We investigate strategic deception in large language models using two complementary testbeds: Secret Agenda (across 38 models) and Insider Trading compliance (via SAE architectures). Secret Agenda reliably induced lying when deception advantaged goal achievement across all model families. Analysis revealed that autolabeled SAE features for "deception" rarely activated during strategic dishonesty, and feature steering experiments across 100+ deception-related features failed to prevent lying. Conversely, insider trading analysis using unlabeled SAE activations separated deceptive versus compliant responses through discriminative patterns in heatmaps and t-SNE visualizations. These findings suggest autolabel-driven interpretability approaches fail to detect or control behavioral deception, while aggregate unlabeled activations provide population-level structure for risk assessment. Results span Llama 8B/70B SAE implementations and GemmaScope under resource constraints, representing preliminary findings that motivate larger studies on feature discovery, labeling methodology, and causal interventions in realistic deception contexts.

The rapid integration of AI-powered coding assistants into developer workflows has raised significant privacy and trust concerns. As developers entrust proprietary code to services like OpenAI's GPT, Google's Gemini, and GitHub Copilot, the unclear data handling practices of these tools create security and compliance risks. This paper addresses this challenge by introducing and applying a novel, expert-validated privacy scorecard. The methodology involves a detailed analysis of four document types; from legal policies to external audits; to score five leading assistants against 14 weighted criteria. A legal expert and a data protection officer refined these criteria and their weighting. The results reveal a distinct hierarchy of privacy protections, with a 20-point gap between the highest- and lowest-ranked tools. The analysis uncovers common industry weaknesses, including the pervasive use of opt-out consent for model training and a near-universal failure to filter secrets from user prompts proactively. The resulting scorecard provides actionable guidance for developers and organizations, enabling evidence-based tool selection. This work establishes a new benchmark for transparency and advocates for a shift towards more user-centric privacy standards in the AI industry.

Fuzzing is effective for vulnerability discovery but struggles with complex targets such as compilers, interpreters, and database engines, which accept textual input that must satisfy intricate syntactic and semantic constraints. Although language models (LMs) have attracted interest for this task due to their vast latent knowledge and reasoning potential, their practical adoption has been limited. The major challenges stem from insufficient exploration of deep program logic among real-world codebases, and the high cost of leveraging larger models. To overcome these challenges, we propose R1-Fuzz, the first framework that leverages reinforcement learning (RL) to specialize cost-efficient LMs and integrate them for complex textual fuzzing input generation. R1-Fuzz introduces two key designs: coverage-slicing-based question construction and a distance-based reward calculation. Through RL-based post-training of a model with our constructed dataset, R1-Fuzz designs a fuzzing workflow that tightly integrates LMs to reason deep program semantics during fuzzing. Evaluations on diverse real-world targets show that our design enables a small model, named R1-Fuzz-7B, to rival or even outperform much larger models in real-world fuzzing. Notably, R1-Fuzz achieves up to 75\% higher coverage than state-of-the-art fuzzers and discovers 29 previously unknown vulnerabilities, demonstrating its practicality.

Retrieval-Augmented Generation (RAG) has significantly improved the performance of large language models (LLMs) on knowledge-intensive tasks in recent years. However, since retrieval systems may return irrelevant content, incorporating such information into the model often leads to hallucinations. Thus, identifying and filtering out unhelpful retrieved content is a key challenge for improving RAG performance.To better integrate the internal knowledge of the model with external knowledge from retrieval, it is essential to understand what the model "knows" and "does not know" (which is also called "self-knowledge"). Based on this insight, we propose SKILL-RAG (Self-Knowledge Induced Learning and Filtering for RAG), a novel method that leverages the model's self-knowledge to determine which retrieved documents are beneficial for answering a given query. We design a reinforcement learning-based training framework to explicitly elicit self-knowledge from the model and employs sentence-level granularity to filter out irrelevant content while preserving useful knowledge.We evaluate SKILL-RAG using Llama2-7B and Qwen3-8B on several question answering benchmarks. Experimental results demonstrate that SKILL-RAG not only improves generation quality but also significantly reduces the number of input documents, validating the importance of self-knowledge in guiding the selection of high-quality retrievals.

Large language models (LLMs) have achieved remarkable performance across a wide range of natural language tasks. Understanding how LLMs internally represent knowledge remains a significant challenge. Despite Sparse Autoencoders (SAEs) have emerged as a promising technique for extracting interpretable features from LLMs, SAE features do not inherently align with human-understandable concepts, making their interpretation cumbersome and labor-intensive. To bridge the gap between SAE features and human concepts, we present ConceptViz, a visual analytics system designed for exploring concepts in LLMs. ConceptViz implements a novel dentification => Interpretation => Validation pipeline, enabling users to query SAEs using concepts of interest, interactively explore concept-to-feature alignments, and validate the correspondences through model behavior verification. We demonstrate the effectiveness of ConceptViz through two usage scenarios and a user study. Our results show that ConceptViz enhances interpretability research by streamlining the discovery and validation of meaningful concept representations in LLMs, ultimately aiding researchers in building more accurate mental models of LLM features. Our code and user guide are publicly available at https://github.com/Happy-Hippo209/ConceptViz.

The rapid adoption of large language models (LLMs) such as ChatGPT has blurred the line between human and AI-generated texts, raising urgent questions about academic integrity, intellectual property, and the spread of misinformation. Thus, reliable AI-text detection is needed for fair assessment to safeguard human authenticity and cultivate trust in digital communication. In this study, we investigate how well current machine learning (ML) approaches can distinguish ChatGPT-3.5-generated texts from human-written texts employing a labeled data set of 250 pairs of abstracts from a wide range of research topics. We test and compare both classical (Logistic Regression armed with classical Bag-of-Words, POS, and TF-IDF features) and transformer-based (BERT augmented with N-grams, DistilBERT, BERT with a lightweight custom classifier, and LSTM-based N-gram models) ML detection techniques. As we aim to assess each model's performance in detecting AI-generated research texts, we also aim to test whether an ensemble of these models can outperform any single detector. Results show DistilBERT achieves the overall best performance, while Logistic Regression and BERT-Custom offer solid, balanced alternatives; LSTM- and BERT-N-gram approaches lag. The max voting ensemble of the three best models fails to surpass DistilBERT itself, highlighting the primacy of a single transformer-based representation over mere model diversity. By comprehensively assessing the strengths and weaknesses of these AI-text detection approaches, this work lays a foundation for more robust transformer frameworks with larger, richer datasets to keep pace with ever-improving generative AI models.

This paper presents VITA (Virtual Teaching Assistants), an adaptive distributed learning (ADL) platform that embeds a large language model (LLM)-powered chatbot (BotCaptain) to provide dialogic support, interoperable analytics, and integrity-aware assessment for workforce preparation in data science. The platform couples context-aware conversational tutoring with formative-assessment patterns designed to promote reflective reasoning. The paper describes an end-to-end data pipeline that transforms chat logs into Experience API (xAPI) statements, instructor dashboards that surface outliers for just-in-time intervention, and an adaptive pathway engine that routes learners among progression, reinforcement, and remediation content. The paper also benchmarks VITA conceptually against emerging tutoring architectures, including retrieval-augmented generation (RAG)--based assistants and Learning Tools Interoperability (LTI)--integrated hubs, highlighting trade-offs among content grounding, interoperability, and deployment complexity. Contributions include a reusable architecture for interoperable conversational analytics, a catalog of patterns for integrity-preserving formative assessment, and a practical blueprint for integrating adaptive pathways into data-science courses. The paper concludes with implementation lessons and a roadmap (RAG integration, hallucination mitigation, and LTI~1.3 / OpenID Connect) to guide multi-course evaluations and broader adoption. In light of growing demand and scalability constraints in traditional instruction, the approach illustrates how conversational AI can support engagement, timely feedback, and personalized learning at scale. Future work will refine the platform's adaptive intelligence and examine applicability across varied educational settings.

This paper presents a temporal expression language for monitoring AI agent behavior, enabling systematic error-detection of LLM-based agentic systems that exhibit variable outputs due to stochastic generation processes. Drawing from temporal logic techniques used in hardware verification, this approach monitors execution traces of agent tool calls and state transitions to detect deviations from expected behavioral patterns. Current error-detection approaches rely primarily on text matching of inputs and outputs, which proves fragile due to the natural language variability inherent in LLM responses. The proposed method instead focuses on the sequence of agent actions -- such as tool invocations and inter-agent communications -- allowing verification of system behavior independent of specific textual outputs. The temporal expression language provides assertions that capture correct behavioral patterns across multiple execution scenarios. These assertions serve dual purposes: validating prompt engineering and guardrail effectiveness during development, and providing regression testing when agents are updated with new LLMs or modified logic. The approach is demonstrated using a three-agent system, where agents coordinate to solve multi-step reasoning tasks. When powered by large, capable models, all temporal assertions were satisfied across many test runs. However, when smaller models were substituted in two of the three agents, executions violated behavioral assertions, primarily due to improper tool sequencing and failed coordination handoffs. The temporal expressions successfully flagged these anomalies, demonstrating the method's effectiveness for detecting behavioral regressions in production agentic systems. This approach provides a foundation for systematic monitoring of AI agent reliability as these systems become increasingly deployed in critical applications.

Implicit Chain-of-Thought (CoT) methods offer a token-efficient alternative to explicit CoT reasoning in Large Language Models (LLMs), but a persistent performance gap has limited their adoption. We identify a core latent instability issue when scaling the computational budget of implicit CoT: as the number of reasoning tokens increases, training often becomes unstable and collapses. Our analysis shows that this instability arises from latent representations becoming homogeneous and losing semantic diversity, caused by insufficient step-level supervision in current implicit CoT methods. To address this, we propose SIM-CoT, a plug-and-play training module that introduces step-level supervision to stabilize and enrich the latent reasoning space. SIM-CoT employs an auxiliary decoder during training to align each implicit token with its corresponding explicit reasoning step, ensuring latent states capture distinct and meaningful information. The auxiliary decoder is removed at inference, preserving the efficiency of implicit CoT with no added overhead. It also provides interpretability by projecting each latent token onto an explicit reasoning vocabulary, enabling per-step visualization and diagnosis. SIM-CoT significantly improves both in-domain accuracy and out-of-domain stability of implicit CoT methods, boosting Coconut by +8.2\% on GPT-2 and CODI by +3.0\% on LLaMA-3.1 8B. It further surpasses the explicit CoT baseline on GPT-2 by 2.1\% with 2.3$\times$ greater token efficiency, while closing the performance gap on larger models like LLaMA-3.1 8B. Code: https://github.com/InternLM/SIM-CoT

This paper examines the integration of emotional intelligence into artificial intelligence systems, with a focus on affective computing and the growing capabilities of Large Language Models (LLMs), such as ChatGPT and Claude, to recognize and respond to human emotions. Drawing on interdisciplinary research that combines computer science, psychology, and neuroscience, the study analyzes foundational neural architectures - CNNs for processing facial expressions and RNNs for sequential data, such as speech and text - that enable emotion recognition. It examines the transformation of human emotional experiences into structured emotional data, addressing the distinction between explicit emotional data collected with informed consent in research settings and implicit data gathered passively through everyday digital interactions. That raises critical concerns about lawful processing, AI transparency, and individual autonomy over emotional expressions in digital environments. The paper explores implications across various domains, including healthcare, education, and customer service, while addressing challenges of cultural variations in emotional expression and potential biases in emotion recognition systems across different demographic groups. From a regulatory perspective, the paper examines emotional data in the context of the GDPR and the EU AI Act frameworks, highlighting how emotional data may be considered sensitive personal data that requires robust safeguards, including purpose limitation, data minimization, and meaningful consent mechanisms.