The requirement for large-scale global simulations of plasma is an ongoing challenge in both space and laboratory plasma physics. Any simulation based on a fluid model inherently requires a closure relation for the high order plasma moments. This review compiles and analyses the recent surge of machine learning approaches developing improved plasma closure models capable of capturing kinetic phenomena within plasma fluid models. We survey two methodological families: neural-network surrogates (from multilayer perceptrons to Fourier neural operators, the latter recently reproducing both linear and non-linear Landau damping online within a fluid solver) and equation-discovery methods such as sparse regression; and organise the studies by whether they are tested offline against reference data or online within a time-evolving solver. We outline the challenges associated with machine-learning closures, including off-diagonal pressure-tensor accuracy, generalisation beyond the training distribution, and stable integration into large-scale simulations, and the directions future research might take to address them.

Animal behavior reflects interactions between the nervous system, body, and environment. Therefore, biomechanics and environmental context must be considered to understand algorithms for behavioral control. Computational models that embed artificial neural controllers within body models in simulated environments, are a powerful tool for this purpose. Here, we review advances in biorealistic neuromechanical models while also highlighting emerging opportunities ahead. We first show how these models enable inference of biophysical variables that are difficult to measure experimentally. Through systematic perturbation, one can generate new experimentally testable hypotheses through these models. We then examine how neuromechanical models facilitate the exchange between neuroscience, robotics, and machine learning, and showcase their applications in healthcare. We envision that coupling experimental studies with active probing of their neuromechanical surrogates will significantly accelerate progress in neuroscience.

Multi-Agent Debate (MAD) is a collaborative framework in which multiple agents iteratively refine solutions through the generation of reasoning and alternating critique cycles. Current work primarily optimizes intra-round topologies and inter-round interactions separately, limiting the adaptation of token costs to task complexity. This work introduces Heterogeneous Consensus-Progressive Reasoning for Efficient Multi-Agent Debate (HCP-MAD), leveraging consensus as a dynamic signal to facilitate progressive reasoning. The core motivation is that a majority of straightforward tasks can be effectively resolved via lightweight pair-agent debates, while complex tasks require expanded collaboration. Firstly, Heterogeneous Consensus Verification conducts rapid consensus verification using a pair of heterogeneous agents for early stopping. Next, Heterogeneous Pair-Agent Debate applies an adaptive stopping criterion to terminate mutual critique of reasoning traces. Finally, the unresolved tasks are addressed through Escalated Collective Voting by aggregating diverse perspectives from additional agents. Experiments across six benchmarks show that HCP-MAD enhances accuracy while substantially reducing token costs. Code is https://github.com/fuyu66/HCP-MAD.

Individualized head-related impulse responses (HRIRs) enable binaural rendering, but dense per-listener measurements are costly. We address HRIR spatial up-sampling from sparse per-listener measurements: given a few measured HRIRs for a listener, predict HRIRs at unmeasured target directions. Prior learning methods often work in the frequency domain, rely on minimum-phase assumptions or separate timing models, and use a fixed direction grid, which can degrade temporal fidelity and spatial continuity. We propose HRIR-Former, a time-domain, grid-free binaural Transformer for reconstructing HRIRs at arbitrary directions from sparse inputs. It uses sinusoidal spatial features, a Conv1D refinement module, and auxiliary interaural time difference (ITD) and interaural level difference (ILD) heads. On SONICOM, it improves normalized mean squared error (NMSE), cosine distance, and ITD/ILD errors over prior methods; ablations validate modules and show minimum-phase preprocessing is unnecessary.

We introduce a new class of deep neural networks (DNNs) with multilayered tree-like architectures. The architectures are codified using numbers from the ring of integers of non-Archimdean local fields. These rings have a natural hierarchical organization as infinite rooted trees. Natural morphisms on these rings allow us to construct finite multilayered architectures. The new DNNs are robust universal approximators of real-valued functions defined on the mentioned rings. We also show that the DNNs are robust universal approximators of real-valued square-integrable functions defined in the unit interval.

As AI systems increasingly mediate negotiations, understanding how the number of negotiated issues impacts human performance is crucial for maintaining human agency. We designed a human-AI negotiation case study in a realistic property rental scenario, varying the number of negotiated issues; empirical findings show that without support, performance stays stable up to three issues but declines as additional issues increase cognitive load. To address this, we introduce a novel uncertainty-based visualization driven by Bayesian estimation of agreement probability. It shows how the space of mutually acceptable agreements narrows as negotiation progresses, helping users identify promising options. In a within-subjects experiment (N=32), it improved human outcomes and efficiency, preserved human control, and avoided redistributing value. Our findings surface practical limits on the complexity people can manage in human-AI negotiation, advance theory on human performance in complex negotiations, and offer validated design guidance for interactive systems.

We present an AI Co-Scientist framework that closes the research loop for the production search-ranking system of a large online travel platform -- pairing LLM agents with direct cloud-compute access so that idea generation, code implementation, GPU experimentation, and result analysis iterate end-to-end with a human scientist in the loop. The framework uses a hybrid agent architecture: single-LLM agents handle routine work, while multi-LLM consensus (GPT-5.2, Gemini Pro 3, Claude Opus 4.5) is invoked for higher-stakes decisions. On the production ranking task, a human-designed transformer baseline (V2) yielded $+0.118\%$ over a pre-transformer baseline (V1); the AI Co-Scientist's automated loop on top of V2 contributed an additional $+0.083\%$, for a combined $+0.201\%$ offline gain delivered in roughly one extra week of wall-clock time (single-run numbers; statistical limits discussed in the paper). The most useful AI proposals -- unified long-sequence layouts, slot-type embeddings, and multi-phase learning-rate schedules -- are standard practice in NLP and Vision but were absent from our production stack, suggesting that LLM agents can serve as cross-disciplinary connectors for ranking teams. We also report deployment context, negative results, and lessons learned.

Recommender agents built on Large Language Models offer a promising paradigm for personalized recommendation. However, existing agents typically suffer from a misalignment between their tool-integrated reasoning trajectories and recommendation feedback, limiting their ability to distinguish fine-grained user preferences. To address these challenges, we propose AgenticRec, an agentic recommendation framework that formulates recommendation as a tool-integrated reasoning process over a recommendation-oriented tool suite. Built upon this framework, we further develop a dedicated two-stage training paradigm tailored for recommender agents. In the first stage, we introduce Recommendation-Oriented Trajectory Activation, optimize the agentic recommendation ability under implicit feedback. In the second stage, Progressive Preference Refinement further refines the agent through bidirectional preference reasoning over self-bootstrapped hard pairs, progressively sharpening preference boundaries. Theoretical analysis and extensive experiments demonstrate the effectiveness of AgenticRec. Our code is available at https://anonymous.4open.science/r/AgenticRec-FB16.

Lifelong interactive agents are expected to assist users over months or years, which requires continually writing long term memories while retrieving the right evidence for each new query under fixed context and latency budgets. Existing memory systems often degrade as histories grow, yielding redundant, outdated, or noisy retrieved contexts. We present \textbf{All-Mem}, an online/offline lifelong memory framework that maintains a topology structured memory bank via explicit, non destructive consolidation, avoiding the irreversible information loss typical of summarization based compression. In online operation, it anchors retrieval on a bounded visible surface to keep coarse search cost bounded. Periodically offline, an LLM diagnoser proposes confidence scored topology edits executed with gating using three operators: Split, Merge, and Update, while preserving immutable evidence for traceability. At query time, typed links enable hop bounded, budgeted expansion from active anchors to archived evidence when needed. Experiments on \textbf{LoCoMo} and \textbf{LongMemEval-s} show improved retrieval and QA over representative baselines. The code is available at https://github.com/LvCan926/All-Mem.

LLM-powered agents execute tasks through a sequential loop of model generation and tool execution. Today's serving systems serialize this loop, leaving tool latency exposed on the task critical path. This paper presents PASTE, a tool-aware agent-serving system that predicts concrete future tool invocations from recurring agent patterns and executes them speculatively while the LLM is still generating. PASTE isolates speculative results until confirmed by the LLM and jointly schedules tool execution and returning LLM sessions to avoid shifting bottlenecks to the GPU. Across deep research, coding, and scientific-agent workloads, PASTE reduces average task completion time by 43.5% and lowers observed tool latency by 1.8x.

With the success of static black-hole imaging, the next frontier is the dynamic and 3D imaging of black holes. Recovering the dynamic 3D gas near a black hole would reveal previously-unseen parts of the universe and inform new physics models. However, only sparse radio measurements from a single viewpoint are possible, making the dynamic 3D reconstruction problem significantly ill-posed. Previously, BH-NeRF addressed the ill-posed problem by assuming Keplerian dynamics of the gas, but this assumption breaks down near the black hole, where the strong gravitational pull of the black hole and increased electromagnetic activity complicate fluid dynamics. To overcome the restrictive assumptions of BH-NeRF, we propose PI-DEF, a physics-informed approach that uses differentiable neural rendering to fit a 4D (time + 3D) emissivity field given EHT measurements. Our approach jointly reconstructs the 3D velocity field with the 4D emissivity field and enforces the velocity as a soft constraint on the dynamics of the emissivity. In experiments on simulated data, we find significantly improved reconstruction accuracy over both BH-NeRF and a physics-agnostic approach. We demonstrate how our method may be used to estimate other physics parameters of the black hole, such as its spin.

Deep learning has achieved recognition for its impact within natural sciences, yet the prohibitive financial and technical cost of training models from scratch inhibit adoption. Following software engineering community guidance, natural scientists are reusing pre-trained deep learning models (PTMs) to amortize these costs. While prior works recommend PTM reuse patterns, we present the first empirical study of PTM reuse patterns in the natural sciences, quantifying the utilization and impact of PTM reuse within the scientific process across 17,718 peer reviewed, open access papers. Our results show that "Biochemistry, Genetics and Molecular Biology" has outpaced other natural scientific fields in PTM reuse, "adaptation" reuse is the most prevalent PTM reuse pattern identified across all natural science fields, and the "testing" stage of the scientific process has been most impacted by PTM integration.

Monotone operator equilibrium networks are implicit-layer models whose output is the unique equilibrium of a monotone operator, guaranteeing existence, uniqueness, and convergence. When deployed on low-precision hardware, weights are quantized, potentially destroying these guarantees. We analyze weight quantization as a spectral perturbation of the underlying monotone inclusion. Convergence of the quantized solver is guaranteed whenever the spectral-norm weight perturbation is smaller than the monotonicity margin; the displacement between quantized and full-precision equilibria is bounded in terms of the perturbation size and margin; and a condition number characterizing the ratio of the operator norm to the margin links quantization precision to forward error. MNIST experiments confirm a phase transition at the predicted threshold: three- and four-bit post-training quantization diverge, while five-bit and above converge. The backward-pass guarantee enables quantization-aware training, which recovers provable convergence at four bits.

Parallel test-time scaling, which generates multiple candidate solutions for a single problem, is a powerful technique for improving large language model performance. However, it is hindered by two key bottlenecks: accurately selecting the correct solution from the candidate pool, and the high inference latency from generating many full solutions. We argue that both challenges are fundamentally linked to verifier calibration, as a well-calibrated verifier improves answer selection and enables early-stopping strategies to reduce latency. However, existing non-generative verifiers are limited as they score each candidate in isolation, overlooking rich contextual information across the set of candidates. To address this, we introduce the Multi-Sequence Verifier (MSV), a lightweight verifier that predicts each candidate's correctness conditioned on the full sampled set. MSV achieves improved calibration, which directly enhances best-of-N selection performance and empowers a novel early-stopping framework. Across challenging mathematical reasoning benchmarks, MSV improves best-of-64 accuracy by up to 6\% relative to strong baselines, and in the early-stopping setting reaches the same accuracy as baselines with less than half the latency.

Speculative decoding (SD) accelerates LLM inference by verifying draft tokens in parallel. However, this method presents a critical trade-off: it improves throughput in low-load, memory-bound systems but degrades performance in high-load, compute-bound environments due to verification overhead. Existing speculative decoding methods use fixed lengths and cannot adapt to workload changes or decide when to stop speculation. The cost of restarting speculative inference also remains unquantified. Under high load, the benefit of speculation diminishes, while retaining the draft model reduces KV cache capacity, limiting batch size and degrading throughput. To overcome this, we propose Nightjar, a resource-aware adaptive speculative framework. It first adjusts to the request load by dynamically selecting the optimal speculative length for different batch sizes. Crucially, Nightjar proactively disables speculative decoding when the MAB planner determines that speculation is no longer beneficial, and during the disabled phase, offloads the draft model to the CPU only under GPU memory pressure. This reclaims memory for the KV cache, thereby facilitating larger batch sizes and maximizing overall system throughput. Experiments show that Nightjar achieves up to 14.76% higher throughput than standard speculative decoding and up to 20.18% lower latency in the main benchmark suite under dynamic request arrival rates for real-time LLM serving scenarios.

Clinical diagnosis is a gradual process of evidence integration, in which physicians move from symptoms and medical history to examinations, competing hypotheses, disease relations, and treatment decisions. Large language models have advanced medical text understanding and generation. Yet their clinical use remains limited by weak evidence grounding, opaque reasoning, and inconsistent links among differential diagnosis, final diagnosis, diagnostic basis, and treatment planning. We introduce MedCollab, a multi-agent framework for full-cycle clinical diagnosis and report generation. MedCollab coordinates specialist and examination agents according to patient records. It structures agent deliberation with an Issue-Based Information System (IBIS) protocol, so that each diagnostic position is supported by patient-specific evidence and medical knowledge. It also builds Hierarchical Disease Relation Chains (HDRC) to connect accepted hypotheses through progression, complication, and comorbidity relations. During multi-round deliberation, a verifier-guided consensus module evaluates evidence support, medical plausibility, and logical conflicts. It then adjusts agent contributions and filters unsupported reasoning. Experiments on ClinicalBench and MIMIC-IV show that MedCollab outperforms leading LLMs and medical multi-agent baselines in diagnostic accuracy, evidence consistency, and clinical reasoning quality. These results indicate that structured and auditable collaboration can produce more faithful and clinically coherent diagnostic reports.

Accurate, low-latency channel modeling is essential for real-time wireless network simulation and digital-twin applications. Traditional modeling methods like ray tracing are however computationally demanding and unsuited to model dynamic conditions. In this paper, we propose AIRMap, a deep-learning framework for ultra-fast radio-map estimation, along with an automated pipeline for creating the largest radio-map dataset to date. AIRMap uses a single-input U-Net autoencoder that processes only a 2D elevation map of terrain and building heights. Trained on 1.2M Boston-area samples and validated across four distinct urban and rural environments with varying terrain and building density, AIRMap predicts path gain with under 4 dB RMSE in 4 ms per inference on an NVIDIA L40S-over 100x faster than GPU-accelerated ray tracing based radio maps. A lightweight calibration using just 20% of field measurements reduces the median error to approximately 5%, significantly outperforming traditional simulators, which exceed 50% error. Integration into the Colosseum emulator and the Sionna SYS platform demonstrate near-zero error in spectral efficiency and block-error rate compared to measurement-based channels. These findings validate AIRMap's potential for scalable, accurate, and real-time radio map estimation in wireless digital twins.

We study one-sided and $α$-correct sequential hypothesis testing for data generated by an ergodic, finite-state Markov chain. The null hypothesis is that the unknown transition matrix belongs to a prescribed set $P$ of stochastic matrices, and the alternative corresponds to a disjoint set $Q$. We establish a non-asymptotic instance-dependent lower bound on the expected stopping time of any valid sequential test under the alternative, which is asymptotically tight. Our novel analysis improves the existing lower bounds, which are either asymptotic or provably sub-optimal in this setting. Our lower bound incorporates both the stationary distribution and the transition structure induced by the unknown Markov chain. We further propose an optimal test whose expected stopping time matches this lower bound asymptotically as $α\to 0$. We illustrate the usefulness of our framework through applications to sequential detection of model misspecification in Markov Chain Monte Carlo and to testing structural properties, such as the linearity of transition dynamics, in Markov decision processes. Our findings yield a sharp and general characterization of optimal sequential testing procedures under Markovian dependence.

In this paper, we investigate discrete optimal transport (DOT) as a black-box attack against modern automatic speaker verification (ASV) and anti-spoofing countermeasure (CM) systems. Our attack operates as a post-processing distribution-alignment step. Frame-level WavLM embeddings of generated speech (or another person speech) are aligned to an unpaired bona fide speech pool using entropic optimal transport and a top-k barycentric projection, followed by neural vocoding. Unlike gradient-based attacks, the proposed method requires no access to model parameters, gradients, or training data. Experiments on ASVspoof2019 and ASVspoof5 demonstrate that DOT attack substantially increases CM EER and substantially degrades ASV performance across multiple spoofing attacks. The attack transfers across datasets and remains effective after CM fine-tuning. Analysis using speaker similarity, Fréchet Audio Distance, and visualization of embedding distributions suggests that DOT succeeds by shifting source speech toward bona fide regions of the representation space rather than by maximizing speaker similarity. These results indicate that optimal-transport-based distribution alignment represents a previously underexplored attack vector for contemporary ASV and anti-spoofing systems.

We present ROSA -- Roundabout Optimized Speed Advisory -- a system that combines multi-agent trajectory prediction with coordinated speed guidance for multimodal, mixed traffic at roundabouts. Using a Transformer-based model, ROSA jointly predicts the future trajectories of vehicles and Vulnerable Road Users (VRUs) at roundabouts. Trained for single-step prediction and deployed autoregressively, it generates deterministic outputs, enabling actionable speed advisories. Incorporating motion dynamics, the model achieves high accuracy (ADE: 1.29m, FDE: 2.99m at a five-second prediction horizon), surpassing prior work. Adding route intention further improves performance (ADE: 1.10m, FDE: 2.36m), demonstrating the value of connected vehicle data. Based on predicted conflicts with VRUs and circulating vehicles, ROSA provides real-time, proactive speed advisories for approaching and entering the roundabout. Despite prediction uncertainty, ROSA significantly improves vehicle efficiency and safety, with positive effects even on perceived safety from a VRU perspective. The source code of this work is available under: github.com/urbanAIthi/ROSA.

Conversational search (CS) requires a complex software engineering pipeline that integrates query reformulation, ranking, and response generation. CS researchers currently face two barriers: the lack of a unified framework for efficiently sharing contributions with the community, and the difficulty of deploying end-to-end prototypes needed for user evaluation. We introduce Orcheo, an open-source platform designed to bridge this gap. Orcheo offers three key advantages: (i) A modular architecture promotes component reuse through single-file node modules, facilitating sharing and reproducibility in CS research; (ii) Production-ready infrastructure bridges the prototype-to-system gap via dual execution modes, secure credential management, and execution telemetry, with built-in AI coding support that lowers the learning curve; (iii) Starter-kit assets include 45+ off-the-shelf components for query understanding, ranking, and response generation, enabling the rapid bootstrapping of complete CS pipelines. We describe the framework architecture and validate Orcheo's utility through case studies that highlight modularity and ease of use. Orcheo is released as open source under the MIT License at https://github.com/AI-Colleagues/orcheo.

Large language model (LLM) based web agents are increasingly deployed to automate complex online tasks by directly interacting with web sites and performing actions on users' behalf. While these agents offer powerful capabilities, their design exposes them to indirect prompt injection attacks embedded in untrusted web content, enabling adversaries to hijack agent behavior and violate user intent. Despite growing awareness of this threat, existing evaluations rely on fixed attack templates, manually selected injection surfaces, or narrowly scoped scenarios, limiting their ability to capture realistic, adaptive attacks encountered in practice. We present MUZZLE, an automated agentic framework for evaluating the security of web agents against indirect prompt injection attacks. MUZZLE utilizes the agent's trajectories to automatically identify high-salience injection surfaces, and adaptively generate context-aware malicious instructions that target violations of confidentiality, integrity, and availability. Unlike prior approaches, MUZZLE adapts its attack strategy based on the agent's observed execution trajectory and iteratively refines attacks using feedback from failed executions. We evaluate MUZZLE across diverse web applications, user tasks, and agent configurations, demonstrating its ability to automatically and adaptively assess the security of web agents with minimal human intervention. Our results show that MUZZLE effectively discovers 44 new attacks on 4 web applications with 10 adversarial objectives that violate confidentiality, availability, or privacy properties across different LLMs and agent scaffolds. MUZZLE also identifies novel attack strategies, including 3 cross-application prompt injection attacks and an agent-tailored phishing scenario.

Agentic systems solve complex tasks by coordinating multiple agents that iteratively reason, invoke tools, and exchange intermediate results. To improve robustness and solution quality, recent approaches deploy multiple agent teams running in parallel to explore diverse reasoning trajectories. However, parallel execution comes at a significant computational cost: when different teams independently reason about similar sub-problems or execute analogous steps, they repeatedly perform substantial overlapping computation. To address these limitations, in this paper, we propose Learning to Share (LTS), a learned shared-memory mechanism for parallel agentic frameworks that enables selective cross-team information reuse while controlling context growth. LTS introduces a global memory bank accessible to all teams and a lightweight controller that decides whether intermediate agent steps should be added to memory or not. The controller is trained using stepwise reinforcement learning with usage-aware credit assignment, allowing it to identify information that is globally useful across parallel executions. Experiments on the AssistantBench and GAIA benchmarks show that LTS significantly reduces overall runtime while matching or improving task performance compared to memory-free parallel baselines, demonstrating that learned memory admission is an effective strategy for improving the efficiency of parallel agentic systems. Project page: https://joefioresi718.github.io/LTS_webpage/

This paper addresses the challenge of audio-visual single-microphone speech separation and enhancement in the presence of real-world environmental noise. Our approach is based on generative inverse sampling, where we model clean speech and ambient noise with dedicated diffusion priors and jointly leverage them to recover all underlying sources. To achieve this, reformulate a recent inverse sampler to match our setting. We evaluate on mixtures of 1, 2, and 3 speakers with noise and show that, despite being entirely unsupervised, our method consistently outperforms leading supervised baselines in WER across all conditions. We further extend our framework to handle off-screen speaker separation. Moreover, the high fidelity of the separated noise component makes it suitable for downstream detection of the acoustic scene. Code and pretrained models will become available upon acceptance. Demo page: https://ssnaps2026.github.io/ssnaps2026/

Governments with limited fiscal and administrative capacity need to know which Sustainable Development Goals (SDGs) propagate progress through the goal system and how quickly. We map the directed interdependence structure of all seventeen goals using a balanced panel of 114 countries observed annually from 2000 to 2024. The goal series are persistent, trending, and cross-sectionally dependent, so we apply two estimators matched to this regime: a Dumitrescu-Hurlin panel Granger non-causality test, run on first-differenced series, to recover the directed interaction network, and panel local projections with Driscoll-Kraay standard errors to measure the dynamic magnitude of 31 theory-derived indicator linkages. Of 272 directed goal pairs, 84 linkages survive false-discovery control (40 synergies, 44 trade-offs; network density 0.31). Synergies and trade-offs occur at comparable strength, so no single goal behaves as a universal accelerator, and the goal-level hierarchy itself is fragile. Driver-receiver rankings correlate weakly across lag orders and centrality metrics, and under a country bootstrap only two roles are distinguishable from zero: peace and strong institutions as the clearest net receiver, and poverty reduction as the most probable effect-size-weighted driver. The supported linkages are dynamic, accruing over four to five years: sanitation and poverty improvements are the strongest predictors of lower child mortality, and the education-child-health association is corroborated in independent World Development Indicators data across 183 countries. These results caution against rankings-based accelerator policy and support adaptive portfolios built on supported, time-lagged linkages monitored through constituent indicators.

Repository-level code completion remains a challenging task for existing code large language models (code LLMs) due to their limited understanding of repository-specific context and domain knowledge. While retrieval-augmented generation (RAG) approaches have shown promise by retrieving relevant code snippets as cross-file context, they suffer from two fundamental problems: misalignment between the query and the target code in the retrieval process, and the inability of existing retrieval methods to effectively utilize the inference information. To address these challenges, we propose AlignCoder, a repository-level code completion framework that introduces a query enhancement mechanism and a reinforcement learning based retriever training method. Our approach generates multiple candidate completions to construct an enhanced query that bridges the semantic gap between the initial query and the target code. Additionally, we employ reinforcement learning to train an AlignRetriever that learns to leverage inference information in the enhanced query for more accurate retrieval. We evaluate AlignCoder on two widely-used benchmarks (CrossCodeEval and RepoEval) across five backbone code LLMs, demonstrating an 18.1% improvement in EM score compared to baselines on the CrossCodeEval benchmark. The results show that our framework achieves superior performance and exhibits high generalizability across various code LLMs and programming languages.

Safety classifiers that screen LLM inputs for jailbreak attempts have become standard deployment components, yet almost all production systems rely on GPU-based models: fine-tuned transformers and LLM-as-a-judge pipelines. These approaches impose significant per-query latency and infrastructure cost. Very little research has asked whether CPU-based classifiers, such as support vector machines and gradient-boosted trees trained on TF-IDF features, can match their accuracy across the conditions that production deployments encounter. We evaluate five CPU classifier families, Mamba-130M as an SSM-based GPU classifier, and transformer-based GPU models (DeBERTa-v3 and Gemma-2B with LoRA) across nine jailbreak sources and three regimes: in-distribution (D1), out-of-distribution (D2), and adversarially obfuscated (D3). On D1, the best CPU classifier matches the best transformer GPU model at roughly one-fifth the deployment cost. On D2, CPU classifiers fail via confident miscalibration, producing high-confidence false negatives that bypass escalation entirely. On D3, CPU classifiers outperform transformer GPU models by more than 26 percentage points in F1. Based on these complementary failure modes, we design GuardChain, a three-stage safety pipeline (Regex -> CPU -> GPU) that routes each prompt to the cheapest stage capable of a confident decision. The CPU stage alone resolves 80\% of in-distribution prompts at near-peak accuracy, and the GPU stage recovers the out-of-distribution failures. For practitioners deploying LLM safety at scale, this work provides evidence that GPU-class infrastructure is unnecessary for the majority of traffic.

Agentic Reinforcement Learning (RL) trains LLMs through multi-turn interactions with environments, producing workloads that mix compute-bound prefill, bandwidth-bound decoding, CPU-heavy environment execution, and bursty reward evaluation. Existing systems either colocate all stages on a single GPU cluster or decouple them only at a coarse granularity, overlooking hardware heterogeneity and incurring substantial synchronization overhead across stages. We present ROLLART, a system for multi-task agentic RL on disaggregated infrastructure. ROLLART maps each pipeline stage to best-fit hardware, routing prefill-heavy tasks to compute-optimized GPUs, decode-heavy tasks to bandwidth-optimized GPUs, and environments to CPU clusters. It decouples rollout at the trajectory level, allowing generation, environment interaction, and reward scoring to proceed independently, so that slow or failed environments never block the others. ROLLART offloads stateless reward computation to serverless infrastructure and overlaps rollout with training via staleness-bounded asynchronous weight synchronization. Our results demonstrate that ROLLART effectively improves training throughput and achieves 1.31--2.05 \(\times\) training time reduction compared to various RL systems. We also evaluated ROLLART by training a hundreds-of-billions-parameter MoE model for Qoder product on an Alibaba cluster with above 3,000 GPUs, demonstrating its stability and scalability.

This paper develops a finite-sample statistical theory for in-context learning (ICL), analyzed within a meta-learning framework that accommodates mixtures of diverse task types. We introduce a principled risk decomposition that separates the total ICL risk into two orthogonal components: Bayes Gap and Posterior Variance. The Bayes Gap quantifies how well the trained model approximates the Bayes-optimal in-context predictor. For a uniform-attention Transformer, we derive a non-asymptotic upper bound on this gap, which explicitly clarifies the dependence on the number of pretraining prompts and their context length. The Posterior Variance is a model-independent risk representing the intrinsic task uncertainty. Our key finding is that this term is determined solely by the difficulty of the true underlying task, while the uncertainty arising from the task mixture vanishes exponentially fast with only a few in-context examples. Together, these results provide a unified view of ICL: the Transformer selects the optimal meta-algorithm during pretraining and rapidly converges to the optimal algorithm for the true task at test time.

Large language models (LLMs) and LLM-based coding agents are now used to generate code from natural-language specifications, yet ensuring such code is both functionally correct and secure remains a challenge. We present DualGauge, the first fully automated framework for jointly evaluating correctness and security of specification-only code generation, supported by DualGauge-Bench, a language-agnostic benchmark of 307 coding tasks each paired with functional and security tests derived from the same specification. Evaluating 10 representative LLMs across Python, C++, and JavaScript, we find that functional correctness substantially overestimates reliable code generation: even the strongest model remains below 15% joint security-functionality success in every language. Common model-side factors--scale, extended thinking, quantization, instruction tuning, and code specialization--do not reliably improve joint performance, suggesting secure-and-correct code generation does not simply emerge from stronger coding capability. Evaluation of 3 leading agentic coding systems (Codex, OpenHands, and Claude Code) shows that iterative scaffolding provides no advantage over direct (LLM-based) generation on specification-only tasks. A qualitative audit reveals failures concentrate at the output contract boundary and in guards that exist but are insufficient--patterns that only joint benchmarking reliably exposes.