Safety alignment instills in Large Language Models (LLMs) a critical capacity to refuse malicious requests. Prior works have modeled this refusal mechanism as a single linear direction in the activation space. We posit that this is an oversimplification that conflates two functionally distinct neural processes: the detection of harm and the execution of a refusal. In this work, we deconstruct this single representation into a Harm Detection Direction and a Refusal Execution Direction. Leveraging this fine-grained model, we introduce Differentiated Bi-Directional Intervention (DBDI), a new white-box framework that precisely neutralizes the safety alignment at critical layer. DBDI applies adaptive projection nullification to the refusal execution direction while suppressing the harm detection direction via direct steering. Extensive experiments demonstrate that DBDI outperforms prominent jailbreaking methods, achieving up to a 97.88\% attack success rate on models such as Llama-2. By providing a more granular and mechanistic framework, our work offers a new direction for the in-depth understanding of LLM safety alignment.

Low-dose chest computed tomography (LDCT) inherently captures both pulmonary and cardiac structures, offering a unique opportunity for joint assessment of lung and cardiovascular health. However, most existing approaches treat these domains as independent tasks, overlooking their physiological interplay and shared imaging biomarkers. We propose an Explainable Cross-Disease Reasoning Framework that enables interpretable cardiopulmonary risk assessment from a single LDCT scan. The framework introduces an agentic reasoning process that emulates clinical diagnostic thinking-first perceiving pulmonary findings, then reasoning through established medical knowledge, and finally deriving a cardiovascular judgment with explanatory rationale. It integrates three synergistic components: a pulmonary perception module that summarizes lung abnormalities, a knowledge-guided reasoning module that infers their cardiovascular implications, and a cardiac representation module that encodes structural biomarkers. Their outputs are fused to produce a holistic cardiovascular risk prediction that is both accurate and physiologically grounded. Experiments on the NLST cohort demonstrate that the proposed framework achieves state-of-the-art performance for CVD screening and mortality prediction, outperforming single-disease and purely image-based baselines. Beyond quantitative gains, the framework provides human-verifiable reasoning that aligns with cardiological understanding, revealing coherent links between pulmonary abnormalities and cardiac stress mechanisms. Overall, this work establishes a unified and explainable paradigm for cardiovascular analysis from LDCT, bridging the gap between image-based prediction and mechanism-based medical interpretation.

On-device AI agents offer the potential for personalized, low-latency assistance, but their deployment is fundamentally constrained by limited memory capacity, which restricts usable context. This reduced practical context window creates a trade-off between supporting rich, stateful interactions with complex tool capabilities and maintaining on-device feasibility. We break this trade-off with a framework for context-efficient on-device agents, driven by three synergistic optimizations (1) a dynamic memory system using specialized LoRA adapters to distill conversational history into a compressed, and structured Context State Object; (2) a minimalist serialization format for tool schemas to minimize token overhead per tool; and (3) a just-in-time schema-passing mechanism that loads full tool definitions only upon tool selection. We instantiate this framework by adapting a 3B parameter SLM to context-efficient trajectories and rigorously evaluate it against a conventional baseline on complex user tasks. Our agent matches, or exceeds, the performance of a conventional baseline while dramatically compressing context, achieving more than a 6-fold reduction in initial system prompt context and a 10- to 25-fold reduction in context growth rate based on the interaction verbosity, demonstrating that strategic context management is key to unlocking capable and persistent on-device AI.

Recovering meaningful concepts from language model activations is a central aim of interpretability. While existing feature extraction methods aim to identify concepts that are independent directions, it is unclear if this assumption can capture the rich temporal structure of language. Specifically, via a Bayesian lens, we demonstrate that Sparse Autoencoders (SAEs) impose priors that assume independence of concepts across time, implying stationarity. Meanwhile, language model representations exhibit rich temporal dynamics, including systematic growth in conceptual dimensionality, context-dependent correlations, and pronounced non-stationarity, in direct conflict with the priors of SAEs. Taking inspiration from computational neuroscience, we introduce a new interpretability objective -- Temporal Feature Analysis -- which possesses a temporal inductive bias to decompose representations at a given time into two parts: a predictable component, which can be inferred from the context, and a residual component, which captures novel information unexplained by the context. Temporal Feature Analyzers correctly parse garden path sentences, identify event boundaries, and more broadly delineate abstract, slow-moving information from novel, fast-moving information, while existing SAEs show significant pitfalls in all the above tasks. Overall, our results underscore the need for inductive biases that match the data in designing robust interpretability tools.

While crossover is a critical and often indispensable component in other forms of Genetic Programming, such as Linear- and Tree-based, it has consistently been claimed that it deteriorates search performance in CGP. As a result, a mutation-alone $(1+λ)$ evolutionary strategy has become the canonical approach for CGP. Although several operators have been developed that demonstrate an increased performance over the canonical method, a general solution to the problem is still lacking. In this paper, we compare basic crossover methods, namely one-point and uniform, to variants in which nodes are ``preserved,'' including the subgraph crossover developed by Roman Kalkreuth, the difference being that when ``node preservation'' is active, crossover is not allowed to break apart instructions. We also compare a node mutation operator to the traditional point mutation; the former simply replaces an entire node with a new one. We find that node preservation in both mutation and crossover improves search using symbolic regression benchmark problems, moving the field towards a general solution to CGP crossover.

Millions of users now design personalized LLM-based chatbots that shape their daily interactions, yet they can only roughly anticipate how their design choices will manifest as behaviors in deployment. This opacity is consequential: seemingly innocuous prompts can trigger excessive sycophancy, toxicity, or other undesirable traits, degrading utility and raising safety concerns. To address this issue, we introduce an interface that enables neural transparency by exposing language model internals during chatbot design. Our approach extracts behavioral trait vectors (empathy, toxicity, sycophancy, etc.) by computing differences in neural activations between contrastive system prompts that elicit opposing behaviors. We predict chatbot behaviors by projecting the system prompt's final token activations onto these trait vectors, normalizing for cross-trait comparability, and visualizing results via an interactive sunburst diagram. To evaluate this approach, we conducted an online user study using Prolific to compare our neural transparency interface against a baseline chatbot interface without any form of transparency. Our analyses suggest that users systematically miscalibrated AI behavior: participants misjudged trait activations for eleven of fifteen analyzable traits, motivating the need for transparency tools in everyday human-AI interaction. While our interface did not change design iteration patterns, it significantly increased user trust and was enthusiastically received. Qualitative analysis revealed nuanced user experiences with the visualization, suggesting interface and interaction improvements for future work. This work offers a path for how mechanistic interpretability can be operationalized for non-technical users, establishing a foundation for safer, more aligned human-AI interactions.

Higher-Order Hypergraph Learning (HOHL) was recently introduced as a principled alternative to classical hypergraph regularization, enforcing higher-order smoothness via powers of multiscale Laplacians induced by the hypergraph structure. Prior work established the well- and ill-posedness of HOHL through an asymptotic consistency analysis in geometric settings. We extend this theoretical foundation by proving the consistency of a truncated version of HOHL and deriving explicit convergence rates when HOHL is used as a regularizer in fully supervised learning. We further demonstrate its strong empirical performance in active learning and in datasets lacking an underlying geometric structure, highlighting HOHL's versatility and robustness across diverse learning settings.

Graph condensation (GC) has gained significant attention for its ability to synthesize smaller yet informative graphs. However, existing studies often overlook the robustness of GC in scenarios where the original graph is corrupted. In such cases, we observe that the performance of GC deteriorates significantly, while existing robust graph learning technologies offer only limited effectiveness. Through both empirical investigation and theoretical analysis, we reveal that GC is inherently an intrinsic-dimension-reducing process, synthesizing a condensed graph with lower classification complexity. Although this property is critical for effective GC performance, it remains highly vulnerable to adversarial perturbations. To tackle this vulnerability and improve GC robustness, we adopt the geometry perspective of graph data manifold and propose a novel Manifold-constrained Robust Graph Condensation framework named MRGC. Specifically, we introduce three graph data manifold learning modules that guide the condensed graph to lie within a smooth, low-dimensional manifold with minimal class ambiguity, thereby preserving the classification complexity reduction capability of GC and ensuring robust performance under universal adversarial attacks. Extensive experiments demonstrate the robustness of \ModelName\ across diverse attack scenarios.

Hypergraphs provide a natural framework for modeling higher-order interactions, yet their theoretical underpinnings in semi-supervised learning remain limited. We provide an asymptotic consistency analysis of variational learning on random geometric hypergraphs, precisely characterizing the conditions ensuring the well-posedness of hypergraph learning as well as showing convergence to a weighted $p$-Laplacian equation. Motivated by this, we propose Higher-Order Hypergraph Learning (HOHL), which regularizes via powers of Laplacians from skeleton graphs for multiscale smoothness. HOHL converges to a higher-order Sobolev seminorm. Empirically, it performs strongly on standard baselines.

The rapid progress of graph generation has raised new security concerns, particularly regarding backdoor vulnerabilities. While prior work has explored backdoor attacks in image diffusion and unconditional graph generation, conditional, especially text-guided graph generation remains largely unexamined. This paper proposes BadGraph, a backdoor attack method against latent diffusion models for text-guided graph generation. BadGraph leverages textual triggers to poison training data, covertly implanting backdoors that induce attacker-specified subgraphs during inference when triggers appear, while preserving normal performance on clean inputs. Extensive experiments on four benchmark datasets (PubChem, ChEBI-20, PCDes, MoMu) demonstrate the effectiveness and stealth of the attack: less than 10% poisoning rate can achieves 50% attack success rate, while 24% suffices for over 80% success rate, with negligible performance degradation on benign samples. Ablation studies further reveal that the backdoor is implanted during VAE and diffusion training rather than pretraining. These findings reveal the security vulnerabilities in latent diffusion models of text-guided graph generation, highlight the serious risks in models' applications such as drug discovery and underscore the need for robust defenses against the backdoor attack in such diffusion models.