arXiv:2605.13230v2 Announce Type: replace
Abstract: On-policy distillation (OPD) has become a promising paradigm for reasoning-oriented post-training of large language models (LLMs), especially when combined with reinforcement learning from verifiable rewards (RLVR). Existing OPD methods rely on reverse KL (RKL)-based teacher supervision over trajectories sampled from the student policy. However, we identify a critical limitation: under large teacher--student policy divergence, RL-driven exploration often produces trajectories outside the teacher distribution, resulting in uninformative negative feedback. To address this, we propose Teacher-Guided Policy Optimization (TGPO), an on-policy reasoning distillation method that remains effective under large policy divergence settings. Rather than relying solely on evaluative supervision, TGPO uses teacher to directly guide token level generation conditioning on student-generated contexts; together with RLVR-style trajectory level rewards, TGPO steers exploration toward improved continuations. Experiments on reasoning benchmarks show that TGPO consistently outperforms existing RKL-based OPD methods and remains robust across different teacher models.
Science Journals
arXiv:2605.11723v2 Announce Type: replace
Abstract: In this paper, we propose Concentrate and Concentrate (CaC), a coarse-to-fine anomaly reward model based on Vision-Language Models. During inference, it first conducts a global temporal scan to anchor anomalous time windows, then performs fine-grained spatial grounding within the localized interval, and finally derives robust judgments via structured spatiotemporal Chain-of-Thought reasoning. To equip the model with these capabilities, we construct the first large-scale generated video anomaly dataset with per-frame bounding-box annotations, temporal anomaly windows, and fine-grained attribution labels. Building on this dataset, we design a three-stage progressive training paradigm. The model initially learns spatial and temporal anchoring through single- and multi-frame supervised fine-tuning, and then is optimized by a reinforcement learning strategy based on two-turn Group Relative Policy Optimization (GRPO). Beyond conventional accuracy rewards, we introduce Temporal and Spatial IoU rewards to supervise the intermediate localization process, effectively guiding the model toward more grounded and interpretable spatiotemporal reasoning. Extensive experiments demonstrate that CaC can stably concentrate on subtle anomalies, achieving a 25.7% accuracy improvement on fine-grained anomaly benchmarks and, when used as a reward signal, CaC reduces generated-video anomalies by 11.7% while improving overall video quality.
arXiv:2604.23862v3 Announce Type: replace
Abstract: We investigate whether the Feed-Forward Network (FFN) sublayer in a decoder-only transformer can be replaced by an explicit learned memory graph while preserving the surrounding autoregressive architecture. The proposed Graph Memory Transformer (GMT) keeps causal self-attention intact, but replaces the usual per-token FFN transformation with a memory cell that routes token representations over a learned bank of centroids connected by a learned directed transition matrix. In the base GMT v7 instantiation studied here, each of 16 transformer blocks contains 128 centroids, a 128 * 128 edge matrix, gravitational source routing, token-conditioned target selection, and a gated displacement readout. The cell therefore returns movement from an estimated source memory state toward a target memory state, rather than a retrieved value. The resulting model is a fully decoder-only language model with 82.2M trainable parameters and no dense FFN sublayers, compared with a 103.0M-parameter dense GPT-style baseline used in the evaluation. The base v7 model trains stably and exposes centroid usage, transition structure, and source-to-target movement as directly inspectable quantities of the forward computation. It remains behind the larger dense baseline in validation loss and perplexity (3.5995/36.58 vs. 3.2903/26.85), while showing close zero-shot benchmark behavior under the evaluated setting. These results are not intended as a state-of-the-art claim; they support the viability and structural interpretability of replacing dense within-token transformation with graph-mediated memory navigation. Broader scaling, optimized kernels, and more extensive benchmark evaluation are left for subsequent work.
arXiv:2605.30258v1 Announce Type: new
Abstract: LLMs are increasingly deployed to simulate social interactions, yet many of the existing simulators remain ad hoc and monolithic. This lack of architectural standardization prevents reproducible research and complicates downstream evaluation. We advance a rigorous science of LLM-based multi-agent simulation by modularizing core components into Environments, Agents, Simulation engines, and Evaluation metrics (EASE). We demonstrate the utility of EASE configuration by wrapping it in an experimental study schema for orchestrating workflows centered around answering explicit research questions in generated scenarios. We contribute SiliSocS, an open-source, research-ready Silicon Society Sandbox implementing a study-structured EASE configuration to enable highly configurable and reproducible LLM-based social simulations. Using SiliSocS and EASE, we present three case studies, showcasing the system's comprehensive assessment of existing questions, ability to dive deeper into complex questions, and elaboration of existing studies, respectively. Together, these case studies highlight the limitations of current modeling approaches and isolate the impacts of design choices on key results.
arXiv:2605.29943v1 Announce Type: new
Abstract: Motor imagery (MI) classification using electroencephalography (EEG) signals is essential for advancing brain-computer interfaces (BCIs). Traditional EEG channel selection methods often face limitations, such as dependency on single-objective criteria and susceptibility to local optima. To address these challenges, this work proposes a multi-objective optimisation framework that employs non-dominated sorting genetic algorithm, multiple-objective particle swarm optimisation, and a multi-objective evolutionary algorithm based on decomposition. Our approach effectively balances spatial relevance, using a Gaussian kernel, and functional discriminability, which assesses intratrial task-related desynchronisation, thereby improving performance. We evaluated this framework on four EEG datasets: Physionet, OpenBMI, HighGamma, and BCIIV-2A. The proposed approach successfully identifies compact, relevant channel subsets concentrated around sensorimotor cortex regions linked to MI activity, addressing the prevalent challenges of dimensionality and complexity inherent to traditional techniques. Furthermore, the framework achieved classification performance of 87%, 71%, 75%, and 65% on the Physionet, OpenBMI, HighGamma, and BCIIV-2A datasets, respectively. By outperforming existing single-objective and accuracy-based methods, and those relying on fixed subsets, these findings demonstrate that this new multi-objective optimisation framework can enhance MI-based BCI performance while facilitating compact channel configurations with reduced computational complexity, making them better suited for wearable, portable, and real-time BCI applications.
arXiv:2605.29126v1 Announce Type: new
Abstract: A linear probe can decode a representation almost perfectly and yet be completely irrelevant to how the model uses it. On calendar-date duration reasoning in language models, a $\sin$/$\cos$ probe recovers day-of-year from a layer's activations, yet ablating its direction has no effect on the model's answers -- while ablating a four-dimensional subspace found by Distributed Alignment Search (DAS) at the same layer collapses performance entirely. We measure the angle between these two subspaces -- the \emph{readout-mediator angle} -- and find it indistinguishable from the angle between two random subspaces (the Haar-uniform null), meaning the probe has learned a direction orthogonal to the model's actual computation. Reverse-engineering the circuit reveals why: attention heads route month-grained context through learned QK offsets at ${\pm}30$ and ${\pm}61$ days, and MLPs then convert \emph{when} (absolute date) into \emph{how long} (duration) -- all downstream of the causal subspace the probe never touches. Sparse-autoencoder decomposition confirms the split: probe-aligned and DAS-aligned features encode semantically disjoint concepts with negligible causal overlap. The dissociation replicates across four scales ($1.5$-$9\,$B) and two model families, with preliminary evidence on two further domains (spatial displacement, symbolic arithmetic), suggesting that readout-mediator orthogonality is a general failure mode of probe-based interpretability. This directly undermines proposals to deploy probes as runtime safety monitors: the probe can report high confidence on a direction the model has silently abandoned.
arXiv:2605.29064v1 Announce Type: new
Abstract: We study how persona prompting shapes language generated by multimodal large language models in an urban perception setting. Using 59,808 annotations from 1,200 persona-conditioned agents and two no-persona settings, we analyze captions, justifications, and perception tags across personas. Results indicate strong convergence in captions for different personas, whereas justifications display systematic variation associated with socioeconomic and political attributes, while perception tags show no statistically significant persona-related differences, though effect trends are observed. Topic analysis further reveals that personas emphasize different evaluative themes when interpreting the same scenes.
arXiv:2605.28909v1 Announce Type: new
Abstract: We develop a comprehensive mathematical and computational framework for sequential surrogate modeling of three-phase black-oil reservoir dynamics using neural operators, with particular emphasis on Fourier Neural Operators (FNO) and their physics-informed variant (PINO). The application focus is the Norne benchmark reservoir, defined on a heterogeneous $46\times112\times22$ grid ($N=113,344$ cells), with a production history spanning $T=30$ timesteps covering 3298 days. Our theoretical contributions are organized around four interlocking problems: (1) functional-analytic formulation in a product-Sobolev-space setting, including well-posedness of the implicit timestep map and sharp local Lipschitz estimates; (2) covariate shift quantification, proving that the Wasserstein-2 distance grows as $W_2 \leq \varepsilon(L^n-1)/(L-1)$, with exponential population-risk discrepancy for $L>1$; (3) physics-constrained spectral stability, showing PINO training with $\lambda_R \geq \lambda^*_R$ reduces the learned Jacobian spectral radius to $\rho_F + C\lambda_R^{-1/2}$, yielding uniform-in-time rollout error $|\delta_n| \leq \varepsilon/(1-\rho)$; and (4) $K$-step TBPTT gradient analysis, deriving geometric bias decay $O(\rho^K)$, optimal window $K^ = O(\log(T/\sigma^2))$, and Adam convergence $O(1/\sqrt{t}) + O(\rho^{K^*})$. Empirical validation confirms all theoretical predictions: autoregressive PINO surrogates sustain $R^2>0.99$ (oil), $R^2>0.90$ (gas), $R^2\approx 0.80$ (pressure), and monotonically improving $R^2$ (water) across the full 3298-day horizon, trained on eight NVIDIA B200 GPUs in under one hour. A 1000-member ensemble runs in under one minute on a single B200 GPU, giving a ${\sim}10^4\times$ wall-clock speedup over the OPM finite-volume simulator.
arXiv:2605.28901v1 Announce Type: new
Abstract: This paper investigates the identification of observable low-frequency (LF) parameters of battery cell's equivalent circuit models (ECMs) using time-domain voltage and current measurements sampled at low frequency by built-in battery management systems (BMS) during operation. Accurate estimation of such parameters is challenging due to measurement resolution available in practical settings. To address this, a modeling and identification framework is proposed in which fractional constant phase element (CPE), commonly used to model LF diffusion phenomena of battery cells, is approximated in the time domain using a high-order RC network with a recursive definition. The parameter estimation problem is formulated as a constrained, non-convex least-squares problem in a discretized state-space representation. To improve robustness, parameter initialization strategies, bounds, and a procedure for selecting the number of RC branches are rigorously derived. The method is evaluated in a numerical study based on a power system application where the battery under the study provides primary frequency control to the grid. Under noise levels representative of typical BMS measurements, the proposed approach achieves, from time-domain measurements, accurate LF parameter estimation (including the CPE), with average errors below 1 %.
arXiv:2605.29971v1 Announce Type: new
Abstract: Causal interventions in language model representations have largely targeted discrete features, like grammatical number. However, language models must also make use of features that are graded. We introduce a method for causal intervention on continuous variables: given activation vectors paired with a graded target variable, we localize a low-dimensional direction for that variable and use this direction to edit a vectors toward counterfactual target values. We apply this method to a continuous feature that is well-studied in psycholinguistics, namely verb bias (which reflects which syntactic structures tend to follow a given verb). We show that verb bias is causally represented in steering vectors extracted from large language models: counterfactual edits to verb bias systematically shift downstream structural preferences. Verb bias has also previously been linked to in-context learning; in further analyses, we find that steering vectors encode error signals that could drive the error-driven update behavior seen in in-context learning but that these aspects of the steering vectors are not causally used in downstream production. Overall, these results show causal interventions can be applied to continuous variables, though connecting continuous variables to in-context learning remains a challenge.
arXiv:2605.29927v1 Announce Type: new
Abstract: Despite recent advances, LLM-based web agents still struggle with limited exploration, omission of critical steps, and sensitivity to task constraints. Prior work suggests that many of these failures stem from weaknesses in planning, yet the impact of alternative natural language plan representation remains unexplored. To address this, we introduce PlanAhead, a static planner-executor framework that evaluates the impact of plan representation in agent performance. We first automatically categorize WebArena tasks into 3 difficulty levels, enabling consistent difficulty grading without human annotation. Then we systematically evaluate 4 different plan representations on the tasks categorized as hard: sequential subgoals, narrative, pseudocode, and checklist; across different families of multimodal LLM powered agents (OpenAI, Alibaba, and Google). To account for stochastic variability, we introduce two novel evaluation metrics: Achievement Rate (AR) and Solved-Task Consistency (STC). Our results show that both, the plan formulation and the underlying LLM generating the plan, significantly influence web-agent robustness and task success.
arXiv:2605.26756v2 Announce Type: replace
Abstract: Diffusion models can unintentionally memorize training samples, raising concerns about privacy and copyright. While recent methods can detect memorization, they often rely on global or model-specific signals and provide limited insight into where memorization appears within a generated image. We provide a geometric characterization of local memorization as a coordinate-wise variance collapse. However, such collapse can also arise from intrinsic data constraints rather than overfitting. To isolate overfitting-driven memorization, we propose curvature-difference methods that subtract the curvature of an underfitted baseline, either the unconditional model or a less-trained version of itself. We further derive a score-difference proxy that provides a geometric explanation for the widely used score-difference-based detection metric. Experiments on Stable Diffusion, evaluated against ground-truth memorization masks, show that our method outperforms the prior attention-based localization method. Code is available at https://github.com/Gwangho99/mem-curv-diff.
arXiv:2605.25376v2 Announce Type: replace
Abstract: KYA (Know Your Agents) is an open-source, framework-agnostic trust and governance layer for autonomous systems, composed of five primitives: (1) a four-gate inbound apply pipeline; (2) an only-tighten composition algebra over a three-channel multi-tenant hierarchy; (3) KYP (Know Your Principal), a schema-level unification of trust scoring across human users, AI agents, and service accounts; (4) auditable interaction-multiplier amplification over an AIVSS-shaped additive baseline; and (5) two-axis delegation attribution: a static premium for risky delegates and a runtime debit for actual delegate misbehavior in multi-agent fan-out. Together these span three pillars (trust, governance, and evidentiary assurance), making an autonomous system's actions authorized, policy-conforming, and post-hoc verifiable: where observability answers how long, how much, and what path, KYA answers was it authorized, did it conform, and can it be verified; it composes with observability rather than replacing it. It ships native adapters for 15+ agent frameworks. On a 4 by 9 cross-backend matrix all 36 cells pass; the pure-function scorer runs sub-millisecond at p99 and the system sustains ~ 1,800 ops/sec at 20 concurrent workers with HMAC chain integrity preserved end-to-end. KYA detects 89% of 1,200 adversarial probes from PyRIT and Garak, including the recently-published topology-guided multi-agent attack. The system is available under Apache 2.0 as the veldt-kya package on PyPI.
arXiv:2605.25303v2 Announce Type: replace
Abstract: The $2 \rightarrow q$ norm of a matrix $X \in \mathbb{R}^{n \times d}$ is defined as $\lVert X \rVert_{2 \rightarrow q} = \sup_{\lVert v \rVert_2 = 1} \lVert Xv \rVert_q$. We give polynomial-time multiplicative approximation algorithms for this norm when $q > 2$ (i.e. in the hypercontractive setting). This problem either directly captures or is closely related to long-standing open problems in combinatorial optimization and hardness of approximation (e.g. Small Set Expansion), quantum information (e.g. Best Separable State), and algorithmic statistics.
Very little is known about what approximation factors we can achieve for this problem in polynomial time, even though such approximations have significant downstream consequences. Barak, Brand\~{a}o, Harrow, Kelner, Steurer, and Zhou showed that no polynomial-time algorithm can achieve an approximation factor better than $2^{\sqrt{\log n}}$, assuming the Exponential Time Hypothesis (FOCS'12). On the other hand, a simple spectral algorithm gives a $d^{1/4}$-approximation as a baseline. We give, to the best of our knowledge, the first polynomial-time approximation algorithm beating this baseline by polynomial factors. For the important special case of $q = 4$ it achieves a $d^{1/8}$-approximation. All previous algorithms required additional assumptions on $X$, or only surpassed the baseline for small values of $n$.
Moreover, we construct sum-of-squares certificates for the $2 \rightarrow q$ norm. This directly implies improved algorithms for robust mean and covariance estimation, robust regression, and clustering, when the data only satisfies a bound on its $q$-th moment.
arXiv:2605.26193v2 Announce Type: replace
Abstract: Time series anomaly detection (TSAD) has long been a hot research topic in data mining due to its various applications. Recent studies challenge the effectiveness of popular deep learning methods for TSAD, suggesting their failure in detecting subtle and prolonged anomalies. Outlier Exposure (OE) and Masked Autoencoder (MAE) emerge as two promising paradigms (classification and reconstruction) for solving the above problems. However, OE-based methods are constrained by poor generalization, while MAE-based methods are limited by masking misalignment issues. To address these limitations, this paper proposes a novel framework, CoAD, which unifies the two paradigms to leverage their complementary strengths while mitigating their respective weaknesses. In this framework, the classification module generates probability-informed soft masks for the reconstruction module, which in turn alleviates the generalization problem of the classification module. This cooperative design enables CoAD to effectively detect subtle and complex anomalies that are often overlooked by existing methods. Additionally, the classification module is carefully designed to resolve issues related to improper classification granularity and the neglect of frequency information. Extensive experiments on high-quality benchmark datasets, conducted under rigorous evaluation protocols, demonstrate that CoAD significantly outperforms both state-of-the-art deep learning and traditional data mining methods, highlighting the potential of deep learning in TSAD. Moreover, CoAD is lightweight and substantially faster than existing SOTA methods, demonstrating its practical value for large-scale, real-time applications.
arXiv:2605.24846v2 Announce Type: replace
Abstract: Large language models (LLMs) display strong comprehensive abilities, yet the internal mechanisms that support these behaviors remain insufficiently understood. In this work, we show that across a wide range of open-weight Transformers, a subset of neurons remains consistently highly activated during inference across tasks of multiple capability dimensions. By probing along the cross-task activation strength, an extremely sparse subset is isolated, whose removal causes a collapse in model behavior, which we term keystone neurons. Our analysis reveals that keystone neurons are a stable and intrinsic neuron subset of the model that is largely established during pretraining. The parameters associated with these neurons are tightly calibrated during the training process, and their precise values are critical for the capabilities of the model. Building on these insights, we propose a supervised fine-tuning approach that updates only keystone neurons, achieving task gains comparable to or even better than full-parameter fine-tuning while better preserving performance in other capability dimensions, despite modifying a much smaller number of parameters.
arXiv:2605.22780v2 Announce Type: replace
Abstract: Plasmonic resonances in metallic nanogaps can confine light into nanometric regions, but reaching modes of volume $\approx 1$ nm$^3$ remains challenging. We present a detailed theoretical analysis of the optical modes of an experimentally-motivated nanoresonator configuration that contains a picopatch formed by the lifting of a few gold atoms in the gap of a Nanoparticle-on-Mirror (NPoM) structure. Classical simulations indicate that the plasmonic modes associated with the picopatch geometry can confine light to extremely small regions and are highly sensitive to the picopatch size and shape, enabling broad tunability. Furthermore, these modes can couple strongly with other nanocavity modes of the structure, as identified from a clear anti-crossing behavior of the polaritonic resonances. Remarkably, up to $\approx 2000$-fold electric field enhancement inside the picopatch and tiny effective mode volumes approaching $\approx 1$ nm$^3$ are obtained. Further, changing the morphology of the picopatch does not modify the qualitative findings, and increasing the absorption losses in the classical simulations, to mimic quantum (non-local) effects in the metal permittivity, decreases the electric field enhancement only moderately. Compared to the standard picocavities formed by single-atom protrusions, picopatches are an intriguing alternative to reach extreme optical field confinement and enhancement in plasmonic cavities.
arXiv:2605.18265v2 Announce Type: replace
Abstract: Resonances from electron attachment to $\text{NO}_2$ have been identified in the total electron scattering cross sections (TCS) measured with a magnetically confined electron transmission apparatus. The corresponding absolute electron attachment cross sections have been derived from the TCS values through a self-consistent scattering channel deconvolution process. The positions of some of these resonances are consistent with previous elastic scattering calculations and dissociative electron attachment experiments. However, both the observed positions and the magnitude of the present resonance cross sections are in contradiction with the available recommended TCS values thus suggesting that electron scattering cross section databases need to be updated.
Keywords: electron scattering resonances, negative ion formation, electron attachment cross sections
arXiv:2605.17286v2 Announce Type: replace
Abstract: While hyperspectral imaging provides rich spatial-spectral information across hundreds of narrow wavelength bands for precise material identification, ground-based hyperspectral pre-trained backbones remain absent, constrained by varying spectral configurations across sensors, the scarcity and inconsistency of labels, and the limited scale and scene diversity of existing datasets. To address these challenges and enable universal perception, we propose HyperVision, the first ground-based hyperspectral pre-trained backbone. First, to handle varying spectral configurations, HyperVision adopts a channel-adaptive dynamic embedding mechanism to map heterogeneous inputs into a unified token space. Second, we develop an unsupervised representation learning framework. Specifically, to address label scarcity and inconsistency, a multi-source pseudo-labeling method is introduced to fuse spatial structures from SAM2 and fine-grained spectral material information from HyperFree. Furthermore, to enrich scene diversity and compensate for limited dataset scale, a cross-modal knowledge distillation mechanism is utilized to transfer rich semantic representations from a pre-trained RGB vision model to our backbone. Pre-trained on a collection of 15k images from 26 diverse ground-based datasets, HyperVision demonstrates exceptional generalization. Requiring only efficient head-only adaptation without adjusting backbone parameters, it achieves state-of-the-art performance compared to task-specific methods across three downstream tasks under varying sensor configurations, yielding up to a 16.3% relative improvement in hyperspectral semantic segmentation $\mathrm{Acc}_{\mathrm{M}}$, a 2.1% relative gain in object tracking AUC, and a 35.5% reduction in salient object detection MAE. The source code and pre-trained model will be publicly available on https://github.com/lronkitty/HyperVision .
arXiv:2605.21739v2 Announce Type: replace
Abstract: Emotional intelligence (EI), the ability to perceive, understand, and respond appropriately to others' emotional states, is central to human communication, and increasingly important to assess as LLMs assume conversational roles in everyday life. Existing EI benchmarks rely on synthetic prompts, single-turn cases, or third-party annotation. These approaches do not directly measure how models infer and respond to a participant's emotional state over the course of a real conversation. We introduce AttuneBench, a benchmark grounded in 200 genuine multi-turn human-model conversations in which participants conversed with anonymized LLMs and provided turn-by-turn annotations of their emotional state, the model's behavior, and their preferred responses. Across 11 evaluated models, we find that model rankings on emotion recognition, behavioral classification, preference prediction, and judged response quality are largely independent, indicating that emotionally intelligent behavior decomposes into separable capabilities. Preference alignment and response-quality judgments are substantially more model-discriminating than emotion-label accuracy. These results indicate that emotionally intelligent behavior requires predicting what kind of response a specific user wants in context, a distinction that aggregate scoring can obscure and that single-turn or synthetic formats cannot directly capture across turns. AttuneBench provides a framework for assessing each of these capabilities and for diagnosing model-specific strengths and failure modes in emotionally salient conversation.
arXiv:2605.05155v3 Announce Type: replace
Abstract: As 3D Gaussian Splatting (3DGS) gains attention in immersive media and digital content creation, assessing the aesthetics of 3D scenes becomes important in helping creators build more visually compelling 3D content. However, existing evaluation methods for 3D scenes primarily emphasize reconstruction fidelity and perceptual realism, largely overlooking higher-level aesthetic attributes such as composition, harmony, and visual appeal. This limitation comes from two key challenges: (1) the absence of general 3DGS datasets with aesthetic annotations, and (2) the intrinsic nature of 3DGS as a low-level primitive representation, which makes it difficult to capture high-level aesthetic features. To address these challenges, we propose Aes3D, the first systematic framework for assessing the aesthetics of 3D neural rendering scenes. Aes3D includes Aesthetic3D, the first dataset dedicated to 3D scene aesthetic assessment, built on our proposed annotation strategy for 3D scene aesthetics. In addition, we present Aes3DGSNet, a lightweight model that directly predicts scene-level aesthetic scores from 3DGS representations. Notably, our model operates solely on 3D Gaussian primitives, eliminating the need for rendering multi-view images and thus reducing computational cost and hardware requirements. Through aesthetics-supervised learning on multi-view 3DGS scene representations, Aes3DGSNet effectively captures high-level aesthetic cues and accurately regresses aesthetic scores. Experimental results demonstrate that our approach achieves strong performance while maintaining a lightweight design, establishing a new benchmark for 3D scene aesthetic assessment. Code and datasets will be made available in a future version.
arXiv:2511.11505v3 Announce Type: replace
Abstract: Blocking communication presents a major hurdle in running MoEs efficiently in distributed settings. To address this, we present FarSkip-Collective which modifies the architecture of modern models to enable overlapping of their computation with communication. Our approach modifies the architecture to skip connections in the model and it is unclear a priori whether the modified model architecture can remain as capable, especially for large state-of-the-art models and while modifying all of the model layers. We answer this question in the affirmative and fully convert a series of state-of-the-art models varying from 16B to 109B parameters to enable overlapping of their communication while achieving accuracy that is comparable with their original open-source releases. For example, we convert Llama 4 Scout (109B) via self-distillation and achieve average accuracy within 1% of its instruction tuned release averaged over a wide range of downstream evaluations. In addition to demonstrating retained accuracy of the large modified models, we realize the benefits of FarSkip-Collective through optimized implementations that explicitly overlap communication with computation, accelerating both training and inference in existing frameworks. For inference, we demonstrate 32.6% speedup in Time To First Token when serving a converted DeepSeek-V3 architecture with expert parallelism in SGLang and achieve 97.3% communication-computation overlap during the prefill stage. During training, our approach enables 88.9% communication overlap of the all-to-all communication collectives when pre-training DeepSeek-V3 MoE layers with expert parallelism.
arXiv:2510.21151v2 Announce Type: replace
Abstract: Multimodal conversational recommendation has recently emerged as a promising paradigm for delivering personalized experiences through natural dialogue enriched by visual and contextual grounding. Yet currently available multimodal conversational recommendation datasets remain limited: existing resources either simulate conversations, omit user history or fail to collect sufficiently detailed feedback, which constrain the types of research and evaluation they support.
To address these gaps we introduce VOGUE, a dataset of 60 human human dialogues containing 2100 granularly labeled utterances in realistic fashion shopping scenarios. Each dialogue is paired with a shared visual catalogue, item metadata, user fashion profiles and post conversation ratings from both users (Seekers) and recommenders (Assistants). This design enables rigorous evaluation of conversational inference, including not only alignment between predicted and ground truth preferences but also calibration against full rating distributions and comparison with explicit and implicit user satisfaction signals.
Our analyses of VOGUE reveal distinctive dynamics of visually grounded dialogue, e.g. recommenders frequently recommend items simultaneously in feature based groups, which creates distinct conversational phases bridged by Seeker critiques and refinements. Benchmarking Multimodal Large Language Models against human Recommenders shows that while MLLMs approach human level alignment in aggregate they exhibit systematic distribution errors in reproducing human ratings and struggle to generalize preference inference beyond explicitly discussed items. These findings establish VOGUE as both a unique resource for studying multimodal conversational systems and a challenge dataset beyond the current recommendation capabilities of existing top tier multimodal foundation models such as GPT-5-mini and Gemini-2.5-Flash.
arXiv:2510.14150v5 Announce Type: replace
Abstract: We introduce CodeEvolve, an open-source framework that couples large language models with island-based evolutionary search for end-to-end algorithmic discovery. CodeEvolve integrates inspiration-based crossover, meta-prompting, and depth-based refinement on top of a CVT-MAP-Elites archive and a weighted LLM ensemble to generate optimized solutions for complex problems. On the AlphaEvolve benchmark suite, CodeEvolve matches or surpasses the reported AlphaEvolve results on 5 of 9 problems and, under matched conditions, outperforms the open-source frameworks OpenEvolve and ShinkaEvolve on 6 of 9. With the open-weight Qwen3-Coder-30B backbone, it surpasses the reported AlphaEvolve score on both CirclePackingSquare instances at roughly an order of magnitude lower cost than a frontier closed-source ensemble, and remains competitive with EoH on heuristic-design tasks without retuning. Ablations show that the interaction between CodeEvolve's components, rather than any single operator, drives these results. We release the framework, experimental data, and practical hyperparameter guidelines at https://github.com/inter-co/science-codeevolve.
arXiv:2510.11499v2 Announce Type: replace
Abstract: Generative models have emerged as a powerful class of policies for offline reinforcement learning (RL) due to their ability to capture complex, multi-modal behaviors. However, existing methods face a stark trade-off: slow, iterative models like diffusion policies are computationally expensive, while fast, single-step models like consistency policies often suffer from degraded performance. In this paper, we demonstrate that it is possible to bridge this gap. The key to moving beyond the limitations of individual methods, we argue, lies in a unifying perspective that views modern generative models, including diffusion, flow matching, and consistency models, as specific instances of learning a continuous-time generative trajectory governed by an Ordinary Differential Equation (ODE). This principled foundation provides a clearer design space for generative policies in RL and allows us to propose Generative Trajectory Policies (GTPs), a new and more general policy paradigm that learns the entire solution map of the underlying ODE. To make this paradigm practical for offline RL, we further introduce two key theoretically principled adaptations. Empirical results demonstrate that GTP achieves state-of-the-art performance on D4RL benchmarks - it significantly outperforms prior generative policies, achieving perfect scores on several notoriously hard AntMaze tasks.