arXiv subjects

Renjie Sun

Publications and source records attributed to Renjie Sun.

5 recordsLinked to original sources

arXiv2026

Forecasting Trajectory-Level Safety Risks in Black-Box Multi-Turn Interactions

As large language models (LLMs) evolve from standalone assistants into autonomous agents, ensuring their safety requires shifting beyond pointwise risk assessment to understand how risks emerge and unfold over long-horizon trajectories. In multi-turn interactions, malicious intent can be decomposed across seemingly harmless turns and gradually reconstructed through interaction trajectories, eventually resulting in safety failures. Existing safeguards remain largely reactive, detecting manifested violations while lacking the ability to predict latent risk evolution and enable preemptive prevention. To address this limitation, we propose Recast, a safety risk forecasting framework that advances LLM safeguarding beyond turn-level violation detection to trajectory-level risk prediction. Recast first retrieves risk-relevant evidence from both short-term dialogue progression and long-term historical context via a dual-scale trajectory view. It then models compositional risk evolution by capturing the current risk configuration and its temporal dynamics. Finally, a causal temporal encoder learns latent risk evolution patterns and predicts the distribution of future risk emergence turns. Extensive experiments across 7 risk categories show that Recast predicts 88.3% of future safety failures with an average lead time of 2.41 turns, while maintaining a false alarm rate of 12.3%, showcasing the effectiveness of trajectory-level forecasting in identifying emerging risks before safety violations occur.

cs.LG↗

arXiv2026

OpenHospital: A Thing-in-itself Arena for Evolving and Benchmarking LLM-based Collective Intelligence

Large Language Model (LLM)-based Collective Intelligence (CI) presents a promising approach to overcoming the data wall and continuously boosting the capabilities of LLM agents. However, there is currently no dedicated arena for evolving and benchmarking LLM-based CI. To address this gap, we introduce OpenHospital, an interactive arena where physician agents can evolve CI through interactions with patient agents. This arena employs a data-in-agent-self paradigm that rapidly enhances agent capabilities and provides robust evaluation metrics for benchmarking both medical proficiency and system efficiency. Experiments demonstrate the effectiveness of OpenHospital in both fostering and quantifying CI.

cs.AI↗

arXiv2025

Enhance Stability of Network by Edge Anchor

With the rapid growth of online social networks, strengthening their stability has emerged as a key research focus. This study aims to identify influential relationships that significantly impact community stability. In this paper, we introduce and explore the anchor trussness reinforcement problem to reinforce the overall user engagement of networks by anchoring some edges. Specifically, for a given graph $G$ and a budget $b$, we aim to identify $b$ edges whose anchoring maximizes the trussness gain, which is the cumulative increment of trussness across all edges in $G$. We establish the NP-hardness of the problem. To address this problem, we introduce a greedy framework that iteratively selects the current best edge. To scale for larger networks, we first propose an upward-route method to constrain potential trussness increment edges. Augmented with a support check strategy, this approach enables the efficient computation of the trussness gain for anchoring one edge. Then, we design a classification tree structure to minimize redundant computations in each iteration by organizing edges based on their trussness. We conduct extensive experiments on 8 real-world networks to validate the efficiency and effectiveness of the proposed model and methods.

cs.SI↗

arXiv2024

Reasoning as a Weapon: Adaptive Dual-Path Jailbreak Attack on Large Language Models

Large language models (LLMs) have achieved remarkable progress across a wide range of applications. Despite advances in safety alignment, they remain vulnerable to jailbreak attacks. Existing methods mainly rely on input-level deception to conceal harmful intent, overlooking a more fundamental attack surface: harmful intent can emerge during reasoning even when absent from the input. To study this gap, we propose ReDPJ, a Reasoning-guided Dual-Path Jailbreak framework that shifts attacks from prompt manipulation to reasoning manipulation. ReDPJ first converts harmful intent into benign dual-modal adversarial anchors, then adaptively guides reasoning to reconstruct latent harmful intent and steer the model toward unsafe trajectories. This design reduces explicit malicious signals in the input while exploiting vulnerabilities in the model's reasoning process. Experiments on state-of-the-art LLMs demonstrate strong effectiveness, achieving ASRs of 82.1% on GPT-4o-2024-11-20, 85.1% on Llama-3-8B-Instruct, and 88.2% on DeepSeek-V3.

cs.CR↗

arXiv2024

Efficient Maximal Frequent Group Enumeration in Temporal Bipartite Graphs

Cohesive subgraph mining is a fundamental problem in bipartite graph analysis. In reality, relationships between two types of entities often occur at some specific timestamps, which can be modeled as a temporal bipartite graph. However, the temporal information is widely neglected by previous studies. Moreover, directly extending the existing models may fail to find some critical groups in temporal bipartite graphs, which appear in a unilateral (i.e., one-layer) form. To fill the gap, in this paper, we propose a novel model, called maximal λ-frequency group (MFG). Given a temporal bipartite graph G=(U,V,E), a vertex set V_S \subseteq V is an MFG if i) there are no less than λtimestamps, at each of which V_S can form a (t_U,t_V)-biclique with some vertices in U at the corresponding snapshot, and ii) it is maximal. To solve the problem, a filter-and-verification (FilterV) method is proposed based on the Bron-Kerbosch framework, incorporating novel filtering techniques to reduce the search space and array-based strategy to accelerate the frequency and maximality verification. Nevertheless, the cost of frequency verification in each valid candidate set computation and maximality check could limit the scalability of FilterV to larger graphs. Therefore, we further develop a novel verification-free (VFree) approach by leveraging the advanced dynamic counting structure proposed. Theoretically, we prove that VFree can reduce the cost of each valid candidate set computation in FilterV by a factor of O(|V|). Furthermore, VFree can avoid the explicit maximality verification because of the developed search paradigm. Finally, comprehensive experiments on 15 real-world graphs are conducted to demonstrate the efficiency and effectiveness of the proposed techniques and model.

cs.DB↗