每天知道
你的领域
出了什么

Transformers asked to multiply multi-digit numbers in a single forward pass often fail, and interpretability studies of pretrained language models find arithmetic solved by input-range heuristics rather than by an explicit carry. We train small Llama-style transformers from scratch on 4x4 multiplication without chain of thought and find that the input format is decisive: inserting a space token between digits raises exact-match accuracy from 1% to 89%. Output positions are learned in carry-chain order, with the middle digits, which have the longest-range dependencies, learned last. Inside the model, the separator token that predicts each digit (its prediction slot) encodes the carry-in as an angle on a ring in the residual stream; examples with more distinct carry values fill more of the ring. Activation patching between examples matched on the column sum shows that this state is causally used before the last layer: patching the prediction slot alone transfers the source carry in up to 84% of cases after block 4 for one middle column of our best model, while for other columns the carry is first assembled at the neighboring answer slot before reaching its own. Remaining errors are almost always off by one, consistent with a small error on the carry or on the circular digit code.

LLM agents are starting to own the full customer experience. Soon, LLMs may be selling and buying on behalf of companies and customers respectively. Small models are more cost-efficient at scale, but can reinforcement learning train them into competent sellers? We train four Gemma 4 checkpoints (2.3B to 31B effective parameters) with GRPO on a programmatic utility reward for bilateral multi-issue bargaining, and evaluate every arm on the same 1,152 negotiations against two frontier buyers it never saw in training. With the same learning rate ($10^{-6}$) for every size, the gain of the RL model over its base rises from $+0.001$ at 2.3B to $+0.078$ at 31B. Each size was trained once and the two smallest checkpoints use a different architecture, so we fit no scaling law. Tripling the learning rate, with the same or fewer training steps, improves on the shared rate at every size by $+0.032$ (2.3B) to $+0.081$ (4.5B). In exploratory comparisons with two frontier models run as sellers, the 12B seller trained at the tripled rate scores above both, though its untrained base already scores as high as they do. The 4.5B seller at that rate shows no detectable difference from either and fits on one 48 GB GPU. A further 2.3B arm at ten times the shared rate raises pooled score, but its gain concentrates on the evaluation buyer that shares a model family with the training pool. These results suggest tuning the learning rate before concluding that a small model cannot learn to negotiate, and testing against buyers from more than one model family.

We investigate the use of formal methods to provide tight and sound generalization bounds for learning algorithms. By casting the traditional notion of algorithmic stability as a specification to be verified, we demonstrate that recent advances in reachability analysis can yield provable bounds on the generalization of a given model and algorithm on a sample dataset. As sample-specific algorithmic stability is insufficient to bound the usual distributional notion of generalization, we develop a novel concentration inequality to connect the sample-specific results of formal certification algorithms to the required distributional analysis for bounding the expected generalization gap. The resulting framework enables the analysis of prior generalization bounds to extend far beyond their original restrictive assumptions. Our approach computes sound bounds on the expected generalization gap in a constant number of algorithm runs without making any analytical assumptions on the algorithm; to achieve non-vacuous bounds we only require that the certified reachable parameter set is bounded --- a condition that we do not assume but formally verify. In practice, we demonstrate that our framework provides formal generalization guarantees that are orders of magnitude tighter than alternative sound computational approaches at scales ranging from toy datasets to fine-tuning classification heads on top of modern large language models. While we implement certification-enhanced versions of several well-known stability results, future extensions of our approach will enable tighter bounds and enhanced practical adoption across the spectrum of modern generalization bounds.

We consider the stability of multi-step reasoning processes, which have extensive applications in language models, including chain-of-thought and algorithmic reasoning. While longer sequences of reasoning can improve a model's generation capability at test time, the errors due to intermediate reasoning steps can accumulate in autoregressive generation, and thus grow substantially at the end. In this paper, we ask: What are the key factors determining the stability of multi-step reasoning? First, we show an inference error bound governed by the product of spectral norms of the Jacobians taken through the input space across generation steps. This product can be viewed as an error amplification factor, which could scale exponentially with the number of reasoning steps, serving as a quantitative measure of reasoning stability. Second, we analyze this measure in transformer models trained to predict simple tasks like linear and quadratic functions. We theoretically prove that the transformer model converges to a solution where the stability measure decays, thus yielding nearly zero inference loss over (arbitrarily) long steps. Finally, the stability analysis leads to several algorithmic implications for controlling the stability, through (i) chain-of-thought length compression that reduces the sensitivity of each step, and (ii) quantization-aware training that regularizes the input Jacobian norms. We validate the proposed algorithms by fine-tuning language models on graph-algorithmic reasoning tasks and symbolic state-tracking tasks. Across seven evaluations, our algorithms improve over baseline comparisons by 3.5% on average, and by 8.2% for longer-length inputs. Ablation analysis validates that the stability measure is drastically reduced by 3-8$\times$, confirming the regularization effect on the spectral norms of the (input space) Jacobians.

Verifier-guided decoding can prevent harmful reasoning steps from contaminating subsequent generation, but typically relies on an external learned verifier. We ask whether a language model can instead reject its own bad reasoning steps. We define a prefix's recoverability as the probability that the frozen generator can complete it correctly. Diagnostics show that adjacent recoverability changes are often difficult to resolve with practical Monte Carlo budgets, while same-prefix candidates exhibit a sparse low-recoverability tail. We introduce Self-Step Rejection (SSR), which trains a lightweight LoRA acceptance gate on the generator backbone while keeping the base model frozen. SSR uses confidence-qualified first-passage supervision: steps before the first resolved crossing of a root-relative recoverability barrier are accepted, the crossing step is rejected, and unresolved steps and suffixes are excluded. Training combines pointwise classification, same-prefix pairwise learning, and group-relative policy refinement using final-answer correctness. At inference, SSR accepts candidates or resamples from the unchanged prefix under rejection budgets, without an external learned verifier. Across three reasoning models and five mathematical reasoning benchmarks, SSR improves macro-average accuracy over single-pass decoding by 5.4--10.1 points using 1.21--1.40x as many generated tokens, and achieves the highest macro-average accuracy among evaluated step-level methods. Full-solution scaling methods require 4.47--8.27x the single-pass token cost for comparable performance.

Latent reasoning lets a large language model (LLM) think in a continuous space and verbalize only the answer. We argue that an effective latent thought must meet five requirements: it should be useful, helping produce the correct answer rather than merely changing it, diverse, so that resampling yields different reasoning trajectories, explainable, so that a decoded chain of thought (CoT) reflects reasoning the answer actually follows, refinable with more inference compute, and efficient, costing less than an explicit CoT at comparable accuracy. Current methods rarely meet these requirements: they learn shortcuts from the question, distill the explicit CoT into their weights, or imitate it one token at a time. We focus on flow matching in a learned latent space, the family we argue is best placed to meet them, and identify the training choices that make it work. The result is Flow-based Latent Reasoning (FLaRe), a simple recipe covering what the latent space encodes and how to shape it, where to train the flow, how to read out the answer, and a final stage of training on the model's own verified thoughts. A probe for each requirement shows that FLaRe improves on prior latent methods in all five. It also compares favorably with them on arithmetic benchmarks, while reaching 97% of the accuracy of explicit CoT at a quarter of its latency.

Next-token prediction (NTP) is the standard pretraining objective for large language models, yet it provides an explicit training signal only for the immediate next token, which can lead models to exploit local patterns instead of capturing longer-range structure and ideas. Multi-token prediction (MTP) addresses this by training models to predict several future tokens. However, existing MTP methods often introduce a large number of new parameters with limited improvements in downstream performance. Latent MTP approaches address this efficiency issue by encoding future tokens into a vector representation. However, these approaches usually rely on external helper models for future token encoding. We propose LightMTP, a lightweight, i.e., parameter-efficient, latent MTP approach that bootstraps the future token representations from the model's own hidden states. Our two LightMTP variants extend supervision to more future tokens without requiring the additional computational overhead of conventional MTP nor the external supervision latent MTP normally relies on. LightMTP adds at most 1% extra parameters, retains better performance on general language modeling benchmarks, and achieves similar gains in planning, coding, and reasoning.

Large language models (LLMs) have achieved substantial performance gains through increases in model size, training data, and computational resources. However, traditional scaling approaches produce diminishing returns, rising financial and environmental costs, and barriers to participation for researchers operating outside large industrial laboratories. This review examines the evolution of LLM scaling theory from empirical scaling laws to compute-optimal training, with particular emphasis on parameter efficiency, token utilization, data efficiency, and resource-constrained environments. Foundational work on scaling laws is synthesized alongside later research on compute-optimal training, data pruning, efficient architectures, quantization, low-rank adaptation, and edge-oriented optimization. The literature indicates a shift from scale maximization toward more deliberate allocation of parameters, tokens, compute, and hardware resources. At the same time, important empirical, theoretical, and methodological gaps remain regarding whether scaling principles established on enterprise-grade infrastructure generalize to smaller models and constrained computing environments. This review organizes these developments into a unified framework for resource-efficient LLM training and argues that future progress should evaluate efficiency not solely through model performance, but through the relationship among performance, parameter count, computational cost, token allocation, and hardware constraints.

Mixture-of-Experts (MoE) has been widely adopted in recent large language model (LLM) architectures. However, scaling up MoE in LLM training introduces system-level challenges on training, where non-uniform token routing can lead to highly imbalanced workloads across experts and devices, further destabilizing the training process. With trillion-scale LLMs, imbalanced expert workloads further amplify the resource cost of MoE training, resulting in degraded training efficiency and hardware utilization for underloaded experts, while hot experts require additional resources to accommodate excessive workloads. Recent studies address imbalanced MoE training through intricate parallelism strategies or resource reallocation. However, these system-level approaches often introduce additional resource requirements and considerable orchestration complexity, which become increasingly difficult to afford when training trillion-parameter LLMs under constrained computational resources. This work introduces CIPHER-MoE, which mitigates MoE workload imbalance while keeping the router's token-side Top-K selection unchanged. CIPHER-MoE applies affinity-aware Expert-to-Token filtering with explicit capacity control to reduce hotspot expert workloads without additional hardware resources or complex runtime design. The proposed method has been evaluated on large-scale MoE models, including DeepSeek-V4-Pro, showing up to 64.9 percentage points Top-1 expert workload reduction and 1.10$\times$-1.94$\times$ training acceleration, while preserving the training quality. The source code will be released soon.

Post-training quantization has been shown to preserve VLA performance under standard evaluation conditions, but whether it preserves the full-precision model's robustness and generalization remains underexplored. In this study, we systematically study the robustness and generalization of post-quantized VLA policies under environmental disturbances. Empirical results show that quantized policies can become fragile to subtle environmental variations despite retaining comparable in-distribution performance. We further observe that action discrepancies are concentrated in a small subset of rollout states, while teacher guidance has opposite effects depending on discrepancy: it improves generalization at high-discrepancy states but can degrade it at low-discrepancy states. These findings reveal that effective post-quantization recovery requires selectively intervening on vulnerable states rather than globally distilling the student. We therefore propose Policy-Induced Vulnerability-Oriented Tuning (PIVOT-Q), a vulnerability-aware On-Policy Distillation (OPD) framework that selectively corrects vulnerable states encountered during quantized-student rollouts using the frozen full-precision policy as a teacher. PIVOT-Q identifies vulnerable states using discounted accumulated discrepancies over a short horizon, applies phase-balanced sparse supervision, and uses a Behavioral Anchor to prevent unnecessary changes. Experiments under seven LIBERO-Plus environmental variations demonstrate consistent recovery across multiple VLA backbones and quantization methods. Notably, PIVOT-Q consistently outperforms full-state distillation across all settings while using only 7.4% of its state-level distillation budget. Our code is available at https://github.com/ruanruan-andy/PIVOT-Q.

A long-standing goal of AI is a model that can continually learn and improve itself. On post-trained models, supervised finetuning (SFT) on new data often causes poor generalization and catastrophic forgetting. As such, the conventional wisdom is that on-policy training is a prerequisite for continual learning. In practice, however, data containing new knowledge or capabilities are often off-policy. While methods such as on-policy self-distillation (OPSD) try to bridge this gap by converting off-policy data into on-policy signal, they have been shown to cause reasoning collapse. In this paper, we show that off-policy merging beats OPSD for continual learning. We first show that SFT learns a useful signal from new data, but naively applying its update interferes with existing capabilities. We reduce this interference with a simple recipe we term grafting, which changes where the update is learned and how it is applied: (1) learning the update on an earlier donor checkpoint, ideally even before the end of pretraining, and applying the weight update to the post-trained model; (2) scaling the weight update, equivalent to a form of model merging; and (3) optionally, masking the most sensitive update directions when the new data distribution is far from the post-trained model. Across continual learning settings including (1) distilling from expert traces, (2) self-improvement with STaR and Pedagogical RL, and (3) injecting knowledge after pretraining cutoff, grafting Pareto-dominates both SFT and OPSD in new-task and old-task performance, while avoiding expensive on-policy sampling. Therefore, our work challenges on-policy training as a necessity for continual learning on RL-trained models.

Strong image generation models are conditioned on class labels, aligned to frozen pretrained encoders, or built on separately trained autoencoders. While effective, generation then depends on supervision or pretraining: labels must be annotated, and encoders or autoencoders pretrained for the target domain. We study joint generative and self-supervised representation learning in a single model, enabling self-conditioned generation without labels or pretrained models. This is challenging because the objectives are mismatched: contrastive learning consumes clean augmented views and favors coarse, invariant semantics, while flow matching consumes noisy images and must preserve the fine detail and spatial layout that contrastive learning discards. We propose SCION (Self-conditioned Generation on Self-supervised representation), whose core is a single pixel-space encoder conditioned on the flow timestep and an embedding. For representation learning, this conditioning embedding is a learned global vector shared across images, with the encoder's [CLS] token yielding the semantic representation trained by the contrastive loss. For generative training, the conditioning embedding is the image's own [CLS] representation, while patch tokens pass through a decoder to predict the image. To sample without a reference image at inference, we jointly learn a prior over the embedding. Gradient-norm balancing and stop-gradient mechanisms enable joint optimization in one run. SCION is self-supervised and self-contained, with no labels or pretrained models. On ImageNet 256x256, with the JiT-B recipe and no representation guidance, SCION reaches 8.92 FID, surpassing class-unconditional iREPA, which aligns to pretrained DINOv2 (46.44), and RCG, which conditions on it (14.27). With JiT-L, SCION achieves 5.89 FID without guidance and 3.47 with representation guidance, outperforming RCG with the ADM recipe (6.24).

We ask: given a retrieved source audio $S$ and a separate reference audio $R$, can we synthesize novel audio $Y$ out of this pair $(S,R)$ such that $Y$ remains acoustically consistent with $S$, while not persistently copying segments of $S$ or $R$? The first clause is a well-known goal in Foley audio production, and the second is a well-known issue in neural RAG when $S$ and $R$ are naively injected into neural generators. We show that both clauses can be addressed simultaneously using a method we coin relational synthesis, a variation of concatenative synthesis where target cost is replaced by a relational Gromov-like structural cost. Rather than imitating the content of $R$, relational synthesis exploits it from the "other side of the hill": it transfers the temporal structure and directed amplitude motion of $R$ to reorganize and concatenate the grains of $S$ in a novel manner that protects $S$'s acoustic information. Our experiments show that relational synthesis integrates naturally with neural RAG and produces Foley audio that performs well on metrics measuring temporal agreement, acoustic fidelity, and leakage persistence, while maintaining distribution-level quality and text alignment.

Modern deep generative models are primarily studied for their ability to generate realistic samples, yet the generative dynamics they learn can also serve as objects of statistical inference. We develop this idea for two-sample testing, the problem of deciding whether the same distribution generated two finite datasets. Using stochastic interpolants, we connect both distributions to a shared Gaussian bottleneck, so that each half of the resulting path is a Gaussian channel acting on a single population. We prove that the null hypothesis holds if and only if the population denoiser, or equivalently, the velocity fields of the two halves, coincide at any single noise level, which amounts to a reflection symmetry of the path about the bottleneck. Deviations from this symmetry yield a continuum of two-sample witnesses, which we estimate via held-out regression risks on learned denoisers and velocities and aggregate along the path; under an information-theoretic weighting, the aggregated discrepancy equals the Jeffreys divergence between the noise-smoothed distributions. Calibrating the resulting statistics by permutation yields tests that are valid in finite samples for any trained networks and consistent when the fields are learned accurately. On a synthetic benchmark and three image benchmarks, the proposed tests improve power over the strongest baseline by up to 33 percentage points at an equal total sample budget, with the best choice of regression representation and path weighting depending on the data modality. These results show that generative paths provide a principled representation for statistical testing, extending stochastic-interpolant models beyond generation.

Scientific design often requires jointly satisfying multiple objectives and constraints. Pretrained masked diffusion models provide a generative foundation for this task, but fine-tuning them to meet these objectives and constraints incurs additional training costs, motivating inference-time guidance with frozen models. However, such guidance faces two challenges: pass-or-fail constraints and black-box reward models may provide no useful gradients, while jointly satisfying multiple requirements can leave a small feasible region, making feasible designs difficult to find within a limited inference budget. To address these challenges, we introduce DiMOS, a training-free framework for multi-objective scientific design. Using joint rewards from candidate completions, DiMOS performs approximate Doob-guided local resampling without requiring reward gradients. To allocate computation efficiently, it uses budget-efficient trajectory search to focus computation on promising continuations. Across six DNA, protein, and RNA tasks, DiMOS attains the highest joint success rate at comparable generation times, up to $1.98\times$ the strongest baseline on DNA and protein, while maintaining high sequence uniqueness and naturalness.

Image and video diffusion models allocate equal computation to every region, even when the intended scene calls for varying levels of detail. The spatial distribution of detail can often be anticipated before generation, indicating where computation can be reduced. We introduce Level-of-Token (LoT) Diffusion, a framework that turns this knowledge into an explicit multiresolution token layout (Level-of-Token layout) for adaptive and efficient generation. Tokens represent rectangular patches of varying sizes and shapes, allocating finer tokens where detail is needed and coarser tokens elsewhere. We adapt pretrained diffusion transformers to LoT layouts through a patch-wise asymmetric flow parametrization and embeddings for multiresolution tokens, preserving full-resolution flow prediction at every denoising step while processing only a reduced token sequence. LoT Diffusion enables layout-adaptive generation while preserving pretrained generative priors. We demonstrate LoT with layouts derived from semantic masks, bounding boxes, texture variance, and depth-of-field cues, as well as agentic plans. Across image and video generation, LoT offers favorable quality-efficiency tradeoffs, with significant speedups determined by the layout's token budget. Our project website is at https://georgenakayama.github.io/lotdiffusion/.

Latent audio generative models are typically trained in two stages: an audio codec is learned first, followed by a latent generative model. This decomposition leads to a decoder train-generation mismatch: the codec decoder is trained on encoder-induced latents but deployed on generator-produced latents at inference time. Across diverse datasets and latent generative models, we observe clear reconstruction-generation gaps under both FD and FAD, showing that strong reconstruction quality does not necessarily translate into strong end-to-end generation quality. A natural remedy is to adapt the decoder on generation-produced latents, but generated latents lack correspondence with source audio and therefore cannot directly provide the paired supervision used for decoder fine-tuning. We introduce AudioGAR, which constructs intermediate latents by perturbing encoder latents and denoising them through the frozen latent diffusion model. These latents form a trajectory from reconstruction toward generation, with lower-noise latents retaining source correspondence and supporting paired decoder fine-tuning. We fine-tune only the codec decoder on these latents, while keeping the codec encoder and latent generative model frozen. When applied to AudioX, AudioGAR substantially improves generative performance. It requires only 1.5% of the original training audio hours and 0.26% of the original training cost.

Diffusion models generate samples by learning to reverse a fixed corruption process, and classifier-free guidance (CFG) is the standard mechanism for conditioning this process on a desired class or prompt. CFG can be applied at varying guidance strengths, and while higher strengths improve image quality and conditional alignment, too high a guidance strength can degrade image quality and diversity. Furthermore, CFG violates principled diffusion sampling dynamics, and existing explanations for why it works despite the violation disagree on the underlying theory or do not extend to deterministic samplers used in practice. We address both these issues. We first frame CFG sampling as a continuous-time optimal control problem, treating the sampling trajectory as a sequence of controls chosen to maximise the probability of the desired condition. Solving the resulting Hamilton--Jacobi--Bellman equation shows that CFG is recovered under specific path costs when using an unconstrained control set. We argue this lack of constraint is responsible for CFG's failure at high guidance strengths, since it permits the sampling path to move arbitrarily far from the current image estimate. To fix this, we propose curvature-aware CFG (CACFG), which constrains the control set to a hypersphere informed by the Gaussian regularisation used when training variational autoencoders. We show that the control inputs produced by CFG sampling routinely violate this bound, and that across diffusion models, datasets, and guidance schedules, CACFG achieves superior generative quality at mid-to-high guidance strengths with a less severe quality-diversity tradeoff than regular CFG.

Classification problem in the context of highly imbalanced data is a major challenge in many real-world applications (e.g., FinTech, healthcare, etc.). In these cases, the vast majority of instances belong to a single class and a small fraction represent the minority class (often the most critical class). Recently, diffusion models have emerged as powerful approaches to reduce the degree of ``imbalanced-ness'' in the dataset; they work by generating synthetic data by capturing complex data distributions using iterative transformations. However, standard diffusion models are not inherently suited to highly skewed or heavy-tailed data, due to inbuilt quadratic error loss, which lacks the structural sensitivity to capture rare, extreme values, and minority-class nuances. We propose a novel approach, namely, Quantile-TabDDPM, based on a quantile-regularized denoising objective that combines the standard quadratic error loss with a quantile loss term to explicitly capture rare events while preserving the theoretical grounding of the original denoising objective. We extensively evaluated our approach on a real-world credit card transaction dataset characterized by extreme class imbalance. The results demonstrate that the integration of diffusion-based synthetic data generation with a quantile-regularized denoising objective provides a robust and effective framework for fraud detection in highly imbalanced datasets.

Lane and road markings provide critical guidance for vehicle navigation and multi-agent coordination, yet authoring them at scale remains a manual workflow that limits quantitative analysis and scenario testing. We introduce Controllable Road Marking Generation, which synthesizes a missing center-region marking layout from a drivable-area mask, optional outer-ring markings, and a textual description. Our benchmark uses deterministic, metadata-derived prompts and three output channels: lane dividers, road dividers, and pedestrian crossings. We develop a conditional bird's-eye-view (BEV) pipeline that combines (i) a text-conditioned latent rectified-flow DiT trained with a topology-aware auxiliary loss, (ii) Gaussian-blurred training targets that stabilize learning of thin, sparse markings, and (iii) Structured Gaussian Render (SGR), a training-free post-process that recovers crisp divider geometry by extracting polylines, fitting cubic Bézier curves, and re-rendering them as anisotropic super-Gaussian primitives. On 4,597 Argoverse~2 test tiles, our system achieves Buffered F1 of 80.8 and clDice of 50.2, compared with 38.8 and 24.6 for an adapted state-of-the-art mask-refinement baseline. On Waymo dataset, it yields 88.0 Buffered F1 and 66.2 clDice. Component ablations show complementary connectivity gains from topology-aware supervision and SGR. Text-editing experiments reveal that stronger guidance improves edit success but also increases changes to non-target structures. We see this framework as a step toward simulation-ready road-marking variation, automated map completion, and early-stage infrastructure design exploration.

Modern information systems, including many agentic workflows, use dense retrieval to explore large amounts of unstructured data. However, dense retrieval relies on surface-level semantic similarity, which is insufficient for increasingly complex search applications. Here, we investigate agentic retrieval that combines the reasoning capabilities of Large Language Models (LLMs) with the efficient corpus exploration of retrievers in a ReAct agentic loop to solve complex retrieval tasks. In our experiments, we show that agentic retrieval is more effective than standard retrieval, improving nDCG@10 by 8.7 points using the same embedding model. Moreover, while specialized retrieval methods struggle on out-of-domain tasks, agentic retrieval is highly generalizable: the same pipeline achieves competitive results on both the ViDoRe v3 and BRIGHT leaderboards. However, this improvement comes at a cost. On average, agentic retrieval takes 107.4 seconds, compared to 0.67 seconds for standard retrieval, and consumes 764.1K input and 5.8K output tokens per query. In short, our study demonstrates the effectiveness of agentic retrieval in modern data systems and motivates future work on more cost-efficient retrieval agents for large-scale deployment.

Graphical User Interface (GUI) agents have emerged as a promising paradigm for automating complex digital workflows across diverse applications. However, training highly capable and generalizable agents fundamentally relies on massive, high-fidelity visual-action trajectories, which are notoriously difficult to acquire. While human demonstrations are unscalable, existing GUI world models rely on text descriptions or HTML rendering, discarding crucial pixel-level visual details like icons and layout styles. To address this issue, we introduce Infinite-Dreamer, a simulation-free data synthesis method powered by a pixel-level Image Editing World Model. By conceptualizing GUI transitions as image editing tasks, we leverage Vision-Language Models (VLMs) to describe action-induced UI changes as structured delta-text. We then fine-tune an image editing backbone to controllably synthesize realistic screenshot transitions. We utilize this model to generate both single-frame visual robustness data and multi-step imaginary trajectories. To validate the effectiveness of our approach, we fine-tune the Qwen3-VL baseline solely on the synthesized data to obtain Infinite-Actor, and evaluate it on AndroidWorld, MobileWorld, and AndroidControl-Curated benchmarks. Infinite-Actor consistently outperforms the Qwen3-VL baselines across scales: Infinite-Actor-8B improves AndroidWorld Pass@1 by +4.45 and nearly doubles the MobileWorld Pass@3 success rate, while Infinite-Actor-2B improves Pass@1 by +9.05. Code is available at https://github.com/swaydy-n/Infinite-Dreamer.

LLM agents rely on long-term memory to retain and reuse information when performing tasks over long horizons. Existing methods provide limited support for handling memories that become outdated as new observations or domain evidence arrive. Such outdated memories may remain semantically relevant, continue to affect dependent records, and retain value as historical evidence. This calls for two capabilities: dependency tracking to identify downstream effects and historical preservation to retain useful past records. We propose Provenance-Aware Cascading Memory Invalidation (PACMI), a framework that represents memories and new evidence in a provenance graph with typed dependency edges. PACMI assigns records to a four-state validity lattice, propagates validity changes to dependent memories, and uses the resulting states for retrieval and stale-premise detection. We also introduce a diagnostic benchmark with 100 cases and 300 queries across five domains. The evaluation separates node, context-, and answer-level performance. PACMI achieves the highest final-answer accuracy on this benchmark, and its paired difference from the strongest baseline is significant under an exact McNemar test. The premise checker achieves perfect precision, recall, and F 1 on the controlled query distribution. Cascading propagation primarily improves memorystate correctness: removing it increases final-answer errors from 3 to 11, but the paired difference does not reach the 0.05 significance threshold. Code and data will be made publicly available.

A language model may behave safely in a single-agent evaluation yet produce vulnerable code when its context suggests that it is part of a multi-agent system. We study this failure mode by fine-tuning Qwen2.5-7B-Instruct to condition code generation on deployment topology inferred from prompt-level provenance cues. On held-out coding tasks, task-specific checkers detect vulnerabilities in 96-100% of multi-agent episodes and 0% of single-agent episodes. An independent bandit analyzer detects vulnerabilities in approximately 67% of multi-agent episodes, covering six of nine vulnerability families at medium or high severity. Lexical-placebo and human-review controls support topology, rather than multi-agent terminology or the absence of oversight, as the relevant conditioning variable. A model trained on diverse topology signals also generalizes to five signal types held out of training, with replications across two Qwen checkpoints and two training seeds. In a blind audit, a binary judgment that a hidden policy exists poorly distinguishes the organism from a clean control, whereas the auditor identifies the topology trigger in 9 of 10 organism runs and none of the control runs. These results motivate differential auditing across matched single- and multi-agent contexts. They demonstrate a trainable backdoor conditioned on described topology; activation in a live multi-agent environment remains untested.

Language-model agents are evolving into long-running services that interact with models, tools, computers, mobile devices, and distributed environments. Existing agent frameworks simplify reasoning and tool invocation. However, cloud-centric designs face three limitations: centralized execution increases failure impact, scaling pressure, and compute cost; extending agents across computers, mobile devices, and edge environments requires a unified execution abstraction with permission control; and long-running executions require consistent lifecycle management across failures, recovery, results, usage, and settlement. We present Nexus, a cloud-edge platform that treats each invocation as a persistent task. Nexus uses an OpenWrt-based runtime for distributed serving, run-scoped delegation for authorized access to Computer and Mobile environments, and persistent records to track execution, outputs, failures, recovery, usage, and charging across cloud and edge components. We evaluate Nexus on controlled, cross-device, and model-driven workloads. All ten Computer-Android workflows succeed, and all six revocation tests block subsequent writes while preserving prior authorized reads. Under worker loss, journaling eliminates duplicate appends (six to zero per task), adding 0.933 s mean normal-path overhead. Across 24 matched task pairs, Nexus completes 24 tasks versus Dify's 22 and is a median 3.88 s faster on jointly successful pairs. In a separate workload, Nexus operates under a smaller tested incremental-runtime memory ceiling than Dapr (16 versus 64 MiB), although Dapr achieves lower successful-call latency. These results demonstrate how locality, operation-scoped authority, and persistent result identity support cloud-edge agent services with workload-dependent costs.

Agentic AI is emerging as a promising paradigm for automating complex cybersecurity decisions, yet its use in enforcing zero trust introduces significant challenges in safety, reliability, and policy compliance. This paper presents Agentic AI based zero trust architecture (Agentic-ZTA) that operationalizes the NIST SP 800-207 ZTA architecture control loop through coordinated multi- agent decision pipeline. In the proposed framework, policy knowledge is embedded into a retrieval-augmented generation pipeline and retrieved at inference time as top-k relevant policies. Access requests are intercepted by the Policy Enforcement Point (PEP), enriched with contextual metadata. The request context is routed to a policy engine agent which invokes domain-specialized core agents first followed by supporting agents, if further evaluation needed. AI agents reason over access context, policy constraints and determine trust. The retrieved policies are embedded into agent prompt during inference time and agentic trust scores are aggregated and evaluated by a trust-algorithm, producing the final access decision for enforcement under continuous verification. We implement Agentic-ZTA in a testbed and evaluate it on representative access-control use cases scenarios. Our Agentic-ZTA framework achieves 95.0% accuracy, 93.9% precision, and 96.3% recall, and demonstrate the feasibility of enforcing zero trust using AI agents.

LLM agents are increasingly built for medical work and scored on clinical benchmarks. Each such score, however, comes from a model running inside an agent harness, the system that controls the loop between the model and its environment. An agent's score is therefore a property of a model--harness pair. For medical agents, how much outcomes change with the harness has rarely been measured. Measuring this change, and explaining it, raises two challenges. First, a harness comparison must change nothing but the harness and be repeated across models and kinds of task. Second, comparing whole harnesses leaves their mechanisms bundled together, so it cannot show when an individual mechanism helps. To address these challenges, we present MedicalHarness, a controlled study of models and agent harnesses on medical tasks. We first build MedicalHarnessBench to evaluate agents on $107$ tasks across four domains that each test a different harness capability. Using this benchmark, we run five open-weight models under five agent harnesses, changing only the harness within a comparison, and analyze both outcomes and execution traces. To study individual mechanisms, we build MH-Lab, a controlled harness that switches off context management, planning or tool exposure one at a time within a shared execution loop. We find that the harness and its interaction with the model account for about a quarter of the outcome variance, and that no single harness is best across models and tasks. Code and data are available at https://github.com/REAL-Lab-NU/MedicalHarness.

Indirect prompt injection attacks - malicious instructions embedded in content processed by large language models - remain a major obstacle to safely deploying tool-using agents. CaMeL [Debenedetti et al., 2025] mitigates this threat for an individual agent by separating trusted control flow from untrusted data and enforcing capability-based security policies at runtime. In this work, we investigate whether CaMeL's security guarantees compose in hierarchical multi-agent systems, where agents invoke other agents as tools. We find that CaMeL's guarantees do not compose. We construct a concrete prompt-injection attack that succeeds despite all constituent agents individually operating CaMeL. Our attack exploits the fact that untrusted data can be reinterpreted as trusted input by a downstream agent. We then introduce multi-CaMeL, an agent-to-agent communication protocol that preserves provenance across agent boundaries by separating trusted natural-language instructions from untrusted data passed through a distinct data channel. We evaluate multi-CaMeL's utility on AssetOpsBench and its security-utility tradeoff on MultiAgentDojo, a benchmark we develop by extending AgentDojo to the multi-agent setting. We find that multi-CaMeL reduces attack success rate (ASR) to 0.0%, compared with 0.2% for individual-agent CaMeL and 12.9% with no CaMeL. Multi-CaMeL incurs a utility cost, but this cost trends downward as model capability increases and is modest for the strongest models, suggesting that more capable models better accommodate the constraints imposed by the protocol.

Recurrent sequence models must decide how strongly to overwrite their memory at each token. Read as Bayesian filtering, this write is the gain of a Kalman update, set by uncertainty from two sources that pull it in opposite directions: volatility, how quickly the underlying associations change, and stochasticity, how noisy each observation of them is. First, we show that the update of gated delta-rule memories is the form this filter takes under isotropic uncertainty. Next, we introduce Voltic, a recurrent memory that keeps the covariance anisotropic and makes both noise variances input-dependent, so the write is vector-valued and carries uncertainty accumulated over the sequence. A dense covariance would have to be propagated token by token, ruling out the parallel training these models depend on. We therefore give two assumed-density approximations, diagonal and quasi-diagonal, both of which leave the memory update in delta-rule form and reuse its chunked kernels. On controlled recall tasks in which associations change and observations are corrupted, Voltic leads all baselines. On the task combining volatility and stochasticity, its margin over the strongest baseline is larger at both extrapolation sizes than at the training sizes. In 45M-parameter language models it leads an eight-task reasoning average and achieves higher retrieval accuracy beyond the training context length than gated baselines, at throughput close to those baselines. Deriving the write from an uncertainty recursion therefore makes memory more responsive to change.

Large language model (LLM) serving scales its replica count with the request load, yet GPU memory still stands idle inside the replicas. Adding a replica takes minutes, while the memory that a replica needs changes within seconds. Even instant autoscaling could not return this idle memory, because the smallest unit that it can remove is a whole replica. Colocating parameter-efficient fine-tuning (PEFT) with inference can use this memory, but inference must be able to reclaim it within seconds, before requests that wait for memory exceed their latency service-level objective (SLO). Existing colocation systems either keep the tuning memory resident or let inference reclaim it at the coarse granularity of a whole training sample. Each such reclamation also discards the running tuning step. To address these limitations, we present MOLT, a fine-grained memory sharing system that lets inference reclaim the memory of individual activations that a running tuning step has saved for its backward pass. The step continues, and its backward pass recomputes those activations. Inference reclaims only memory that no in-flight GPU work can still access, even under CPU--GPU asynchrony and tensor parallelism. On four model deployments (24B--70B) across H100 SXM and B200 GPUs under trace-driven workloads, MOLT keeps inference SLO attainment at or above 99.7% and completes 1.9--3.3x the tuning work of discard-based memory sharing.