Large language model (LLM) safety has received extensive attention since LLMs are vulnerable to manipulation. To understand and mitigate the risk, this paper studies the problem of jailbreak attacks on LLMs, which aim to deliberately break the safety guard of LLMs for harmful or unethical responses. Current black-box jailbreak attacks are limited by either a reliance on human expertise for manual prompt design or by intricate workflows in automated approaches. Towards this end, we propose a novel approach named Reward-guided Meta-prompt Evolving with reflection (ROOT) for automatic jailbreak attack generation. The core idea of our ROOT is to optimize a meta-prompt using attack rewards as jailbreak guidance. In particular, our ROOT feeds a meta-prompt with toxic questions into LLMs to generate prompts for jailbreak attempts. The responses from these attempts are further evaluated using a judge model, which further summarizes the reflection of both successful and unsuccessful jailbreak attempts into different meta-prompt optimization strategies. To reduce the noise, we estimate their reward scores to select high-quality strategies, which can optimize our meta-prompt for better attack generation ability. Extensive experiments show that ROOT demonstrates strong generalizability and broad adaptability by achieving high jailbreak success rates (above 90%) across both multiple LLMs and various categories of harmful tasks.

Retrieval-augmented generation (RAG) enables LLMs to access external knowledge for answering knowledge-intensive questions. For complex multi-hop questions, multi-turn retrieval-augmented reasoning extends RAG into an iterative process that repeatedly searches for and integrates evidence across documents. However, existing reinforcement-learning (RL) approaches for agentic RAG are typically optimized with final-answer rewards, which provide sparse supervision and overlook whether the model actually retrieves the required evidence chain. We present GTA-RAG, a graph-trajectory-augmented RL framework for multi-turn retrieval-augmented reasoning. From an entity–document graph, we sample connected document paths, synthesize multi-hop QA trajectories, and validate them with the deployed retriever to obtain executable trajectory-level supervision. We then optimize the retrieval policy with Group Relative Policy Optimization (GRPO) and a trajectory-guided reward that encourages both accurate answers and acquisition of target evidence documents, followed by answer-reward training on natural QA instances. Experiments on three multi-hop and two simple QA benchmarks show that GTA-RAG consistently outperforms RL-based RAG baselines with both Qwen2.5-3B and Qwen2.5-7B backbones, while substantially improving evidence-chain coverage. Our code is available at https://github.com/cjcj46262/GTA-RAG

Coordinating heterogeneous large language models (LLMs) can improve instruction following, yet current multi-agent systems often depend on manual role choices or repeated calls to one backbone. This makes it difficult to separate complementary model behavior from model bias and verbosity. We introduce Role-Specialized Multi-Agent Collaboration (RS-MA), a prompting-only framework that assigns candidate LLMs to four interpretable roles: Judge, Summarizer, Ideator, and Reasoner. RS-MA estimates role fit with audition prompts, aggregates role-specific rankings with Borda scoring, and uses the selected assignment as an inspectable role map. A refinement phase records judgment gaps and targeted edits before synthesis. On AlpacaEval 2.0, audition-selected RS-MA reaches an 88.86% length-controlled win rate, and a validated role map reaches 90.72%. A full 805-example ablation shows that single-model reuse attains high raw preference through much longer outputs, whereas the validated map improves length-controlled quality with moderate length growth. MT-Bench and FLASK further indicate preserved dialogue quality and balanced fine-grained skills. The results support explicit role assignment as a practical mechanism for combining heterogeneous LLMs when quality, completeness, and inspectable evidence are jointly required.

This paper studies the problem of uncertainty-aware reasoning over knowledge graphs (KGs), which aims to incorporate uncertainty quantification into unseen knowledge predictions over KGs. Previous works typically incorporate uncertainty measurement tools such as conformal predictions into graph representation learning. Despite the progress, the performance of these approaches remains limited since they usually neglect both structural information in KGs and intrinsic asymmetry between over-estimation and under-estimation. Towards this end, we propose a novel approach named Context-anchored Asymmetric Conformal Prediction for uncertainty-aware REasoning over KGs (CARE). The core of our CARE is to identify the context of each query using anchors, which guide the construction of triple-specific asymmetry prediction intervals. In particular, we first take the subset of the calibration set as a reference pool. For each triple query, we summarize the KG structural statistics into a vector, which guides the identification of its anchor samples from the pool. The confidence residuals between query and anchors are utilized to quantify over-estimation and under-estimation, producing contextual nonconformity scores. Finally, we utilize the unselected calibration dataset to determine cutoff levels for constructing asymmetric prediction intervals. Extensive experiments on benchmark datasets validate the effectiveness of CARE. The code can be found at https://github.com/fanboyang/CARE.

This paper studies the problem of test-time adaptation for vision-language models (VLMs). Recent approaches typically measure the prediction entropy to store a confident cache for logit refinement. However, these confident samples tend to approach prototypes with limited coverage of data distribution, which could result in biased predictions as the distribution evolves. Towards this end, we propose a novel approach named Diversity-attended Dynamic Caching with Asymmetric Quantization (DANCE) for test-time adaptation of VLMs. The core of our DANCE is to maintain a dynamic cache to store diversity-aware test samples, which support efficient logit adjustment via asymmetric quantization. In particular, we first generate multiple augmented views of each sample and aggregate their outputs from pre-trained VLMs via a consistency-aware mechanism. More importantly, we construct a dynamic cache, which stores the most reliable and diverse samples to cover evolving test distributions. To measure the diversity efficiently, we quantize cached samples and compute the asymmetric similarity across query samples and memory samples, which guide the cache updating via replacing samples with the lowest scores iteratively. Finally, we leverage the asymmetric similarity between the quantized prototype representations from the dynamic cache to update logits under distribution shifts. Extensive experiments on various benchmark datasets validate the superiority of the proposed DANCE in different settings.

This paper studies the problem of misclassification detection for vision-language models (VLMs), which aims to identify samples wrongly assigned by the pre-trained VLM. Previous works usually focus on exploring the latent space and measuring pairwise similarity scores for confidence estimation. However, these similarity scores can be biased for boundary points with different fine-grained features. To address this issue, this paper proposes an approach named Distribution-aware Prototypical Calibration with Fine-grained Voting (DECENT). The core idea of DECENT is to utilize probabilistic distributions to characterize both the semantics of the test image and its corresponding class prototype. Specifically, we first conduct clustering within each class in the latent space to construct a distribution of fine grained prototypes. More importantly, multiple augmented views of the test image are considered as samples from its underlying semantic distribution. Then, we measure the distribution distance between the image’s underlying semantic distribution and its neighboring fine grained prototype distributions, which is aggregated via fine-grained voting into a distributional score. This score can be fused with the VLM’s semantic alignment score to significantly enhance misclassification detection sensitivity, or to improve overall classification robustness. Extensive experiments on benchmark datasets validate the dual effectiveness of the proposed DECENT.

Efficient long-context inference remains a major challenge for large language models (LLMs), as the cost of attention computation during auto-regressive decoding grows linearly with the context length. Recent sparse attention methods attempt to reduce the computational burden by selecting a subset of tokens at each step, while most rely on static importance scores that are repeatedly computed over the entire cache, overlooking the relational dynamics of the decoding process. In this work, we revisit sparse attention in LLMs and propose to model token importance as a dynamic process that evolves over decoding steps and propagates through model layers. To efficiently measure token importance, we propose two lightweight mechanisms: (1) Cross-Step Accumulation, which incrementally maintains long-term, query-agnostic importance via decayed accumulation of sparse attention scores, avoiding recomputing the importance of decoded tokens; and (2) Cross-Layer Propagation, which leverages the model’s intrinsic Retrieval Heads to compute query-aware indices and efficiently propagate them across layers; Together, these mechanisms preserve both stable context memory and adaptive query relevance while reduce redundant computation. We evaluate our approach on PG-19, RULER, LongBench, and mathematical reasoning benchmarks using models employing Multi-Head and Grouped-Query Attention. Under varying KV cache budgets, our method consistently outperforms prior sparse attention baselines, approaches full attention performance in most settings, and achieves speedups of up to 5.36× for attention latency and 2.33× for end-to-end decoding. Our code is available at: https://github.com/iLearn-Lab/ACL26-EvoSparse.

This paper investigates the problem of safe decoding for Large Language Models (LLMs) during inference, particularly under jailbreak attacks. Previous approaches typically either detect malicious content or regulate the decoding alignment of LLMs to mitigate such attacks. Although effective in defending against attacks, these methods often over-reject benign content, limiting their generalizability in real-world scenarios where harmful and benign information coexist. Towards this end, we propose an innovative framework named Sequence-level risk Accumulation for calibrating test-time alignment (SEAT). Specifically, SEAT introduces a reward-guided branch decoding paradigm to incorporate safety awareness during generation. To balance the detection of harmful content with the accurate response to benign information, SEAT employs a sequence-level risk monitor that smooths risk signals over the entire sequence, preventing over-confident refusals for certain tokens. Furthermore, we conduct extensive experiments on four attack benchmarks and two neutral datasets, comparing SEAT with eight state-of-the-art baselines. Consequently, the results demonstrate that SEAT achieves superior performance both in defending against jailbreak attacks and in generating high-quality responses on neutral datasets. Our code is available at https://github.com/ShanwenTan/SEAT.

Temporal Knowledge Graphs (TKGs) incorporate the temporal feature to express the transience of knowledge by describing when facts occur. TKG extrapolation aims to infer possible future facts based on known history, which has garnered significant attention in recent years. Some existing methods treat TKG as a sequence of independent subgraphs to model temporal evolution patterns, demonstrating impressive reasoning performance. However, they still have limitations: 1) In modeling subgraph semantic evolution, they usually neglect the internal structural interactions between subgraphs, which are actually crucial for encoding TKGs. 2) They overlook the potential smooth features that do not lead to semantic changes, which should be distinguished from the semantic evolution process. Therefore, we propose Disentangled Multi-span Evolutionary Network (DiMNet) for TKG reasoning. Specifically, we design a multi-span evolution strategy that captures local neighbor features while perceiving historical neighbor semantic information, thus enabling internal interactions between subgraphs during the evolution process. To maximize the capture of semantic change patterns, we design a disentangle component that adaptively separates nodes’ active and stable features, used to dynamically control the influence of historical semantics on future evolution. Extensive experiments demonstrate that DiMNet achieves substantial performance in TKG reasoning, outperforming the state-of-the-art up to 22.7% in MRR.

The recent advancements in language models have significantly catalyzed progress in computational biology. A growing body of research strives to construct unified foundation models for single-cell biology, with language models serving as the cornerstone. In this paper, we systematically review the developments in foundation language models designed specifically for single-cell biology. Our survey offers a thorough analysis of various incarnations of single-cell foundation language models, viewed through the lens of both pre-trained language models (PLMs) and large language models (LLMs). This includes an exploration of data tokenization strategies, pre-training/tuning paradigms, and downstream single-cell data analysis tasks. Additionally, we discuss the current challenges faced by these pioneering works and speculate on future research directions. Overall, this survey provides a comprehensive overview of the existing single-cell foundation language models, paving the way for future research endeavors.