Papers

LLM-as-judge is essential for evaluating open-ended text and steering post-training, yet improving the judge itself typically relies on expensive annotations, reward models, or distillation from stronger teachers. In this work, we eliminate external gold supervision from the RL training reward: the model's own evaluative capability generates learning signals for its optimization – a closed-loop setting of bounded recursive self-improvement (RSI) termed Recursive Self-Evaluation (RecurSE). We study two central questions: when can self-improvement occur, and when must it stop? First, RecurSE pairs a trainable judge evaluating candidate responses under per-rule rubrics (Pass 1) with a synchronized policy-copy checker that audits the judge's reasoning against meta-rubrics to supply a scalar process reward (Pass 2). To enable learning, interface decoupling structurally isolates the checker's scalar score from the judge's verdict tokens, eliminating a degenerative token-copying shortcut that inflates self-assigned rewards. Second, because unanchored recursive learning is inherently bounded, Pairwise Advantage Validity (PAV) serves as an unbiased validation monitor that jointly tracks judge accuracy and checker fidelity to reliably identify the optimal early-stopping window. Across Qwen3.5-9B, Gemma-4-E4B-it, and Qwen3.6-27B, RecurSE achieves consistent generalization gains across held-out medical, pairwise, summarization, and professional benchmarks. Ablations demonstrate that synchronized judge-checker co-evolution outperforms frozen checkers, external meta-judges, self-consistency, and scaled teacher distillation. Furthermore, preference pairs curated by our judge effectively enhance downstream policy alignment. Bounded RSI for LLM-as-judge is thus viable when self-produced reward validity is explicitly decoupled and monitored.

Quantum noise is expected to degrade quantum machine learning by driving circuits away from their noiseless implementations. Yet recent studies show moderate noise can reduce testing error, a behavior unexplained by weak-noise perturbative error accumulation or strong-noise trainability collapse. Here we develop a statistical learning theory connecting microscopic noise processes to macroscopic learning performance. At its heart is a noise-order purity parameter, derived from a surrogate model analysis, that predicts the noise-induced reduction in model complexity and the consequent reduction in the generalization gap. Noise simultaneously increases prediction bias. Their competition explains the intermediate-noise regime left open between these limits. It produces a finite-noise optimum whose location depends on the learning setup and can disappear in the large-sample limit. Numerical experiments validate these predictions. Noise programming can move a model towards this optimum. These results make the non-monotonic effect of noise predictable and provide a route to harness it.

Many LLM applications are most useful when they provide several candidate outputs for comparison, validation, or combination. Predominant evaluation settings, however, still focus on individual outputs or reduce multiple samples to a single success or selected answer. This can miss whether the outputs include several genuinely different useful results. We introduce VTC-Bench, a five-domain benchmark for this setting, together with Validated Task Coverage (VTC) as its core evaluation quantity. The benchmark is built from carefully selected real-data tasks where both output quality and task-relevant distinctness can be checked automatically and reproducibly, without model-based judges. VTC measures how many distinct useful results are obtained within $k$ attempts. Across multiple models and inference settings, the benchmark leads to different conclusions from conventional evaluation: configurations that look strongest from single-draw quality are not necessarily those with the best coverage, and simple measures of output variation do not reliably recover task-relevant coverage. These results show that finite candidate sets can be evaluated directly as objects of interest, revealing differences in model behavior that are not apparent from conventional per-output evaluation.

Effective human-AI interaction requires systems that dynamically adapt to a user's behavior and evolving understanding. When users interact with Large Language Models (LLMs), these models typically respond to prompts without sensing the user's immediate reactions. This lack of communicative synchrony can lead to information overload or leave confusion unresolved in real time. In this paper, we introduce Aura, a framework that enables LLM systems to dynamically modulate output based on a user's evolving emotions. Aura's Perception Module continuously estimates the user's emotional state from facial expressions. Our Policy Module then selects interventions through a probabilistic belief model. Finally, Aura's Generation Module uses parameter-efficient Low-Rank Adaptation (LoRA) adapters to produce contextually tailored responses mid-turn during response generation. We evaluated Aura in a within-subjects user study (N=20) on information-seeking tasks, where it achieved statistically significantly higher normalized perceived learning gains than a Llama-3 baseline and reduced interaction time by 21% relative to existing LLM baselines (GPT-4o, Llama-3). Our results indicate that real-time, context-sensitive interventions can improve learning efficiency and user satisfaction without observable degradation in factual accuracy. Aura thus supports the potential for more responsive and effective human-AI interaction.

Event-based motion estimation is central to tasks that demand high temporal resolution and robustness to fast motion. Existing methods typically rely on iterative optimization or repeated hypothesis comparison, offsetting the sensor's low-latency advantage. We propose Oriented Distance Field Motion Estimation (ODF Motion Estimation), which replaces this optimization with a single averaging step over a precomputed field of event distance vectors, combined with an adaptive event-count selection strategy and a parameter-free trail filter. On public and self-collected datasets, ODF motion estimation reaches sub-pixel accuracy at the lowest latency among compared methods. We validate its generality on two downstream applications rather than treating them as separate contributions. First, the estimated trajectory is converted into a blur kernel and paired with a compact iterative-unfolding network, trained on simulated motion-estimation noise, for real-time non-blind image deblurring, attaining competitive or superior PSNR/SSIM with under 1M parameters. Second, the same precomputed field is repurposed for directional event filtering in a low-power asynchronous pupil and glint tracker, sustaining stable tracking for tens of seconds while lowering a near-eye module's power draw.

The digital transformation of the Dutch labour market is reshaping occupational language, career pathways, and job-related skills. Addressing these changes requires granular labour market intelligence. This paper develops an AI-based methodology to analyse digitalisation using data covering millions of Dutch job profiles. The methodology combines embedding-based similarity search and large language model classification to map unstructured job information to harmonised ESCO occupations. We also introduce a Digital Semantic Score that measures how strongly job titles and skills are associated with digital concepts relative to a non-digital reference. Using embeddings and cosine similarity to transparent digital and non-digital anchor groups, this indicator moves beyond keyword-based approaches by capturing broader digital meanings in occupational language and worker skill profiles. It enables analysis across occupations, career transitions, emerging job-title vocabulary, and skill digitality. The findings reveal that digitalisation is unevenly distributed across the labour market. Digital job-title language is most prominent among managerial, professional and ICT-related occupations, but is increasingly visible in hybrid business, marketing and automation-related roles. Career-transition analyses show that movement toward digital work is pathway-dependent, while skill analyses highlight the multidimensional nature of digital capability, encompassing technical, hybrid and business-systems skills. By combining profile data, AI-supported occupational classification and semantic scoring, this study advances AI-driven labour market analytics and provides a scalable framework for monitoring digital labour market change. The methodology helps identify emerging skill needs, support reskilling strategies, and inform policies addressing skills mismatches and labour shortages in the Netherlands.

Answering developer questions about a software repository is a critical yet under-explored problem in software engineering. While existing repository understanding methods have advanced the field, they predominantly rely on surface-level code retrieval and lack the ability for deep reasoning over multiple files, complex software architectures, and grounding answers in long-range code dependencies. To address these limitations, we propose DeepRepoQA, a novel question answering (QA) framework for repository-level code understanding. DeepRepoQA builds on an agentic framework where LLM agents find answers through a systematic tree search over the repository structure. A Monte-Carlo Tree Search (MCTS) mechanism is employed to empower agents to dynamically search, navigate, and inspect code, enabling effective multi-hop reasoning over long-range code dependencies. Comprehensive experiments on the SWE-QA benchmark demonstrate substantial performance gains over strong baselines, validating the effectiveness of systematic MCTS-guided exploration for multi-hop repository reasoning.

Facial attractiveness has been linked to statistical regularities such as symmetry and averageness, suggesting that beauty may depend on the ease with which a face is perceived. We empirically test this hypothesis by training variational autoencoders on four face datasets without attractiveness supervision and evaluating their representations on the 597 faces from the Chicago Face Database. We report three main novel findings. First, we identify that human attractiveness ratings closely aligns with the direction defined by the VAE evidence lower bound (ELBO) in rate-distortion space. Second, we find that independently learned latent spaces contain an attractiveness direction that transfers strongly across random initializations and training data. Finally, we observe that attractive faces are more prototypical in both shape and latent space. Our work provides the first explicit connection between classic accounts of aesthetics with learned generative models and empirical support for a variational interpretation of the processing fluency theory of aesthetic pleasure.

Enterprise entity alignment must handle semi-structured records, implicit attributes, and unit or granularity mismatches. Manual matching is still common in practice, but does not scale as schemas and providers evolve. LLM-only matching improves semantic recall, yet can violate structural and physical invariants, producing fluent yet operationally invalid correspondences. We propose constraint-guided mapping (CGM), a neuro-symbolic method with three stages: (i) schema-grounded admissibility constraints with metadata mc = <tau_c, delta_c>, where tau_c denotes the constraint type and delta_c provides executable relation and normalization logic; (ii) constraint-restricted candidate generation with cascade relaxation to guarantee a nonempty feasible set under noise; and (iii) neural ranking with bounded LLM disambiguation restricted to that feasible set. Methodologically, constraints operate as hypothesis-space operators rather than post-hoc validators, enabling controlled degradation under relaxation and auditable, human-guidable decisions. On a controlled structural-decoy benchmark, hard admissibility shrinks the candidate space by ~480x without dropping the GT, and a layer-by-layer ablation shows this gate, not the LLM, is the decisive lift (F1 0.08 to 0.66). The benefit is model-independent and adds no extra inference cost: a small model with constraints matches a frontier LLM used without them at ~28x lower cost. The method, not a single tuned configuration, transfers across seven enterprise makes (macro F1 0.70), each under its own automatically discovered, expert-refinable constraints, and lowers expert effort by ~7x versus spreadsheet workflows. Public Valentine results add an external ranking sanity check and mark the boundary: constraints should be hard only where structural invariants are match-determining.

We propose CARO, a contact-agnostic residual observation framework for policy adaptation. CARO embeds a fixed-base Euler–Lagrange model into the reinforcement learning control loop and constructs a torque-level residual observation without requiring torque sensors, explicit contact estimation, or vision-based measurements of the floating-base position and linear velocity. A disturbance observer extracts a structured signal representing dynamics mismatch, while the policy learns to exploit this feedback for online adaptation. CARO is trained under the same terrain, command, and domain-randomization conditions as the nominal policy, without specialized disturbance curricula or additional adaptation supervision. Nevertheless, it achieves substantially improved zero-shot robustness in simulation and sim-to-real transfer tasks involving out-of-distribution payloads, center-of-mass shifts, terrain geometries, abrupt dynamics changes, and elevated-platform landings.

Large language model (LLM)-based multi-agent social simulation has demonstrated compelling results, but Agentopia was evaluated with 100 agents over 10 simulated years using Qwen3.5-397B-A17B, leaving the behavior of reduced-scale deployments on consumer hardware unclear. In this paper, we implement and evaluate a reduced-scale Agentopia port on a single NVIDIA RTX 5070 Ti(12 GB VRAM) using Qwen3-8B-AWQ, a 4-bit quantized model. We introduce three structural adaptations for this setting: (1) system-managed layered memory compression, (2) four activity blocks per simulated day, and (3) explicit physical- and mental-health state variables. Across three independent stochastic runs, two runs completed 52 weeks and the third completed 50 weeks before reaching the context limit, totaling 154 system-weeks (770 agent-weeks). No agent died,and no threshold-based health warning was logged; activity records containing at least one NO_RESPONSE field occurred at rates of 10.15-10.29% across runs. A 52-week memory-off run tied L2/L3 artifact production to layered memory; a separate 10-week comparison associated four daily time blocks with 2.72 times more finalized records and lower lexical duplication, but a higher missing-field rate. These comparisons do not support causal behavioral claims. We release validated configurations, derived audits, analysis scripts, aggregate figure data, and our implementation changes in a public fork; raw runs and initial persona data are excluded because their redistribution provenance is not fully resolved.

Agentic retrieval-augmented generation (RAG) requires language models to decide when to continue searching and when to answer. Existing RL-based methods rely on external supervision and overlook the agent's internal belief about whether the current evidence is sufficient. To address this problem, we reformulate the search decision quality as belief-action alignment and propose MetaRAG, a belief-action aligned policy optimization framework for agentic RAG. MetaRAG uses Verify-first Action Generation to elicit an explicit verification process before each actual action, and Internal Belief Probing to estimate the policy model's own answerability belief from the same question-history context. Based on these, MetaRAG derives a consistency reward that is further gated by answer correctness, avoiding reinforcement of internally consistent but incorrect trajectories. The belief probe is used only during training and introduces no inference-time overhead. Experiments on seven public QA benchmarks show that MetaRAG consistently improves the accuracy-efficiency trade-off over strong RL-based agentic RAG baselines, with gains that transfer to deep research settings, different optimizers, and multiple model backbones.

The advancement of Embodied AI necessitates high-quality simulation assets that faithfully mirror the real world. However, transforming raw visual observations into simulation-ready scenes remains challenging due to the lack of physical grounding and scene-level interactivity in current image-to-URDF methods. We propose NeoWorld-Pro, a framework that reformulates monocular scene reconstruction as procedural programming for interactive 3D environments. Leveraging the zero-shot reasoning and code synthesis capabilities of MLLMs, NeoWorld-Pro converts a single RGB image into executable programs specifying object geometry, articulation, and physical properties. A physics-in-the-loop mechanism then iteratively refines the generated programs by validating their execution in a physics engine, enforcing physically plausible articulations, valid object compositions and interactions, and accurate spatial relationships. Experiments show that NeoWorld-Pro outperforms open-loop and prior monocular reconstruction methods, while enabling complex downstream tasks such as stable stacking and fine-grained manipulation.

Compressive sensing predicts that sparse vectors (signals) can be recovered from a small number of linear measurements via efficient algorithms. This finding, which dates back two decades, has triggered a paradigm shift in signal processing and initiated many developments both in practical signal-processing applications, such as medical imaging, radar, and astronomy, and on the theoretical side. More recently, seminal connections to the field of deep learning have led to further advances in the field, such as the use of unrolled neural networks for sparse recovery and the discovery that common training algorithms (variants of gradient descent) favor sparsity in overparameterized scenarios – the so-called implicit bias phenomenon. This article gives an introduction to compressive sensing and outlines connections to deep learning. In particular, we will discuss generalization for neural networks generated by unrolling sparse recovery algorithms and implicit regularization for gradient descent applied to learning simplified linear neural networks.

Training neural networks requires balancing the trade-off between fitting the training data and achieving robust performance on unseen inputs. This ability, commonly referred to as generalizability, is determined by the gap between the empirical risk on the training set (“empirical loss”) and the expected risk over the data distribution (“generalization error”). Existing approaches typically estimate the generalization error numerically, requiring gradient descent training and an “early stopping” strategy. In this work, we introduce an analytic framework that estimates the optimal time of early stopping without the need for training. Several works in the literature also give such analytical estimations, but they are generally based on random matrix theory and often make assumptions on the distribution of the data or the eigenvalue distribution of the covariance matrix. In contrast, our work is based on Rademacher complexity (RC) without needing such probabilistic assumptions. For both theoretical and numerical reasons, it is more relevant to express RC with the L1- norm rather than with the L2-norm. We focus on the case of linear models and the problem of linear regression. Thanks to the “linear probing” method, our results can, however, be successfully applied to nonlinear neural networks, as illustrated in the classification MNIST example.

Inspired by human multi-modal perception, Audio-Visual Multi-Modal Learning (AVMML) integrates auditory and visual information to leverage complementary cross-modal cues, enabling more robust and comprehensive scene perception. Existing studies predominantly tackle each AVMML task in isolation, which stands in stark contrast to humans' unified cognitive capacity for handling versatile perception. However, naive joint training across multiple AVMML tasks often suffers from mutual interference, arising from the intricate inter-task relationships. To address this, we propose a unified framework that simultaneously accommodates versatile AVMML tasks. Specifically, benefiting from powerful representation and generalization capabilities of large language models, we design a task-disentangled Low-Rank Adaptation (LoRA) mechanism that enables dynamic integration of both task-specific and task-shared knowledge, thereby facilitating effective multi-task collaboration. The proposed task-disentangled LoRA comprises three components: a task-general low-rank matrix, task-specific modulation matrices, and cross-task collaboration experts, which respectively capture universal audio-visual knowledge, decouple task-specific pattern, and exploit inherent inter-task correlations. By unifying explicit collaboration from both model and task perspectives, our approach not only surpasses existing unified audio-visual models across multiple AVMML tasks, but also outperforms most task-specific models on certain AVMML tasks.

Semantic identifiers (SIDs) represent entities as hierarchical token sequences for generative retrieval and recommendation. Residual-quantization tokenizers construct these sequences by selecting a codeword at each level and passing a residual to the next. We view this process as progressive commonality removal: each token captures a component shared within its group, while later tokens should model the remaining differences. This view reveals three limitations: a corpus-wide shared component can consume first-level capacity, hard assignment ignores graded similarities to nearby codewords, and full-codeword subtraction can leave variation along the selected-codeword direction in the next residual. We therefore develop our solution in the post-hoc setting, where residual construction is not constrained by input reconstruction. Specifically, we propose PRQ-KMeans, which removes the global-mean component, refines centroids with Top-k similarity-weighted updates, and replaces full-codeword subtraction with a projection residual that removes each representation's selected-centroid component. Experiments on a large-scale industrial search dataset and four public recommendation benchmarks show that PRQ-KMeans achieves the strongest overall performance among the evaluated tokenizers, including gains of up to 7.4% in HitRate and 11.8% in MRR on the industrial dataset.

With the rise of scientific machine learning, physics-informed neural networks (PINNs) have been extensively applied to a wide range of problems. Nevertheless, most theoretical analyses of PINNs remain confined to linear equations, and a substantial gap persists between PINNs and classical numerical analysis for nonlinear problems. To address this issue, we propose a robust training framework for PINNs solving nonlinear partial differential equations, together with a rigorous error analysis. Specifically, for the $p$-Laplace equation, we introduce a novel loss formulation that combines a dual residual loss measured in $W^{-1,p'}$ with a boundary loss measured in a fractional Sobolev norm $W^{1-\frac{1}{p},p}$. This formulation is designed to accommodate the limited regularity of weak solutions and enables us to establish rigorous a priori and a posteriori error estimates. Moreover, the proposed framework and its analysis are extended to a parametric setting in which the exponent, the source term, and the boundary condition may all vary with the parameters. Finally, we present numerical experiments that substantiate our theoretical findings.

LLM inference applications are gaining significant traction. The demand for inference is growing exponentially, and the GPU usage of inference is increasingly surpassing that of training. Due to the mobility penalty, edge-side inference fails to deliver satisfactory performance. Consequently, most inference service providers currently rely on cloud-based inference, which incurs substantial, not sustainable costs for enterprises, and is even increasing in the agentic paradigm. Therefore, our goal is to enable powerful computing capabilities as server-grade GPUs on mobile devices. We propose Wireless GPU Computing Infrastructure (WiCi) in this paper. Through WiCi, mobile devices can wirelessly access server-grade GPUs, running inference tasks on mobile clients but offloading GPU-related computations to a nearby GPU via WiFi. WiCi introduces a series of designs to make sure the infrastructure is scalable with different applications, compatible with different mobile devices, and has comparable performance to running on a physical GPU. We test WiCi from mobile devices and find that WiCi can reduce time to first token by up to 90%, improve the token rate by approximately 39x compared to local inference on mobile devices for the same model, and support much larger models. WiCi also achieves up to nearly 80% of the native performance of the server-grade GPU across different applications.

We develop eigenvalue bounds for symmetric, block-tridiagonal multiple saddle-point linear systems, preconditioned with block-triangular matrices, based on approximate Schur complements. Irrespective on the number of blocks, we prove that all complex eigenvalues, with nontrivial imaginary part, are strictly contained in a circle within the complex plane, with center 1. The real and positive eigenvalues are bounded in terms of the extremal roots of a sequences of parametric polynomials. Numerical results reveal that the bounds describe very well the eigenvalue distribution of the preconditioned matrix.

Generative simulation for surgical robotics still lacks real-time interaction. Physical-robot experiments, often involving animal or cadaver labs, are time-consuming, costly, and difficult to reproduce, while classical simulators struggle to capture photorealistic appearance and deformable-tissue dynamics. We address this gap with Cosmos-H-Dreams, an integrated real-time surgical world-model system combining an action-conditioned generative model, a teacher-to-student distillation recipe, and a deployment stack built on the NVIDIA FlashDreams streaming-inference library. Starting from Cosmos-H-Surgical-Simulator, a multi-embodiment action-conditioned surgical video world model fine-tuned on the large-scale Open-H-Embodiment corpus, we post-train this checkpoint on embodiment- and procedure-specific data. By distilling the resulting bidirectional teacher into a causal, few-step student with Self Forcing, we turn a passive video generator into a controllable surgical simulator that streams at $\sim$160 inference FPS on a single NVIDIA RTX PRO 6000 Blackwell workstation GPU. Crucially, Cosmos-H-Dreams is controller-agnostic: any interface that emits a stream of robot kinematics can drive it. We demonstrate live control through a browser keyboard over WebRTC, a Meta Quest headset over WebXR, a commercial surgical robot console such as CMR Surgical's Versius, and learned policies operating in closed loop. To our knowledge, this is the first interactive surgical world model supporting live human and policy control. Human operators and policies alike can act inside the synthesized world and observe the consequences in real time. We release Cosmos-H-Dreams as an open surgical simulation system, providing a common foundation for surgical education, scalable synthetic data generation, and future intraoperative decision support.

This paper develops a fully discrete finite element method for three-dimensional incompressible magnetohydrodynamic (MHD) flows with nonlinear $p$-Laplace viscosity. The scheme combines spatial finite elements with a semi-implicit Euler time discretisation. Convergence to a weak solution is proved using a time-translation compactness argument together with Minty's monotonicity method. Under additional regularity assumptions, we derive unconditional error estimates for both the velocity and magnetic field, with no coupling restriction between the time step and mesh size. The framework extends finite element analysis of incompressible MHD systems to non-Newtonian shear-thickening fluids ($p>2$), while recovering the classical Newtonian case ($p=2$). Finally, numerical simulations are provided to validate the theoretical convergence rates and demonstrate the robustness of the proposed method.

Mixture-of-Experts (MoE) models provide a flexible framework for partitioning complex prediction problems into simpler local learning tasks through an input-dependent gating mechanism. Existing interpretable MoE approaches, such as Mixture of Decision Trees (MoDT), achieve transparency by employing homogeneous decision-tree experts, but this restricts the model to a single inductive bias across all regions of the feature space. We extend the MoDT framework by introducing heterogeneous expert families comprising decision trees, linear support vector machines, and quadratic discriminant analysis under a common probabilistic gating mechanism. To ensure coherent likelihood-based inference, non-probabilistic experts are calibrated to produce conditional class probabilities, allowing parameter estimation within the generalized Expectation-Maximization framework of MoDT. We further establish theoretical monotone ascent guarantees for the proposed heterogeneous gating updates, providing a justification for the optimization procedure. Experiments on a diverse collection of synthetic and real-world benchmark datasets demonstrate that the proposed framework adaptively specializes experts according to local data geometry, yielding interpretable expert assignments while achieving predictive performance competitive with homogeneous MoDT and Random Forests. The proposed approach combines interpretability, adaptive inductive bias selection, and probabilistic coherence within a unified mixture-of-experts framework.

In truthful interval covering, each agent has a private interval of unit length, and the goal is to decide where to place a public unit interval so as to minimize the total uncovered length while incentivizing the agents to be truthful. Previous work proposed a universally strategyproof lottery over order-statistic mechanisms with approximation ratio at most $5/3$, and established a lower bound of $3/2-o(1)$ for this class of mechanisms. We close this gap by characterizing the approximation ratio of every report-independent lottery over order-statistics. In particular, we show that, for every $n\ge2$, the optimal approximation ratio for this class is $\frac32-\frac{1}{2\lfloor n/2\rfloor}$. Beyond rank lotteries, we show that every universally strategyproof randomized mechanism has approximation ratio at least $9/8$.

Aligning large language models to human preferences is crucial for real-world deployment but frequently incurs an alignment tax, leading to the catastrophic forgetting of pre-trained general capabilities. While previous works primarily frame this problem as an optimization or architectural challenge, the inherent characteristics of preference data that drive this degradation remain largely underexplored. In this paper, we propose BALIGN, a balanced data selection strategy that explicitly mitigates catastrophic forgetting while optimizing alignment efficacy. Through theoretical and empirical analyses of the preference optimization gradient, we identify three key data-centric features that dictate parameter drift: the reference model's log-probability margin, the token length difference between chosen and rejected responses, and the TF-IDF similarity to general capability corpora. By aggregating these orthogonal features into a unified composite risk score, BALIGN systematically filters out high-risk preference samples that disrupt intrinsic model parameters or provide minimal alignment utility. Extensive experiments on standard human preference datasets demonstrate that BALIGN strongly preserves foundational capabilities without compromising alignment gains, consistently achieving the optimal Pareto frontier with minimal computational overhead.

Urdu, the world's tenth most spoken language with 246 million speakers, remains almost entirely absent from mainstream LLM safety evaluation and nine years of WOAH proceedings. To investigate whether this absence has measurable consequences for content moderation reliability, five large language models, GPT-4o, Claude Sonnet 4.5, Gemini 2.5 Flash, Qwen-2.5, and Llama-3.1, were tested across six datasets spanning Nastaliq Urdu, Roman Urdu, English, and code-switched Urdu-English. Across the five Urdu-script datasets, label instability between original-script and English-translation classification ranged from 15.9% (Gemini 2.5 Flash) to 31.6% (Qwen-2.5), with a 'Missed-in-Urdu' rate, content flagged as harmful in English translation but passed as normal in the original script, ranging from 2.4% to 9.9% (median 4.3%). A complete enumeration of all 205 papers across nine ALW/WOAH editions via the ACL Anthology API confirms zero dedicated Urdu papers across the entire period. Results indicate that current LLMs provide uneven safety assurance across Urdu's script varieties, with smaller open-weight models showing substantially higher instability and missed-harm rates than frontier closed models.

Memory systems for conversational LLMs are conventionally evaluated by direct, fact-seeking questions about prior dialogue (Direct QA): can the model recall fact X from a prior conversation? We tested whether higher Direct QA accuracy correlates with higher user satisfaction in a 4-month deployment (40 users, 1,872 sessions, 7 memory conditions). Existing-benchmark Direct QA varies from 19.7% to 70.1% across the 7 conditions, but satisfaction does not change. We hypothesize that existing benchmarks and user satisfaction are tracking different capabilities: benchmarks measure elicited retrieval (recall when asked), while conversation requires natural integration (detecting relevance and naturally weaving prior context into a response). To examine this, we introduce MemUse, a set of real user-cued memory moments drawn from the deployment, scored by an integration-aware judgment of the natural conversational response. Holding the model and context fixed, the same system that scores 78.8% on Direct QA references only 7.9% of those facts in conversation – a 71-point gap. Within these moments, Natural Integration is associated with satisfaction, whereas Direct QA is not. We release the deployment corpus and MemUse together with all judgments and scoring prompts at https://github.com/ryuichi-sumida/memuse.

Coding agents re-send large file reads and tool outputs to a frontier LLM every turn, and this context dominates their token bill. General-purpose prompt compressors are trained on prose and suit code poorly: they paraphrase identifiers and drop the exact spans an agent needs to edit. We present Paritok-4B, a 4B LoRA compressor for coding-agent trajectories built on two commitments. It is extractive: it selects spans rather than rewriting them, and 96.0% of the identifiers, paths, and numbers it emits already appear in its input, holding at 96.2% on held-out SWE-bench Lite output. It is intent-conditioned: told the agent's current task, it acts chiefly inside a retained segment, selecting which lines survive (retained lines are +0.067 more intent-relevant than removed ones, paired 95% CI [+0.056, +0.078]) rather than changing how much is retained. We distil a gpt-4.1-mini teacher over 67,074 real OpenHands trajectories into 40,606 validated examples and fine-tune Qwen3-4B. On all 300 SWE-bench Lite instances, Paritok-4B compresses agent context to 25.7% of its size, 2.0x harder than a gpt-4.1-mini compressor (50.2%) and 2.4x harder than gpt-5 (61.9%), while retaining 86.5% of uncompressed single-shot solve quality. Fed the cat -n line-numbered input real agents produce, it compresses slightly less (27.8%) and retains more (89.3%); there the paired test is informative, with 30 instances solved only uncompressed and 17 only compressed, an exact McNemar p=0.079, so at this sample size compressing context to roughly a quarter of its size does not significantly reduce the solve rate. The model is a 264 MB adapter that self-hosts on one 24 GB GPU with no per-token compressor fee, which at list prices decides the economics: gpt-5 as a compressor is net-negative, costing more than the downstream tokens it saves. Weights, data, and evaluation scripts are open (Apache 2.0).

This paper studies a class of multi-cluster aggregative games characterized by the coexistence of cooperation and competition, where each agent's cost function depends on its own strategy and the aggregate of all agents' strategies. To address the Nash equilibrium seeking problem for such games in the non-Euclidean setting, a distributed mirror descent algorithm with gradient extrapolation is proposed over time-varying intra-cluster and inter-cluster networks. The mirror descent framework employs a general Bregman divergence as the distance measure, providing greater flexibility than Euclidean-based methods, while gradient extrapolation exploits historical gradient information to improve convergence performance. Under the restricted strong monotonicity characterized by the Bregman divergence, the convergence of the proposed algorithm is established, and it achieves the $\mathcal{O}(1/k)$ convergence rate with the appropriately selected step-size and parameters. Finally, the effectiveness of the proposed algorithm is verified by an example on the demand response of energy systems.

In this paper, we develop an algebraic framework for merge-regime convertible codes that works directly with the polynomial-evaluation structure of the underlying codes. We characterize the minimal skew polynomials associated with unions of conjugacy classes, establish an evaluation-compatible product rule, and derive a special Chinese Remainder Theorem (sCRT) tailored to these evaluation sets. Based on this machinery, we propose two conversion templates for skew polynomial evaluation codes (PECs). The first applies to distinct initial PECs, whose different evaluation sets naturally provide the moduli required by sCRT. The second treats identical initial PECs, for which distinct auxiliary evaluation sets and explicit algebraic compatibility conditions are introduced to preserve unchanged symbols and to generate written symbols from designated read symbols. When instantiated with the standard basis $1,x,\ldots,x^{k-1}$, both constructions yield merge conversions for linearized Reed–Solomon codes whose final codes are equivalent to linearized Reed–Solomon codes and achieve per-symbol access-optimal cost. The framework further specializes to the commutative ring $\mathbb{F}_q[x]$, yielding corresponding conversion constructions for ordinary PECs and recovering the known polynomial-form constructions for Reed–Solomon and Tamo–Barg codes as special cases. As a further application, to the best of our knowledge, this specialization gives the first merge-regime convertible construction with Gabidulin initial codes and a final code equivalent to a Gabidulin code, while attaining per-symbol optimal access under the symbol-access model considered in this paper.