Group Relative Policy Optimization (GRPO) is an attractive reinforcement learning paradigm for Large Reasoning Models because its groupwise relative advantages eliminate the need for a value network, but its efficiency is severely limited by zero-variance queries whose rollouts are uniformly correct or uniformly incorrect and therefore consume expensive Chain-of-Thought generation without providing useful learning signals. Existing remedies typically rely on offline difficulty labels or external capability trackers, which either fail to follow the model’s rapidly evolving competence or compromise the end-to-end simplicity of GRPO. We propose Active-GRPO, a prompt-free extension that equips the policy with intrinsic fuzzy metacognition via a lightweight confidence head attached to the final prefill representation. The head predicts ordinal confidence bins and is jointly calibrated against empirical rollout outcomes through a dual-objective optimization framework. It further enables a multi-criterion voting mechanism that filters uninformative queries before rollout generation, thereby concentrating compute on samples near the learning boundary with negligible overhead and without perturbing the reasoning distribution. Experiments on mathematical and logical reasoning benchmarks show that Active-GRPO reduces training computation while improving both reasoning accuracy and intrinsic calibration quality.

As large language models (LLMs) are integrated into numerous applications, LLMs’ safety becomes critical for both application developers and intended users. Currently, great efforts have been made to develop safety benchmarks with fine-grained taxonomies. However, these benchmarks’ taxonomies are disparate with different safety policies. Thus, existing safeguards trained on these benchmarks are either coarse-grained to only distinguish between "safe” and "unsafe,” or constrained by the specified narrow risk taxonomies. To leverage these fine-grained safety policies across multiple safety taxonomies, we propose GSPR, a Generalizable Safety Policy Reasoner to identify unsafe inputs and outputs with violated safety taxonomies and concise explanations. Unlike prior safeguards which only cover a fixed set of risk factors, GSPR incentivizes its reasoning capability with varied safety taxonomies through reinforcement learning. Our GSPR can be trained across multiple safety benchmarks with distinct taxonomies and naturally exhibits powerful generalization ability. We conduct extensive experiments to show that GSPR significantly improves existing safety guardrails’ reasoning capabilities for both safety and category prediction tasks. Moreover, GSPR also achieves the least inference token costs with explanations.

The democratization of LLMs has accelerated the generation and circulation of highly fluent disinformation, making traditional syntax-semantic verification increasingly insufficient. Such deception often combines persuasive rhetoric, emotional manipulation, and narrative role construction to shape readers’ interpretations. However, existing detectors typically emphasize isolated signals—such as syntax, external knowledge, persuasion, or affective cues—and thus struggle to capture multi-faceted manipulation or provide structured, taxonomy-grounded evidence. To address this gap, we present ExTax, a taxonomy-aligned framework for interpretable disinformation detection. ExTax unifies persuasive rhetoric, emotional manipulation, and narrative roles into a 17-dimensional taxonomic space comprising 6 persuasive-rhetoric strategies, 5 emotional-manipulation methods, and 6 narrative-role categories. It elicits attributes from multiple frontier LLMs, reconciles their disagreements through Entropy-driven Dynamic Label Smoothing, and fuses the resulting taxonomic representations with contextual encodings via Heterogeneous Multi-Head Attention. Across five cross-domain and cross-genre benchmarks, ExTax achieves an overall Macro F₁ of 0.8456, outperforming strong deep learning and LLM-based baselines. It also remains robust under severe genre imbalance, where the strongest deep baseline drops from 0.9454 to 0.6194 across genres. Our code is available at https://github.com/owlsan49/ExTax.

Zero-shot fine-grained emotion recognition (FER) is pivotal for empathetic AI but remains constrained by two inherent limitations in retrieval-augmented in-context learning (ICL): 1) subtle emotional distinctions inject semantic noise that blurs decision boundaries, and 2) naive similarity-based retrieval often surfaces counterproductive or antonymous demonstrations. To address these, we propose EmoDap, a dual-agent prompt retrieval framework for zero-shot FER. EmoDap first constructs a high-quality demonstration pool via self-consistency pseudo-labeling and derives explicit emotion prototypes through contrastive learning. Subsequently, a demonstration retriever and an emotion predictor are jointly optimized via Group Relative Policy Optimization (GRPO): the retriever learns to select examples that maximize the predictor’s accuracy, while the predictor leverages prototype guidance to dynamically filter candidate emotions. Extensive experiments across four benchmarks demonstrate that EmoDap consistently outperforms state-of-the-art baselines, establishing a robust paradigm for zero-shot FER driven by open-source LLM backbones.

As Large Language Models (LLMs) become indispensable assistants, they remain vulnerable to misuse. Jailbreaking is an essential adversarial technique for red-teaming models to uncover and patch security flaws. However, existing jailbreak methods suffer from significant limitations. Token-level jailbreak attacks often produce incoherent or unreadable inputs and exhibit poor transferability, while prompt-level attacks lack scalability and rely heavily on manual effort and human ingenuity. We propose AGILE, a concise and effective two-stage framework that combines the advantages of these approaches. The first stage performs a one-shot, scenario-based generation of context and rephrases the original malicious query to obscure its harmful intent. The second stage utilizes information from the model’s hidden states to guide fine-grained edits, effectively steering the model’s internal representation of the input from a malicious one toward a benign one. Extensive experiments demonstrate that AGILE achieves state-of-the-art Attack Success Rate, with gains of up to 37.74% over the strongest baseline, and AGILE exhibits excellent transferability to black-box and large-scale models. Our code is available at https://github.com/SELGroup/AGILE.

Long Chain-of-Thought (CoT) reasoning has significantly advanced the capabilities of Large Language Models (LLMs), but this progress is accompanied by substantial memory and latency overhead from the extensive Key-Value (KV) cache. Although KV cache quantization is a promising compression technique, existing low-bit quantization methods often exhibit severe performance degradation on complex reasoning tasks. Fixed-precision quantization struggles to handle outlier channels in the key cache, while current mixed-precision strategies fail to accurately identify components requiring high-precision representation. We find that an effective low-bit KV cache quantization strategy must consider two factors: a key channel’s intrinsic quantization difficulty and its relevance to the query. Based on this insight, we propose MixKVQ, a novel plug-and-play method that introduces a lightweight, query-aware algorithm to identify and preserve critical key channels that need higher precision, while applying per-token quantization for value cache. Experiments on complex reasoning datasets demonstrate that our approach significantly outperforms existing low-bit methods, achieving performance comparable to a full-precision baseline at a substantially reduced memory footprint.

Current Emotion Recognition in Conversation (ERC) research follows a closed-domain assumption. However, there is no clear consensus on emotion classification in psychology, which presents a challenge for models when it comes to recognizing previously unseen emotions in real-world applications. To bridge this gap, we introduce the Unseen Emotion Recognition in Conversation (UERC) task for the first time and propose ProEmoTrans, a solid prototype-based emotion transfer framework. This prototype-based approach shows promise but still faces key challenges: First, implicit expressions complicate emotion definition, which we address by proposing an LLM-enhanced description approach. Second, utterance encoding in long conversations is difficult, which we tackle with a proposed parameter-free mechanism for efficient encoding and overfitting prevention. Finally, the Markovian flow nature of emotions is hard to transfer, which we address with an improved Attention Viterbi Decoding (AVD) method to transfer seen emotion transitions to unseen emotions. Extensive experiments on three datasets show that our method serves as a strong baseline for preliminary exploration in this new area.

Multimodal Large Language Models (MLLMs) have serious security vulnerabilities. While safety alignment using multimodal datasets consisting of text and data of additional modalities can effectively enhance MLLM’s security, it is costly to construct these datasets. Existing low-resource security alignment methods, including textual alignment, have been found to struggle with the security risks posed by additional modalities. To address this, we propose Synthetic Embedding augmented safety Alignment (SEA), which optimizes embeddings of additional modality through gradient updates to expand textual datasets. This enables multimodal safety alignment training even when only textual data is available. Extensive experiments on image, video, and audio-based MLLMs demonstrate that SEA can synthesize a high-quality embedding on a single RTX3090 GPU within 24 seconds. SEA significantly improves the security of MLLMs when faced with threats from additional modalities. To assess the security risks introduced by video and audio, we also introduced a new benchmark called VA-SafetyBench. High attack success rates across multiple MLLMs validate its challenge. Our code and data will be available at https://github.com/ZeroNLP/SEA.

Large Language Models (LLMs) have revolutionized open-domain dialogue agents but encounter challenges in multi-character role-playing (MCRP) scenarios. To address the issue, we present Neeko, an innovative framework designed for efficient multiple characters imitation. Neeko employs a dynamic low-rank adapter (LoRA) strategy, enabling it to adapt seamlessly to diverse characters. Our framework breaks down the role-playing process into agent pre-training, multiple characters playing, and character incremental learning, effectively handling both seen and unseen roles. This dynamic approach, coupled with distinct LoRA blocks for each character, enhances Neeko’s adaptability to unique attributes, personalities, and speaking patterns. As a result, Neeko demonstrates superior performance in MCRP over most existing methods, offering more engaging and versatile user interaction experiences.