arXiv:2601.19966v2 Announce Type: replace-cross
Abstract: We introduce ELECTRAFI, a fast, end-to-end differentiable model for predicting periodic charge densities in crystalline materials. ELECTRAFI constructs anisotropic Gaussians in real space and exploits their closed-form Fourier transforms to analytically evaluate plane-wave coefficients via the Poisson summation formula. This formulation delegates non-local and periodic behavior to analytic transforms, enabling reconstruction of the full periodic charge density with a single inverse FFT. By avoiding explicit real-space grid probing, periodic image summation, and spherical harmonic expansions, ELECTRAFI matches or exceeds state-of-the-art accuracy across periodic benchmarks while being up to $633 \times$ faster than the strongest competing method, reconstructing crystal charge densities in a fraction of a second. When used to initialize DFT calculations, ELECTRAFI reduces total DFT compute cost by up to ~20%, whereas slower charge density models negate savings due to high inference times. Our results show that accuracy and inference cost jointly determine end-to-end DFT speedups, and motivate our focus on efficiency.
Science Journals
arXiv:2601.20076v2 Announce Type: replace-cross
Abstract: We consider minimizing an objective function subject to constraints defined by the intersection of lower-level sets of convex functions. We study two cases: (i) strongly convex and Lipschitz-smooth objective function and (ii) convex but possibly nonsmooth objective function. To deal with the constraints that are not easy to project on, we use a randomized feasibility algorithm with Polyak steps and a random number of sampled constraints per iteration, while taking (sub)gradient steps to minimize the objective function. For case (i), we prove linear convergence in expectation of the objective function values to any prescribed tolerance using an adaptive stepsize. For case (ii), we develop a fully problem parameter-free and adaptive stepsize scheme that yields an $O(1/\sqrt{T})$ worst-case rate in expectation. The infeasibility of the iterates decreases geometrically with the number of feasibility updates almost surely, while for the averaged iterates, we establish an expected lower bound on the function values relative to the optimal value that depends on the distribution for the random number of sampled constraints. For certain choices of sample-size growth, optimal rates are achieved. Finally, simulations on a Quadratically Constrained Quadratic Programming (QCQP) problem, Support Vector Machines (SVM), and logistic regression with group fairness constraints demonstrate the computational efficiency of our algorithm compared to other state-of-the-art methods.
arXiv:2605.30039v2 Announce Type: replace
Abstract: Large Language Models have demonstrated remarkable progress in general-purpose capabilities and can achieve strong performance in specific domains through fine-tuning on domain-specific data. However, acquiring high-quality data for target domains remains a significant challenge. Existing data synthesis approaches follow a deductive paradigm, heavily relying on explicit domain descriptions expressed in natural language and careful prompt engineering, limiting their applicability in real-world scenarios where domains are difficult to describe or formally articulate. In this work, we tackle the underexplored problem of domain-specific data synthesis through an inductive paradigm, where the target domain is defined only through a set of reference examples, particularly when domain characteristics are difficult to articulate in natural language. We propose a novel framework, DOMINO, that learns a minimal sufficient domain representation from reference samples and leverages it to guide the generation of domain-aligned synthetic data. DOMINO integrates prompt tuning with a contrastive disentanglement objective to separate domain-level patterns from sample-specific noise, mitigating overfitting while preserving core domain characteristics. Theoretically, we prove that DOMINO expands the support of the synthetic data distribution, ensuring greater diversity. Empirically, on challenging coding benchmarks where domain definitions are implicit, fine-tuning on data synthesized by DOMINO improves Pass@1 accuracy by up to 4.63\% over strong, instruction-tuned backbones, demonstrating its effectiveness and robustness. This work establishes a new paradigm for domain-specific data synthesis, enabling practical and scalable domain adaptation without manual prompt design or natural language domain specifications.
arXiv:2501.02672v3 Announce Type: replace-cross
Abstract: Characterising cause-effect relationships in complex systems is fundamental to understanding their underlying mechanisms. Granger causality (GC) remains a widely used computational tool for identifying causal relationships in time series data. However, like other causal discovery methods, GC has limitations and has been criticised for lacking a rigorous causal foundation. In this work, we present a fix to this criticism by reinterpreting GC through the lenses of Reichenbach's principles and causal Bayesian networks. This reinterpretation was implemented as an algorithm we call causalized Granger causality (c-GC). We demonstrate, both theoretically and graphically, that this reformulation endows GC with a robust causal interpretation under specific assumptions. c-GC yields satisfactory results on synthetic data, offering a more principled framework for causal discovery in observational datasets.
Broken Memories: Detecting and Mitigating Memorization in Diffusion Models with Degraded Generations
arXiv:2605.22050v4 Announce Type: replace
Abstract: While diffusion models excel at generating high-quality images, their tendency to memorize training data poses significant privacy and copyright risks. In this work, we for the first time identify that memorization induces internal numerical instability, often manifesting as visually ``broken'' artifacts. Inspired by stability analysis in numerical methods, we introduce empirical stability regions based on latent update norms to quantitatively characterize stable behavior during generation. Leveraging this, we propose a principled, on-the-fly framework for step-wise detection and adaptive mitigation. Our approach suppresses memorization without altering prompts or guidance, thereby preserving semantic fidelity and image quality. Extensive experiments on Stable Diffusion 1.4 demonstrate that our method achieves an AUC $>0.999$ detection performance and a $0.0\%$ memorization rate after mitigation with negligible overhead ($\approx0.01$s per image).
arXiv:2605.26371v2 Announce Type: replace
Abstract: Hierarchical Reinforcement Learning (HRL) promises to solve long-horizon Reinforcement Learning (RL) tasks more efficiently than non-hierarchical counterparts by discovering and reusing temporally-extended skills. However, obtaining skills that are actually reusable remains an open challenge. Towards this end, we focus on abstractions that exploit the intuition of local dynamics: local transitions in different global contexts require similar kinds of action sequences. By aligning these contexts with the action sequences they require, we are able to learn which skills to reuse and where to reuse them. In principle, this information should benefit many HRL algorithms, where high-level policies have to reason about the low-level skills they use. The resulting algorithm CARL (Contrastive Action-based Representations for Reusable Local Control) shows both qualitative clustering of meaningful skills in complex humanoid environments and improved downstream performance on the OGBench benchmark when integrated with HIQL.
arXiv:2605.26711v2 Announce Type: replace
Abstract: We construct a binary mixed-regime process with one deterministic textual regime and one random regime governed by an unobserved latent state. Even an ideal infinite-capacity sequence predictor that exactly recovers the text-only marginal law can become overconfident when the observed prefix is compatible with the wrong latent regime. The resulting entropy difference is not an ordinary optimization error; it is a sufficiency gap caused by marginalization over an unobserved state. We then formalize retrieval, tool use, and external grounding through an auxiliary binary signal with fidelity $\gamma \in [1/2,1]$. The resulting Bayesian update yields a contextual dominance threshold: a corrective signal reverses the posterior odds induced by the textual history exactly when its fidelity exceeds the text-only posterior weight assigned to the misleading regime. This threshold reduces, but does not generally eliminate, the sufficiency gap; complete closure requires perfect revelation of the relevant latent state or an equivalent verification mechanism. The analysis clarifies why temperature scaling cannot restore missing context, why grounding mechanisms must be both informative and learnably usable by the model, and why autonomous sequence models require structurally decoupled observers or verifiers in high-stakes domains.
arXiv:2605.31316v1 Announce Type: new
Abstract: Multi-stakeholder platforms (MSPs) coordinate diverse stakeholder groups, often with competing or conflicting requirements. As these platforms increasingly take digital form, engineers building them make architectural decisions about data visibility, service decomposition, and algorithm design that directly determine which stakeholder requirements are prioritized when conflicts arise. Software architecture literature provides patterns for data isolation and access control among tenants but does not address how architectural decisions resolve conflicts among stakeholders with structurally divergent interests. MSP governance literature identifies the principles at stake but treats technology as neutral infrastructure. Neither addresses the translation between governance principles and architectural decision spaces. This paper proposes a governance-architecture correspondence framework that surfaces implicit governance decisions, making them explicit and debatable before deployment. The framework maps five MSP governance principles to the architectural decision spaces where they must be addressed, identifying for each the governance-aware design choice and the technically convenient default it overrides. We illustrate the framework in a constructed knowledge platform for pig farming in Rwanda, where five stakeholder types present structurally conflicting requirements. As work in progress, the framework is proposed but not yet empirically validated; a planned pre/post judgment study with platform users across all stakeholder types will test falsifiable predictions about governance outcomes.
arXiv:2605.28245v2 Announce Type: replace
Abstract: Estimating time-resolved velocity and pressure fields from Particle Image Velocimetry (PIV) remains challenging due to its limited temporal resolution in many applications. Data-driven approaches that combine snapshot PIV with high-frequency probe data have shown great promise in reconstructing the flow dynamics for advection-dominated flows; however, they typically exploit only the probe measurements directly synchronized with the PIV frames, leaving a large volume of probe-only data acquired between snapshots unused. In this work, we propose a framework that enriches the original PIV training dataset by time-marching a simple advection model and then exploits unlabelled probe data through a semi-supervised learning strategy. Two neural networks are trained to predict the temporal coefficients of Proper Orthogonal Decomposition (POD) modes of the flow fields, and their temporal derivatives, respectively. Unlabelled probe samples are leveraged to enforce temporal consistency and expand the coverage of flow scenarios beyond those captured by snapshot PIV, which is crucial for obtaining physically consistent temporal gradients required for pressure field reconstruction. A least-squares regularization step is further employed to reconcile the predictions and enforce consistency between temporal coefficients and their derivatives. The proposed approach is validated on both synthetic turbulent channel flow data and experimental PIV measurements of an airfoil wake. Results demonstrate that incorporating unlabelled probe data significantly improves the accuracy and temporal smoothness of velocity reconstruction, leading to more reliable pressure estimation via the Navier-Stokes equations, without increasing the experimental cost.
arXiv:2605.02125v3 Announce Type: replace
Abstract: Federated learning (FL) across multiple HPC facilities faces stochastic admission delays from batch schedulers that dominate wall-clock time. Synchronous FL suffers from severe stragglers, while asynchronous FL accumulates stale updates when queues spike. We propose FedQueue, a queue-aware FL protocol that incorporates scheduler delays directly into training and aggregation, which (i) predicts per-facility queue delays online to budget local work, (ii) applies cutoff-based admission that buffers late arrivals to bound staleness, and (iii) performs staleness-aware aggregation to stabilize heterogeneous local workloads. We prove the convergence for non-convex objectives at rate $\mathcal{O}(1/\sqrt{R})$ under bounded staleness, and show that the admission controls yield bounded staleness with high probability under queue-prediction error. Real-world cross-facility deployment of FedQueue shows 20.5% improvement over baseline algorithms. Controlled queue simulations demonstrate robust improvement over the baselines; in particular, up to 60% reduction in time to reach a target accuracy level under high queue variance and non-IID partitions.
arXiv:2605.30248v2 Announce Type: replace
Abstract: Image generation models have evolved from text-conditioned pixel synthesis toward multimodal agents endowed with visual comprehension and tool invocation capabilities. Yet, existing agents remain at the mercy of underlying black-box image models. Their workflow is trapped in a repetitive cycle of prompt rewriting for generation refinement, leaving them with no mechanism to directly manipulate the canvas. In essence, the potential of LLMs to serve as a genuine "brush" for precise visual construction remains largely untapped. In this paper, we propose GenClaw, a code-driven agentic image generation paradigm that empowers the agent to create like a human artist: first conceptualizing, then sketching, and finally coloring. Specifically, the agent first constructs the conceptual knowledge and context through search and reasoning. It then utilizes code (e.g., SVG, HTML, ThreeJS) to render executable visual sketches. Finally, it employs an image generation model to supplement textures, materials, and photorealism. In this workflow, code serves as a controllable intermediate canvas bridging linguistic reasoning and pixel synthesis, seamlessly integrating programmatic logic with the visual expressiveness of generative models. By transforming image generation from a black-box paradigm into a staged process akin to authentic human creation, GenClaw offers a step toward for highly controllable and interpretable visual generation systems.
arXiv:2510.10988v4 Announce Type: replace-cross
Abstract: Learning-to-Defer (L2D) enables hybrid decision-making by routing inputs either to a predictor or to external experts. While promising, L2D is highly vulnerable to adversarial perturbations, which can not only flip predictions but also manipulate deferral decisions. Prior robustness analyses focus solely on two-stage settings, leaving open the end-to-end (one-stage) case where predictor and allocation are trained jointly. We introduce the first framework for adversarial robustness in one-stage L2D, covering both classification and regression. Our approach formalizes attacks, proposes cost-sensitive adversarial surrogate losses, and establishes theoretical guarantees including $\mathcal{H}$, $(\mathcal{R }, \mathcal{F})$, and Bayes consistency. Experiments on benchmark datasets confirm that our methods improve robustness against untargeted and targeted attacks while preserving clean performance.
arXiv:2509.06856v2 Announce Type: replace-cross
Abstract: We propose a novel randomized framework for the estimation problem of large-scale linear statistical models, namely Sequential Least-Squares Estimators with Fast Randomized Sketching (SLSE-FRS), which integrates Sketch-and-Solve and Iterative-Sketching methods for the first time. By iteratively constructing and solving sketched least-squares (LS) subproblems with increasing sketch sizes to achieve better precisions, SLSE-FRS gradually refines the estimators of the true parameter vector, ultimately producing high-precision estimators. We analyze the convergence properties of SLSE-FRS, and provide its efficient implementation. Numerical experiments show that SLSE-FRS outperforms the state-of-the-art methods, namely the Preconditioned Conjugate Gradient (PCG) method, and the Iterative Double Sketching (IDS) method.
arXiv:2605.06137v2 Announce Type: replace
Abstract: In this work, we propose Prologue, an approach to bridging the reconstruction-generation gap in autoregressive (AR) image generation. Instead of modifying visual tokens to satisfy both reconstruction and generation, Prologue generates a small set of prologue tokens prepended to the visual token sequence. These prologue tokens are trained exclusively with the AR cross-entropy (CE) loss, while visual tokens remain dedicated to reconstruction. This decoupled design lets us optimize generation through the AR model's true distribution without affecting reconstruction quality, which we further formalize from an ELBO perspective. On ImageNet 256x256, Prologue-Base reduces gFID from 21.01 to 10.75 without classifier-free guidance while keeping reconstruction almost unchanged; Prologue-Large reaches a competitive rFID of 0.99 and gFID of 1.46 using a standard AR model without auxiliary semantic supervision. Interestingly, driven only by AR gradients, prologue tokens exhibit emergent semantic structure: linear probing on 16 prologue tokens reaches 35.88% Top-1, far above the 23.71% of the first 16 tokens from a standard tokenizer; resampling with fixed prologue tokens preserves a similar high-level semantic layout. Our results suggest a new direction: generation quality can be improved by introducing a separate learned generative representation while leaving the original representation intact.
arXiv:2605.31321v1 Announce Type: new
Abstract: Diffusion-based imitation learning methods have driven rapid progress in robot dexterous manipulation tasks. However, they have limitations when applied to tasks that involve complex free-form surface constraints because of their lack of explicit surface geometry constraint modeling and the dynamic feasibility issue, resulting in stochastic action generation that fails to achieve reliable surface alignment and maintain stable contact. To address these limitations, we propose a novel surface constraint policy (SCP) for generating robot actions that satisfy free-form surface constraints on the basis of human demonstrations and real-time visual observations. First, the surface geometry constraint is encoded using a two-dimensional weighted Gaussian kernel function that is derived from demonstrations. Building on the encoded surface geometry constraints, the diffusion-based policy is used to infer task-level action intentions from multimodal sensory inputs, including visual observations and robot state feedback. These intentions are further transformed into surface-constrained dynamic movement primitives (DMPs) through a similarity-based action mapping method, thereby enabling smooth and compliant motion execution. The SCP achieves generation of structured surface geometric intent and dynamically admissible actions. The proposed method is validated on multiple surface manipulation tasks and compared with existing techniques. The experimental results demonstrate superior task success rates and contact stability under surface constraints.
arXiv:2605.31410v1 Announce Type: new
Abstract: Food-as-Medicine requires models to reason beyond what a dish is or what nutrition it contains: they must decide whether a concrete food choice is appropriate for a specific health condition. Existing food AI benchmarks primarily evaluate dish recognition, recipe understanding, nutrient estimation, or general nutrition question answering, leaving this health-aware decision layer largely untested. We introduce FAM-Bench, a multi-modal Food-as-Medicine benchmark with 2500 nutrition-expert-verified instances across 13 diet-related health conditions. The benchmark contains two complementary tasks: dish-level suitability assessment, where models judge whether a dish is suitable for a condition from its image and ingredient list, and comparative dish analysis, where models rank four candidate dishes by condition-specific suitability. Both tasks require integrating ingredient evidence, visual preparation cues, and clinical nutrition constraints, providing a standardized testbed for grounded health-aware reasoning in language and vision-language models.
Beyond Instance-Level Alignment and Uniformity: Semantic Factor Learning for Collaborative Filtering
arXiv:2605.31414v1 Announce Type: new
Abstract: Collaborative filtering (CF) is widely used in recommender systems (RecSys) due to its simplicity and efficiency. However, existing CF methods follow an instance-level learning paradigm. During the instance learning stage, a large number of uninteracted user-item instances, of which items are potential interested by the user, are incorrectly treated as true negative samples resulting in a severe limitation to the generalization and scalability of models. Moreover, mainstream graph convolutional networks (GCNs) inherently suffer from high computational cost and over-smoothing issues, which limit the ability in capturing higher-order connectivity and lead to a poor generalization under sparse supervision signals. To address the above limitations, we propose Semantic Factor enhanced Alignment and Uniformity (SaFeAU), a novel framework that augments interacted instances with semantic factors, thereby mitigating false negative labeling and enabling matrix factorization (MF) to capture high-order CF signals without graph neighborhood aggregation. Specifically, SaFeAU consists of three tightly coupled components. First, Semantic Factor Routing (SFR) disentangles item representations into independent and global semantic factors. Building on these factors, Semantic Factor Matching (SFM) identifies uninteracted items, which share the same semantic factors with interacted ones, as potential positive pairs for enriching sparse supervision signals. Finally, Semantic Pairs Alignment (SPA) aligns both observed and potential positive pairs while promoting uniformity of user and item representations. Extensive experiments on four sparse real-world datasets show that SaFeAU consistently outperforms GCN-based and MF-based state-of-the-art CF methods in both recommendation accuracy and computational efficiency, confirming the effectiveness of the proposed semantic enhanced learning paradigm.
arXiv:2507.04545v2 Announce Type: replace-cross
Abstract: Network effects -- the utility gains from additional consumers of a good -- are widely regarded as critical to the digital economy. Yet recent theory and evidence suggest that local network effects -- the economic value created by specific social network connections -- drive value in networked online platforms. Using incentive-compatible online choice experiments with 19,923 Facebook, Instagram, LinkedIn, and X users in the United States, we provide the first large-scale empirical measurement of local network effects in the digital economy and measure heterogeneity in connection value across platforms. Platform value ranges from \$78 to \$101 per consumer per month, with 8.1-23.7% explained by local network effects. We find that 1) stronger ties are more valuable on Facebook and Instagram, while weaker ties are more valuable on LinkedIn and X; 2) work connections are most valuable on LinkedIn and least on Facebook, and job-seekers value LinkedIn significantly more and Facebook significantly less; 3) men value connections to women significantly more than to other men, particularly on Instagram, Facebook, and X, while women value connections to men and women equally across platforms; 4) consumers value connections on any platform more if they are also connected on other platforms, suggesting that platforms are complements, not substitutes; 5) white consumers disproportionately value same-race connections on Facebook while, on Instagram, connections to alters eighteen or younger are valued significantly more than any other age group -- two patterns not seen on other platforms. Each platform generates between \$53B and \$215B in annual US consumer surplus. These results suggest that social media generates significant value, that local network effects drive a substantial fraction of it, and that the sources and contours of these effects vary across platforms, consumers, and connections.
Position-Blind Ptychography: Viability of image reconstruction via data-driven variational inference
arXiv:2509.25269v3 Announce Type: replace-cross
Abstract: In this work, we present and investigate the novel blind inverse problem of position-blind ptychography, i.e., ptychographic phase retrieval without any knowledge of scan positions, which then must be recovered jointly with the image. The motivation for this problem comes from single-particle diffractive X-ray imaging, where particles in random orientations are illuminated and a set of diffraction patterns is collected. If one uses a highly focused X-ray beam, the measurements would also become sensitive to the beam positions relative to each particle and therefore ptychographic, but these positions are also unknown. We investigate the viability of image reconstruction in a simulated, simplified 2-D variant of this difficult problem, using variational inference with modern data-driven image priors in the form of score-based diffusion models. We find that, with the right illumination structure and a strong prior, one can achieve reliable and successful image reconstructions even under measurement noise, in all except the most difficult evaluated imaging scenario.
arXiv:2510.02168v2 Announce Type: replace-cross
Abstract: A novel anomaly detection algorithm is presented. The Wasserstein normalized autoencoder (WNAE) is a normalized probabilistic model that minimizes the Wasserstein distance between the learned probability distribution--a Boltzmann distribution where the energy is the reconstruction error of the autoencoder--and the distribution of the training data. This algorithm has been developed and applied to the identification of semivisible jets--conical sprays of visible standard model particles and invisible dark matter states--with the CMS experiment at the CERN LHC. Trained on jets of particles from simulated standard model processes, the WNAE is shown to learn the probability distribution of the input data in a fully unsupervised fashion, such that it effectively identifies new physics jets as anomalies. The model exhibits stable, convergent training and recovers strong classification performance for a wide range of signals against the selected background process, for which a standard autoencoder fails because of outlier reconstruction. In addition, the model improves upon standard normalized autoencoders while remaining fully agnostic to the signal. The WNAE directly tackles the problem of outlier reconstruction, a common failure mode of autoencoders in anomaly detection tasks.
arXiv:2605.31417v1 Announce Type: new
Abstract: We revisit the Binary Closest String problem, which asks, given a set of binary strings $X \subseteq \{0, 1\}^n$, to compute a string minimizing the maximum Hamming distance to $X$. A long line of work has focused on parameterized algorithms with respect to the optimal distance $d$, yielding a sequence of improvements from $O^*(d^d)$ through $O^*(16^d)$, $O^*(9.513^d)$, $O^*(8^d)$, $O^*(6.731^d)$ to the current best-known running time of $O^*(5^d)$ [Chen, Ma, Wang; Algorithmica '16].
We present a faster randomized algorithm running in time $O^*(4^d)$. Our result matches a recent fine-grained lower bound [Abboud, Fischer, Goldenberg, Karthik C.S., Safier; ESA '23], and is therefore conditionally optimal. As an extra benefit, our algorithm is remarkably simple.
arXiv:2605.31545v1 Announce Type: new
Abstract: As Large Language Models (LLMs) evolve from general-purpose assistants to user-centric agents, personalization has become central to aligning model behavior with individual preferences, making the evaluation of personalized alignment a critical bottleneck. Existing evaluation methods-ranging from automatic metrics to LLM-as-a-judge approaches-fail to capture subjective, user-specific preferences embedded in long-term interaction histories. We identify three essential principles for reliable and effective personalized evaluation: Representativeness, User-Consistency, and Discriminativeness. To address these principles, we introduce Personalized Evaluation as Learning, a paradigm that formulates personalized evaluation as a learning problem rather than a static judgment. Under this paradigm, we propose PARL (Preference-Aware Rubric Learning for Personalized Evaluation), a framework that learns to induce preference-aware evaluation rubrics directly from raw user histories and performs a self-validation mechanism to ensure consistency with the user's preferences. PARL integrates rubric induction with a discriminative reinforcement learning objective that contrasts user-authored responses against competitive personalized model outputs, enabling the learned rubrics to capture precise, user-specific decision boundaries. Experiments on real-world personalized text generation tasks show that PARL consistently induces high-fidelity rubrics that reliably identify user-aligned responses and generalize across users and tasks, while capturing stable stylistic preferences and fine-grained evaluative patterns. To ensure reproducibility, our code is available at https://github.com/SnowCharmQ/PARL.
arXiv:2605.25773v2 Announce Type: replace-cross
Abstract: Efficient benchmarking techniques aim to lower the computational cost of evaluating LLMs by predicting full benchmark scores using only a subset of a benchmark's questions. By reframing this problem as an instance of multiple regression with feature selection, we find that existing efficient benchmarking methods can be greatly improved by simply using kernel ridge regression at the prediction stage. Additionally, using an information-theoretic feature-selection algorithm called minimum redundancy maximum relevance (mRMR), we can further improve upon these methods by selecting question subsets that will be maximally useful for prediction. Except in very data-poor settings, these approaches consistently achieve smaller prediction errors (in both MAE and RMSE), and greater ranking correlation between predicted and true scores (in both Spearman $\rho$ and Kendall $\tau$) across a range of benchmarks using both binary and continuous metrics. Furthermore, mRMR subsampling is much faster than competitor methods (which often involve fitting probabilistic models or running clustering algorithms), and is more likely to select the same questions under different random seeds or training data splits. Tutorial code can be found at https://github.com/sambowyer/mrmr_eval .
SEMA-RAG: A Self-Evolving Multi-Agent Retrieval-Augmented Generation Framework for Medical Reasoning
arXiv:2605.17101v2 Announce Type: replace
Abstract: Retrieval-Augmented Generation (RAG) is widely employed to mitigate risks such as hallucinations and knowledge obsolescence in medical question answering, yet its predominantly single-round, static retrieval paradigm misaligns with the multi-stage process of clinical reasoning. This compressed workflow induces two structural deficiencies: question-to-query translation often lacks clinically grounded semantic interpretation, and retrieval lacks iterative sufficiency feedback, making it difficult to form reliable evidence chains. We argue that both issues stem from a deeper cause: overloading a single reasoning chain with heterogeneous tasks of interpretation, exploration, and adjudication. The remedy is to reconstruct the workflow via task decoupling and dynamic multi-round exploration. To this end, we propose SEMA-RAG, a Self-Evolving Multi-Agent RAG framework for medical question answering, which assigns these roles to three specialist agents: the Interpreter Agent for clinical schema interpretation, the Explorer Agent for sufficiency-driven self-evolving retrieval, and the Arbiter Agent for evidence adjudication and answer selection. Across five benchmarks and five LLM backbones, SEMA-RAG improves the strongest baseline by +6.46 accuracy points on average, measured per backbone.
arXiv:2605.31126v1 Announce Type: new
Abstract: Can a language model improve from plain text sampled from itself, with no prompts, no teacher, no verifier, and no reward model? Yes, but only when the synthetic corpus is compatible with the student, a relational property of the source-student pair rather than an intrinsic property of the data. We call this the latent capability resurfacing hypothesis: weak self-training can amplify capabilities already present in the pretrained model, but only under this compatibility condition. We study this in the minimal setting of prompt-free unconditional self-training, where base language models are fine-tuned on text generated from the BOS token alone, with no task specification or external supervision. We report three findings. First, synthetic utility is relational rather than intrinsic: self-generated data is the most effective source, same-lineage transfer outperforms stronger but differently trained sources, and cross-family transfer is substantially weaker. Second, common intrinsic proxies fail: neither benchmark-level semantic similarity nor average per-token likelihood under the student predicts which corpora help. Third, this regime produces a surprising byproduct. In controlled Pythia experiments, capability and verbatim memorization decouple: benchmark utility is preserved or improved while held-out exact-match extraction drops by over 95 percent, with no forget set, privacy objective, or targeted unlearning. Together, these results suggest that prompt-free self-training works by amplifying what the student already knows, not by importing structure from the data. They also reveal a regime in which capability and verbatim memorization can be separated without any explicit unlearning objective.