Representation learning is fundamental to modern machine learning, powering applications such as text retrieval and multimodal understanding. However, learning robust and generalizable representations remains challenging. While prior work has demonstrated that active noise injection, a form of data augmentation, can enhance encoding performance, most existing methods rely on heuristic or static noise, overlooking the dynamic nature of feature distributions during training. In this work, we systematically study the role of noise in representation learning from both gradient-based and feature distribution perspectives, using InfoNCE loss as a representative example. Focusing on multimodal representation learning, we propose FANoise, a novel feature-adaptive noise injection strategy. By leveraging the dynamics of contrastive learning, FANoise effectively mitigates the negative impacts of noise while preserving its benefits. Under this theoretically grounded framework, comprehensive experiments demonstrate that FANoise consistently improves overall performance on multimodal tasks across various base VLM models.

Autonomous agents choose actions using scores that may not reflect experimental success. We developed OPERA, an operator-residual framework for optical experiments. It represents experimental actions as optical operators and evaluates their outcomes using physically interpretable residuals. Operators specify executable changes to measurement, control or reconstruction, while residuals report departures from specified physical conditions. The agent uses both to select, combine or generate operators, and physical performance is evaluated independently against a withheld reference. Across three optical tasks, score-only feedback produced score increases without physical improvement in 23.6–39.0% of decisions, compared with 0.9–1.9% for operator-residual feedback. Operator-residual feedback increased the probability of reaching and maintaining task targets and reduced experimental budgets. Protocols selected in digital twins were transferred to three optical instruments, and repeated experiments showed a lower projection budget in structured-light reconstruction. Together, operators and residuals guide autonomous decisions using measurable physical evidence.

We propose ControlMLLM++, a novel test-time adaptation framework that injects learnable visual prompts into frozen multimodal large language models (MLLMs) to enable fine-grained region-based visual reasoning without any model retraining or fine-tuning. Leveraging the insight that cross-modal attention maps intrinsically encode semantic correspondences between textual tokens and visual regions, ControlMLLM++ optimizes a latent visual token modifier during inference via a task-specific energy function to steer model attention towards user-specified areas. To enhance optimization stability and mitigate language prompt biases, ControlMLLM++ incorporates an improved optimization strategy (Optim++) and a prompt debiasing mechanism (PromptDebias). Supporting diverse visual prompt types including bounding boxes, masks, scribbles, and points, our method demonstrates strong out-of-domain generalization and interpretability. The code is available at https://github.com/mrwu-mac/ControlMLLM.

Large Language Models (LLMs) have demonstrated strong performance across a wide range of NLP tasks. However, they often exhibit suboptimal behaviors and inconsistencies when exposed to unfamiliar external information, underscoring their limitations in effectively leveraging such knowledge. Inspired by constructivist learning theory, we propose ThinkNote, a novel framework that enhances the external knowledge utilization of LLMs through a two-stage constructivist cognitive modeling process. Specifically, ThinkNote performs knowledge assimilation to align new information with the model's parametric memory, forming a coherent internal representation. It then applies thought accommodation to adapt internal reasoning, thereby promoting more consistent and reliable outputs. Extensive experimental results demonstrate that ThinkNote achieves a 10

While Mixture-of-Experts (MoE) scales model capacity without proportionally increasing computation, its massive total parameter footprint creates significant storage and memory-access bottlenecks, which hinder efficient end-side deployment that simultaneously requires high performance, low computational cost, and small storage overhead. To achieve these properties, we present DECO, a sparse MoE architecture designed to match the performance of dense Transformers under identical total parameter budgets and training tokens. DECO utilizes the differentiable and flexible ReLU-based routing enhanced by learnable expert-wise scaling, which adaptively balances the contributions of routed and shared experts. Furthermore, we introduce NormSiLU, an activation function that normalizes inputs prior to SiLU operators, producing a more stable trend of routed-expert activation ratio and a higher intrinsic sparsity level. We also identify an empirical advantage in using non-gated MLP experts with ReLU-based routing, indicating the possibility of MoE architecture simplification. Experiments demonstrate that DECO, activating only 20

On modern GPUs the fastest stencil kernel depends on the stencil's shape, precision, and host application, and a kernel tuned for one case is rarely fastest for another. Stencil DSLs, code generators, and autotuners instead pursued generality: a single human-authored method reused across cases and validated mainly on microbenchmarks, because per-case specialization was too costly to scale. ForgeStencil starts from the opposite assumption. Code-synthesis agents have reduced that cost enough to build a fresh solution for each case and deploy it end-to-end in real software. A Kernel Agent synthesizes CUDA and forges a per-configuration matrix of specialized operators that matches or exceeds the strongest publicly available state-of-the-art (SOTA) baseline for each case. An App Agent extends the principle to whole applications: it locates hotspots, rewrites application structure, and validates and integrates each change across 100+ real industrial and scientific codes. Most of the measured speedup comes from structural and host-side rewrites, with pure stencil replacement in the minority; the gain also correlates negatively with how well the baseline was already tuned, consistent with gains coming from specialization rather than generic reuse. Every result is checked by a measurement-integrity harness that turns an overstated speedup into a system-level error. The forged kernels reach a same-precision f32 geometric mean of 2.35x against the per-case SOTA baselines (fp16 gains, 1.95x, disclosed separately; A100), and the end-to-end application median is 1.41x across 100 codes, all against same-architecture GPU baselines with program-provided validation and timing. Of 116 candidates, every one that failed the correctness, measurement, or speedup criteria was recorded as rejected or downgraded instead of written up as a speedup.

Unsupervised Reinforcement Learning with Verifiable Rewards (URLVR) offers a pathway for Large Language Models (LLMs) to improve without human supervision. Particularly, many works use model intrinsic information as rewards for URLVR, showing promising improvements, yet their potential and limitations remain unclear. In this work, we revisit URLVR through the lens of intrinsic rewards. We present a unified theoretical framework showing that intrinsic reward methods share a core mechanism: they trade uncertainty for performance by leveraging the model’s prior knowledge to sharpen output distributions. Empirical analysis confirms this tradeoff, revealing distinct failure modes and showing that collapse is not inevitable in small, domain-specific regimes such as test-time training. Beyond these findings, early intrinsic reward dynamics also provide a lightweight indicator of model-task priors, complementing $pass@k$ in assessing RL trainability. These insights highlight both the promise and pitfalls of URLVR, motivating future directions such as external rewards and hybrid supervision strategies.

LLM-based multi-agent systems are increasingly used for strategic decision-making tasks. In such settings, performance depends not only on individual model capabilities, but also on the policies by which agents interact and adapt. Multi-agent reinforcement learning can optimise these interaction policies, but its reward design often remains task-specific and weakly grounded in interaction structure. To address this gap, we propose GARL, a GAme-theoretic Reinforcement Learning framework for multi-agent strategic prioritisation. GARL formalises strategic prioritisation as a two-stage game: competing agents first allocate strategic resources over a shared candidate set, and a higher-level arbiter then produces the final ranking. The resulting game-theoretic utilities are converted into role-specific reinforcement signals, allowing policy optimisation to be guided by structured interaction. We instantiate GARL on issues-in-dispute ranking, where the goal is to prioritise core issues in legal proceedings. Experiments show that GARL improves ranking performance, enables small open-source LLMs to become competitive with a strong closed-source LLM under the same candidate-ranking setting, and yields gains in legal-domain competence and broader strategic decision-making. Overall, GARL demonstrates how game-theoretic interaction structure can be turned into reinforcement-learning objectives, providing a principled approach to policy optimisation in multi-agent strategic prioritisation.

Hybrid Transformer architectures, which combine softmax attention blocks and recurrent neural networks (RNNs), have shown a desirable performance-throughput tradeoff for long-context modeling, but their adoption and studies are hindered by the prohibitive cost of large-scale pre-training from scratch. Some recent studies have shown that pre-trained softmax attention blocks can be converted into RNN blocks through parameter transfer and knowledge distillation. However, these transfer methods require substantial amounts of training data (more than 10B tokens), and the resulting hybrid models also exhibit poor long-context performance, which is the scenario where hybrid models enjoy significant inference speedups over Transformer-based models. In this paper, we present HALO (Hybrid Attention via Layer Optimization), a pipeline for distilling Transformer models into RNN-attention hybrid models. We then present HypeNet, a hybrid architecture with superior length generalization enabled by a novel position encoding scheme (named HyPE) and various architectural modifications. We convert the Qwen3 series into HypeNet using HALO, achieving performance comparable to the original Transformer models while enjoying superior long-context performance and efficiency. The conversion requires just 2.3B tokens, less than 0.01

Modern language models increasingly adopt hybrid architectures that combine full attention with efficient attention modules, such as sliding-window attention (SWA) and recurrent sequence mixers. However, how these efficient modules shape model capabilities remains poorly understood. To address this gap, we conduct a systematic analysis across hybrid architectures from three perspectives: scaling behavior, mechanism analysis, and architecture design. First, from a scaling perspective, we find that efficient-attention design primarily affects how fast long-context capability emerges, while different hybrids eventually converge to comparable long-context performance under sufficient training. Second, mechanistically, we show that long-range retrieval is mainly carried by full attention, whereas efficient attention shapes its optimization trajectory. This explains a counter-intuitive phenomenon we call Large-Window Laziness: larger SWA windows can delay the formation of retrieval heads in full-attention layers. Third, guided by this mechanism, we show that applying NoPE to only the full-attention layers of a small-window SWA hybrid substantially improves long-context performance with negligible impact on short-context performance.

Large Language Models (LLMs) have exhibited impressive capabilities in reasoning and language understanding. However, their reliance on memorized knowledge and tendency to generate hallucinated content limit their reliability in real-world applications. Retrieval-Augmented Generation (RAG) mitigates these issues by integrating a retrieval module that supplements LLMs with relevant external knowledge. This paradigm bridges parametric memory and explicit retrieval, offering a principled way to ground generation in factual evidence. Despite substantial progress, most prior work has focused on optimizing isolated components, either retrieval or generation, while overlooking the agentic perspective, in which LLMs act as autonomous agents capable of actively acquiring and strategically utilizing knowledge. In this perspectives paper, we argue for reinterpreting RAG as a collaborative knowledge process among agents with distinct yet complementary roles. We categorize knowledge-intensive agents into two primary roles: knowledge acquisition (e.g., routing, query reformulation) and knowledge utilization (e.g., knowledge refinement, response generation). From this viewpoint, RAG becomes a dynamic system in which knowledge is continuously transmitted, transformed, and aligned across agent roles. To fully realize this paradigm, we advocate a joint optimization framework for knowledge-intensive agents within RAG systems. This framework explicitly models the dynamics of knowledge flow in multi-agent settings, aligning knowledge supply with knowledge demand through LLM-driven data synthesis, feedback, and evaluation. By fostering adaptive and targeted knowledge exchange, the framework mitigates conflicts between parametric and retrieved knowledge, thereby enhancing both coherence and factuality. We argue that this multi-agent joint optimization paradigm improves RAG systems in scalability, reliability, and adaptability, unlocking the potential for next-generation knowledge-intensive LLMs that reason, retrieve, and collaborate across deep retrieval processes and diverse vertical domains.

Does reinforcement learning (RL) teach large language models (LLMs) genuinely new skills, or does it merely activate existing ones? This question lies at the core of ongoing debates about the role of RL in LLM post-training. On one side, strong empirical results can be achieved with RL alone even without preceding supervised finetuning; on the other, critics argue that RL contributes little beyond reweighting existing reasoning strategies. This work provides concrete evidence that LLMs can acquire genuinely new skills during RL by composing existing ones, mirroring one of the central mechanisms by which humans acquire new cognitive skills \citep{Anderson1982Acquisition}. To mitigate data contamination and other confounding factors and to allow precise control over task complexity, we develop a synthetic framework for our investigation. Specifically, we define a skill as the ability to infer the output of a string transformation function $f(x)$ given $x$. Once an LLM has already learned $f$ and $g$ prior to RL, our experiments reveal that RL enables it to learn unseen compositions of them $h(x)=g(f(x))$. Further, this compositional ability generalizes to more difficult problems such as compositions of $>2$ functions unseen during training. Our experiments provide surprising evidence that this compositional ability, acquired on the source task, transfers to a different target task. This transfer occurs even though the model has never trained with RL on any compositional problems in the target task, as long as it has acquired the target task's atomic skills prior to RL on the source task. Our qualitative analysis shows that RL fundamentally changes the reasoning behaviors of the models. In contrast, neither of the findings is observed in next-token prediction training with the same data. Our systematic experiments provide fresh insights into the learning behaviors of widely-used post-training approaches for LLMs. They suggest the value of building base models with the necessary basic skills, followed by RL with appropriate incentivization to acquire more advanced skills that generalize better to complex and out-of-domain problems.

Generative models for speech synthesis face a fundamental trade-off: discrete tokens ensure stability but sacrifice expressivity, while continuous signals retain acoustic richness but suffer from error accumulation due to task entanglement. This challenge has driven the field towards multi-stage pipelines that rely on pre-trained discrete speech tokenizers, but these create a semantic-acoustic divide, limiting holistic and expressive speech generation. We resolve these dilemma through hierarchical semantic-acoustic modeling with semi-discrete residual representations.Our framework introduces a differentiable quantization bottleneck that induces natural specialization: a Text-Semantic Language Model (TSLM) generates semantic-prosodic plans, while a Residual Acoustic Model (RALM) recovers fine-grained acoustic details.This hierarchical semantic-acoustic representation guides a local diffusion-based decoder to generate high-fidelity speech latents. Critically, the entire architecture is trained end-to-end under a simple diffusion objective, eliminating dependency on external discrete speech tokenizers. Trained on over 1 million hours of speech, our 0.5B-parameter model achieves state-of-the-art zero-shot TTS performance among open-source systems, demonstrating that our approach delivers expressive and stable synthesis. Audio samples are available at: https://voxcpm.github.io/VoxCPM-demopage/.

We present VoxCPM2, a https://info.arxiv.org/help/prep#abstractsfully open-source multilingual and controllable speech generation foundation model that extends the hierarchical diffusion-autoregressive modeling paradigm of VoxCPM. VoxCPM2 advances the framework in three key dimensions: (i) capability, by unifying 30 languages, 9 Chinese dialects, natural-language voice design, style-controllable voice cloning, and high-fidelity continuation cloning within a single backbone; (ii) quality, through an asymmetric AudioVAE that encodes at 16 kHz and reconstructs at 48 kHz, enabling implicit super-resolution with high encoding efficiency; and (iii) scale, by jointly scaling the model to 2B parameters and the training data to over 2 million hours of multilingual speech. To support these diverse capabilities within one model, we introduce a unified sequence organization that expresses all generation modes through different arrangements of the same input building blocks, allowing joint training under a single set of parameters and objective. VoxCPM2 achieves state-of-the-art or competitive performance on public zero-shot and instruction-following TTS benchmarks. On our internal 30-language evaluation set, it attains an average WER of 1.68

A recent cutting-edge topic in multimodal modeling is to unify visual comprehension and generation within a single model. However, the two tasks demand mismatched decoding regimes and visual representations, making it non-trivial to jointly optimize within a shared feature space. In this work, we present Cheers, a unified multimodal model that decouples patch-level details from semantic representations, thereby stabilizing semantics for multimodal understanding and improving fidelity for image generation via gated detail residuals. Cheers includes three key components: (i) a unified vision tokenizer that encodes and compresses image latent states into semantic tokens for efficient LLM conditioning, (ii) an LLM-based Transformer that unifies autoregressive decoding for text generation and diffusion decoding for image generation, and (iii) a cascaded flow matching head that decodes visual semantics first and then injects semantically gated detail residuals from the vision tokenizer to refine high-frequency content. Experiments on popular benchmarks demonstrate that Cheers matches or surpasses advanced UMMs in both visual understanding and generation. Cheers also achieves 4x token compression, enabling more efficient high-resolution image encoding and generation. Notably, Cheers outperforms the Tar-1.5B on the popular benchmarks GenEval and MMBench, while requiring only 20

Large language models (LLMs) are increasingly used in medicine and clinical workflows, yet we know little about their decision and affective profiles. Taking a historically informed outlook on the future, we treated successive OpenAI models as an evolving lineage and compared them with humans in a gambling task with repeated happiness ratings. Computational analyses showed that some aspects became more human-like: newer models took more risks and displayed more human-like patterns of Pavlovian approach and avoidance. At the same time, distinctly non-human signatures emerged: loss aversion dropped below neutral levels, choices became more deterministic than in humans, affective decay increased across versions and exceeded human levels, and baseline mood remained chronically higher than in humans. These "developmental" trajectories reveal an emerging psychology of machines and have direct implications for AI ethics and for thinking about how LLMs might be integrated into clinical decision support and other high-stakes domains.

Existing methods for Large Language Models (LLMs) personalization typically rely on extensive pre-collected user data. However, practical personalization for LLMs-based chatbots often starts from a “cold-start” scenario with limited interaction history, where forming an accurate initial impression is crucial for user acquisition and retention. This critical challenge of “cold-start” personalization is further compounded by the absence of dedicated benchmarks for its evaluation. To address this gap, we introduce ColdChat, the first benchmark designed to assess LLM personalization using brief interaction histories. The collection process for ColdChat involved tasking human annotators with (1) engage in multi-session open-domain dialogues with LLM, (2) annotate personalized user profiles based on their dialogue history, and (3) label a user-specific test set for the evaluation of personalization. Our experiment on ColdChat reveals that state-of-the-art LLMs struggle to be personalized in this “cold-start” setup. To enhance LLM personalization in this scenario, we propose EPIC, a novel framework for Extracting user Profile from Interaction Context. Experimental results on ColdChat reveal that the profiles extracted by EPIC yield substantial improvements for personalization, yielding a +27.2