Large language models (LLMs) often generate natural language rationales -- free-form explanations that help improve performance on complex reasoning tasks and enhance interpretability for human users. However, evaluating these rationales remains challenging. While recent work has relied on binary preference judgments from humans or LLM judges, such evaluations are often opaque and coarse-grained, offering limited insight into what makes one rationale better than another. In this work, we rethink preference evaluation for LLM-generated rationales by asking: (1) What attributes define good rationales? (2) Can human preferences be explained by these attributes? (3) Can attribute-based evaluation overcome the limitations of binary comparisons? We identify a set of key rationale attributes from prior literature and assess them using automatic metrics, LLM judgments, and human annotations. We then analyze two standard human preference datasets MT Bench and Chatbot Arena using SHAP to identify which attributes best explain human preference outcomes. Finally, we re-evaluate model-generated rationales using attribute-specific ELO scores, revealing more nuanced model comparisons and insights. Our findings suggest that fine-grained attribute evaluations can better characterize rationale quality and guide future research toward more interpretable and reliable evaluation practices.

Computational cognitive architectures are broadly scoped models of the human mind that combine different psychological functionalities (as well as often different computational methods for these different functionalities) into one unified framework. They structure them in a psychologically plausible and validated way. However, such models thus far have only limited computational capabilities, mostly limited by the computational tools and techniques that were adopted. More recently, LLMs have proved to be more capable computationally than any other tools. Thus, in order to deal with both real-world complexity and psychological realism at the same time, incorporating LLMs into cognitive architectures naturally becomes an important task. In the present article, a synergistic combination of the Clarion cognitive architecture and LLMs is discussed as a case study. The implicit-explicit dichotomy that is fundamental to Clarion is leveraged for a seamless integration of Clarion and LLMs. As a result, computational power of LLMs is combined with psychological nicety of Clarion.

We introduce PHLoRA (Pronounced "flora"). (Post-hoc LoRA), a simple yet powerful method to extract low-rank adaptation adapters from full-rank fine-tuned models without requiring access to training data or gradients. By computing the low-rank decomposition of weight differences between a base model and its fine-tuned counterpart, our method reconstructs adapter modules that can be merged or dynamically routed at inference time via S-LoRA, or served in scalable, industry settings using platforms like NVIDIA NIM. This approach amortizes latency overhead across requests and yields substantial cost savings. Unlike prior work that trains each adapter explicitly, our approach decouples fine-tuning from adapter generation, allowing adapter extraction from existing full-rank models or third-party checkpoints. Experiments on text, image, and video benchmarks using the Amazon Nova model family demonstrate that extracted adapters preserve high energy from the full weight delta, can be pruned safely, and yield negligible degradation in downstream task performance when re-merged. Overall, PHLoRA provides a practical path for making all existing full-rank checkpoints adapter-ready, democratizing scalable inference for all models.

Given an input query, generative models such as large language models produce a random response drawn from a response distribution. Given two input queries, it is natural to ask if their response distributions are the same. While traditional statistical hypothesis testing is designed to address this question, the response distribution induced by an input query is often sensitive to semantically irrelevant perturbations to the query, so much so that a traditional test of equality might indicate that two semantically equivalent queries induce statistically different response distributions. As a result, the outcome of the statistical test may not align with the user's requirements. In this paper, we address this misalignment by incorporating into the testing procedure consideration of a collection of semantically similar queries. In our setting, the mapping from the collection of user-defined semantically similar queries to the corresponding collection of response distributions is not known a priori and must be estimated, with a fixed budget. Although the problem we address is quite general, we focus our analysis on the setting where the responses are binary, show that the proposed test is asymptotically valid and consistent, and discuss important practical considerations with respect to power and computation.

Abstract--Large Language Models (LLMs) are increasingly used to automate software generation in embedded machine learning workflows, yet their outputs often fail silently or behave unpredictably. This article presents an empirical investigation of failure modes in LLM-powered ML pipelines, based on an autopilot framework that orchestrates data preprocessing, model conversion, and on-device inference code generation. We show how prompt format, model behavior, and structural assumptions influence both success rates and failure characteristics, often in ways that standard validation pipelines fail to detect. Our analysis reveals a diverse set of error-prone behaviors, including format-induced misinterpretations and runtime-disruptive code that compiles but breaks downstream. We derive a taxonomy of failure categories and analyze errors across multiple LLMs, highlighting common root causes and systemic fragilities. Though grounded in specific devices, our study reveals broader challenges in LLMbased code generation. We conclude by discussing directions for improving reliability and traceability in LLM-powered embedded ML systems. Index T erms--Large Language Models, Embedded Machine Learning, Code Generation, AI Failure Modes, Prompt Engineering.

The proliferation of clickbait headlines poses significant challenges to the credibility of information and user trust in digital media. While recent advances in machine learning have improved the detection of manipulative content, the lack of explainability limits their practical adoption. This paper presents a model for explainable clickbait detection that not only identifies clickbait titles but also attributes them to specific linguistic manipulation strategies. We introduce a synthetic dataset generated by systematically augmenting real news headlines using a predefined catalogue of clickbait strategies. This dataset enables controlled experimentation and detailed analysis of model behaviour. We present a two - stage framework for automatic clickbait analysis comprising detection and tactic attribution. In the first stage, we compare a fine - tuned BERT classifier with large language models (LLMs), specifically GPT - 4.0 and Gemini 2.4 Flash, under both zero - shot prompting and few - shot prompting enriched with illustrative clickbait headlines and their associated persuasive tactics. In the second stage, a dedicated BERT - based classifier predicts the specific clickbait strategies present in each headline. We share the dataset with the research community at https://github.com/LLM - HITCS25S/ClickbaitTacticsDetection The widespread use of clickbait headlines in digital media has become a pervasive challenge, undermining the credibility of information and exploiting user attention through manipulative linguistic techniques. While automated systems for detecting clickbait have improved in recent years, their focus has remained mainly on binary classification, simply labelling content as clickbait or not. However, effective mitigation of such content requires going beyond detection to understanding how and why certain headlines manipulate readers. Specifically, it is crucial to evaluate whether current AI models can accurately recognize and distinguish the diverse linguistic styles and persuasive strategies commonly employed in clickbait.

In-context learning (ICL) in Large Language Models (LLMs) has shown remarkable performance across various tasks without requiring fine-tuning. However, recent studies have highlighted the risk of private data leakage through the prompt in ICL, especially when LLMs are exposed to malicious attacks. While differential privacy (DP) provides strong privacy guarantees, it often significantly reduces the utility of in-context learning (ICL). To address this challenge, we incorporate task-related public data into the ICL framework while maintaining the DP guarantee. Based on this approach, we propose a private in-context learning algorithm that effectively balances privacy protection and model utility. Through experiments, we demonstrate that our approach significantly improves the utility of private ICL with the assistance of public data. Additionally, we show that our method is robust against membership inference attacks, demonstrating empirical privacy protection.

Large Language Models (LLMs) have demonstrated remarkable capabilities across diverse tasks, but their potential misuse for harmful purposes remains a significant concern. To strengthen defenses against such vulnerabilities, it is essential to investigate universal jailbreak attacks that exploit intrinsic weaknesses in the architecture and learning paradigms of LLMs. In response, we propose \textbf{H}armful \textbf{P}rompt \textbf{La}undering (HaPLa), a novel and broadly applicable jailbreaking technique that requires only black-box access to target models. HaPLa incorporates two primary strategies: 1) \textit{abductive framing}, which instructs LLMs to infer plausible intermediate steps toward harmful activities, rather than directly responding to explicit harmful queries; and 2) \textit{symbolic encoding}, a lightweight and flexible approach designed to obfuscate harmful content, given that current LLMs remain sensitive primarily to explicit harmful keywords. Experimental results show that HaPLa achieves over 95% attack success rate on GPT-series models and 70% across all targets. Further analysis with diverse symbolic encoding rules also reveals a fundamental challenge: it remains difficult to safely tune LLMs without significantly diminishing their helpfulness in responding to benign queries.

Cultural competence, defined as the ability to understand and adapt to multicultural contexts, is increasingly vital for large language models (LLMs) in global environments. While several cultural benchmarks exist to assess LLMs' cultural competence, current evaluations suffer from fragmented taxonomies, domain specificity, and heavy reliance on manual data annotation. To address these limitations, we introduce CultureSynth, a novel framework comprising (1) a comprehensive hierarchical multilingual cultural taxonomy covering 12 primary and 130 secondary topics, and (2) a Retrieval-Augmented Generation (RAG)-based methodology leveraging factual knowledge to synthesize culturally relevant question-answer pairs. The CultureSynth-7 synthetic benchmark contains 19,360 entries and 4,149 manually verified entries across 7 languages. Evaluation of 14 prevalent LLMs of different sizes reveals clear performance stratification led by ChatGPT-4o-Latest and Qwen2.5-72B-Instruct. The results demonstrate that a 3B-parameter threshold is necessary for achieving basic cultural competence, models display varying architectural biases in knowledge processing, and significant geographic disparities exist across models. We believe that CultureSynth offers a scalable framework for developing culturally aware AI systems while reducing reliance on manual annotation\footnote{Benchmark is available at https://github.com/Eyr3/CultureSynth.}.

"The division of labor between System 1 (fast) and System 2 (slow) is highly efficient: it minimizes effort and optimizes performance. Nav-R1 is an embodied foundation model that integrates dialogue, reasoning, planning, and navigation capabilities to enable intelligent interaction and task execution in 3D environments. Abstract-- Embodied navigation requires agents to integrate perception, reasoning, and action for robust interaction in complex 3D environments. Existing approaches often suffer from incoherent and unstable reasoning traces that hinder generalization across diverse environments, and difficulty balancing long-horizon semantic reasoning with low-latency control for real-time navigation. T o address these challenges, we propose Nav-R1, an embodied foundation model that unifies reasoning in embodied environments. We first construct Nav-CoT -110K, a large-scale dataset of step-by-step Chains-of-Thought (CoT) for embodied tasks, which enables cold-start initialization with structured reasoning. Building on this foundation, we design a GRPO-based reinforcement learning framework with three complementary rewards: format, understanding, and navigation, to improve structural adherence, semantic grounding, and path fidelity. Furthermore, we introduce a Fast-in-Slow reasoning paradigm, decoupling deliberate semantic reasoning from low-latency reactive control for efficient yet coherent navigation.