Large Language Models (LLMs) employ multi-turn interaction as a fundamental paradigm for completing complex tasks. However, their performance often degrades in extended interactions, as they are typically trained on static, single-turn data, which hinders their ability to adapt to real-time user feedback. To address this limitation, we first propose a new paradigm: Test-Time Policy Adaptation for Multi-Turn Interactions (T2PAM), which utilizes user feedback from the ongoing interaction as a reward signal to estimate a latent optimal policy aligned with user preferences, then updates a small subset of parameters to steer the model toward this policy, ultimately enabling efficient in-conversation self-correction. We then introduce Optimum-Referenced One-Step Adaptation (ROSA), a lightweight algorithm that operationalizes T2PAM. ROSA guides the model parameters toward a theoretical optimal policy in a single, efficient update step, avoiding costly iterative gradient-based optimization and minimizing computational overhead. We provide a rigorous theoretical analysis guaranteeing that the policy of ROSA converges to the preference of user as the number of interactions increases. Extensive experiments on challenging benchmark demonstrate that ROSA achieves significant improvements in both task effectiveness and efficiency.

Test-time scaling via solution sampling and aggregation has become a key paradigm for improving the reasoning performance of Large Language Models (LLMs). While reward model selection is commonly employed in this approach, it often fails to identify minority-yet-correct answers, which limits its effectiveness beyond that of simple majority voting. We argue that this limitation stems from a lack of informative critique signals during verifier training. To bridge this gap, we introduce Mirror-Critique, a framework that trains a verifier with informative critiques. Our key insight is to leverage the rich critique signal by contrasting model-generated solutions with ground-truth solutions. We deploy a small instruction-tuned model to synthesize high-quality critique data with rejection sampling that teaches the verifier not only what is wrong, but also why. The synthetic data is used to cold-start the LLMs in the RL VR process to further improve the verification ability. The resulting Mirror-V erifier is deployed to evaluate candidate solutions by generating multiple critiques per solution, aggregating them into a verify score used for weighted voting or selective abstention. The experimental results show that our Mirror-V erifier significantly outperforms majority voting in terms of solution accuracy and also improves the solver's honesty to recognize and abstain from answering beyond its capability boundaries. Reinforcement Learning with V erifiable Reward (RL VR) has emerged as a powerful method for training Large Language Models (LLMs) to perform complex reasoning tasks, enabling significant improvements in domains such as mathematics, code generation, and scientific problem-solving.

Tree search has recently emerged as a powerful framework for aligning generative models with task-specific rewards at test time. Applying tree search to Masked Diffusion Language Models, however, introduces two key challenges: (i) parallel unmasking yields highly correlated branches, limiting exploration, and (ii) reward evaluation via sampled completions produces high-variance estimates, making pruning unstable. Theoretically, we quantify branching efficiency gains in NFEs (number of function evaluations), show that the scoring rule approximates the true reward with error bounded by predictive uncertainty, and prove improvements with larger tree widths. Masked Diffusion Language Models (MDLMs) (Nie et al., 2025; Sahoo et al., 2024; Shi et al., 2024; Y ang et al., 2025b) have emerged as a compelling alternative to autoregressive models (Brown et al., 2020; Radford et al., 2019; Touvron et al., 2023). They start with all-mask tokens and gradually reveal tokens through a sequence of discrete denoising steps. At each step, the model predicts token distributions for masked positions, conditioned on the current partially masked sequence and the diffusion timestep. This formulation enables flexible sampling schedules and broad conditioning patterns, making MDLMs well-suited for controllable generation tasks.Figure 1: Conceptual illustration of TR However, this flexibility is not fully realized without mechanisms to align the model's outputs with user-defined objectives. Test-Time Alignment (TT A) enables guiding language model outputs toward task-specific goals without retraining. In applications such as toxicity avoidance (Logacheva et al., 2022), sentiment control (Barbieri et al., 2020), or enforcing linguistic acceptability (Warstadt et al., 2019), aligning generation with external reward functions at test time offers a flexible and training-free alternative to supervised fine-tuning.

Transformer-based architectures dominate time series modeling by enabling global attention over all timestamps, yet their rigid 'one-size-fits-all' context aggregation fails to address two critical challenges in real-world data: (1) inherent lag effects, where the relevance of historical timestamps to a query varies dynamically; (2) anomalous segments, which introduce noisy signals that degrade forecasting accuracy. To resolve these problems, we propose the Temporal Mix of Experts (TMOE), a novel attention-level mechanism that reimagines key-value (K-V) pairs as local experts (each specialized in a distinct temporal context) and performs adaptive expert selection for each query via localized filtering of irrelevant timestamps. Complementing this local adaptation, a shared global expert preserves the Transformer's strength in capturing long-range dependencies. We then replace the vanilla attention mechanism in popular time-series Transformer frameworks (i.e., PatchTST and Timer) with TMOE, without extra structural modifications, yielding our specific version TimeExpert and general version TimeExpert-G. Extensive experiments on seven real-world long-term forecasting benchmarks demonstrate that TimeExpert and TimeExpert-G outperform state-of-the-art methods. Code is available at https://github.com/xwmaxwma/TimeExpert.

Recent advancements have endowed Large Language Models (LLMs) with impressive general reasoning capabilities, yet they often struggle with personalization reasoning - the crucial ability to analyze user history, infer unique preferences, and generate tailored responses. To address this limitation, we introduce TagPR, a novel training framework that significantly enhances an LLM's intrinsic capacity for personalization reasoning through a tagging the thought approach. Our method first develops a data-driven pipeline to automatically generate and semantically label reasoning chains, creating a structured dataset that fosters interpretable reasoning. We then propose a synergistic training strategy that begins with Supervised Fine-Tuning (SFT) on this tagged data to establish foundational reasoning patterns, followed by a multi-stage reinforcement learning (RL) process. This RL phase is guided by a unique composite reward signal, which integrates tag-based constraints and a novel Personalization Reward Model with User Embeddings (PRMU) to achieve fine-grained alignment with user-specific logic. Extensive experiments on the public LaMP benchmark and a self-constructed dataset demonstrate that our approach achieves state-of-the-art results, delivering an average improvement of 32.65% over the base model across all tasks. Our work validates that structured, interpretable reasoning is a highly effective pathway to unlocking genuine personalization capabilities in LLMs.

Implicit Neural Representations (INRs) have emerged as a transformative paradigm in signal processing and computer vision, excelling in tasks from image reconstruction to 3D shape modeling. Yet their effectiveness is fundamentally limited by the absence of principled strategies for optimal configuration - spanning activation selection, initialization scales, layer-wise adaptation, and their intricate interdependencies. These choices dictate performance, stability, and generalization, but current practice relies on ad-hoc heuristics, brute-force grid searches, or task-specific tuning, often leading to inconsistent results across modalities. This work introduces OptiINR, the first unified framework that formulates INR configuration as a rigorous optimization problem. Leveraging Bayesian optimization, OptiINR efficiently explores the joint space of discrete activation families - such as sinusoidal (SIREN), wavelet-based (WIRE), and variable-periodic (FINER) - and their associated continuous initialization parameters. This systematic approach replaces fragmented manual tuning with a coherent, data-driven optimization process. By delivering globally optimal configurations, OptiINR establishes a principled foundation for INR design, consistently maximizing performance across diverse signal processing applications.

Conventional Wi-Fi received signal strength indicator (RSSI) fingerprinting cannot meet the growing demand for accurate indoor localization and navigation due to its lower accuracy, while solutions based on light detection and ranging (LiDAR) can provide better localization performance but is limited by their higher deployment cost and complexity. To address these issues, we propose a novel indoor localization and navigation framework integrating Wi-Fi RSSI fingerprinting, LiDAR-based simultaneous localization and mapping (SLAM), and inertial measurement unit (IMU) navigation based on an extended Kalman filter (EKF). Specifically, coarse localization by deep neural network (DNN)-based Wi-Fi RSSI fingerprinting is refined by IMU-based dynamic positioning using a Gmapping-based SLAM to generate an occupancy grid map and output high-frequency attitude estimates, which is followed by EKF prediction-update integrating sensor information while effectively suppressing Wi-Fi-induced noise and IMU drift errors. Multi-group real-world experiments conducted on the IR building at Xi'an Jiaotong-Liverpool University demonstrates that the proposed multi-sensor fusion framework suppresses the instability caused by individual approaches and thereby provides stable accuracy across all path configurations with mean two-dimensional (2D) errors ranging from 0.2449 m to 0.3781 m. In contrast, the mean 2D errors of Wi-Fi RSSI fingerprinting reach up to 1.3404 m in areas with severe signal interference, and those of LiDAR/IMU localization are between 0.6233 m and 2.8803 m due to cumulative drift.

Large Language Model (LLM)-based systems present new opportunities for autonomous health monitoring in sensor-rich industrial environments. This study explores the potential of LLMs to detect and classify faults directly from sensor data, while producing inherently explainable outputs through natural language reasoning. We systematically evaluate how LLM-system architecture (single-LLM vs. multi-LLM), input representations (raw vs. descriptive statistics), and context window size affect diagnostic performance. Our findings show that LLM systems perform most effectively when provided with summarized statistical inputs, and that systems with multiple LLMs using specialized prompts offer improved sensitivity for fault classification compared to single-LLM systems. While LLMs can produce detailed and human-readable justifications for their decisions, we observe limitations in their ability to adapt over time in continual learning settings, often struggling to calibrate predictions during repeated fault cycles. These insights point to both the promise and the current boundaries of LLM-based systems as transparent, adaptive diagnostic tools in complex environments.

A fundamental reason for the dominance of attention over RNNs and LSTMs in LLMs is its ability to capture long-range dependencies by modeling direct interactions between all tokens, overcoming the sequential limitations of recurrent architectures. Similarly, a key reason why today's vision language models (VLMs) hallucinate and underperform pure language models is that they rely on direct concatenation of image and text tokens with a modality-blinded positional encoding, which conveniently adopts the pretrained LLM backbone but forces unnecessary long-distance attention between semantically related tokens across modalities. This underscores the urgent need for mechanisms that efficiently enhance token locality and cross-modal alignment. In response, we propose Attention Anchor, a parameter-free framework that efficiently groups semantically similar tokens across modalities, improving cross-modal locality. By inserting text tokens near relevant visual patches, we create semantic signposts that reveal true content-based cross-modal attention scores, guiding the model to focus on the correct image regions for tasks such as VQA, MMBench and POPE. This improves answer accuracy and reduces hallucinations without disrupting the prompt's semantic flow. AttAnchor achieves improvements across 13 out of 15 different metrics and benchmarks, including up to 32% gains on reasoning tasks and up to 15% improvements on hallucination benchmarks. AttAnchor enables TinyLLaVA 1B to outperform much larger models like LLaVA 7B and QwenVL 3B on POPE with only 0.1% inference time overhead. To the best of our knowledge, this work is among the first to investigate mixed-modal token grouping, where text and image tokens are clustered jointly into shared groups rather than being grouped within a single modality or merely aligned post-hoc with additional alignment losses.

The Muon optimizer, based on matrix orthogonalization, has recently shown faster convergence and up to 2x computational efficiency over AdamW in LLM pretraining. Like AdamW, Muon is stateful, requiring storage of both model weights and accumulated gradients. While 8-bit AdamW variants mitigate this overhead using blockwise quantization, they are typically stable only under dynamic quantization - which improves stability on linear quantization for extreme values. In this paper, we introduce the 8-bit Muon optimizer using blockwise quantization, supporting both linear and dynamic schemes. We demonstrate that 8-bit Muon maintains stability under both, while delivering $\sim$74\% reduction in memory footprint compared to full-precision Muon. In extensive experiments, 8-bit Muon closely matches the performance of Muon while outperforming AdamW and 8-bit AdamW in pre-training a 1.6B model on 4B FineWeb tokens. It also shows competitive results when fine-tuning the Llama 3.2 3B model on post-training data. We also provide a theoretical perspective to help explain this robustness under quantization.