Flatness of the loss landscape has been widely studied as an important perspective for understanding the behavior and generalization of deep learning algorithms. Motivated by this view, we propose Flatness-Aware Stochastic Gradient Langevin Dynamics (fSGLD), a first-order optimization method that biases learning its dynamics toward flat basins while retaining the computational and memory efficiency of SGD and SGLD. We provide a non-asymptotic theoretical analysis showing that fSGLD targets a flatness-biased Gibbs distribution under a theoretically prescribed coupling between the noise scale $σ$ and the inverse temperature $β$, together with explicit excess risk guarantees. We empirically evaluate fSGLD across standard optimizer benchmarks, Bayesian image classification, uncertainty quantification, and out-of-distribution detection, demonstrating consistently strong performance and reliable uncertainty estimates. Additional experiments confirm the effectiveness of the theoretically prescribed $β$-$σ$ coupling compared to decoupled choices.

In the era of increasingly complex AI models for time series forecasting, progress is often measured by marginal improvements on benchmark leaderboards. However, this approach suffers from a fundamental flaw: standard evaluation metrics conflate a model's performance with the data's intrinsic unpredictability. To address this pressing challenge, we introduce a novel, predictability-aligned diagnostic framework grounded in spectral coherence. Our framework makes two primary contributions: the Spectral Coherence Predictability (SCP), a computationally efficient ($O(N\log N)$) and task-aligned score that quantifies the inherent difficulty of a given forecasting instance, and the Linear Utilization Ratio (LUR), a frequency-resolved diagnostic tool that precisely measures how effectively a model exploits the linearly predictable information within the data. We validate our framework's effectiveness and leverage it to reveal two core insights. First, we provide the first systematic evidence of "predictability drift", demonstrating that a task's forecasting difficulty varies sharply over time. Second, our evaluation reveals a key architectural trade-off: complex models are superior for low-predictability data, whereas linear models are highly effective on more predictable tasks. We advocate for a paradigm shift, moving beyond simplistic aggregate scores toward a more insightful, predictability-aware evaluation that fosters fairer model comparisons and a deeper understanding of model behavior.

Large Language Models (LLMs) often assign disproportionate attention to the first token, a phenomenon known as the attention sink. Several recent approaches aim to address this issue, including Sink Attention in GPT-OSS and Gated Attention in Qwen3-Next. However, a comprehensive analysis of the relationship among these attention mechanisms is lacking. In this work, we provide both theoretical and empirical evidence demonstrating that the sink in Vanilla Attention and Sink Attention naturally construct a Mixture-of-Experts (MoE) mechanism within attention layers. This insight explains the head collapse phenomenon observed in prior work, where only a fixed subset of attention heads contributes to generation. To mitigate head collapse, we propose a sink-aware training algorithm with an auxiliary load balancing loss designed for attention layers. Extensive experiments show that our method achieves effective head load balancing and improves model performance across Vanilla Attention, Sink Attention, and Gated Attention. We hope this study offers a new perspective on attention mechanisms and encourages further exploration of the inherent MoE structure within attention layers.

Interest in utilizing autonomous uncrewed aerial vehicles (UAVs) for under-canopy forest remote sensing has increased in recent years, resulting in the publication of numerous autonomous flight algorithms in the scientific literature. To support the selection and development of such algorithms, a reliable comparison of existing approaches based on published studies is essential. However, reliable comparisons are currently challenging due to widely varying experimental setups and incomplete reporting practices. This study proposes a standardized experimental setup for evaluating autonomous under-canopy UAV systems to fill this gap. The proposed setup emphasizes quantitative reporting of forest complexity, visual representation of test environments, execution of multiple repeated flights, and reporting of flight success rates alongside qualitative flight results. In addition, flights at multiple target speeds are encouraged, with reporting of realized flight speed, mission completion time, and point-to-point flight distance. The proposed setup is demonstrated using a lightweight lidar-based quadrotor employing state-of-the-art open-source algorithms, evaluated through extensive experiments in two natural boreal forest environments. Based on a systematic evaluation of the original system, several improvements were introduced. The same experimental protocol was then repeated with the optimized system, resulting in a total of 93 real-world flights. The optimized system achieved success rates of 12/15 and 15/15 at target flight speeds of 1 m/s and 2 m/s, respectively, in a medium-difficulty forest, and 12/15 and 5/15 in a difficult forest. Adoption of the proposed experimental setup would facilitate the literature-based comparison of autonomous under-canopy flight systems and support systematic performance improvement of future UAV-based forest robotics solutions.

We introduce a guided stochastic sampling method that augments sampling from diffusion models with physics-based guidance derived from partial differential equation (PDE) residuals and observational constraints, ensuring generated samples remain physically admissible. We embed this sampling procedure within a new Sequential Monte Carlo (SMC) framework, yielding a scalable generative PDE solver. Across multiple benchmark PDE systems as well as multiphysics and interacting PDE systems, our method produces solution fields with lower numerical error than existing state-of-the-art generative methods.

Reasoning large language models exhibit complex reasoning behaviors via extended chain-of-thought generation that are highly fragile to information loss during decoding, creating critical challenges for KV cache compression. Existing token-dropping methods directly disrupt reasoning chains by removing intermediate steps, while head-reallocation methods, designed for retrieval tasks, fail to preserve the heads essential for generative reasoning. However, no existing method can identify which attention heads genuinely maintain reasoning consistency and control generation termination. To address this, we propose RLKV, which uses reinforcement learning as a probe to discover which heads contribute to reasoning quality by directly optimizing their cache usage against actual generation outcomes. This discovery naturally leads to an efficient compression strategy: we allocate full KV cache to reasoning-critical heads while aggressively compressing others with constant-size KV cache. Experiments reveal that a fraction of heads proves essential for reasoning, enabling 20--60% cache reduction with near-lossless performance across diverse tasks and models, and up to 2.06x end-to-end speedup at 60% reduction.

In-Context Learning has shown great potential for aligning Large Language Models (LLMs) with human values, helping reduce harmful outputs and accommodate diverse preferences without costly post-training, known as In-Context Alignment (ICA). However, LLMs' comprehension of input prompts remains agnostic, limiting ICA's ability to address value tensions--human values are inherently pluralistic, often imposing conflicting demands, e.g., stimulation vs. tradition. Current ICA methods therefore face the Instruction Bottleneck challenge, where LLMs struggle to reconcile multiple intended values within a single prompt, leading to incomplete or biased alignment. To address this, we propose PICACO, a novel pluralistic ICA method. Without fine-tuning, PICACO optimizes a meta-instruction that incorporates multiple values to better elicit LLMs' understanding of them and improve alignment. This is achieved by maximizing the total correlation between specified values and LLM responses, which theoretically reinforces value conformity and reduces distractive noise, resulting in more effective instructions. Extensive experiments on five value sets show that PICACO works well with both black-box and open-source LLMs, outperforms several recent strong baselines, and achieves a better balance across up to 8 distinct values.

Multimodal Large Language Models (MLLMs) are a major focus of recent AI research. However, most prior work focuses on static image understanding, while their ability to process sequential audio-video data remains underexplored. This gap highlights the need for a high-quality benchmark to systematically evaluate MLLM performance in a real-world setting. We introduce SONIC-O1, a comprehensive, fully human-verified benchmark of 60 hours (231 clips) spanning 13 real-world conversational domains with 4,958 annotations and demographic metadata. SONIC-O1 evaluates three capabilities: open-ended summarization, multiple-choice question (MCQ) answering, and temporal localization with supporting rationales (reasoning). Across closed- and open-source models, we find that the MCQ accuracy shows the smallest gap between model families, but the best closed-source model outperforms the best open-source model by 22.6% on temporal localization. We further observe accuracy gaps of up to 21.4% on temporal localization across demographic groups, indicating persistent disparities in model behaviour. SONIC-O1 provides an open evaluation suite for temporally grounded and demographically robust multimodal understanding. SONIC-O1 is publicly available for research: Project page (https://vectorinstitute.github.io/sonic-o1/), Dataset (https://huggingface.co/datasets/vector-institute/sonic-o1), GitHub (https://github.com/vectorinstitute/sonic-o1), Leaderboard (https://huggingface.co/spaces/vector-institute/sonic-o1-leaderboard).

We study stochastic logistic bandits with $d$-dimensional action features under the simple-regret objective, where a learner uses $T$ rounds of exploration to output a single final action. The logistic structure is essential here: because the informativeness of an action depends on the local curvature of the sigmoid, actions that are best for immediate reward need not be the most useful for identifying the best final recommendation. We show that the first-order minimax difficulty is governed by $κ_*$, the inverse slope of the sigmoid at the optimal action. The lower bound is realized by a shifted saturated hard family in which saturation simultaneously limits the information available about the final decision and controls the value loss from a wrong recommendation. This reveals a hard mechanism distinct from cumulative-regret constructions, even though online-to-batch reductions recover the same leading order in expectation. We then develop two curvature-aware algorithms: \MULog, a pure-exploration method whose final recommendation satisfies a high-probability upper bound of order $\tilde O(d/\sqrt{κ_* T})$, matching the lower bound up to logarithmic factors, and \THATS, a Thompson-sampling-style method that provides a computationally lighter alternative. Experiments on both hard and easy geometries support the same picture: informative low-reward actions can make instances substantially easier, and the curvature-aware methods exploit this structure especially effectively.

Efficient compression of language model weights is increasingly critical as model scale and deployment grow. Yet, most existing methods rely on handcrafted transforms and heuristics, reflecting the limited understanding of weights as a data modality. To move beyond this paradigm, we formulate weight compression as neural codec learning and propose Neural Weight Compression (NWC), a framework for training neural codecs on pretrained weight datasets. NWC addresses challenges intrinsic to weight compression, including tensor heterogeneity and the mismatch between reconstruction losses and downstream performance. Experiments show that NWC achieves highly competitive accuracy-compression tradeoffs, with particularly strong results in the 4-6 bit regime, without relying on rigid handcrafted components such as the Hadamard transform. These gains extend to across diverse architectures, e.g., vision encoders. Our analysis highlights the roles of entropy-constrained quantization and learned transforms in adapting compression to weight data and downstream tasks.

We present a systematic review of 337 articles evaluating the syntactic abilities of Transformer-based language models (TLMs), reporting on over 3,000 datapoints spanning a wide range of syntactic phenomena, languages, models, and methods. We take the data to collectively show that TLMs encode a non-trivial amount of syntactic knowledge. Behavioral evidence shows strong performance on formal syntactic phenomena, but weaker and more variable performance on phenomena at the syntax-semantics interface. Performance is also consistently lower for languages with less digital support. Probing and mechanistic studies further support the presence of syntactic knowledge in TLMs. Yet, because most work remains observational and current approaches are methodologically heterogeneous, insight into the detailed computational mechanisms underlying syntactic processing remains limited. At the same time, the literature remains heavily concentrated on English and BERT-like models. We discuss the implications of our results and provide recommendations for future research.

Cross-layer reuse of early attention projections can improve optimization and data efficiency, but it creates a structural conflict: the first layer must simultaneously act as a stable, reusable anchor for all deeper layers and as an effective computational block. We demonstrate that this tension constrains the performance of internal-anchor designs. We propose ExoFormer, which resolves the conflict by learning exogenous anchor projections outside the sequential layer stack. We introduce a unified normalized mixing framework that mixes queries, keys, values, and gate logits using learnable coefficients (exploring coefficient granularities: elementwise, headwise, and scalar), and we show that normalizing anchor sources is key to stable reuse. ExoFormer variants consistently outperform their internal-anchor counterparts, and the dynamic variant yields 1.5x downstream accuracy points while matching validation loss using 1.5x fewer tokens than Gated Attention. We explain this efficacy via an Offloading Hypothesis: external anchors preserve essential token identity, allowing layers to specialize exclusively in feature transformation. We release code and models to facilitate future research.

Jailbreak attacks on Large Language Models (LLMs) have demonstrated various successful methods whereby attackers manipulate models into generating harmful responses that they are designed to avoid. Among these, Greedy Coordinate Gradient (GCG) has emerged as a general and effective approach that optimizes the tokens in a suffix to generate jailbreakable prompts. While several improved variants of GCG have been proposed, they all rely on fixed-length suffixes. However, the potential redundancy within these suffixes remains unexplored. In this work, we propose Mask-GCG, a plug-and-play method that employs learnable token masking to identify impactful tokens within the suffix. Our approach increases the update probability for tokens at high-impact positions while pruning those at low-impact positions. This pruning not only reduces redundancy but also decreases the size of the gradient space, thereby lowering computational overhead and shortening the time required to achieve successful attacks compared to GCG. We evaluate Mask-GCG by applying it to the original GCG and several improved variants. Experimental results show that most tokens in the suffix contribute significantly to attack success, and pruning a minority of low-impact tokens does not affect the loss values or compromise the attack success rate (ASR), thereby revealing token redundancy in LLM prompts. Our findings provide insights for developing efficient and interpretable LLMs from the perspective of jailbreak attacks.

Recent advances in video generation have produced models capable of synthesizing stunning visual content from simple text prompts. However, these models struggle to generate long-form, coherent narratives from high-level concepts like dialogue, revealing a ``semantic gap'' between a creative idea and its cinematic execution. To bridge this gap, we introduce a novel, end-to-end agentic framework for dialogue-to-cinematic-video generation. Central to our framework is ScripterAgent, a model trained to translate coarse dialogue into a fine-grained, executable cinematic script. To enable this, we construct ScriptBench, a new large-scale benchmark with rich multimodal context, annotated via an expert-guided pipeline. The generated script then guides DirectorAgent, which orchestrates state-of-the-art video models using a cross-scene continuous generation strategy to ensure long-horizon coherence. Our comprehensive evaluation, featuring an AI-powered CriticAgent and a new Visual-Script Alignment (VSA) metric, shows our framework significantly improves script faithfulness and temporal fidelity across all tested video models. Furthermore, our analysis uncovers a crucial trade-off in current SOTA models between visual spectacle and strict script adherence, providing valuable insights for the future of automated filmmaking.

We argue that multi-document reasoning is constrained not only by how much text a model can read, but also by how limited query-time evidence budget is allocated across documents and semantic granularities. Full-context inference exposes the model to broad evidence non-selectively and at high per-query cost, while flat chunk retrieval often returns locally relevant passages that are weakly organized for cross-document synthesis. We present \textbf{BEAR}, a framework for structured evidence allocation that builds hierarchical semantic indices offline and performs coarse-to-fine evidence access at query time through complementary \emph{exploration} and \emph{recovery} paths. This coarse-to-fine design can be viewed as structured evidence allocation under a fixed evidence-context budget. Across synthetic and real-world benchmarks, BEAR performs particularly strongly on DragonBall, remains competitive with strong retrieval-based baselines on HotpotQA, and yields the best retrieval-based result on 2Wiki under our evaluated protocol, while operating under substantially smaller \emph{query-time evidence budgets} than the reported long-context references. Additional analyses suggest that the gains are associated with hierarchy as an allocation substrate together with complementary exploration and recovery, rather than semantic chunking alone.

Low dimensional structures appear ubiquitously in the eigenspectra of deep learning matrices in classification networks trained in the overparameterized regime. While theoretical advances have aimed to explain this phenomenology, they typically succeed only in capturing subsets of the full behavior or rely on assumptions that cannot hold in practice. In this work, we provide an analytic explanation for the bulk plus outlier structure of several canonical deep learning matrices, including the Hessian, gradients, and weights. We achieve this using unconstrained feature models (UFMs), a now-common tool for studying the emergence of deep neural collapse (DNC). We show that DNC is the source of these low dimensional eigenspectra, in each case, the eigenvalues and eigenvectors can be constructed from feature means, the characterizing objects of DNC. This provides a unifying analytic explanation for a wide range of spectral phenomena in deep learning and goes beyond empirical characterizations, which typically focus on eigenvalues, by providing a detailed analysis of eigenvectors. We prove that our results hold for both linear and ReLU networks and provide numerical validation in both the modeling context and standard deep-network architectures on canonical datasets.

We present PEAR (Pairwise Evaluation for Automatic Relative Scoring), a supervised quality estimation (QE) metric family that reframes reference-free machine translation (MT) evaluation as a graded pairwise comparison. Given a source segment and two candidate translations, PEAR predicts the direction and magnitude of their quality difference. The metrics are trained using pairwise supervision derived from differences in human judgments, with an additional regularization term that encourages sign inversion under candidate order reversal. On the WMT24 meta-evaluation benchmark, PEAR outperforms strictly matched single-candidate QE baselines trained with the same data and backbones, isolating the benefit of the proposed pairwise formulation. Despite using substantially fewer parameters than recent large metrics, PEAR surpasses far larger QE models and reference-based metrics. Our analysis further indicates that PEAR yields a less redundant evaluation signal relative to other top metrics. Finally, we show that PEAR is an effective utility function for minimum Bayes risk (MBR) decoding, reducing pairwise scoring cost at negligible impact.

Semi-supervised regression (SSR), which aims to predict continuous scores for samples while reducing the reliance on large-scale labeled data, has recently attracted considerable attention across various applications, including computer vision, natural language processing, audio analysis, and medical analysis. Existing SSR methods typically train models with scarce labeled data by introducing constraint-based regularization or ordinal ranking to mitigate overfitting. However, these approaches often fail to fully exploit the abundance of unlabeled samples. Although consistency-driven pseudo-labeling methods attempt to incorporate unlabeled data, their performance is highly sensitive to pseudo-label quality and noisy predictions. To address these challenges, we propose a Dual-stream Knowledge Distillation framework (DKD), which is specifically designed for SSR to distill both continuous-valued knowledge and distributional information. This design better preserves regression magnitude information and improves sample efficiency. Specifically, in DKD, the teacher is optimized solely with ground-truth labels for label distribution estimation, while the student learns from a mixture of real labels and teacher-generated pseudo targets on unlabeled data. The distillation process enables effective supervision transfer, allowing the student to leverage pseudo labels more robustly. Furthermore, we introduce a Decoupled Distribution Alignment (DDA) module, which separately aligns the target and non-target distributions between the teacher and student. To improve the reliability of non-target knowledge transfer, DDA incorporates a variance-guided non-target distribution alignment strategy that adaptively downweights uncertain teacher predictions, thereby enhancing the student's ability to mitigate noise in pseudo-label supervision and learn a better-calibrated regression predictor.

Reinforcement learning (RL) has shown promising results in active flow control (AFC), yet progress in the field remains difficult to assess as existing studies rely on heterogeneous observation and actuation schemes, numerical setups, and evaluation protocols. Current AFC benchmarks attempt to address these issues but heavily rely on external computational fluid dynamics (CFD) solvers, are not fully differentiable, and provide limited 3D and multi-agent support. To overcome these limitations, we introduce FluidGym, the first standalone, fully differentiable benchmark suite for RL in AFC. Built entirely in PyTorch on top of the GPU-accelerated PICT solver, FluidGym runs in a single Python stack, requires no external CFD software, and provides standardized evaluation protocols. We present baseline results with PPO, SAC, DPC, and TD-MPC, and release all environments, datasets, and trained models as public resources. FluidGym enables systematic comparison of control methods, establishes a scalable foundation for future research in learning-based flow control, and is available at github.com/safe-autonomous-systems/fluidgym.

E-commerce platforms increasingly rely on Large Language Models (LLMs) and Vision Language Models (VLMs) to detect illicit or misleading product content. However, these models remain vulnerable to evasive content, which refers to inputs that have been deliberately modified through techniques such as word splitting, euphemistic language, or image cropping to conceal policy violations while still conveying prohibited claims. Crucially, detecting such content requires a model to simultaneously master two capabilities: accurately comprehending complex rules, and correctly inferring the true intent behind deliberately obfuscated multimodal inputs. While prior work has separately explored LLM reasoning over complex rules and LLM-based detection of evasive content, no existing benchmark combines both within a unified evaluation framework. This gap is particularly consequential in e-commerce, where accurate moderation demands that both capabilities operate in concert. To address this gap, we introduce EVADE-Bench, the first expert-curated Chinese multimodal benchmark specifically designed to evaluate LLMs and VLMs on evasive content detection in real-world e-commerce scenarios. Our comprehensive evaluation of 26 open- and closed-source LLMs and VLMs reveals that even state-of-the-art models frequently misclassify evasive samples. We further demonstrate that clearer rule categorization significantly improves model prediction consistency and reduces false predictions, highlighting the critical role of benchmark design in enabling reliable evaluation. To explore paths for performance improvement, we investigate the feasibility of multi-agent decomposition for multimodal reasoning, wherein visual description and logical inference are decoupled into separate agents, and find that this strategy yields notable accuracy gains.

This study investigates the use of neural topic modeling and LLMs to uncover meaningful themes from patient storytelling data, to offer insights that could contribute to more patient-oriented healthcare practices. We analyze a collection of transcribed interviews with cancer patients (132,722 words in 13 interviews). We first evaluate BERTopic and Top2Vec for individual interview summarization by using similar preprocessing, chunking, and clustering configurations to ensure a fair comparison on Keyword Extraction. LLMs (GPT4) are then used for the next step topic labeling. Their outputs for a single interview (I0) are rated through a small-scale human evaluation, focusing on {coherence}, {clarity}, and {relevance}. Based on the preliminary results and evaluation, BERTopic shows stronger performance and is selected for further experimentation using three {clinically oriented embedding} models. We then analyzed the full interview collection with the best model setting. Results show that domain-specific embeddings improved topic \textit{precision} and \textit{interpretability}, with BioClinicalBERT producing the most consistent results across transcripts. The global analysis of the full dataset of 13 interviews, using the BioClinicalBERT embedding model, reveals the most dominant topics throughout all 13 interviews, namely ``Coordination and Communication in Cancer Care Management" and ``Patient Decision-Making in Cancer Treatment Journey''. Although the interviews are machine translations from Dutch to English, and clinical professionals are not involved in this evaluation, the findings suggest that neural topic modeling, particularly BERTopic, can help provide useful feedback to clinicians from patient interviews. This pipeline could support more efficient document navigation and strengthen the role of patients' voices in healthcare workflows.

We introduce Alterbute, a diffusion-based method for editing an object's intrinsic attributes in an image. We allow changing color, texture, material, and even the shape of an object, while preserving its perceived identity and scene context. Existing approaches either rely on unsupervised priors that often fail to preserve identity or use overly restrictive supervision that prevents meaningful intrinsic variations. Our method relies on: (i) a relaxed training objective that allows the model to change both intrinsic and extrinsic attributes conditioned on an identity reference image, a textual prompt describing the target intrinsic attributes, and a background image and object mask defining the extrinsic context. At inference, we restrict extrinsic changes by reusing the original background and object mask, thereby ensuring that only the desired intrinsic attributes are altered; (ii) Visual Named Entities (VNEs) - fine-grained visual identity categories (e.g., ''Porsche 911 Carrera'') that group objects sharing identity-defining features while allowing variation in intrinsic attributes. We use a vision-language model to automatically extract VNE labels and intrinsic attribute descriptions from a large public image dataset, enabling scalable, identity-preserving supervision. Alterbute outperforms existing methods on identity-preserving object intrinsic attribute editing.

Supervised convolutional neural networks (CNNs) are widely used to solve imaging inverse problems, achieving state-of-the-art performance in numerous applications. However, despite their empirical success, these methods are poorly understood from a theoretical perspective and often treated as black boxes. To bridge this gap, we analyze trained neural networks through the lens of the Minimum Mean Square Error (MMSE) estimator, incorporating functional constraints that capture two fundamental inductive biases of CNNs: translation equivariance and locality via finite receptive fields. Under the empirical training distribution, we derive an analytic, interpretable, and tractable formula for this constrained variant, termed Local-Equivariant MMSE (LE-MMSE). Through extensive numerical experiments across various inverse problems (denoising, inpainting, deconvolution), datasets (FFHQ, CIFAR-10, FashionMNIST), and architectures (U-Net, ResNet, PatchMLP), we demonstrate that our theory matches the neural networks outputs (PSNR $\gtrsim25$dB). Furthermore, we provide insights into the differences between \emph{physics-aware} and \emph{physics-agnostic} estimators, the impact of high-density regions in the training (patch) distribution, and the influence of other factors (dataset size, patch size, etc).

Therapeutic dialogue is not a sequence of isolated responses: client goals, motivation, resistance, and therapeutic alliance evolve over time. Yet current LLM-based mental health dialogue systems often lack explicit mechanisms for tracking these dynamics across extended interactions, which can lead to poorly timed interventions or premature goal resolution. We introduce CALM-IT, a framework for generating and evaluating long-form Motivational Interviewing dialogues through explicit modeling of evolving client and counselor states, guiding both counseling strategy selection and utterance generation. We evaluate CALM-IT on a large-scale corpus of 8,232 synthetic dialogues spanning multiple dialogue lengths and frameworks. Compared with all baselines, CALM-IT achieves the best performance on most MITI 4.2 global ratings, including Empathy, Partnership, and Softening Sustain Talk, as well as on other key performance metrics while exhibiting minimal performance degradation as dialogue length increases. Notably, although CALM-IT initiates fewer change-directed prompts, it produces the highest client acceptance rate (64.3%) on average across different length conditions. We release a reproducible generation framework, a MITI-grounded process-level evaluation protocol, and a large-scale synthetic corpus for studying therapeutic LLMs under realistic long-form interaction conditions.

With the increasing adoption of vision-language models (VLMs) in critical decision-making systems such as healthcare or autonomous driving, the calibration of their uncertainty estimates becomes paramount. Yet, this dimension has been largely underexplored in the VLM test-time prompt-tuning (TPT) literature, which has predominantly focused on improving their discriminative performance. Recent state-of-the-art advocates for enforcing full orthogonality over pairs of text prompt embeddings to enhance separability, and therefore calibration. Nevertheless, as we theoretically show in this work, the inherent gradients from fully orthogonal constraints will strongly push semantically related classes away, ultimately making the model overconfident. Based on our findings, we propose Semantic Orthogonal Calibration (SoC), a Huber-based regularizer that enforces smooth prototype separation while preserving semantic proximity, thereby improving calibration compared to prior orthogonality-based approaches. Across a comprehensive empirical validation, we demonstrate that SoC consistently improves calibration performance, while also maintaining competitive discriminative capabilities.

As Natural Language Generation (NLG) dominates modern NLP, scalable evaluation remains a critical bottleneck. Consequently, LLM-as-a-judge (LaaJ) adoption has accelerated rapidly, appearing in more papers than human evaluation in 2025. This pivotal shift motivates a critical analysis of current evaluation practices. Overcoming the limits of rigid keyword filtering and manual review, we employ a multi-LLM information extraction pipeline to gather structured metadata from 14,171 papers across four major NLP conferences (2020-2025). Analyzing 3,334 filtered NLG papers, we identify three systemic challenges. (1) Metric inertia: despite the shift toward open-ended generation, legacy lexical metrics (BLEU, ROUGE) persist as primary indicators, typically used alongside rather than replaced by semantic alternatives. (2) Metric-criteria mapping problem: our paper-level co-occurrence data reveals that general-purpose automatic metrics are applied as broad proxies for quality, without specifying which dimension of text generation they are intended to evaluate. (3) Validation gap: LaaJ has grown rapidly without commensurate human validation (fewer than 8% of papers). Crucially, while LaaJ correlates with aggregate quality, alignment collapses on fine-grained criteria like fluency. To address these gaps, we distill our findings into a minimal Evaluation Checklist to guide metric selection, construct validity, and LaaJ deployment.

Generative spoken language models pretrained on large-scale raw audio can continue a speech prompt with appropriate content while preserving attributes like speaker and emotion, serving as foundation models for spoken dialogue. In prior literature, these models are often evaluated using ``global token perplexity'', which directly applies the text perplexity formulation to speech tokens. However, this practice overlooks fundamental differences between speech and text modalities, possibly leading to an underestimation of the speech characteristics. In this work, we propose a variety of likelihood- and generative-based evaluation methods that serve in place of naive global token perplexity. We demonstrate that the proposed evaluations more faithfully reflect perceived generation quality, as evidenced by stronger correlations with human-rated mean opinion scores (MOS). When assessed under the new metrics, the relative performance landscape of spoken language models is reshaped, revealing a significantly reduced gap between the best-performing model and the human topline. Together, these results suggest that appropriate evaluation is critical for accurately assessing progress in spoken language modeling.

Cheating in chess, by using advice from powerful software, has become a major problem, reaching the highest levels. As opposed to the large majority of previous work, which concerned {\em detection} of cheating, here we try to evaluate the possible gain in performance, obtained by cheating a limited number of times during a game. We develop threshold-based and Bellman-style intervention policies, and test them in a controlled engine-vs-engine setting using Stockfish. A judicious choice of 1 or 2 cheats yields average scores of 0.71 and 0.82, respectively, compared to 0.51 with no cheats. We also introduce a fast, engine-free simulator that enables hyperparameter optimization without running games, closely matching the engine-based optimum. The goal of this work is not to assist cheaters, but to measure the effectiveness of cheating -- which is crucial as part of the effort to contain and detect it.

Although Audio Large Language Models (ALLMs) have witnessed substantial advancements, their long audio understanding capabilities remain unexplored. A plethora of benchmarks have been proposed for general audio tasks, they predominantly focus on short-form clips, leaving without a consensus on evaluating ALLMs over extended durations. This paper proposes ChronosAudio, the first multi-task benchmark tailored for long-audio understanding in ALLMs. It encompasses six major task categories and comprises 36,000 test instances totaling over 200 hours audio, stratified into short, middle, and long-form categories to comprehensively evaluate length generalization. Extensive experiments on 16 state-of-the-art models using ChronosAudio yield three critical findings: 1.Precipitous Long-Context Collapse: ALLMs exhibit a severe inability to sustain performance, with the transition from short to long contexts triggering a staggering performance degradation of over 90% in specific tasks. 2.Structural Attention Dilution: Performance degradation stems from a fundamental failure in maintaining temporal locality; attention mechanisms suffer from significant diffusion in later sequences. 3.Restorative Ceiling of Mitigation: Current strategies only offer 50% recovery. These findings reveal significant challenges in long-audio, underscoring the urgent need for approaches to achieve robust, document-level audio reasoning.

Sign Language Translation (SLT) is a challenging cross-modal task requiring joint modeling of manual articulations and non-manual signals. Existing gloss-free SLT methods effectively capture gestural dynamics but often underutilize facial expressions, which play crucial grammatical and disambiguating roles. This limitation can cause semantic degradation when distinct concepts share similar manual configurations. To address this issue, we propose FEA-SLT (**F**acial-**E**xpression-**A**ware **S**ign **L**anguage **T**ranslation), a gloss-free end-to-end framework that uses facial dynamics as semantic anchors for resolving manual ambiguity. FEA-SLT employs a domain-transferred facial encoder to extract expression-sensitive representations and integrates them with manual features through a linguistically constrained *Facial-Expression-Aware Fusion* (FEAF) module. FEAF captures reciprocal dependencies between manual and facial channels via bidirectional modulation, enhancing syntactic fidelity. Experiments on PHOENIX14T and CSL-Daily show that FEA-SLT achieves state-of-the-art BLEU performance among gloss-free methods, while targeted analyses confirm improved translation of facial-sensitive utterances. Code is available at [https://github.com/TuGuobin/FEA-SLT](https://github.com/TuGuobin/FEA-SLT).