arXiv:2605.29708v1 Announce Type: new
Abstract: Mixture-of-Experts (MoE) LLMs rely on sparse, router-driven expert activation, yet how safety alignment interacts with routed expert specialization remains underexplored. A common intuition is that safety behavior may be controlled by routing harmful requests to distinct refusal-oriented experts. In this work, we provide empirical evidence for a different picture: routing patterns in aligned MoE LLMs are largely topic-driven, while safety behavior can be altered with little change to the model's intrinsic routing path.
Motivated by this observation, we present **RASET** (**R**outer-**A**gnostic **S**afety-critical **E**xpert **T**uning), a red-teaming framework that probes safety enforcement that is localized in a small subset of experts while preserving the model's intrinsic routing behavior. **RASET** identifies safety-critical experts via a contrastive routing-sensitivity criterion and applies parameter-efficient tuning only to the selected experts, minimizing semantic disruption relative to router-steering interventions. These results reveal a distinct MoE safety risk, highlighting the need for expert-aware alignment mechanisms.
Science Journals
arXiv:2605.29353v1 Announce Type: new
Abstract: The proliferation of AI-generated synthetic media poses a critical threat to the integrity of digital evidence in legal and forensic contexts. Existing deepfake detection systems typically address a single modality and provide no mechanism for tamper-proof evidence preservation. We present DeepFake Forensics AI, a unified platform that detects synthetic media across image, video, and audio modalities, identifies generative architecture fingerprints, and anchors forensic evidence immutably on the Ethereum blockchain. Our system trains four independent neural networks from scratch: an EfficientNet-B4 image detector (AUC = 0.9868), a Bidirectional LSTM video detector (AUC= 0.9628), an ECAPA-TDNN audio detector (EER = 18.63%), and a novel GAN fingerprinting module (accuracy = 99.88%) that identifies the generative architecture behind a fake image. Evidence files are hashed with SHA-256, stored on IPFS via Pinata, and registered on-chain via a Solidity smart contract with role-based access control. The platform provides a React frontend and FastAPI backend suitable for deployment in forensic and legal workflows. To our knowledge, this is the first system to unify multi-modal deepfake detection with blockchain-based chain-of custody management.
arXiv:2605.27809v2 Announce Type: replace
Abstract: Despite recent progress in backdoor attacks, existing methods remain susceptible to post-training defenses that erase the backdoor through fine-tuning or pruning. We revisit the core objectives of backdoor attacks and derive principled criteria characterizing optimal sample-specific trigger construction under a Bayes-optimal model of the victim's training. Our analysis reveals that both attack success and clean-accuracy preservation are simultaneously optimized when triggered samples are steered into low-density regions of the clean data distribution, a distributional condition that controls all moments of the poisoned distribution at once rather than a handful of input-space summary statistics. We introduce a bilevel optimization framework that estimates density ratios via conditional time-score matching and optimizes a mixture-model objective to place triggered samples in these sparse regions. Extensive evaluations on MNIST, CIFAR-10, GTSRB, and TinyImageNet demonstrate that our method achieves above 99\% attack success rate before defense and retains 50--85 percentage points higher post-defense ASR than the strongest baselines under fine-tuning defenses. Against neuron-pruning defenses, the method exhibits complete immunity, with zero neurons identified for removal across all pruning thresholds. These results expose a fundamental gap in current defense paradigms and underscore the need for defenses that operate beyond the support of the clean distribution.
arXiv:2605.27755v2 Announce Type: replace
Abstract: In this paper, we present an open-source measurement platform designed to characterize the performance of commercial cellular (Verizon, a major US provider) and LEO satellite (Starlink) networks through real-world flight tests in rural environments. We implement a comprehensive multi-layer measurement approach spanning physical layer signal metrics, multi-cell network topology, and end-to-end (E2E) application performance. Through an extensive flight campaign with more than $10$ flight tests, $4.5$+ hours of flight time resulting in more than $18$K samples, we present the first detailed, open-source dataset analyzing dual cellular and Starlink performance for low-altitude UAV operations. Our cellular-Starlink comparative results, which are collected \emph{simultaneously at the same time and location}, demonstrate significant performance differences between the two technologies: the LEO satellite link achieves superior latency performance with $95\%$ of Round-Trip Time (RTT) measurements below $50$ ms compared to $80\%$ under $150$ ms for cellular, and exceptional downlink capacity with $95\%$ exceeding $25$ Mbps versus only $5$ Mbps for cellular. Our analysis on cellular network performance demonstrates that while higher altitudes (e.g., $330+$ m above the sea level) improve signal power by $15-20$ dB via line-of-sight (LOS) propagation, it causes a $3-4$ $\times$ increase in handover rates, which is due to excessive multi-cell visibility rather than signal degradation. Furthermore, we observe asymmetric impacts on the RTT performance due to handovers such that $53.5$\% of handovers improve RTT, but worst-case degradation ($275$ ms) is $2$ $\times$ larger than best-case improvement ($137$ ms).
arXiv:2605.27580v2 Announce Type: replace
Abstract: A central puzzle for the behavioural sciences and for human-facing artificial intelligence is the persistence of within-person variability. The same individual, presented with the same observable input, produces different outcomes on different occasions, and different individuals produce divergent outcomes that no observable covariate fully predicts. We argue that this variability belongs in the dynamic latent state of the person, and that human outcomes are controllable in a precise and operational sense through interventions that target the state and its weighting at the moment a decision is being formed.
We define a state as the time-indexed weighting vector over the dimensions that govern how an individual's biology, physiology, and neuropsychology process the next event into a decision and an outcome. The relationship between state, decision, and outcome is causal rather than correlational. The weighting vector is dynamic at sub-daily timescales. The conscious channel through which outcomes are reportable is a narrow attentional bottleneck whose contents are themselves state-dependent. Taken together, these claims imply that the outcome of a given event is controllable, conditionally, on the state-trajectory at the time of intervention.
We motivate the framework with six strands of established evidence (causal inference, predictive processing, allostasis, attentional bottleneck, chronobiology, computational psychiatry) and a 24-month observational base from a deployed behavioural platform spanning more than 200,000 consented users across four occupational personas (research period 2023 to 2026). We derive seven testable predictions, list six operational requirements for state-aware systems, and discuss implications for digital health, education, AI personalisation, and personal agency.
arXiv:2605.25090v2 Announce Type: replace
Abstract: We improve upon the Johnson-type bound of Hayashi and Yasunaga for insertion-deletion codes by encoding each local list into a binary constant-weight code. The resulting local list-size bound is tight for sufficiently large alphabets. Applying the McEliece--Rodemich--Rumsey--Welch bound to this constant-weight formulation yields an asymptotic rate bound that strictly improves on Yasunaga's Elias-type bound in the nontrivial range.
arXiv:2605.29933v1 Announce Type: new
Abstract: Clustering is a fundamental problem in data science with a long-standing research history, yielding numerous insightful algorithms. Despite this progress, a systematic and large-scale empirical evaluation that jointly considers conventional algorithms, deep learning-based methods, and recent foundation model-based clustering remains largely absent, leading to limited guidance on algorithm selection and deployment. To address this gap, we introduce CLUBench, a comprehensive clustering benchmark comprising 24 algorithms of diverse principles evaluated on 131 datasets across tabular, text, and image data, involving 178,815 experiments. Importantly, our analyses of (i) the impact of hyperparameter tuning,(ii) the impact of data types and characteristics,(iii) the impact of pretrained embeddings,(iv) large language model-based clustering,(v) the similarity of algorithms, and (vi) the low-rank structures of performance matrices, yield meaningful insights and promising pathways for clustering research. For instance, our study reveals that: 1) All evaluated deep clustering methods do not exhibit a significant advantage compared with the top-performing conventional clustering algorithms (e.g., KMeans, SpeClu) in terms of average performance; 2) For image and text clustering tasks, combining pretrained embeddings with conventional clustering algorithms (e.g., KMeans, SpeClu) offers effective and efficient clustering; 3) Clustering remains a challenging and nontrivial problem, even in the era of increasingly dominant foundation models. Moreover, we propose to use the low-rank structure in cross-model performance matrices to efficiently approximate the overall performance evaluation in practical applications. We further demonstrate the feasibility of model selection based on the performance matrices across all hyperparameter configurations.
arXiv:2605.30333v1 Announce Type: new
Abstract: A plausible future mathematical claim must satisfy two constraints: it should follow the direction of prior work and respect the formal dependencies that constrain what can validly follow. Existing approaches typically model only one of these sources, producing claims that are either weakly grounded or insufficiently motivated. We introduce grounded future mathematical generation, where the goal is to generate a plausible future theorem-like claim for an anchor paper using two complementary sources of context: its scientific citation graph and aligned formal theorem dependency graph. To address this setting, we propose COMPOSE, a dual-graph framework that conditions a language model on both scientific citation context and formal theorem structure. To support this setting, we construct a dataset of 108K paired scientific-formal graph examples from arXiv and Mathlib, together with a benchmark of 47K future papers from 2024--2025. Experiments show that COMPOSE outperforms strong baselines on retrieval to real future papers and achieves the best overall performance under LLM-judge evaluation, producing more grounded and mathematically richer outputs. These results show that future mathematical generation benefits from combining scientific context with formal structure. Project page is available at https://david-busbib.github.io/COMPOSE-page/.
arXiv:2605.29334v1 Announce Type: new
Abstract: This work presents a weighted quadrature (WQ) method to fast assemble Galerkin matrices based on unstructured spline surfaces. The method is developed upon a particular variant of unstructured splines, namely the bicubic analysis-suitable unstructured T-splines (ASUTS). While existing WQ approaches have significant speedup for structured splines (e.g., B-splines), their extension to unstructured splines faces several challenges: (1) lack of a global parametric domain for defining quadrature points, (2) a varying number of basis functions across elements that complicates the determination of the optimal number of quadrature points, and (3) ill-conditioned underdetermined linear systems that must be solved to find the quadrature weights. To solve these issues, we first define the WQ rule directly in the physical domain. Second, we specify the number of quadrature points function-wise (rather than element-wise), which naturally satisfies the well-posedness condition, namely the number of unknown weights no less than that of exactness constraints. Third, we employ the truncated Singular Value Decomposition to improve the conditioning of the underdetermined systems by discarding extremely small singular values, which are caused by the splines around extraordinary points. Several different model problems are studied, such as Poisson's problem, the biharmonic problem, and the nonlinear heat transfer problem. In the end, a variety of numerical tests are performed to demonstrate the accuracy and efficiency of the proposed method.
arXiv:2605.29728v1 Announce Type: new
Abstract: Sparse tensors are the most used representation of sparse multidimensional data. Operations that decompose them, selecting their most important features while reducing their dimension, have become prevalent procedures in machine learning. One of the most used tensor decomposition algorithms is the Alternating Least Squares Canonical Polyadic Decomposition (CP-ALS), where the most time-consuming operation is the Sparse Matricized Tensor Times Khatri-Rao Product (spMTTKRP). This operation is strongly memory-bound, making it hard to implement efficiently on general-purpose processors. This work proposes PRISM, the first approach to tackle this operation using Processing-In-Memory (PIM) technology. We extensively characterize different partitioning strategies, number formats, and kernel optimizations that efficiently adapt this operation to UPMEM PIM, which is further boosted by heterogeneous collaboration with the CPU. The experimental results show that the proposed PIM-based and heterogeneous approaches achieve up to 2.37x and 2.64x speedup compared to state-of-the-art CPU implementations, respectively. However, the UPMEM distributed memory system can significantly hinder performance on certain workloads. Nonetheless, the efficiency of resource consumption for this approach, measured by peak performance fraction usage, is significantly higher than for both CPU and GPU.
arXiv:2605.29731v1 Announce Type: new
Abstract: High-density electroencephalography (HD-EEG) enables fine-grained measurement of cortical activity but requires expensive hardware and lengthy setup times, limiting its clinical and research accessibility. We propose EMAG (EEG Mixture of Anisotropic Gaussians), a differentiable framework that reconstructs HD-EEG signals from a sparse subset of low-density (LD) electrodes by representing brain electrical sources as a mixture of anisotropic 4D space-time Gaussians. EMAG places a mixture of multiple Gaussians at each point of a spherical brain grid, each parameterized by a full 4 x 4 precision matrix, enabling anisotropic spatial spreads and explicit coupling between spatial and temporal dimensions. The forward model renders scalp EEG via differentiable Gaussian field contributions at electrode locations, enabling end-to-end training without explicit source localization supervision. We evaluate EMAG on three public EEG benchmarks (Localize-MI, SEED, and SEED-IV) at super-resolution factors of 2x through 8/16x. EMAG outperforms the current state-of-the-art EEG super-resolution method at most super-resolution factors on three standard benchmarks (Localize-MI, SEED, SEED-IV). The explicit Gaussian parameterization further enables direct visualization and interpretability of learned brain source configurations, potentially opening avenues for clinical and neuroscientific applications, such as source localization or biomarker discovery.
arXiv:2605.30058v1 Announce Type: new
Abstract: While LLM agents have demonstrated remarkable task-oriented abilities such as planning, reasoning, and action, few works have treated them as complete human personalities where emotional dimensions hold equal importance. In this paper, we introduce a novel benchmark to systematically assess whether LLM agents can simulate coherent, human-like psychology. Specifically, our benchmark constructs 11 diverse human characters grounded in orthogonal Big Five personality traits, with each profile deeply integrated with 1,000 structured autobiographical-style episodic memories distributed across theory-grounded developmental life stages. To rigorously evaluate the psychological manifestations of LLMs, we designed a curated suite of 64 decision-making scenarios, guided by the DIAMONDS taxonomy, a psychological framework that characterizes situations along eight dimensions: Duty, Intellect, Adversity, Mating, pOsitivity, Negativity, Deception, and Sociality. By subjecting agents to varying scenarios, the benchmark evaluates whether they can consolidate their innate personality traits and autobiographical memories to make behavioral decisions that are consistent with their specific psychological profiles. After systematic human validation and filtering, we obtained a benchmark consisting of 673 multiple-choice questions (MCQs). We believe this benchmark provides a principled and scalable testbed for studying human-like emotions, personality consistency, and value-consistent behavioural decision-making in LLM-based agents.
arXiv:2605.29259v1 Announce Type: new
Abstract: Given the wide range of deployment targets, flexible model selection is essential for optimizing performance within a given compute budget. Recent work demonstrates that stitching pretrained models within a model family enables cost-effective interpolation of the accuracy-efficiency tradeoff space. Stitching transforms intermediate activations from one pretrained model into another, producing a new interpolated stitched network. Such networks provide a pool of deployment options along the accuracy-efficiency spectrum. However, existing stitching approaches often yield suboptimal tradeoffs and lack generalizability, as they primarily rely on heuristics to select stitch configurations. We argue that constructing improved accuracy-efficiency tradeoffs requires explicitly capturing and leveraging the similarity between pretrained models being stitched. To this end, we introduce KLAS, a novel stitch selection framework that automates and generalizes stitch selection across model families by leveraging KL divergence between intermediate representations. KLAS identifies the most promising binary stitches from the $O(k^2n^2)$ possibilities for $k$ pretrained models of depth $n$. Through comprehensive experiments, we demonstrate that KLAS improves the accuracy-efficiency curve of stitched models at the same finetuning cost as baselines. KLAS achieves up to $1.21\%$ higher ImageNet-1K top-1 accuracy at the same computational cost, or maintains accuracy with a $1.33\times$ reduction in FLOPs.
arXiv:2605.29761v1 Announce Type: new
Abstract: Compositional implicit surface representations model scenes as collections of objects, each encoded by a Signed Distance Field (SDF). A fundamental limitation of this approach is that multiple SDFs can produce geometries that interpenetrate, violating physical plausibility. Existing mitigation strategies rely on soft penalty terms that reduce but do not eliminate intersections, and require careful loss weighting. To truly prevent interpenetration, we propose a hard constraint on vector-valued SDFs and introduce S2MDF, a lightweight plug-and-play module that enforces the constraint on any object-compositional SDF representation without architectural modifications. It introduces negligible computational overhead and is compatible with linearly-interpolated standard meshing algorithms such as Marching Cubes. It can be applied during training or as a post-processing step. Experiments on multiple state-of-the-art compositional methods show that S2MDF reduces intersections to numerical precision while preserving reconstruction quality, outperforming existing mitigation strategies.
arXiv:2605.29351v1 Announce Type: new
Abstract: We study minimal attention-only transformers under all-token corruption and show they admit a two-stage empirical Bayes interpretation. A single attention step computes a kernel-weighted posterior mean with respect to the empirical distribution defined by the context. Depth refines this distribution through particle dynamics (Stage 1), while a long-range skip-connection carries the noisy input as a query for posterior inference (Stage 2), revealing distinct statistical roles for depth and attention residuals. The framework isolates a minimal setting in which the context itself induces a depth-dependent energy landscape governing in-context inference. We show that effective denoising can emerge without an explicit noise schedule: a fixed kernel bandwidth and finite integration horizon suffice, yielding a principled depth-noise relationship. We further establish a posterior-mean recovery guarantee for a class of well-behaved priors, where the empirical estimator converges to the Bayes-optimal predictor under asymptotic conditions. Connecting these dynamics to reverse-diffusion limits, our results provide a statistical interpretation of attention as in-context inference via sample-based posterior estimation, without explicit density modeling.
arXiv:2605.29220v1 Announce Type: new
Abstract: Tracking the dynamics of non-canonical biological systems in microscopy videos remains a persistent challenge. Both classical and learning-based trackers depend on expert-reviewed data to be evaluated and adapted, yet exhaustive manual annotation rarely scales to the videos where these tools are needed most. We developed RIPPLE (Refinement Interpolation Platform for Point Location Estimation), which recasts annotation as sparse correction: a user clicks a starting point, RIPPLE proposes a full trajectory, and the user intervenes only where the trajectory drifts. We tested RIPPLE on five challenging microscopy datasets from our laboratories, four from the transparent jellyfish Clytia hemisphaerica and one tracking landmarks on rapidly moving sperm. Across these, RIPPLE matched the quality of exhaustive manual annotation while reducing manual clicks by 3 to 25 times across datasets. RIPPLE thereby fills a missing layer between manual annotation and fully automated tracking, enabling immediate quantification of biological dynamics, method benchmarking, and the production of the gold-standard data needed to adapt future automated microscopy trackers.
arXiv:2605.29270v1 Announce Type: new
Abstract: The era of the Internet of Agents (IoA) is taking shape: LLM agents are expected to fulfill user goals by orchestrating fast-growing populations of Model Context Protocol (MCP) servers, Agent-to-Agent (A2A) endpoints, reusable skills, and other LLM-callable services. Yet LLMs face a structural mismatch with this regime: effective context is a scarce resource that does not scale with the number of services. Concatenating thousands of service descriptions into a prompt overflows the context window, and even when the window is large enough, models systematically under-attend to information in the middle of long inputs, the well-documented Lost-in-the-Middle phenomenon. This is fundamentally a question of context management for service discovery. To address this, we propose an LLM-native progressive-disclosure scheme and its concrete instantiation, A2X (Agent-to-Anything service discovery): an LLM-driven pipeline that automatically organizes the registered services into a hierarchical taxonomy and walks it layer by layer at query time, so that every LLM call sees only a small candidate set highly relevant to the user query. This decouples effective-context scarcity from registry size and significantly reduces token consumption while improving retrieval accuracy. Compared to full-context dumping, A2X achieves a 6.2-point Hit Rate gain at one-ninth the prompt-token cost; compared to the state-of-the-art open-source embedding-based baseline, A2X improves Hit Rate by more than 20 points.
arXiv:2605.29739v1 Announce Type: new
Abstract: Network psychometrics conceptualises psychological constructs as emergent properties of systems of interacting variables. Energy-based probabilistic models have gained popularity as models of these interactions, but their psychometric application has so far been limited, since most implementations assume binary or ternary responses and rely on limiting inference assumptions. We infer and analyse three generalised spin models of ordinal questionnaire data: the generalised Ising, Blume-Capel (BC), and Blume-Emery-Griffiths (BEG) models. We prove the concavity of the maximum likelihood estimation of the parameters, as well as the gauge invariance of the Ising and BC models. Afterwards, we propose an inference protocol of approximated likelihood maximisation, based on the Monte Carlo estimation of the likelihood gradients. We apply this procedure to eleven psychometric and sociological questionnaires, comparing the inferred spin models against the multivariate Gaussian. We then assess whether the inferred models reproduce the empirical features of the data in terms of principal-component histograms, and histograms of Euclidean and Mahalanobis distances to the mean answer. The multi-modality observed in the histograms of principal components is partially captured by the spin models. This trait of polarisation can be understood, in the light of mean-field theory, as coexistence of stable and metastable phases of the spin models. The BEG model systematically outperforms the other models in capturing the distribution of distances to the mean, while all models underestimate the heavy tails of the Mahalanobis distance. Overall, the analysis witnesses the predictive power of the BEG model, able to account better than others for the abundance of outliers and mean responders, and reveals highly non-linear features of questionnaire data that both Gaussian and spin models fail to account for.
arXiv:2605.29350v1 Announce Type: new
Abstract: Mixture-of-Experts (MoE) language models reduce per-token computation but still require storing and serving all experts, making deployment memory-intensive. Existing post-training compression methods mainly shrink this cost by pruning experts or merging their weights. We formulate post-training MoE compression as expert-pool consolidation: retaining a smaller set of pretrained experts as reusable prototypes and deterministically remapping each original expert reference to one selected prototype. This view separates the reduced expert pool from the reuse structure that represents the original expert slots, and allows prototype sharing within local layer scopes while preserving the original router interface. We propose ConMoE, a train-free prototype remapping framework that selects retained experts using calibration-based contribution and replaceability signals, then redirects original expert calls to the selected prototypes without weight updates or post-compression fine-tuning. Experiments on three pretrained MoE language models show that ConMoE matches or outperforms strong pruning and merging baselines in several settings, achieving the best average score on deepseek-moe-16b-base at both 25% and 50% routed-expert reduction, while remaining competitive on Qwen3-30B-A3B and OLMoE-1B-7B-0125. Ablations indicate that deterministic reassignment is the most stable component, whereas broader cross-layer sharing and post-hoc weight fusion are model-dependent.
arXiv:2605.29888v1 Announce Type: new
Abstract: Reinforcement learning (RL) post-training has shown to improve reasoning in large language models (LLMs). However, there has been little exploration on the problem of data contamination in RL post-training, potentially undermining generalization and evaluation reliability of the training process itself. Existing detection methods primarily rely on output-level signals such as likelihood or entropy, which become unreliable for RL-trained models since RL shapes behavior through trajectory-level rewards rather than token likelihoods. We propose LaRA, a layer-wise representation analysis framework for detecting contamination in RL post-trained LLMs. LaRA introduces three complementary metrics, measuring perturbation sensitivity, directional collapse, and local representation rigidity under controlled perturbations. We find that contamination produces progressive geometric deviations across layers, including amplified perturbation sensitivity, stronger directional collapse, and enhanced local rigidity. Based on our findings, we also develop a contamination detection protocol that aggregates representation-level deviations across layers and metrics. Experiments on RL-trained reasoning models show that our protocol outperforms existing output-level baselines for contamination detection.
arXiv:2605.29272v1 Announce Type: new
Abstract: Fraud detection models in payment networks train on chargeback labels that are systematically biased. Every label must survive three sequential gates: authorization (declined transactions generate no labels), issuer reporting (unreported fraud is invisible), and delay (pending chargebacks are missing at training time). Labels that do arrive may be corrupted by first-party misuse or issuer misclassification. A companion paper [arXiv:2605.27557] proved that these four impairments impose a minimax lower bound on detection performance. This paper asks: can that bound be achieved?
We formalize the observation pipeline as a sequential missing-data problem with three propensity stages and a corruption layer, and construct the Sequential Triply Robust (STR) estimator. The STR corrects for all four impairments simultaneously and achieves the semiparametric efficiency bound -- no estimator can have lower asymptotic variance. It is sequentially triply robust: at each gate, consistency requires only that either the propensity model or the outcome regression is correctly specified, not both. We provide corruption correction via noise-rate-adjusted pseudo-labels, empirical Bayes shrinkage to stabilize inverse-propensity weights for small issuers, a plug-in variance estimator yielding valid confidence intervals, and a Bernstein concentration inequality for finite-sample guarantees.
On the operational side, we derive the optimal training delay -- the maturity window that minimizes the sum of label-quality loss and model staleness -- and prove that the STR permits training on data that is days old rather than months old, decoupling model freshness from the chargeback maturity cycle. The STR provably dominates naive chargeback-based training in mean squared error for any sample size.
arXiv:2605.29414v1 Announce Type: new
Abstract: Recent studies have shown that code-switching data (CSD), in which multiple languages are mixed within the same context, can improve cross-lingual transfer and multilingual alignment in large language models (LLMs). However, existing studies primarily focus on bilingual transfer between English and a target language, leaving multilingual settings involving three or more languages largely unexplored. In this work, we investigate multilingual code-switching instruction tuning across four languages: English, Japanese, Korean, and Chinese. We evaluate multilingual understanding on Belebele. Our experiments show that simple sentence-level multilingual CSD consistently improves average multilingual performance across all four languages, indicating that multilingual code-switching can be effective beyond bilingual transfer settings.
arXiv:2605.29766v1 Announce Type: new
Abstract: Imitation learning has become a cornerstone for solving complex robotic manipulation tasks. In particular, multimodality, which enables robots to capture diverse yet valid behavioral patterns, has driven the rapid emergence of generative policies as a dominant paradigm in robot learning. However, achieving such multimodality typically relies on stochastic noise initialization and iterative denoising procedures, resulting in substantial training complexity and low inference efficiency. Meanwhile, not all phases of a robotic task inherently require behavioral diversity. Motivated by this insight, we propose the Modality-Adaptive Robot Sampling (MARS) policy, which adaptively invokes tailored stochasticity only when it is truly beneficial, while reverting to an efficient deterministic learning during single-modal phases. In other words, the proper amount of noise is injected only at the proper time. By selectively activating multimodal generation, MARS policy bridges the gap between the multimodal capability of generative policies and the superior training and inference efficiency of deterministic models. Empirical studies across 8 simulated and 4 real-world tasks demonstrate that MARS exhibits robust multimodal expressivity and high efficiency, with a 16.67% success rate improvement and an 83.20% inference latency reduction in real-world tests. Counterintuitively, MARS also outpaces deterministic policies in training efficiency on near-deterministic tasks by more effectively modeling nuanced action diversity.
arXiv:2605.29549v1 Announce Type: new
Abstract: 3D Gaussian Splatting (3DGS) is a recent approach for scene rendering. Although primarily designed for view synthesis, its potential for scene understanding tasks remains underexplored. In this work, we conduct a comparative evaluation of various geometric deep learning architectures for the classification of 3D scenes represented using Gaussian Splatting. We benchmark point-based and graph-based models across both traditional point cloud datasets and dedicated Gaussian Splatting datasets. Scenes are embedded into latent representations, which are evaluated through end-to-end classification, linear probing, and clustering analysis. Our study provides insight into the suitability of different geometry-aware architectures and input feature configurations for learning effective 3D Gaussian Splat representations. The results highlight consistent differences between architectural families and reveal the impact of Gaussian-specific attributes on the quality of representation.
arXiv:2605.29889v1 Announce Type: new
Abstract: Patient-voiced clinical-triage benchmarks report high under-triage rates for consumer LLMs for constrained multiple-choice output, yet the same cases score differently with free-text. We ask whether output format changes the model's \emph{clinical representation} or only the mapping from a preserved representation to an answer. Using sparse-autoencoder (SAE) features in Gemma 3 4B/12B IT and Qwen3-8B, we find the same medical features fire on the shared clinical narrative under both formats but go {silent} at the multiple-choice decision token in all the cases at every model. Three independent methods (natural-language autoencoder verbalization, decision-token logit attribution, and top-feature characterization) agree that scaffold and format features, but not medical features, drive the decision logits. Behaviorally, the multiple-choice penalty inverts under both structured and natural-language input, option-order shuffle rules out positional bias, and the gap is dominated by off-by-one decision (the model picks an adjacent acuity letter to the gold answer) rather than knowledge failure. Thus, the failure originates in the output format and not in the clinical representation.