcement learning; continual fine-tuning possible Context Handling Static or manually coded states and rules Dynamic context injection via agent protocols (e.g., MCP, A2A) and runtime awareness Communication Message-passing via ACL or KQML Real-time, event-driven collaboration; natural language interfacesTool Use Limited or predefined tools and actions Dynamic tool invocation, chaining, and API calling based on contextMemory Optional, often hardcoded or task-specific Integrated memory systems supporting long-and short-term information retention

Recent advancements in reasoning-reinforced Large Language Models (LLMs) have shown remarkable capabilities in complex reasoning tasks. However, the mechanism underlying their utilization of different human reasoning skills remains poorly investigated, especially for multilingual commonsense reasoning that involves everyday knowledge across different languages and cultures. To address this gap, we propose a \textbf{M}ultilingual and Scalable Benchmark for \textbf{S}kill-based \textbf{Co}mmonsense \textbf{Re}asoning (\textbf{mSCoRe}). Our benchmark incorporates three key components that are designed to systematically evaluate LLM's reasoning capabilities, including: (1) a novel taxonomy of reasoning skills that enables fine-grained analysis of models' reasoning processes, (2) a robust data synthesis pipeline tailored specifically for commonsense reasoning evaluation, and (3) a complexity scaling framework allowing task difficulty to scale dynamically alongside future improvements in LLM abilities. Extensive experiments on eights state-of-the-art LLMs of varying sizes and training approaches demonstrate that \textbf{mSCoRe} remains significantly challenging for current models, particularly at higher complexity levels. Our results reveal the limitations of such reasoning-reinforced models when confronted with nuanced multilingual general and cultural commonsense. We further provide detailed analysis on the models' reasoning processes, suggesting future directions for improving multilingual commonsense reasoning capabilities.

For large language models (LLMs), reasoning over graphs could help solve many problems. Prior work has tried to improve LLM graph reasoning by examining how best to serialize graphs as text and by combining GNNs and LLMs. However, the merits of such approaches remain unclear, so we empirically answer the following research questions: (1) Can LLMs learn to solve fundamental graph tasks without specialized graph encoding models?, (2) Can LLMs generalize learned solutions to unseen graph structures or tasks?, and (3) What are the merits of competing approaches to learn graph tasks? We show that even small LLMs can learn to solve graph tasks by training them with instructive chain-of-thought solutions, and this training generalizes, without specialized graph encoders, to new tasks and graph structures.