Reinforcement learning (RL) has catalyzed the emergence of Large Reasoning Models (LRMs) that have pushed reasoning capabilities to new heights. While their performance has garnered significant excitement, exploring the internal mechanisms driving these behaviors has become an equally critical research frontier. This paper provides a comprehensive survey of the mechanistic understanding of LRMs, organizing recent findings into three core dimensions: 1) training dynamics, 2) reasoning mechanisms, and 3) unintended behaviors. By synthesizing these insights, we aim to bridge the gap between black-box performance and mechanistic transparency. Finally, we discuss under-explored challenges to outline a roadmap for future mechanistic studies, including the need for applied interpretability, improved methodologies, and a unified theoretical framework.

Large Language Models (LLMs)-based question answering (QA) systems play a critical role in modern AI, demonstrating strong performance across various tasks. However, LLM-generated responses often suffer from hallucinations, unfaithful statements lacking reliable references. Retrieval-Augmented Generation (RAG) frameworks enhance LLM responses by incorporating external references but also introduce new forms of hallucination due to complex interactions between the retriever and generator. To address these challenges, researchers have explored attribution-based techniques that ensure responses are verifiably supported by retrieved content. Despite progress, a unified pipeline for these techniques, along with a clear taxonomy and systematic comparison of their strengths and weaknesses, remains lacking. A well-defined taxonomy is essential for identifying specific failure modes within RAG systems, while comparative analysis helps practitioners choose appropriate solutions based on hallucination types and application context. This survey investigates how attribution-based techniques are used within RAG systems to mitigate hallucinations and addresses the gap by: (i) outlining a taxonomy of hallucination types in RAG systems, (ii) presenting a unified pipeline for attribution techniques, (iii) reviewing techniques based on the hallucinations they target, and (iv) discussing strengths and weaknesses with practical guidelines. This work offers insights for future research and practical use of attribution techniques in RAG systems.

Introduction: Large language models (LLMs) can process requests and generate texts, but their feasibility for assessing complex academic content needs further investigation. To explore LLM's potential in assisting scientific review, this study examined ChatGPT-5, Gemini-3-Pro, and Claude-Sonnet-4.5's consistency and reliability in evaluating abstracts compared to one another and to human reviewers. Methods: 160 abstracts from a local conference were graded by human reviewers and three LLMs using one rubric. Composite score distributions across three LLMs and fourteen reviewers were examined. Inter-rater reliability was calculated using intraclass correlation coefficients (ICCs) for within-AI reliability and AI-human concordance. Bland-Altman plots were examined for visual agreement patterns and systematic bias. Results: LLMs achieved good-to-excellent agreement with each other (ICCs: 0.59-0.87). ChatGPT and Claude reached moderate agreement with human reviewers on overall quality and content-specific criteria, with ICCs ~.45-.60 for composite, impression, clarity, objective, and results. They exhibited fair agreement on subjective dimensions, with ICC ranging from 0.23-0.38 for impact, engagement, and applicability. Gemini showed fair agreement on half criteria and no reliability on impact and applicability. Three LLMs showed acceptable or negligible mean difference (ChatGPT=0.24, Gemini=0.42, Claude=-0.02) from the human mean composite scores. Discussion: LLMs could process abstracts in batches with moderate agreement with human experts on overall quality and objective criteria. With appropriate process architecture, they can apply a rubric consistently across volumes of abstracts exceeding feasibility for a human rater. The weaker performance on subjective dimensions indicates that AI should serve a complementary role in evaluation, while human expertise remains essential.

Hallucinations in Large Language Models (LLMs), i.e., the tendency to generate plausible but non-factual content, pose a significant challenge for their reliable deployment in high-stakes environments. However, existing hallucination detection methods generally operate under unrealistic assumptions, i.e., either requiring expensive intensive sampling strategies for consistency checks or white-box LLM states, which are unavailable or inefficient in common API-based scenarios. To this end, we propose a novel efficient zero-shot metric called Lowest Span Confidence (LSC) for hallucination detection under minimal resource assumptions, only requiring a single forward with output probabilities. Concretely, LSC evaluates the joint likelihood of semantically coherent spans via a sliding window mechanism. By identifying regions of lowest marginal confidence across variable-length n-grams, LSC could well capture local uncertainty patterns strongly correlated with factual inconsistency. Importantly, LSC can mitigate the dilution effect of perplexity and the noise sensitivity of minimum token probability, offering a more robust estimate of factual uncertainty. Extensive experiments across multiple state-of-the-art (SOTA) LLMs and diverse benchmarks show that LSC consistently outperforms existing zero-shot baselines, delivering strong detection performance even under resource-constrained conditions.

Large language models can generate fluent peer reviews, yet their assessments often lack sufficient critical rigor when substantive issues are subtle and distributed across a paper. In this paper, we introduce PaperAudit-Bench, which consists of two components: (1) PaperAudit-Dataset, an error dataset covering both errors identifiable within individual sections and those requiring cross-section reasoning, designed for controlled evaluation under long-context settings; and (2) PaperAudit-Review, an automated review framework that integrates structured error detection with evidence-aware review generation to support critical assessment. Experiments on PaperAudit-Bench reveal large variability in error detectability across models and detection depths, highlighting the difficulty of identifying such errors under long-context settings. Relative to representative automated reviewing baselines, incorporating explicit error detection into the review workflow produces systematically stricter and more discriminative evaluations, demonstrating its suitability for peer review. Finally, we show that the dataset supports training lightweight LLM detectors via SFT and RL, enabling effective error detection at reduced computational cost.

We introduce the \emph{Arrow Language Model}, a neural architecture derived from an intuitionistic-logic interpretation of next-token prediction. Instead of representing tokens as additive embeddings mixed by attention, we encode a prefix as a \emph{left-nested implication chain} whose structure preserves order through non-commutative composition. Next-token prediction corresponds to \emph{modus ponens}, and sequence processing becomes constructive proof extension under the Curry--Howard correspondence. Our Prolog-based specialized theorem provers validate fundamental properties of the neural models, among which relations between commutative vs. non-commutative sequencing and single-token vs. multi-token prediction choices. We show that a neural architecture equivalent to multiplicative RNNs arises naturally from a proof-theoretic interpretation of next-token prediction as nested intuitionistic implication, we present a practical low-rank neural realization and position the model relative to Transformers and state-space models. Keywords: logic-based derivation of neural architectures, intuitionistic implicational logic, token-as-operator neural models, state-space models, alternatives to transformer-based foundational models.

3D scene generation is a core technology for gaming, film/VFX, and VR/AR. Growing demand for rapid iteration, high-fidelity detail, and accessible content creation has further increased interest in this area. Existing methods broadly follow two paradigms - indirect 2D-to-3D reconstruction and direct 3D generation - but both are limited by weak structural modeling and heavy reliance on large-scale ground-truth supervision, often producing structural artifacts, geometric inconsistencies, and degraded high-frequency details in complex scenes. We propose GeoDiff3D, an efficient self-supervised framework that uses coarse geometry as a structural anchor and a geometry-constrained 2D diffusion model to provide texture-rich reference images. Importantly, GeoDiff3D does not require strict multi-view consistency of the diffusion-generated references and remains robust to the resulting noisy, inconsistent guidance. We further introduce voxel-aligned 3D feature aggregation and dual self-supervision to maintain scene coherence and fine details while substantially reducing dependence on labeled data. GeoDiff3D also trains with low computational cost and enables fast, high-quality 3D scene generation. Extensive experiments on challenging scenes show improved generalization and generation quality over existing baselines, offering a practical solution for accessible and efficient 3D scene construction.

Large reasoning models (LRMs) aim to solve diverse and complex problems through structured reasoning. Recent advances in group-based policy optimization methods have shown promise in enabling stable advantage estimation without reliance on process-level annotations. However, these methods rely on advantage gaps induced by high-quality samples within the same batch, which makes the training process fragile and inefficient when intra-group advantages collapse under challenging tasks. To address these problems, we propose a reinforcement learning mechanism named \emph{\textbf{R^3}} that along three directions: (1) a \emph{cross-context \underline{\textbf{R}}eplay} strategy that maintains the intra-group advantage by recalling valuable examples from historical trajectories of the same query, (2) an \emph{in-context self-\underline{\textbf{R}}eflection} mechanism enabling models to refine outputs by leveraging past failures, and (3) a \emph{structural entropy \underline{\textbf{R}}anking reward}, which assigns relative rewards to truncated or failed samples by ranking responses based on token-level entropy patterns, capturing both local exploration and global stability. We implement our method on Deepseek-R1-Distill-Qwen-1.5B and train it on the DeepscaleR-40k in the math domain. Experiments demonstrate our method achieves SoTA performance on several math benchmarks, representing significant improvements and fewer reasoning tokens over the base models. Code and model will be released.

We present a fast 3DGS reconstruction pipeline designed to converge within one minute, developed for the SIGGRAPH Asia 3DGS Fast Reconstruction Challenge. The challenge consists of an initial round using SLAM-generated camera poses (with noisy trajectories) and a final round using COLMAP poses (highly accurate). To robustly handle these heterogeneous settings, we develop a two-stage solution. In the first round, we use reverse per-Gaussian parallel optimization and compact forward splatting based on Taming-GS and Speedy-splat, load-balanced tiling, an anchor-based Neural-Gaussian representation enabling rapid convergence with fewer learnable parameters, initialization from monocular depth and partially from feed-forward 3DGS models, and a global pose refinement module for noisy SLAM trajectories. In the final round, the accurate COLMAP poses change the optimization landscape; we disable pose refinement, revert from Neural-Gaussians back to standard 3DGS to eliminate MLP inference overhead, introduce multi-view consistency-guided Gaussian splitting inspired by Fast-GS, and introduce a depth estimator to supervise the rendered depth. Together, these techniques enable high-fidelity reconstruction under a strict one-minute budget. Our method achieved the top performance with a PSNR of 28.43 and ranked first in the competition.

Multi-Agent Path Finding (MAPF) is an NP-hard problem with applications in warehouse automation and multi-robot coordination. Learning-based MAPF solvers offer fast and scalable planning but often produce feasible trajectories that contain unnecessary or oscillatory movements. We propose Judgelight, a post-optimization layer that improves trajectory quality after a MAPF solver generates a feasible schedule. Judgelight collapses closed subwalks in agents' trajectories to remove redundant movements while preserving all feasibility constraints. We formalize this process as MAPF-Collapse, prove that it is NP-hard, and present an exact optimization approach by formulating it as integer linear programming (ILP) problem. Experimental results show Judgelight consistently reduces solution cost by around 20%, particularly for learning-based solvers, producing trajectories that are better suited for real-world deployment.

Using multiple open-access models trained on public datasets, we developed Tri-Reader, a comprehensive, freely available pipeline that integrates lung segmentation, nodule detection, and malignancy classification into a unified tri-stage workflow. The pipeline is designed to prioritize sensitivity while reducing the candidate burden for annotators. To ensure accuracy and generalizability across diverse practices, we evaluated Tri-Reader on multiple internal and external datasets as compared with expert annotations and dataset-provided reference standards.

Large Language Models (LLMs) have recently demonstrated remarkable reasoning abilities, yet hallucinate on knowledge-intensive tasks. Retrieval-augmented generation (RAG) mitigates this issue by grounding answers in external sources, e.g., knowledge graphs (KGs). However, existing KG-based RAG approaches rely on semantics-unaware path sampling and are weakly aligned with KG reasoning objectives, which limits further accuracy gains. They also feed retrieved paths directly into the reasoner without organizing them into answer-centered reasoning paths, hindering small LLMs' ability to leverage the retrieved knowledge. Furthermore, prior works predominantly rely on large LLMs (e.g., ChatGPT/GPT-4) or assume backbones above 7B parameters, leaving sub-7B models underexplored. We address this gap with RPO-RAG, the first KG-based RAG framework specifically designed for small LLMs, to the best of our knowledge. RPO-RAG introduces three key innovations: (1) a query-path semantic sampling strategy that provides informative supervisory signals; (2) a relation-aware preference optimization that aligns training with intermediate KG reasoning signals (e.g., relation); and (3) an answer-centered prompt design that organizes entities and reasoning paths in an interpretable format. Extensive experiments on two benchmark Knowledge Graph Question Answering (KGQA) datasets, WebQSP and CWQ, demonstrate that RPO-RAG effectively bridges the performance gap between small and large language models. On WebQSP, it improves F1 by up to 8.8%, reflecting enhanced answer precision, while on CWQ it achieves new state-of-the-art results among models under 8B parameters in both Hit and F1. Overall, RPO-RAG substantially improves the reasoning capability of small LLMs, even under 3B parameters-highlighting their potential for resource-efficient and practical on-device KGQA applications.

Improper exposure often leads to severe loss of details, color distortion, and reduced contrast. Exposure correction still faces two critical challenges: (1) the ignorance of object-wise regional semantic information causes the color shift artifacts; (2) real-world exposure images generally have no ground-truth labels, and its labeling entails massive manual editing. To tackle the challenges, we propose a new unsupervised semantic-aware exposure correction network. It contains an adaptive semantic-aware fusion module, which effectively fuses the semantic information extracted from a pre-trained Fast Segment Anything Model into a shared image feature space. Then the fused features are used by our multi-scale residual spatial mamba group to restore the details and adjust the exposure. To avoid manual editing, we propose a pseudo-ground truth generator guided by CLIP, which is fine-tuned to automatically identify exposure situations and instruct the tailored corrections. Also, we leverage the rich priors from the FastSAM and CLIP to develop a semantic-prompt consistency loss to enforce semantic consistency and image-prompt alignment for unsupervised training. Comprehensive experimental results illustrate the effectiveness of our method in correcting real-world exposure images and outperforms state-of-the-art unsupervised methods both numerically and visually.

This paper presents a Quantum Reinforcement Learning (QRL) solution to the dynamic portfolio optimization problem based on Variational Quantum Circuits. The implemented QRL approaches are quantum analogues of the classical neural-network-based Deep Deterministic Policy Gradient and Deep Q-Network algorithms. Through an empirical evaluation on real-world financial data, we show that our quantum agents achieve risk-adjusted performance comparable to, and in some cases exceeding, that of classical Deep RL models with several orders of magnitude more parameters. However, while quantum circuit execution is inherently fast at the hardware level, practical deployment on cloud-based quantum systems introduces substantial latency, making end-to-end runtime currently dominated by infrastructural overhead and limiting practical applicability. Taken together, our results suggest that QRL is theoretically competitive with state-of-the-art classical reinforcement learning and may become practically advantageous as deployment overheads diminish. This positions QRL as a promising paradigm for dynamic decision-making in complex, high-dimensional, and non-stationary environments such as financial markets. The complete codebase is released as open source at: https://github.com/VincentGurgul/qrl-dpo-public

Triad magnetometer components change with sensor attitude even when the IGRF total intensity target stays invariant. NavFormer forecasts this invariant target with rotation invariant scalar features and a Canonical SPD module that stabilizes the spectrum of window level second moments of the triads without sign discontinuities. The module builds a canonical frame from a Gram matrix per window and applies state dependent spectral scaling in the original coordinates. Experiments across five flights show lower error than strong baselines in standard training, few shot training, and zero shot transfer. The code is available at: https://anonymous.4open.science/r/NavFormer-Robust-IGRF-Forecasting-for-Autonomous-Navigators-0765

When humans face problems beyond their immediate capabilities, they rely on tools, providing a promising paradigm for improving visual reasoning in multimodal large language models (MLLMs). Effective reasoning, therefore, hinges on knowing which tools to use, when to invoke them, and how to compose them over multiple steps, even when faced with new tools or new tasks. We introduce \textbf{AdaReasoner}, a family of multimodal models that learn tool use as a general reasoning skill rather than as tool-specific or explicitly supervised behavior. AdaReasoner is enabled by (i) a scalable data curation pipeline exposing models to long-horizon, multi-step tool interactions; (ii) Tool-GRPO, a reinforcement learning algorithm that optimizes tool selection and sequencing based on end-task success; and (iii) an adaptive learning mechanism that dynamically regulates tool usage. Together, these components allow models to infer tool utility from task context and intermediate outcomes, enabling coordination of multiple tools and generalization to unseen tools. Empirically, AdaReasoner exhibits strong tool-adaptive and generalization behaviors: it autonomously adopts beneficial tools, suppresses irrelevant ones, and adjusts tool usage frequency based on task demands, despite never being explicitly trained to do so. These capabilities translate into state-of-the-art performance across challenging benchmarks, improving the 7B base model by +24.9\% on average and surpassing strong proprietary systems such as GPT-5 on multiple tasks, including VSP and Jigsaw.

We introduce GenAgent, unifying visual understanding and generation through an agentic multimodal model. Unlike unified models that face expensive training costs and understanding-generation trade-offs, GenAgent decouples these capabilities through an agentic framework: understanding is handled by the multimodal model itself, while generation is achieved by treating image generation models as invokable tools. Crucially, unlike existing modular systems constrained by static pipelines, this design enables autonomous multi-turn interactions where the agent generates multimodal chains-of-thought encompassing reasoning, tool invocation, judgment, and reflection to iteratively refine outputs. We employ a two-stage training strategy: first, cold-start with supervised fine-tuning on high-quality tool invocation and reflection data to bootstrap agent behaviors; second, end-to-end agentic reinforcement learning combining pointwise rewards (final image quality) and pairwise rewards (reflection accuracy), with trajectory resampling for enhanced multi-turn exploration. GenAgent significantly boosts base generator(FLUX.1-dev) performance on GenEval++ (+23.6\%) and WISE (+14\%). Beyond performance gains, our framework demonstrates three key properties: 1) cross-tool generalization to generators with varying capabilities, 2) test-time scaling with consistent improvements across interaction rounds, and 3) task-adaptive reasoning that automatically adjusts to different tasks. Our code will be available at \href{https://github.com/deep-kaixun/GenAgent}{this url}.

Large language models are increasingly applied to multi-document and long-form inputs, yet long-context inference remains memory- and noise-inefficient. Key-value (KV) caching scales linearly with context length, while external retrieval methods often return lexically similar but causally irrelevant passages. We present S3-Attention, a memory-first inference-time framework that treats long-context processing as attention-aligned endogenous retrieval. S3-Attention decodes transient key and query projections into top-k sparse feature identifiers using lightweight sparse autoencoders, and constructs a CPU-based inverted index mapping features to token positions or spans during a single streaming scan. This design allows the KV cache to be discarded entirely and bounds GPU memory usage by the scan chunk size. At generation time, feature co-activation is used to retrieve compact evidence spans, optionally fused with BM25 for exact lexical matching. Under a unified LongBench evaluation protocol with fixed prompting, decoding, and matched token budgets, S3-Hybrid closely matches full-context inference across multiple model families and improves robustness in several information-dense settings. We also report an engineering limitation of the current prototype, which incurs higher wall-clock latency than optimized full-KV baselines, motivating future kernel-level optimization.

Learning systems acquire structured internal representations from data, yet classical information-theoretic results state that deterministic transformations do not increase information. This raises a fundamental question: how can learning produce abstraction and insight without violating information-theoretic limits? We argue that learning is inherently an irreversible process when performed over finite time, and that the realization of epistemic structure necessarily incurs entropy production. To formalize this perspective, we model learning as a transport process in the space of probability distributions over model configurations and introduce an epistemic free-energy framework. Within this framework, we define the free-energy reduction as a bookkeeping quantity that records the total reduction of epistemic free energy along a learning trajectory. This formulation highlights that realizing such a reduction over finite time necessarily incurs irreversible entropy production. We then derive the Epistemic Speed Limit (ESL), a finite-time inequality that lower-bounds the minimal entropy production required by any learning process to realize a given distributional transformation. This bound depends only on the Wasserstein distance between initial and final ensemble distributions and is independent of the specific learning algorithm.

Audio codecs power discrete music generative modelling, music streaming and immersive media by shrinking PCM audio to bandwidth-friendly bit-rates. Recent works have gravitated towards processing in the spectral domain; however, spectrogram-domains typically struggle with phase modeling which is naturally complex-valued. Most frequency-domain neural codecs either disregard phase information or encode it as two separate real-valued channels, limiting spatial fidelity. This entails the need to introduce adversarial discriminators at the expense of convergence speed and training stability to compensate for the inadequate representation power of the audio signal. In this work we introduce an end-to-end complex-valued RVQ-VAE audio codec that preserves magnitude-phase coupling across the entire analysis-quantization-synthesis pipeline and removes adversarial discriminators and diffusion post-filters. Without GANs or diffusion we match or surpass much longer-trained baselines in-domain and reach SOTA out-of-domain performance. Compared to standard baselines that train for hundreds of thousands of steps, our model reducing training budget by an order of magnitude is markedly more compute-efficient while preserving high perceptual quality.

The quadratic complexity of standard attention mechanisms poses a significant scalability bottleneck for large language models (LLMs) in long-context scenarios. While hybrid attention strategies that combine sparse and full attention within a single model offer a viable solution, they typically employ static computation ratios (i.e., fixed proportions of sparse versus full attention) and fail to adapt to the varying sparsity sensitivities of downstream tasks during inference. To address this issue, we propose Elastic Attention, which allows the model to dynamically adjust its overall sparsity based on the input. This is achieved by integrating a lightweight Attention Router into the existing pretrained model, which dynamically assigns each attention head to different computation modes. Within only 12 hours of training on 8xA800 GPUs, our method enables models to achieve both strong performance and efficient inference. Experiments across three long-context benchmarks on widely-used LLMs demonstrate the superiority of our method.

Intelligent Transportation Systems (ITS) demand real-time collision prediction to ensure road safety and reduce accident severity. Conventional approaches rely on transmitting raw video or high-dimensional sensory data from roadside units (RSUs) to vehicles, which is impractical under vehicular communication bandwidth and latency constraints. In this work, we propose a semantic V2X framework in which RSU-mounted cameras generate spatiotemporal semantic embeddings of future frames using the Video Joint Embedding Predictive Architecture (V-JEPA). To evaluate the system, we construct a digital twin of an urban traffic environment enabling the generation of d verse traffic scenarios with both safe and collision events. These embeddings of the future frame, extracted from V-JEPA, capture task-relevant traffic dynamics and are transmitted via V2X links to vehicles, where a lightweight attentive probe and classifier decode them to predict imminent collisions. By transmitting only semantic embeddings instead of raw frames, the proposed system significantly reduces communication overhead while maintaining predictive accuracy. Experimental results demonstrate that the framework with an appropriate processing method achieves a 10% F1-score improvement for collision prediction while reducing transmission requirements by four orders of magnitude compared to raw video. This validates the potential of semantic V2X communication to enable cooperative, real-time collision prediction in ITS.

We present a novel approach for extracting 3D atomic-level information from transmission electron microscopy (TEM) images affected by significant noise. The approach is based on formulating depth estimation as a semantic segmentation problem. We address the resulting segmentation problem by training a deep convolutional neural network to generate pixel-wise depth segmentation maps using simulated data corrupted by synthetic noise. The proposed method was applied to estimate the depth of atomic columns in CeO2 nanoparticles from simulated images and real-world TEM data. Our experiments show that the resulting depth estimates are accurate, calibrated and robust to noise.

Adapting large pre-trained language models to downstream tasks often entails fine-tuning millions of parameters or deploying costly dense weight updates, which hinders their use in resource-constrained environments. Low-rank Adaptation (LoRA) reduces trainable parameters by factorizing weight updates, yet the underlying dense weights still impose high storage and computation costs. Magnitude-based pruning can yield sparse models but typically degrades LoRA's performance when applied naively. In this paper, we introduce SALR (Sparsity-Aware Low-Rank Representation), a novel fine-tuning paradigm that unifies low-rank adaptation with sparse pruning under a rigorous mean-squared-error framework. We prove that statically pruning only the frozen base weights minimizes the pruning error bound, and we recover the discarded residual information via a truncated-SVD low-rank adapter, which provably reduces per-entry MSE by a factor of $(1 - r/\min(d,k))$. To maximize hardware efficiency, we fuse multiple low-rank adapters into a single concatenated GEMM, and we adopt a bitmap-based encoding with a two-stage pipelined decoding + GEMM design to achieve true model compression and speedup. Empirically, SALR attains 50\% sparsity on various LLMs while matching the performance of LoRA on GSM8K and MMLU, reduces model size by $2\times$, and delivers up to a $1.7\times$ inference speedup.

As large language models (LLMs) are increasingly applied to legal domain-specific tasks, evaluating their ability to perform legal work in real-world settings has become essential. However, existing legal benchmarks rely on simplified and highly standardized tasks, failing to capture the ambiguity, complexity, and reasoning demands of real legal practice. Moreover, prior evaluations often adopt coarse, single-dimensional metrics and do not explicitly assess fine-grained legal reasoning. To address these limitations, we introduce PLawBench, a Practical Law Benchmark designed to evaluate LLMs in realistic legal practice scenarios. Grounded in real-world legal workflows, PLawBench models the core processes of legal practitioners through three task categories: public legal consultation, practical case analysis, and legal document generation. These tasks assess a model's ability to identify legal issues and key facts, perform structured legal reasoning, and generate legally coherent documents. PLawBench comprises 850 questions across 13 practical legal scenarios, with each question accompanied by expert-designed evaluation rubrics, resulting in approximately 12,500 rubric items for fine-grained assessment. Using an LLM-based evaluator aligned with human expert judgments, we evaluate 10 state-of-the-art LLMs. Experimental results show that none achieves strong performance on PLawBench, revealing substantial limitations in the fine-grained legal reasoning capabilities of current LLMs and highlighting important directions for future evaluation and development of legal LLMs. Data is available at: https://github.com/skylenage/PLawbench.

Many spoken languages, including English, exhibit wide variation in dialects and accents, making accent control an important capability for flexible text-to-speech (TTS) models. Current TTS systems typically generate accented speech by conditioning on speaker embeddings associated with specific accents. While effective, this approach offers limited interpretability and controllability, as embeddings also encode traits such as timbre and emotion. In this study, we analyze the interaction between speaker embeddings and linguistically motivated phonological rules in accented speech synthesis. Using American and British English as a case study, we implement rules for flapping, rhoticity, and vowel correspondences. We propose the phoneme shift rate (PSR), a novel metric quantifying how strongly embeddings preserve or override rule-based transformations. Experiments show that combining rules with embeddings yields more authentic accents, while embeddings can attenuate or overwrite rules, revealing entanglement between accent and speaker identity. Our findings highlight rules as a lever for accent control and a framework for evaluating disentanglement in speech generation.

Time series data play a critical role in various fields, including finance, healthcare, marketing, and engineering. A wide range of techniques (from classical statistical models to neural network-based approaches such as Long Short-Term Memory (LSTM)) have been employed to address time series forecasting challenges. In this paper, we reframe time series forecasting as a two-part task: (1) predicting the trend (directional movement) of the time series at the next time step, and (2) forecasting the quantitative value at the next time step. The trend can be predicted using a binary classifier, while quantitative values can be forecasted using models such as LSTM and Bidirectional Long Short-Term Memory (Bi-LSTM). Building on this reframing, we propose the Trend-Adjusted Time Series (TATS) model, which adjusts the forecasted values based on the predicted trend provided by the binary classifier. We validate the proposed approach through both theoretical analysis and empirical evaluation. The TATS model is applied to a volatile financial time series (the daily gold price) with the objective of forecasting the next days price. Experimental results demonstrate that TATS consistently outperforms standard LSTM and Bi-LSTM models by achieving significantly lower forecasting error. In addition, our results indicate that commonly used metrics such as MSE and MAE are insufficient for fully assessing time series model performance. Therefore, we also incorporate trend detection accuracy, which measures how effectively a model captures trends in a time series.

Constant rescaling of a Riemannian metric appears in many computational settings, often through a global scale parameter that is introduced either explicitly or implicitly. Although this operation is elementary, its consequences are not always made clear in practice and may be confused with changes in curvature, manifold structure, or coordinate representation. In this note we provide a short, self-contained account of constant metric scaling on arbitrary Riemannian manifolds. We distinguish between quantities that change under such a scaling, including norms, distances, volume elements, and gradient magnitudes, and geometric objects that remain invariant, such as the Levi--Civita connection, geodesics, exponential and logarithmic maps, and parallel transport. We also discuss implications for Riemannian optimization, where constant metric scaling can often be interpreted as a global rescaling of step sizes rather than a modification of the underlying geometry. The goal of this note is purely expository and is intended to clarify how a global metric scale parameter can be introduced in Riemannian computation without altering the geometric structures on which these methods rely.

Large Language Models (LLMs) have achieved remarkable strides in multilingual translation but are hindered by a systemic cross-lingual verbosity bias, rendering them unsuitable for strict time-constrained tasks like subtitling and dubbing. Current prompt-engineering approaches struggle to resolve this conflict between semantic fidelity and rigid temporal feasibility. To bridge this gap, we first introduce Sand-Glass, a benchmark specifically designed to evaluate translation under syllable-level duration constraints. Furthermore, we propose HOMURA, a reinforcement learning framework that explicitly optimizes the trade-off between semantic preservation and temporal compliance. By employing a KL-regularized objective with a novel dynamic syllable-ratio reward, HOMURA effectively "tames" the output length. Experimental results demonstrate that our method significantly outperforms strong LLM baselines, achieving precise length control that respects linguistic density hierarchies without compromising semantic adequacy.

In this work, we provide a new convergence theory for plug-and-play proximal gradient descent (PnP-PGD) under prior mismatch where the denoiser is trained on a different data distribution to the inference task at hand. To the best of our knowledge, this is the first convergence proof of PnP-PGD under prior mismatch. Compared with the existing theoretical results for PnP algorithms, our new results removed the need for several restrictive and unverifiable assumptions. Moreover, we derive the convergence theory for equivariant PnP (EPnP) under the prior mismatch setting, proving that EPnP reduces error variance and explicitly tightens the convergence bound.