Think Through Uncertainty: Improving Long-Form Generation Factuality via Reasoning Calibration

Large language models (LLMs) often hallucinate in long-form generation. Existing approaches mainly improve factuality through post-hoc revision or reinforcement learning (RL) with correctness-based rewards, but they do not teach the model to estimate which parts of its generation are reliable. As a result, models may still state incorrect claims confidently in their responses. Recent advances in reasoning have significantly improved LLM performance, and have been leveraged to estimate confidence by incorporating calibration into RL objectives. However, existing approaches remain limited to a single scalar confidence for the entire response, which is insufficient for long-form generation where uncertainty varies across individual claims. To mitigate this problem, we propose CURE, a framework that improves long-form factuality by teaching LLMs to reason about uncertainty at the claim level. We first introduce a Claim-Aware Reasoning Protocol, which structures outputs into atomic claims paired with explicit confidence estimates. We then develop a multi-stage training pipeline that aligns model confidence with claims' correctness and then optimizes on factuality. The resulting calibrated confidence further enables selective prediction, allowing the model to abstain from uncertain claims at inference time. Experiments on four long-form factuality benchmarks show that CURE consistently improves factual accuracy over competitive supervised and RL baselines, while maintaining factual recall. In particular, it improves claim-level accuracy by up to 39.9% on Biography generation. These gains are accompanied by improved calibration, as reflected by a 16.0% increase in AUROC on FactBench.

TSUBASA: Improving Long-Horizon Personalization via Evolving Memory and Self-Learning with Context Distillation

Personalized large language models (PLLMs) have garnered significant attention for their ability to align outputs with individual's needs and preferences. However, they still struggle with long-horizon tasks, such as tracking a user's extensive history of conversations or activities. Existing memory mechanisms often fail to capture evolving behaviors, and RAG paradigms are trapped by a quality-efficiency tradeoff. Meanwhile, parametric adaptation is bottlenecked by train-inference gap due to the scarcity of labeled data. To enhance the long-horizon capabilities of PLLMs, we introduce TSUBASA, a two-pronged approach designed to improve memory writing via dynamic memory evolution, and memory reading via self-learning with a context distillation objective to internalize user experiences. Extensive evaluations on long-horizon benchmarks using the Qwen-3 model family (4B to 32B) validate the effectiveness of TSUBASA, surpassing competitive memory-augmented systems that rely primarily on memory writing, such as Mem0 and Memory-R1. Our analyses further confirms that TSUBASA breaks the quality-efficiency barrier to achieve Pareto improvements, delivering robust, high-fidelity personalization with a reduced token budget.

MiroEval: Benchmarking Multimodal Deep Research Agents in Process and Outcome

Recent progress in deep research systems has been impressive, but evaluation still lags behind real user needs. Existing benchmarks predominantly assess final reports using fixed rubrics, failing to evaluate the underlying research process. Most also offer limited multimodal coverage, rely on synthetic tasks that do not reflect real-world query complexity, and cannot be refreshed as knowledge evolves. To address these gaps, we introduce MiroEval, a benchmark and evaluation framework for deep research systems. The benchmark comprises 100 tasks (70 text-only, 30 multimodal), all grounded in real user needs and constructed via a dual-path pipeline that supports periodic updates, enabling a live and evolving setting. The proposed evaluation suite assesses deep research systems along three complementary dimensions: adaptive synthesis quality evaluation with task-specific rubrics, agentic factuality verification via active retrieval and reasoning over both web sources and multimodal attachments, and process-centric evaluation audits how the system searches, reasons, and refines throughout its investigation. Evaluation across 13 systems yields three principal findings: the three evaluation dimensions capture complementary aspects of system capability, with each revealing distinct strengths and weaknesses across systems; process quality serves as a reliable predictor of overall outcome while revealing weaknesses invisible to output-level metrics; and multimodal tasks pose substantially greater challenges, with most systems declining by 3 to 10 points. The MiroThinker series achieves the most balanced performance, with MiroThinker-H1 ranking the highest overall in both settings. Human verification and robustness results confirm the reliability of the benchmark and evaluation framework. MiroEval provides a holistic diagnostic tool for the next generation of deep research agents.

Wi-Spike: A Low-power WiFi Human Multi-action Recognition Model with Spiking Neural Networks

WiFi-based human action recognition (HAR) has gained significant attention due to its non-intrusive and privacy-preserving nature. However, most existing WiFi sensing models predominantly focus on improving recognition accuracy, while issues of power consumption and energy efficiency remain insufficiently discussed. In this work, we present Wi-Spike, a bio-inspired spiking neural network (SNN) framework for efficient and accurate action recognition using WiFi channel state information (CSI) signals. Specifically, leveraging the event-driven and low-power characteristics of SNNs, Wi-Spike introduces spiking convolutional layers for spatio-temporal feature extraction and a novel temporal attention mechanism to enhance discriminative representation. The extracted features are subsequently encoded and classified through spiking fully connected layers and a voting layer. Comprehensive experiments on three benchmark datasets (NTU-Fi-HAR, NTU-Fi-HumanID, and UT-HAR) demonstrate that Wi-Spike achieves competitive accuracy in single-action recognition and superior performance in multi-action recognition tasks. As for energy consumption, Wi-Spike reduces the energy cost by at least half compared with other methods, while still achieving 95.83% recognition accuracy in human activity recognition. More importantly, Wi-Spike establishes a new state-of-the-art in WiFi-based multi-action HAR, offering a promising solution for real-time, energy-efficient edge sensing applications.

Think While Watching: Online Streaming Segment-Level Memory for Multi-Turn Video Reasoning in Multimodal Large Language Models

Multimodal large language models (MLLMs) have shown strong performance on offline video understanding, but most are limited to offline inference or have weak online reasoning, making multi-turn interaction over continuously arriving video streams difficult. Existing streaming methods typically use an interleaved perception-generation paradigm, which prevents concurrent perception and generation and leads to early memory decay as streams grow, hurting long-range dependency modeling. We propose Think While Watching, a memory-anchored streaming video reasoning framework that preserves continuous segment-level memory during multi-turn interaction. We build a three-stage, multi-round chain-of-thought dataset and adopt a stage-matched training strategy, while enforcing strict causality through a segment-level streaming causal mask and streaming positional encoding. During inference, we introduce an efficient pipeline that overlaps watching and thinking and adaptively selects the best attention backend. Under both single-round and multi-round streaming input protocols, our method achieves strong results. Built on Qwen3-VL, it improves single-round accuracy by 2.6% on StreamingBench and by 3.79% on OVO-Bench. In the multi-round setting, it maintains performance while reducing output tokens by 56%. Code is available at: https://github.com/wl666hhh/Think_While_Watching/

Federated Heterogeneous Language Model Optimization for Hybrid Automatic Speech Recognition

Training automatic speech recognition (ASR) models increasingly relies on decentralized federated learning to ensure data privacy and accessibility, producing multiple local models that require effective merging. In hybrid ASR systems, while acoustic models can be merged using established methods, the language model (LM) for rescoring the N-best speech recognition list faces challenges due to the heterogeneity of non-neural n-gram models and neural network models. This paper proposes a heterogeneous LM optimization task and introduces a match-and-merge paradigm with two algorithms: the Genetic Match-and-Merge Algorithm (GMMA), using genetic operations to evolve and pair LMs, and the Reinforced Match-and-Merge Algorithm (RMMA), leveraging reinforcement learning for efficient convergence. Experiments on seven OpenSLR datasets show RMMA achieves the lowest average Character Error Rate and better generalization than baselines, converging up to seven times faster than GMMA, highlighting the paradigm's potential for scalable, privacy-preserving ASR systems.

Neural Image Space Tessellation

We present Neural Image-Space Tessellation (NIST), a lightweight screen-space post-processing approach that produces the visual effect of tessellated geometry while rendering only the original low-polygon meshes. Inspired by our observation from Phong tessellation, NIST leverages the discrepancy between geometric normals and shading normals as a minimal, view-dependent cue for silhouette refinement. At its core, NIST performs multi-scale neural tessellation by progressively deforming image-space contours with convolutional operators, while jointly reassigning appearance information through an implicit warping mechanism to preserve texture coherence and visual fidelity. Experiments demonstrate that our approach produces smooth, visually coherent silhouettes comparable to geometric tessellation, while incurring a constant per-frame cost and fully decoupled from geometric complexity, making it well-suited for large-scale real-time rendering scenarios. To the best of our knowledge, our NIST is the first work to reformulate tessellation as a post-processing operation, shifting it from a pre-rendering geometry pipeline to a screen space neural post-processing stage.

An LLM-Enabled Frequency-Aware Flow Diffusion Model for Natural-Language-Guided Power System Scenario Generation

Diverse and controllable scenario generation (e.g., wind, solar, load, etc.) is critical for robust power system planning and operation. As AI-based scenario generation methods are becoming the mainstream, existing methods (e.g., Conditional Generative Adversarial Nets) mainly rely on a fixed-length numerical conditioning vector to control the generation results, facing challenges in user conveniency and generation flexibility. In this paper, a natural-language-guided scenario generation framework, named LLM-enabled Frequency-aware Flow Diffusion (LFFD), is proposed to enable users to generate desired scenarios using plain human language. First, a pretrained LLM module is introduced to convert generation requests described by unstructured natural languages into ordered semantic space. Second, instead of using standard diffusion models, a flow diffusion model employing a rectified flow matching objective is introduced to achieve efficient and high-quality scenario generation, taking the LLM output as the model input. During the model training process, a frequency-aware multi-objective optimization algorithm is introduced to mitigate the frequency-bias issue. Meanwhile, a dual-agent framework is designed to create text-scenario training sample pairs as well as to standardize semantic evaluation. Experiments based on large-scale photovoltaic and load datasets demonstrate the effectiveness of the proposed method.

Mitigating the Safety-utility Trade-off in LLM Alignment via Adaptive Safe Context Learning

While reasoning models have achieved remarkable success in complex reasoning tasks, their increasing power necessitates stringent safety measures. For safety alignment, the core challenge lies in the inherent trade-off between safety and utility. However, prevailing alignment strategies typically construct CoT training data with explicit safety rules via context distillation. This approach inadvertently limits reasoning capabilities by creating a rigid association between rule memorization and refusal. To mitigate the safety-utility trade-off, we propose the Adaptive Safe Context Learning (ASCL) framework to improve the reasoning given proper context. ASCL formulates safety alignment as a multi-turn tool-use process, empowering the model to autonomously decide when to consult safety rules and how to generate the ongoing reasoning. Furthermore, to counteract the preference for rule consultation during RL, we introduce Inverse Frequency Policy Optimization (IFPO) to rebalance advantage estimates. By decoupling rule retrieval and subsequent reasoning, our method achieves higher overall performance compared to baselines.

Learning Ordinal Probabilistic Reward from Preferences

Reward models are crucial for aligning large language models (LLMs) with human values and intentions. Existing approaches follow either Generative (GRMs) or Discriminative (DRMs) paradigms, yet both suffer from limitations: GRMs typically demand costly point-wise supervision, while DRMs produce uncalibrated relative scores that lack probabilistic interpretation. To address these challenges, we introduce a novel reward modeling paradigm: Probabilistic Reward Model (PRM). Instead of modeling reward as a deterministic scalar, our approach treats it as a random variable, learning a full probability distribution for the quality of each response. To make this paradigm practical, we present its closed-form, discrete realization: the Ordinal Probabilistic Reward Model (OPRM), which discretizes the quality score into a finite set of ordinal ratings. Building on OPRM, we propose a data-efficient training strategy called Region Flooding Tuning (RgFT). It enables rewards to better reflect absolute text quality by incorporating quality-level annotations, which guide the model to concentrate the probability mass within corresponding rating sub-regions. Experiments on various reward model benchmarks show that our method improves accuracy by $\textbf{2.9%}\sim\textbf{7.4%}$ compared to prior reward models, demonstrating strong performance and data efficiency. Analysis of the score distribution provides evidence that our method captures not only relative rankings but also absolute quality.