arXiv:2603.07211v2 Announce Type: replace
Abstract: Direct Preference Optimization (DPO) has become a standard framework for safety alignment, but its reliance on pairwise preference updates makes training sensitive to imperfect supervision. Existing robust DPO methods often address this sensitivity through global loss corrections or external data-level interventions, while largely overlooking how unreliable comparisons distort batch-level optimization dynamics. We propose CompassDPO, a reward-free DPO framework that stabilizes preference optimization through dynamics control. Using the implicit DPO reward margin as a training-time compass, CompassDPO regulates sample influence along two complementary axes: update direction and update magnitude. For directional control, it applies sparse, budgeted, and warm-up delayed loss mixing to attenuate update components that conflict with the emerging preference direction. For magnitude control, it adaptively soft-winsorizes high-loss tail contributions, reducing tail dominance while preserving useful gradients from hard examples. Both mechanisms use only signals available during standard DPO training and require no external reward model or additional supervision. Experiments on PKU-SafeRLHF across four backbones and multiple out-of-distribution safety benchmarks show that CompassDPO consistently improves robustness over vanilla DPO and strong DPO-family baselines, especially under controlled label-flip noise. Code is available at https://anonymous.4open.science/r/CompassDPO-4D00
Science Journals
arXiv:2603.08413v2 Announce Type: replace
Abstract: Deep neural networks for image classification often exhibit overconfidence on out-of-distribution (OOD) samples. To address this, we introduce Geometrically Constrained Outlier Synthesis (GCOS), a training-time regularization framework aimed at improving OOD robustness during inference. GCOS addresses a limitation of prior synthesis methods by generating virtual outliers in the hidden feature space that respect the learned manifold structure of in-distribution (ID) data. The synthesis proceeds in two stages: (i) a dominant-variance subspace extracted from the training features identifies geometrically informed, off-manifold directions; (ii) a conformally-inspired shell, defined by the empirical quantiles of a nonconformity score from a calibration set, adaptively controls the synthesis magnitude to produce boundary samples. The shell ensures that generated outliers are neither trivially detectable nor indistinguishable from in-distribution data, facilitating smoother learning of robust features. This is combined with a contrastive regularization objective that promotes separability of ID and OOD samples in a chosen score space, such as Mahalanobis or energy-based. Experiments demonstrate that GCOS outperforms state-of-the-art methods using standard energy-based inference on near-OOD benchmarks, defined as tasks where outliers share the same semantic domain as in-distribution data. As an exploratory extension, the framework naturally transitions to conformal OOD inference, which translates uncertainty scores into statistically valid p-values and enables thresholds with formal error guarantees, providing a pathway toward more predictable and reliable OOD detection.
arXiv:2603.08671v2 Announce Type: replace
Abstract: We present the first simultaneous, spatially-resolved measurement of all three velocity components (radial, azimuthal, and axial) in an annular magnetized argon gas-puff Z-pinch, performed using Collective Thomson Scattering along three orthogonal lines of sight from the same scattering volume at each time step. Measurements were carried out on the Llamp{\"u}dke{\~n} pulse-power generator (400 kA peak, 200 ns rise time), for applied axial fields ranging from 0.04 to 0.26 T using two coil configurations: a double coil with negligible initial radial field at the probed plane ($z = 8$ mm), and a single coil that imposes a finite initial radial field. Three principal results are reported. First, the axial velocity component, which had not previously been measured experimentally in this configuration, reaches 60--70~km\,s$^{-1}$ near the axis at low applied fields ($B_{z0} < 0.1$ T) and is suppressed to within $\pm 20$~km\,s$^{-1}$ for stronger applied fields, in correlation with the reduction of the zippering angle, with direct implications for the implosion energy balance. Second, the self-generated rotation extends across the full plasma diameter, not only at the periphery, and the diametrical profile of the azimuthal velocity decreases toward the axis with an exponential-like shape consistent with the underlying current density distribution; this feature was not visible in previous edge-localized measurements. Third, rotation persists in the double-coil case ($B_{r0} \approx 0$) and is enhanced in the single-coil case (finite $B_{r0}$), supporting the interpretation that $B_r$ develops self-consistently during the implosion and drives the rotation through a $J_z \times B_r$ torque. These results constrain the role of each magnetic-field component and motivate direct measurement of $B_r$ and the current density distribution as the next step.
arXiv:2603.09551v2 Announce Type: replace
Abstract: While Vision-Language Models (VLMs) have significantly advanced remote sensing interpretation, enabling them to perform complex, step-by-step reasoning remains highly challenging. Recent efforts to introduce Chain-of-Thought (CoT) reasoning to this domain have shown promise, yet ensuring the visual faithfulness of these intermediate steps remains a critical bottleneck. To address this, we introduce GeoSolver, a novel framework that transitions remote sensing reasoning toward verifiable, process-supervised reinforcement learning. We first construct Geo-PRM-2M, a large-scale, token-level process supervision dataset synthesized via entropy-guided Monte Carlo Tree Search (MCTS) and targeted visual hallucination injection. Building upon this dataset, we train GeoPRM, a token-level process reward model (PRM) that provides granular faithfulness feedback. To effectively leverage these verification signals, we propose Process-Aware Tree-GRPO, a reinforcement learning algorithm that integrates tree-structured exploration with a faithfulness-weighted reward mechanism to precisely assign credit to intermediate steps. Extensive experiments demonstrate that our resulting model, GeoSolver-9B, achieves state-of-the-art performance across diverse remote sensing benchmarks. Crucially, GeoPRM unlocks robust Test-Time Scaling (TTS). Serving as a universal geospatial verifier, it seamlessly scales the performance of GeoSolver-9B and directly enhances general-purpose VLMs, highlighting its remarkable cross-model generalization.
arXiv:2604.10102v2 Announce Type: replace
Abstract: AI-generated image detectors suffer significant performance degradation under real-world image corruptions such as JPEG compression, Gaussian blur, and resolution downsampling. We observe that state-of-the-art methods, including B-Free, treat degradation robustness as a byproduct of data augmentation rather than an explicit training objective. In this work, we propose Degradation-Consistent Paired Training (DCPT), a simple yet effective training strategy that explicitly enforces robustness through paired consistency constraints. For each training image, we construct a clean view and a degraded view, then impose two constraints: a feature consistency loss that minimizes the cosine distance between clean and degraded representations, and a prediction consistency loss based on symmetric KL divergence that aligns output distributions across views. DCPT adds zero additional parameters and zero inference overhead. Experiments on the Synthbuster benchmark (9 generators, 8 degradation conditions) demonstrate that DCPT improves the degraded-condition average accuracy by 9.1 percentage points compared to an identical baseline without paired training, while sacrificing only 0.9% clean accuracy. The improvement is most pronounced under JPEG compression (+15.7% to +17.9%). Ablation further reveals that adding architectural components leads to overfitting on limited training data, confirming that training objective improvement is more effective than architectural augmentation for degradation robustness.
arXiv:2604.08999v2 Announce Type: replace
Abstract: Table serialization remains a critical bottleneck for Large Language Models (LLMs) in complex table question answering, hindered by challenges such as structural neglect, representation gaps, and reasoning opacity. Existing serialization methods fail to capture explicit hierarchies and lack schema flexibility, while current tree-based approaches suffer from limited semantic adaptability. To address these limitations, we propose ASTRA (Adaptive Semantic Tree Reasoning Architecture) including two main modules, AdaSTR and DuTR. First, we introduce AdaSTR, which leverages the global semantic awareness of LLMs to reconstruct tables into Logical Semantic Trees. This serialization explicitly models hierarchical dependencies and employs an adaptive mechanism to optimize construction strategies based on table scale. Second, building on this structure, we present DuTR, a dual-mode reasoning framework that integrates tree-search-based textual navigation for linguistic alignment and symbolic code execution for precise verification. Experiments on complex table benchmarks demonstrate that our method achieves state-of-the-art (SOTA) performance.
arXiv:2603.13282v2 Announce Type: replace
Abstract: Federated Learning (FL) with Low-Rank Adaptation (LoRA) has become a standard for privacy-preserving LLM fine-tuning. However, existing personalized methods predominantly operated under a restrictive Flat-Model Assumption: they addressed client-side \textit{statistical heterogeneity} but treated the model as a monolithic block, ignoring the \textit{functional heterogeneity} across LLM layers. We argue that these two statistical (horizontal) and functional (vertical) dimensions, are \textit{orthogonal in source yet coupled in interaction}, implying that the optimal depth of parameter sharing is functionally dependent on client similarity. To address this, we propose \textbf{FedTreeLoRA}, a framework employing tree-structured aggregation for fine-grained, layer-wise alignment. By dynamically constructing an aggregation hierarchy, FedTreeLoRA allows clients to share broad consensus on shallow `trunks' while progressively specializing on deep `branches'. Experiments on NLU and NLG benchmarks demonstrate that FedTreeLoRA significantly outperforms state-of-the-art methods by effectively reconciling generalization and personalization.
arXiv:2602.17335v3 Announce Type: replace
Abstract: Parquet is the de facto columnar file format in modern analytical systems, yet its configuration guidelines have largely been shaped by CPU-centric execution models. As GPU-accelerated data processing becomes increasingly prevalent, Parquet files generated with CPU-oriented defaults can severely underutilize GPU parallelism, turning GPU scans into a performance bottleneck. In this work, we systematically study how Parquet configurations affect GPU scan performance. We show that Parquet's poor GPU performance is not inherent to the format itself but rather a consequence of suboptimal configuration choices. By applying GPU-aware configurations, we increase effective read bandwidth up to 125 GB/s without modifying the Parquet specification.
arXiv:2602.09878v2 Announce Type: replace
Abstract: World-model-based imagine-then-act becomes a promising paradigm for robotic manipulation, yet existing approaches typically support either purely image-based forecasting or reasoning over partial 3D geometry, limiting their ability to predict complete 4D scene dynamics. This work proposes a novel embodied 4D world model that enables geometrically consistent, arbitrary-view RGBD generation: given only a single-view RGBD observation as input, the model imagines the remaining viewpoints, which can then be back-projected and fused to assemble a more complete 3D structure across time. To efficiently learn the multi-view, cross-modality generation, we explicitly design cross-view and cross-modality feature fusion that jointly encourage consistency between RGB and depth and enforce geometric alignment across views. Beyond prediction, converting generated futures into actions is often handled by inverse dynamics, which is ill-posed because multiple actions can explain the same transition. We address this with a test-time action optimization strategy that backpropagates through the generative model to infer a trajectory-level latent best matching the predicted future, and a residual inverse dynamics model that turns this trajectory prior into accurate executable actions. Experiments on three datasets demonstrate strong performance on both 4D scene generation and downstream manipulation, and ablations provide practical insights into the key design choices.
arXiv:2603.16870v2 Announce Type: replace
Abstract: Recent advances in video generation have revealed an unexpected phenomenon: diffusion-based video models exhibit non-trivial reasoning capabilities. Prior work attributes this to a Chain-of-Frames (CoF) mechanism, where reasoning is assumed to unfold sequentially across video frames. In this work, we challenge this assumption and uncover a fundamentally different mechanism. We show that reasoning in video models instead primarily emerges along the diffusion denoising steps. Through qualitative analysis and targeted probing experiments, we find that models explore multiple candidate solutions in early denoising steps and progressively converge to a final answer, a process we term Chain-of-Steps (CoS). Beyond this core mechanism, we identify several emergent reasoning behaviors critical to model performance: (1) working memory, enabling persistent reference; (2) self-correction and enhancement, allowing recovery from incorrect intermediate solutions; and (3) perception before action, where early steps establish semantic grounding and later steps perform structured manipulation. During a diffusion step, we further uncover self-evolved functional specialization within Diffusion Transformers, where early layers encode dense perceptual structure, middle layers execute reasoning, and later layers consolidate latent representations. Motivated by these insights, we present a simple training-free strategy as a proof-of-concept, demonstrating how reasoning can be improved by ensembling latent trajectories from identical models with different random seeds. Overall, our work provides a systematic understanding of how reasoning emerges in video generation models, offering a foundation to guide future research in better exploiting the inherent reasoning dynamics of video models as a new substrate for intelligence.
arXiv:2603.23985v3 Announce Type: replace
Abstract: Large language models (LLMs) have demonstrated remarkable capabilities, but their massive scale poses significant challenges for practical deployment. Structured pruning offers a promising solution by removing entire dimensions or layers, yet existing methods face critical trade-offs: task-agnostic approaches cannot adapt to task-specific requirements, while task-aware methods require costly training to learn task adaptability. We propose DIET (Dimension-wise global pruning of LLMs via merging Task-wise importance scores), a training-free structured pruning method that combines dimension-level granularity with task-aware selection. DIET profiles activation magnitudes across tasks using only 100 samples per task, then applies majority voting to construct a single global mask. DIET does not require large costs from pre-computation or training. Experiments on seven zero-shot benchmarks using Gemma-2 2B and 9B models demonstrate the effectiveness of DIET; for example, at 20% sparsity on Gemma-2 2B, DIET achieves near 10% average accuracy improvement, compared to previous state-of-the-art structured pruning methods. This advantage persists across various sparsity levels and model scales, positioning DIET as a practical and robust choice for structured LLM pruning.
arXiv:2604.13502v2 Announce Type: replace
Abstract: Social Determinants of Health (SDOH) refer to environmental, behavioral, and social conditions that influence how individuals live, work, and age. SDOH have a significant impact on personal health outcomes, and their systematic identification and management can yield substantial improvements in patient care. However, SDOH information is predominantly captured in unstructured clinical notes within electronic health records, which limits its direct use as machine-readable entities. To address this issue, researchers have employed Natural Language Processing (NLP) techniques using pre-trained BERT-based models, demonstrating promising performance but requiring sophisticated implementation and extensive computational resources. In this study, we investigated prompt engineering strategies for extracting structured SDOH events utilizing LLMs with advanced reasoning capabilities. Our method consisted of four modules: 1) developing concise and descriptive prompts integrated with established guidelines, 2) applying few-shot learning with carefully curated examples, 3) using a self-consistency mechanism to ensure robust outputs, and 4) post-processing for quality control. Our approach achieved a micro-F1 score of 0.866, demonstrating competitive performance compared to the leading models. The results demonstrated that LLMs with reasoning capabilities are effective solutions for SDOH event extraction, offering both implementation simplicity and strong performance.
arXiv:2604.14684v2 Announce Type: replace
Abstract: Visual prompted object detection enables interactive and flexible definition of target categories, thereby facilitating open-vocabulary detection. Since visual prompts are derived directly from image features, they often outperform text prompts in recognizing rare categories. Nevertheless, research on visual prompted detection has been largely overlooked, and it is typically treated as a byproduct of training text prompted detectors, which hinders its development. To fully unlock the potential of visual-prompted detection, we investigate the reasons why its performance is suboptimal and reveal that the underlying issue lies in the absence of global discriminability in visual prompts. Motivated by these observations, we propose DETR-ViP, a robust object detection framework that yields class-distinguishable visual prompts. On top of basic image-text contrastive learning, DETR-ViP incorporates global prompt integration and visual-textual prompt relation distillation to learn more discriminative prompt representations. In addition, DETR-ViP employs a selective fusion strategy that ensures stable and robust detection. Extensive experiments on COCO, LVIS, ODinW, and Roboflow100 demonstrate that DETR-ViP achieves substantially higher performance in visual prompt detection compared to other state-of-the-art counterparts. A series of ablation studies and analyses further validate the effectiveness of the proposed improvements and shed light on the underlying reasons for the enhanced detection capability of visual prompts.
arXiv:2503.21002v5 Announce Type: replace-cross
Abstract: We determine the covert capacity for entanglement generation over a noisy quantum channel. While secrecy guarantees that the transmitted information remains inaccessible to an adversary, covert communication ensures that the transmission itself remains undetectable. The entanglement dimension follows a square root law (SRL) in the covert setting, i.e., $O(\sqrt{n})$ Einstein-Podolsky-Rosen (EPR) pairs can be distributed covertly and reliably over $n$ channel uses. We begin with covert communication of classical information under a secrecy constraint. We then leverage this result to construct a coding scheme for covert entanglement generation. Single-letter expressions are derived for the covert key-assisted and unassisted secrecy capacities, as well as for the covert entanglement-generation capacity.
arXiv:2604.26102v2 Announce Type: replace
Abstract: Large language model agents have made strong progress on software engineering, yet current systems suffer from a context coupling problem: the standard code editing interface conflates code inspection, modification planning, and edit execution within a single context window, forcing agents to interleave exploratory viewing with strictly formatted edit generation. Irrelevant context accumulates and edit reliability degrades. We propose SWE-Edit, which decomposes the editing interface into two specialized subagents: a Viewer that extracts task-relevant code on demand, and an Editor that executes modifications from high-level natural language plans -- letting the main agent focus on reasoning while delegating context-intensive operations to clean context windows. On SWE-Bench Verified, this decomposition raises resolve rate by 2.1 pp and cuts inference cost by 17.9%, with consistent gains across multiple reasoning-model families (Kimi-K2, MiniMax-M2.1, GLM-4.7). We further show that effective edit-format selection can be trained into a small model rather than requiring frontier-scale capacity: GRPO training on Qwen3-8B with an adaptive find-replace/whole-file-rewrite policy improves edit success by 12.5 pp and brings an 8B open-source editor to parity with GPT-5-nano on downstream SWE-Bench resolve rate. To enable rapid editor iteration, we release PR-Edit, a lightweight evaluation whose scores correlate strongly with SWE-Bench resolve rate. We release our code at https://github.com/microsoft/SWE-Edit.
arXiv:2602.08586v3 Announce Type: replace
Abstract: Multi-agent LLM systems consistently outperform single-agent baselines, yet practitioners still cannot predict which design works for a new task or diagnose why one fails. We argue this gap persists largely because the field lacks a diagnostic framework with measurable primitives and testable predictions. We introduce \textbf{DIANOIA}, a three-channel decomposition of multi-agent reasoning gain into coverage, fidelity, and synthesis, each of which is empirically measurable. From this decomposition, we derive a diagnostic protocol that identifies the bottleneck channels for any given task. We instantiate the protocol as a multi-agent system whose three components mirror the channels: role-diverse proposers for coverage, execution-grounded verification for fidelity, and iterative synthesis. On GSM8K, AIME-2025, MBPP, and BFCL-SP, our method outperforms strong multi-agent baselines under matched token budgets, dominating the Pareto frontier on MBPP at $\sim$$5{\times}$ token savings and reaching $+4.6$pp at matched cost. On every benchmark, the protocol picks the right bottleneck channels; the system we built around it leads across models. We release code, adapters, diagnostic metrics, and a Claude Code skill at https://anonymous.4open.science/r/DIANOIA4MAS. DIANOIA reframes multi-agent design as channel-aware resource allocation: diagnose which channel is the bottleneck for your task, then invest tokens accordingly.
arXiv:2602.07120v2 Announce Type: replace
Abstract: Language models (LMs) tend to memorize portions of their training data and emit verbatim spans. When the underlying sources are sensitive or copyright-protected, such reproduction raises issues of consent and compensation for creators and compliance risks for developers. We propose Anchored Decoding, a plug-and-play inference-time method for suppressing verbatim copying: it enables decoding from any risky LM trained on mixed-license data by keeping generation in bounded proximity to a permissively trained safe LM. Anchored Decoding adaptively allocates a user-chosen information budget over the generation trajectory and enforces per-step constraints that yield a sequence-level guarantee, enabling a tunable risk-utility trade-off. To make Anchored Decoding practically useful, we introduce a new permissively trained safe model (TinyComma 1.8B), as well as Anchored$_{\mathrm{Byte}}$ Decoding, a byte-level variant of our method that enables cross-vocabulary fusion via the ByteSampler framework (Hayase et al., 2025). Across six model pairs on long-form metrics for copying risk and utility, Anchored and Anchored$_{\mathrm{Byte}}$ Decoding define a new Pareto frontier, preserving near-original fluency and factuality while closing up to 75% of the measurable copying gap between the risky baseline and a safe reference, at a modest inference overhead.
arXiv:2602.04990v3 Announce Type: replace
Abstract: The allocation of scarce donor organs constitutes one of the most consequential algorithmic challenges in healthcare. While the field is rapidly transitioning from rigid, rule-based systems to machine learning and data-driven optimization, we argue that current approaches often overlook a fundamental barrier: incentives. In this position paper, we highlight that organ allocation is not merely an optimization problem, but rather a complex game involving organ procurement organizations, transplant centers, clinicians, patients, and regulators. Focusing on US adult heart transplant allocation, we identify critical incentive misalignments across the decision-making pipeline, and present data showing that they are having adverse consequences today. Our main position is that the next generation of allocation policies should be incentive aware. We outline a research agenda for the machine learning community, calling for the integration of mechanism design, strategic classification, causal inference, and social choice to ensure robustness, efficiency, fairness, and trust in the face of strategic behavior from the various constituent groups.
arXiv:2602.04931v2 Announce Type: replace
Abstract: Geometric analyses of large language model (LLM) representations reveal structured variation across depth but remain fundamentally correlational with respect to token prediction formation. Meanwhile, causal interventions expose depth-dependent efficacy profiles without a unifying account of their representational dynamics. A complete account of LLM function requires explaining how representational structure evolves across depth to causally produce predictions. We synthesize these perspectives by combining geometric analysis with mechanistic interventions, explicitly centralizing depth-wise dynamics as the organizing axis for interpreting LLM function. In decoder-only LLMs, we identify a sharp transition from context-processing to prediction-forming computation, accompanied by a more gradual reorganization of representational geometry across layers. This synthesis reveals a late-layer geometric code in which angular structure parameterizes next-token distributional similarity and enables selective causal control over predictions, while representation norms encode information largely decoupled from prediction. Together, our results provide a synthesis of causal and geometric perspectives, yielding a mechanistic account of how control-relevant geometric dynamics across depth transform context into prediction in language models. This perspective reconciles previously puzzling findings and implies that layer-wise function cannot be understood or effectively intervened upon in isolation, but only within the emergent global dynamical structure of the network.
arXiv:2602.04599v2 Announce Type: replace
Abstract: We propose stochastic decision horizons (SDH), a theoretically grounded framework for solving constrained RL problems with every-step constraint satisfaction, a desirable property in many real-world applications. In SDH, a constraint violation yields an effective shortening of horizon via a state-action continuation probability. Using Control as Inference, we develop the first off-policy and regularized algorithms for RL with instantaneous constraints. We identify two principled semantics for what counts as a decision after a violation. Absorbing-state semantics end the decision process, so only surviving decisions pay entropy cost, yielding max-entropy AS-SAC. Virtual-termination keeps the decision process alive while stopping reward credit, yielding KL-regularized VT-MPO. To connect SDH with CMDPs, we track how violations accumulate along trajectories (their violation-depth profile). SDH effectively weights each trajectory by the exponential of its total violation; this matches an additive CMDP budget exactly when violations occur at a single characteristic scale, and we pinpoint where it cannot: when rare, deep violations mix with frequent, shallow ones. Experiments validate the theory. On the 90-muscle H2190 humanoid (Hyfydy), VT-MPO matches state-of-the-art gait realism with $4\times$ fewer environment steps and substantially more stable training. On Safety Gymnasium, violation-depth profiles correctly identify the regimes in which SDH delivers strong reward-violation trade-offs. Experiments validate the theory. On the 90-muscle H2190 humanoid (Hyfydy), VT-MPO matches state-of-the-art gait realism with 4x fewer environment steps and substantially more stable training. On Safety Gymnasium, violation-depth profiles correctly identify the regimes in which SDH delivers strong reward-violation trade-offs.
arXiv:2602.03517v2 Announce Type: replace
Abstract: Many decision-making problems require ranking individuals by their treatment effects rather than estimating the exact effect magnitudes. Examples include prioritizing patients for preventive care interventions, or ranking customers by the expected incremental impact of an advertisement. Surprisingly, while causal effect estimation has received substantial attention in the literature, the problem of directly learning rankings of treatment effects has largely remained unexplored. In this paper, we introduce Rank-Learner, a novel two-stage learner that directly learns the ranking of treatment effects from observational data. We first show that naive approaches based on precise treatment effect estimation solve a harder problem than necessary for ranking, while our Rank-Learner optimizes a pairwise learning objective that recovers the true treatment effect ordering, without explicit CATE estimation. We further show that our Rank-Learner is Neyman-orthogonal and thus comes with strong theoretical guarantees, including robustness to estimation errors in the nuisance functions. In addition, our Rank-Learner is model-agnostic, and can be instantiated with arbitrary machine learning models (e.g., neural networks). We demonstrate the effectiveness of our method through extensive experiments where Rank-Learner consistently outperforms standard CATE estimators and non-orthogonal ranking methods. Overall, we provide practitioners with a new, orthogonal two-stage learner for ranking individuals by their treatment effects.
arXiv:2602.02192v5 Announce Type: replace
Abstract: Reinforcement learning (RL) is a critical stage in post-training large language models (LLMs), involving repeated interaction between rollout generation, reward evaluation, and centralized learning. Distributing rollout execution offers opportunities to leverage more cost-efficient inference resources, but introduces challenges in wide-area coordination and policy dissemination. We present ECHO-2, a distributed RL framework for post-training with remote inference workers and non-negligible dissemination latency. ECHO-2 combines centralized learning with distributed rollouts and treats bounded policy staleness as a user-controlled parameter, enabling rollout generation, dissemination, and training to overlap. We introduce an overlap-based capacity model that relates training time, dissemination latency, and rollout throughput, yielding a practical provisioning rule for sustaining learner utilization. To mitigate dissemination bottlenecks and lower cost, ECHO-2 employs peer-assisted pipelined broadcast and cost-aware activation of heterogeneous workers. Experiments on GRPO post-training of LLMs ranging from 4B to 32B parameters under real wide-area bandwidth regimes show that ECHO-2 significantly improves cost efficiency while preserving RL reward comparable to strong baselines.
arXiv:2602.01518v2 Announce Type: replace
Abstract: Despite their importance in model sampling, efficient implementation of Top-k and Top-p algorithms for large vocabularies remains a significant challenge. Existing approaches often rely on sorting, which incurs significant computation and memory overhead on GPUs, or on stochastic approaches that alter the algorithm's output. In this work, we propose Qrita, an efficient Top-k and Top-p algorithm based on a pivot-based truncation and selection. Qrita leverages pivot-based search for both Top-k and Top-p with two key techniques: 1. Gaussian-based sigma-truncation, which greatly reduces the search space of the vocabulary, and 2. Quaternary pivot search with duplication handling, which halves the number of pivot search iterations and guarantees deterministic output. We implement Qrita using Triton and evaluate its performance against the Top-k and Top-p kernels of high-performance LLM execution engines such as SGLang and FlashInfer, improving end-to-end serving throughput up to 1.4 times with half the memory usage, while providing the same output as the sorting-based algorithms. Qrita is now the default Top-k and Top-p sampler for the GPU execution path of vLLM, and a ternary implementation of Qrita is available at https://github.com/vllm-project/vllm/blob/main/vllm/v1/sample/ops/topk_topp_triton.py.
arXiv:2602.00827v2 Announce Type: replace
Abstract: Feature learning strength (FLS), i.e., the inverse of the effective output scaling of a model, plays a critical role in shaping the optimization dynamics of neural nets. While its impact has been extensively studied under the asymptotic regimes -- both in training time and FLS -- existing theory offers limited insight into how FLS affects generalization in practical settings, such as when training is stopped upon reaching a target training risk. In this work, we investigate the impact of FLS on generalization in deep networks under such practical conditions. Through empirical studies, we first uncover the emergence of an $\textit{optimal FLS}$ -- neither too small nor too large -- that yields substantial generalization gains. This finding runs counter to the prevailing intuition that stronger feature learning universally improves generalization. To explain this phenomenon, we develop a theoretical analysis of gradient flow dynamics in two-layer ReLU nets trained with logistic loss, where FLS is controlled via initialization scale. Our main theoretical result establishes the existence of an optimal FLS arising from a trade-off between two competing effects: An excessively large FLS induces an $\textit{over-alignment}$ phenomenon that degrades generalization, while an overly small FLS leads to $\textit{over-fitting}$.
arXiv:2605.01489v2 Announce Type: replace
Abstract: Frontier scientific reasoning is rapidly emerging as a key foundation for advancing AI agents in automated scientific discovery. Deep research agents offer a promising approach to this challenge. These models develop robust problem-solving capabilities through post-training on information-seeking tasks, which are typically curated via knowledge graph construction or iterative web browsing. However, these strategies face inherent limitations in frontier science, where domain-specific knowledge is scattered across sparse and heterogeneous academic sources, and problem solving requires sophisticated computation and reasoning far beyond factual recall. To bridge this gap, we introduce SciResearcher, a fully automated agentic framework for frontier-science data construction. SciResearcher synthesizes diverse conceptual and computational tasks grounded in academic evidence, while eliciting information acquisition, tool-integrated reasoning, and long-horizon capabilities. Leveraging the curated data for supervised fine-tuning and agentic reinforcement learning, we develop SciResearcher-8B, an agent foundation model that achieves 19.46% on the HLE-Bio/Chem-Gold benchmark, establishing a new state of the art at its parameter scale and surpassing several larger proprietary agents. It further achieves 13-15% absolute gains on SuperGPQA-Hard-Biology and TRQA-Literature benchmarks. Overall, SciResearcher introduces a new paradigm for automated data construction for frontier scientific reasoning and offers a scalable path toward future scientific agents.