As large language models continue to develop, the feasibility and significance of text-based symbolic music tasks have become increasingly prominent. While symbolic music has been widely used in generation tasks, LLM capabilities in understanding and reasoning about symbolic music remain largely underexplored. To address this gap, we propose ABC-Eval, the first open-source benchmark dedicated to the understanding and instruction-following capabilities in text-based ABC notation scores. It comprises 1,086 test samples spanning 10 sub-tasks, covering scenarios from basic musical syntax comprehension to complex sequence-level reasoning. Such a diverse scope poses substantial challenges to models' ability to handle symbolic music tasks. We evaluated seven state-of-the-art LLMs on ABC-Eval, and the results reveal notable limitations in existing models' symbolic music processing capabilities. Furthermore, the consistent performance of individual baselines across different sub-tasks supports the reliability of our benchmark.

Prompt evolution algorithms offer a powerful paradigm for enhancing AI systems based on closed-source models, while few work explores whether aggregating results from multiple prompts to reach a consensus can further advance the system capability boundary. In this paper, we introduce Consensus-Evolve (C-Evolve), an evolutionary algorithm that discovers a group of prompts whose aggregated outputs after majority voting achieve optimal performance. More specifically, C-Evolve employs an island-based evolutionary algorithm to maintain population diversity, and prompts from distinct islands are selected to form groups to aggregate their outputs. The key difference from single individual evolution is a voting score, which evaluates each individual prompt's contribution within groups. We take this as the fitness score for evolution instead of individual performance. Consequently, C-Evolve is more likely to produce and maintain prompts with higher potential to form a high-performing group and eliminate low-performing ones, gradually improving the group performance after reaching consensus. Our method achieves state-of-the-art performance across a wide range of tasks, including both open-ended tasks like HotpotQA and closed-ended tasks like MATH. On Qwen3-8B, C-Evolve achieves 70.67% on HotpotQA and 43.88% on IFBench, which are 4.95% and 2.73% higher than GEPA, respectively. For GPT-4.1-mini, the accuracy on IFBench is further improved to 47.96% and reaches 95.33% in the MATH benchmark. These results demonstrate the C-Evolve's competitive performance.

Large language models (LLMs) have achieved significant advancements in reasoning capabilities through reinforcement learning (RL) via environmental exploration. As the intrinsic properties of the environment determine the abilities that LLMs can learn, the environment plays a important role in the RL finetuning process. An ideal LLM reasoning environment should possess three core characteristics: scalability, generalizable reasoning, and verifiability. However, existing mathematical and coding environments are difficult to scale due to heavy reliance on expert annotation, while the skills learned in game-based environments are too specialized to generalize. To bridge this gap, we introduce the Structured In-context Environment (SIE) framework. SIE achieves scalability by automatically constructing reasoning environments from large-scale structured data, where the rich compositional patterns naturally support generalizable reasoning. Moreover, the explicit schemas and reasoning chains in structured data provide a foundation for rule-based verifiability. Experimental results show that SIE framework not only achieves substantial improvements in in-domain structured reasoning, but also enables the learned compositional reasoning skills to generalize effectively to out-of-domain mathematical and logical reasoning tasks. We further explored learning in information-limited partial SIEs and found that LLMs can infer the missing information through exploring the environment, leading to robust reasoning improvements and generalization performance. Fine-tuning large language models (LLMs) with reinforcement learning (RL) has emerged as a dominant post-training paradigm for eliciting complex reasoning capabilities (Jaech et al., 2024; Guo et al., 2025; Team et al., 2025; Comanici et al., 2025). This mechanism of learning from environmental feedback enables LLMs to acquire crucial reasoning strategies such as self-reflection, backtracking, and chain-of-thought. RL fine-tuning has shown significant progress in math reasoning and code generation (Zeng et al., 2025; Hu et al., 2025b; Chen et al., 2025), and is gradually being extended to more challenging applications, such as interacting with search engines and building deep research agents (Jin et al., 2025; Zheng et al., 2025b; Li et al., 2025; Team, 2025).

Deploying Large Language Models (LLMs) on mobile devices faces the challenge of insufficient performance in smaller models and excessive resource consumption in larger ones. This paper highlights that mobile Neural Processing Units (NPUs) have underutilized computational resources, particularly their matrix multiplication units, during typical LLM inference. To leverage this wasted compute capacity, we propose applying parallel test-time scaling techniques on mobile NPUs to enhance the performance of smaller LLMs. However, this approach confronts inherent NPU challenges, including inadequate hardware support for fine-grained quantization and low efficiency in general-purpose computations. To overcome these, we introduce two key techniques: a hardware-aware tile quantization scheme that aligns group quantization with NPU memory access patterns, and efficient LUT-based replacements for complex operations such as Softmax and dequantization. We design and implement an end-to-end inference system that leverages the NPU's compute capability to support test-time scaling on Qualcomm Snapdragon platforms. Experiments show our approach brings significant speedups: up to 19.0 for mixed-precision GEMM and 2.2 for Softmax. More importantly, we demonstrate that smaller models using test-time scaling can match or exceed the accuracy of larger models, achieving a new performance-cost Pareto frontier.

Policy compliance assessment is a fundamental task of evaluating whether an input case strictly complies with a set of human-defined rules, more generally known as policies. In practice, human experts follow a systematic, step-by-step process to identify violations with respect to specific stipulations outlined in the policy. However, such documentation of gold-standard, expert-level reasoning processes is costly to acquire. In this paper, we introduce Policy Reasoning Traces (PRT), a form of specialized generated reasoning chains that serve as a reasoning bridge to improve an LLM's policy compliance assessment capabilities. Our empirical evaluations demonstrate that the use of PRTs for both inference-time and training-time scenarios significantly enhances the performance of open-weight and commercial models, setting a new state-of-the-art for HIPAA and GDPR policies. Beyond accuracy gains, we also highlight how PRTs can improve an LLM's ability to accurately cite policy clauses, as well as influence compliance decisions through their high utilization from the raw chains of thought.

Diffusion large language models (dLLMs) enable any-order generation, but this flexibility enlarges the attack surface: harmful spans may appear at arbitrary positions, and template-based prefilling attacks such as DIJA bypass response-level refusals. We introduce A2D (Any-Order, Any-Step Defense), a token-level alignment method that aligns dLLMs to emit an [EOS] refusal signal whenever harmful content arises. By aligning safety directly at the token-level under randomized masking, A2D achieves robustness to both any-decoding-order and any-step prefilling attacks under various conditions. It also enables real-time monitoring: dLLMs may begin a response but automatically terminate if unsafe continuation emerges. On safety benchmarks, A2D consistently prevents the generation of harmful outputs, slashing DIJA success rates from over 80% to near-zero (1.3% on LLaDA-8B-Instruct, 0.0% on Dream-v0-Instruct-7B), and thresholded [EOS] probabilities allow early rejection, yielding up to 19.3x faster safe termination.

Abstract-- Large Language Models (LLMs) and Vision-Language Models (VLMs) are increasingly deployed in robotic environments but remain vulnerable to jailbreaking attacks that bypass safety mechanisms and drive unsafe or physically harmful behaviors in the real world. Data-driven defenses such as jailbreak classifiers show promise, yet they struggle to generalize in domains where specialized datasets are scarce, limiting their effectiveness in robotics and other safety-critical contexts. T o address this gap, we introduce J-DAPT, a lightweight framework for multimodal jailbreak detection through attention-based fusion and domain adaptation. J-DAPT integrates textual and visual embeddings to capture both semantic intent and environmental grounding, while aligning general-purpose jailbreak datasets with domain-specific reference data. Evaluations across autonomous driving, maritime robotics, and quadruped navigation show that J-DAPT boosts detection accuracy to nearly 100% with minimal overhead. These results demonstrate that J-DAPT provides a practical defense for securing VLMs in robotic applications. Additional materials are made available at: https://j-dapt.github.io. The recent surge in capabilities of Large Language Models (LLMs) and Vision-Language Models (VLMs) has enabled significant advances in robotics, from sophisticated in-the-wild reasoning and decision-making [5] to intuitive language-guided manipulation [12] and embodied navigation tasks [26]. These models are increasingly integrated into safety-critical systems, from autonomous driving to robotic healthcare, where adversarial failures could lead to real-world harm.

Financial dialogue transcripts pose a unique challenge for sentence-level information extraction due to their informal structure, domain-specific vocabulary, and variable intent density. We introduce Fin-ExBERT, a lightweight and modular framework for extracting user intent-relevant sentences from annotated financial service calls. Our approach builds on a domain-adapted BERT (Bidirectional Encoder Representations from Transformers) backbone enhanced with LoRA (Low-Rank Adaptation) adapters, enabling efficient fine-tuning using limited labeled data. We propose a two-stage training strategy with progressive unfreezing: initially training a classifier head while freezing the backbone, followed by gradual fine-tuning of the entire model with differential learning rates. To ensure robust extraction under uncertainty, we adopt a dynamic thresholding strategy based on probability curvature (elbow detection), avoiding fixed cutoff heuristics. Empirical results show strong precision and F1 performance on real-world transcripts, with interpretable output suitable for downstream auditing and question-answering workflows. The full framework supports batched evaluation, visualization, and calibrated export, offering a deployable solution for financial dialogue mining.

Abstract--The rapid advancement of large language models (LLMs) has enabled an emergence of agentic artificial intelligence (AI) with powerful reasoning and autonomous decision-making capabilities. This integration with edge computing has led to the development of Mobile Edge General Intelligence (MEGI), which brings real-time, privacy-preserving reasoning to the network edge. However, deploying LLM-based agentic AI reasoning in MEGI environments poses significant challenges due to the high computational demands of reasoning and the limited resources of edge devices. T o address these challenges, we propose a joint optimization framework for efficient LLM reasoning deployment in MEGI. First, we review methods that enhance LLM reasoning capabilities, such as Chain-of-Thought (CoT) prompting, Supervised Fine-T uning (SFT), and Mixture of Experts (MoE). Next, we present a distributed framework that addresses two correlated aspects: reasoning enhancement through adaptive CoT prompting and scalable deployment through distributed MoE architecture. The framework dynamically activates expert networks and adjusts reasoning depth based on task complexity and device capabilities. Experimental results demonstrate the framework's effectiveness in balancing reasoning quality with resource efficiency, validating the practical viability of deploying sophisticated LLM reasoning capabilities in resource-constrained MEGI environments. Recent advances in artificial intelligence (AI) demonstrate that reasoning is fundamental to intelligence.

Large language models (LLMs) frequently memorize sensitive or personal information, raising significant privacy concerns. Existing variants of differential privacy stochastic gradient descent (DPSGD) inject uniform noise into every gradient step, significantly extending training time and reducing model accuracy. We propose that concentrating noise primarily on gradients associated with sensitive tokens can substantially decrease DP training time, strengthen the protection of sensitive information, and simultaneously preserve the model's performance on non-sensitive data. We operationalize this insight through Adaptive Token-Weighted Differential Privacy (ATDP), a modification of vanilla DP-SGD that adaptively assigns different gradient weights to sensitive and non-sensitive tokens. By employing a larger noise scale at the early stage of training, ATDP rapidly disrupts memorization of sensitive content. As a result, ATDP only requires a few additional epochs of lightweight post-processing following standard fine-tuning, injecting targeted noise primarily on parameters corresponding to sensitive tokens, thus minimally affecting the model's general capabilities. ATDP can be seamlessly integrated into any existing DP-based fine-tuning pipeline or directly applied to non-private models as a fast privacy-enhancing measure. Additionally, combined with an initial redacted fine-tuning phase, ATDP forms a streamlined DP pipeline that achieves comparable canary protection to state-of-the-art DP-SGD methods, significantly reduces the computational overhead of DP fine-tuning, shortening training time by approximately 90 percent, while achieving comparable or superior privacy protection and minimal accuracy degradation.