Audio Language Models (ALMs) have made significant progress recently. These models integrate the audio modality directly into the model, rather than converting speech into text and inputting text to Large Language Models (LLMs). While jailbreak attacks on LLMs have been extensively studied, the security of ALMs with audio modalities remains largely unexplored. Currently, there is a lack of an adversarial audio dataset and a unified framework specifically designed to evaluate and compare attacks and ALMs. In this paper, we present JALMBench, a comprehensive benchmark to assess the safety of ALMs against jailbreak attacks. JALMBench includes a dataset containing 11,316 text samples and 245,355 audio samples with over 1,000 hours. It supports 12 mainstream ALMs, 4 text-transferred and 4 audio-originated attack methods, and 5 defense methods. Using JALMBench, we provide an in-depth analysis of attack efficiency, topic sensitivity, voice diversity, and architecture. Additionally, we explore mitigation strategies for the attacks at both the prompt level and the response level.

Obtaining large-scale, high-quality reasoning data is crucial for improving the geometric reasoning capabilities of multi-modal large language models (MLLMs). However, existing data generation methods, whether based on predefined tem plates or constrained symbolic provers, inevitably face diversity and numerical generalization limitations. To address these limitations, we propose NeSyGeo, a novel neuro-symbolic framework for generating geometric reasoning data. First, we propose a domain-specific language grounded in the entity-attributes-relations paradigm to comprehensively represent all components of plane geometry, along with generative actions defined within this symbolic space. We then design a symbolic-visual-text pipeline that synthesizes symbolic sequences, maps them to visual and textual representations and generates reasoning path with reverse search and forward validation. Based on this framework, we construct NeSyGeo CoT and NeSyGeo-Caption datasets, containing 100k samples, and release a new benchmark NeSyGeo-Test for evaluating geometric reasoning abilities in MLLMs. Experiments demonstrate that the proposal significantly and consistently improves the performance of multiple MLLMs under both reinforcement and supervised fine-tuning. With only 4k samples and two epochs of reinforcement fine-tuning, base models achieve improvements of up to +15.8% on MathVision, +8.4% on MathVerse, and +7.3% on GeoQA. Notably, a 4B model can be improved to outperform an 8B model from the same series on geometric reasoning tasks.s

Neural language models are black-boxes--both linguistic patterns and factual knowledge are distributed across billions of opaque parameters. This entangled encoding makes it difficult to reliably inspect, verify, or update specific facts. We introduce Limited Memory Language Models (LMLM), a new class of language models that externalizes factual knowledge to external database during pre-training rather than memorizing them. Our pre-training approach strategically masks externally retrieved factual values from the training loss, thereby teaching the model to perform targeted lookups rather than relying on memorization in model weights. Our experiments demonstrate that LMLMs achieve competitive performance compared to significantly larger LLMs on standard benchmarks, while offering the advantages of explicit, editable, and verifiable knowledge bases.

Low-Rank Adaptation (LoRA) has emerged as a highly efficient framework for finetuning the weights of large foundation models, and has become the go-to method for data-driven customization of LLMs. Despite the promise of highly customized behaviors and capabilities, switching between relevant LoRAs in a multiturn setting is inefficient, as the key-value (KV) cache of the entire turn history must be recomputed with the LoRA weights before generation can begin. To address this problem, we propose Activated LoRA (aLoRA), an adapter architecture which modifies the LoRA framework to only adapt weights for the tokens in the sequence after the aLoRA is invoked. This change crucially allows aLoRA to accept the base model's KV cache of the input string, meaning that aLoRA can be instantly activated whenever needed in a chain without recomputing the prior keys and values. This enables building what we call intrinsics, i.e. specialized models invoked to perform well-defined operations on portions of an input chain or conversation that otherwise uses the base model by default. We train a set of aLoRA-based intrinsics models, demonstrating competitive accuracy with standard LoRA while significantly improving inference efficiency. We contributed our Activated LoRA implementation to the Huggingface PEFT library https://github.com/huggingface/peft.

The emergence of reinforcement learning in post-training of large language models has sparked significant interest in reward models. Reward models assess the quality of sampled model outputs to generate training signals. This task is also performed by evaluation metrics that monitor the performance of an AI model. We find that the two research areas are mostly separate, leading to redundant terminology and repeated pitfalls. Common challenges include susceptibility to spurious correlations, impact on downstream reward hacking, methods to improve data quality, and approaches to meta-evaluation. Our position paper argues that a closer collaboration between the fields can help overcome these issues. To that end, we show how metrics outperform reward models on specific tasks and provide an extensive survey of the two areas. Grounded in this survey, we point to multiple research topics in which closer alignment can improve reward models and metrics in areas such as preference elicitation methods, avoidance of spurious correlations and reward hacking, and calibration-aware meta-evaluation.

To solve complex reasoning tasks for Large Language Models (LLMs), prompting-based methods offer a lightweight alternative to fine-tuning and reinforcement learning. However, as reasoning chains extend, critical intermediate steps and the original prompt will be buried in the context, receiving insufficient attention and leading to errors. In this paper, we propose Self-Anchor, a novel pipeline that leverages the inherent structure of reasoning to steer LLM attention. Self-Anchor decomposes reasoning trajectories into structured plans and automatically aligns the model's attention to the most relevant inference steps, allowing the model to maintain focus throughout generation. Our experiment shows that Self-Anchor outperforms SOTA prompting methods across six benchmarks. Notably, Self-Anchor significantly reduces the performance gap between ``non-reasoning'' models and specialized reasoning models, with the potential to enable most LLMs to tackle complex reasoning tasks without retraining.

Multi-LLM systems harness the complementary strengths of diverse Large Language Models, achieving performance and efficiency gains unattainable by a single model. In existing designs, LLMs communicate through text, forcing internal representations to be transformed into output token sequences. This process both loses rich semantic information and incurs token-by-token generation latency. Motivated by these limitations, we ask: Can LLMs communicate beyond text? Oracle experiments show that enriching the KV-Cache semantics can improve response quality without increasing cache size, supporting KV-Cache as an effective medium for inter-model communication. Thus, we propose Cache-to-Cache (C2C), a new paradigm for direct semantic communication between LLMs. C2C uses a neural network to project and fuse the source model's KV-cache with that of the target model to enable direct semantic transfer. A learnable gating mechanism selects the target layers that benefit from cache communication. Compared with text communication, C2C utilizes the deep, specialized semantics from both models, while avoiding explicit intermediate text generation. Experiments show that C2C achieves 8.5-10.5% higher average accuracy than individual models. It further outperforms the text communication paradigm by approximately 3.0-5.0%, while delivering an average 2.0x speedup in latency. Our code is available at https://github.com/thu-nics/C2C.

Diffusion language models, especially masked discrete diffusion models, have achieved great success recently. While there are some theoretical and primary empirical results showing the advantages of latent reasoning with looped transformers or continuous chain-of-thoughts, continuous diffusion models typically underperform their discrete counterparts. In this paper, we argue that diffusion language models do not necessarily need to be in the discrete space. In particular, we prove that continuous diffusion models have stronger expressivity than discrete diffusions and looped transformers. We attribute the contradiction between the theoretical expressiveness and empirical performance to their practical trainability: while continuous diffusion provides intermediate supervision that looped transformers lack, they introduce additional difficulty decoding tokens into the discrete token space from the continuous representation space. We therefore propose Coevolutionary Continuous Discrete Diffusion (CCDD), which defines a joint multimodal diffusion process on the union of a continuous representation space and a discrete token space, leveraging a single model to simultaneously denoise in the joint space. By combining two modalities, CCDD is expressive with rich semantics in the latent space, as well as good trainability and sample quality with the help of explicit discrete tokens. We also propose effective architectures and advanced training/sampling techniques for CCDD, which reveals strong empirical performance in extensive language modeling experiments on real-world tasks.

Web agents powered by large language models (LLMs) must process lengthy web page observations to complete user goals; these pages often exceed tens of thousands of tokens. This saturates context limits and increases computational cost processing; moreover, processing full pages exposes agents to security risks such as prompt injection. Existing pruning strategies either discard relevant content or retain irrelevant context, leading to suboptimal action prediction. We introduce FocusAgent, a simple yet effective approach that leverages a lightweight LLM retriever to extract the most relevant lines from accessibility tree (AxTree) observations, guided by task goals. By pruning noisy and irrelevant content, FocusAgent enables efficient reasoning while reducing vulnerability to injection attacks. Experiments on WorkArena and WebArena benchmarks show that FocusAgent matches the performance of strong baselines, while reducing observation size by over 50%. Furthermore, a variant of FocusAgent significantly reduces the success rate of prompt-injection attacks, including banner and pop-up attacks, while maintaining task success performance in attack-free settings. Our results highlight that targeted LLM-based retrieval is a practical and robust strategy for building web agents that are efficient, effective, and secure.

Large Language Models (LLMs) achieve strong results on code tasks, but how they derive program meaning remains unclear. We argue that code communicates through two channels: structural semantics, which define formal behavior, and human-interpretable naming, which conveys intent. Surprisingly, we also observe consistent reductions on execution tasks that should depend only on structure, revealing that current benchmarks reward memorization of naming patterns rather than genuine semantic reasoning. To disentangle these effects, we introduce a suite of semantics-preserving obfuscations and show that they expose identifier leakage across both summarization and execution. Large language models (LLMs) now achieve striking results across code intelligence--program synthesis, repair, summarization, and test generation. Y et how these models derive meaning from source code remains unclear. If an LLM truly understands a program's intent, its behavior should remain stable when human-interpretable names are perturbed while semantics stay fixed; conversely, strong performance drops would indicate an overreliance on surface cues rather than semantic reasoning.