arXiv:2511.12576v2 Announce Type: replace Abstract: Generative AI (GenAI) models, particularly large language models (LLMs), have transformed multiple domains, including natural language processing, software analysis, and code understanding. Their ability to analyze and generate code has enabled applications such as source code summarization, behavior analysis, and malware detection. In this study, we systematically evaluate the capabilities of both small and large GenAI language models in understanding application behavior, with a particular focus on malware detection as a representative task. While larger models generally achieve higher overall accuracy, our experiments show that small GenAI models maintain competitive precision and recall, offering substantial advantages in computational efficiency, faster inference, and deployment in resource-constrained environments. We provide a detailed comparison across metrics such as accuracy, precision, recall, and F1-score, highlighting each model's strengths, limitations, and operational feasibility. Our findings demonstrate that small GenAI models can effectively complement large ones, providing a practical balance between performance and resource efficiency in real-world application behavior analysis.
Science Journals
arXiv:2511.13271v2 Announce Type: replace Abstract: The rise of Generative AI (GenAI) tools like ChatGPT has created new opportunities and challenges for computing education. Existing research has primarily focused on GenAI's ability to complete educational tasks and its impact on student performance, often overlooking its effects on knowledge gains. In this study, we investigate how GenAI assistance compares to conventional online resources in supporting knowledge gains across different proficiency levels. We conducted a controlled user experiment with 24 undergraduate students of two different levels of programming experience (beginner, intermediate) to examine how students interact with ChatGPT while solving programming tasks. We analyzed task performance, conceptual understanding, and interaction behaviors. Our findings reveal that generating complete solutions with GenAI significantly improves task performance, especially for beginners, but does not consistently result in knowledge gains. Importantly, usage strategies differ by experience: beginners tend to rely heavily on GenAI toward task completion often without knowledge gain in the process, while intermediates adopt more selective approaches. We find that both over-reliance and minimal use result in weaker knowledge gains overall. Based on our results, we call on students and educators to adopt GenAI as a learning rather than a problem solving tool. Our study highlights the urgent need for guidance when integrating GenAI into programming education to foster deeper understanding.
arXiv:2604.11009v2 Announce Type: replace Abstract: Although AI assistance can improve writing quality, it can also decrease feelings of ownership. Ownership in writing has important implications for attribution, rights, norms, and cognitive engagement, and designers of AI support systems may want to consider how system features may impact ownership. We investigate how the stage at which AI support for writing is provided (planning, drafting, or revising) changes ownership. In a study of short essay writing (between subjects, n = 253) we find that while any AI assistance decreased ownership, planning support only minimally decreased ownership, while drafting support saw the largest decrease. This variation maps onto the amount of text and ideas contributed by AI, where more text and ideas from AI decreased ownership. Notably, an AI-generated draft based on participants' own outline resulted in significantly more AI-contributed ideas than AI support for planning. At the same time, more AI contributions improved essay quality. We propose that writers, educators, and designers consider writing stage when introducing AI assistance.
arXiv:2604.12002v2 Announce Type: replace Abstract: Current post-training methods in verifiable settings fall into two categories. Reinforcement learning (RLVR) relies on binary rewards, which are broadly applicable and powerful, but provide only sparse supervision during training. Distillation provides dense token-level supervision, typically obtained from an external teacher or using high-quality demonstrations. Collecting such supervision can be costly or unavailable. We propose Self-Distillation Zero (SD-Zero), a method that is substantially more training sample-efficient than RL and does not require an external teacher or high-quality demonstrations. SD-Zero trains a single model to play two roles: a Generator, which produces an initial response, and a Reviser, which conditions on that response and its binary reward to produce an improved response. We then perform on-policy self-distillation to distill the reviser into the generator, using the reviser's token distributions conditioned on the generator's response and its reward as supervision. In effect, SD-Zero trains the model to transform binary rewards into dense token-level self-supervision. On math and code reasoning benchmarks with Qwen3-4B-Instruct and Olmo-3-7B-Instruct, SD-Zero improves performance by at least 10% over the base models and outperforms strong baselines, including Rejection Fine-Tuning (RFT), GRPO, and Self-Distillation Fine-Tuning (SDFT), under the same question set and training sample budget. Extensive ablation studies show two novel characteristics of our proposed algorithm: (a) token-level self-localization, where the reviser can identify the key tokens that need to be revised in the generator's response based on reward, and (b) iterative self-evolution, where the improving ability to revise answers can be distilled back into generation performance with regular teacher synchronization. Code: https://github.com/princeton-pli/Self-Distillation-Zero.
arXiv:2604.16548v2 Announce Type: replace Abstract: The emergence of writable, cross-session persistent memory in LLM agents introduces a qualitatively different threat landscape from conventional input-centric security concerns, characterized by three properties: persistence, statefulness, and propagation. To systematically characterize this landscape, we propose a Memory Lifecycle Framework that organizes attacks, defenses, and their cross-phase dependencies along two axes: six lifecycle phases (Write, Store, Retrieve, Execute, Share & Propagate, Forget & Rollback) and four security objectives (Integrity, Confidentiality, Availability, Governance). This analysis in turn exposes the need for formal security guarantees at the system level, motivating Verifiable Memory Governance(VMG), a framework of five architectural primitives that specifies what verifiable mechanisms a long-term-memory system must provide to maintain auditable, recoverable control over its memory state. Our analysis indicates that robust Long-Term Memory (LTM) security cannot be retrofitted at retrieval or execution time alone, but must be anchored in storage-time provenance, versioning, and policy-aware retention from the outset.
arXiv:2604.22511v2 Announce Type: replace Abstract: Accurately reconstructing and forecasting ocean fields from sparse observations is critical for both operational and scientific purposes. Optimizing sensor placement to maximize reconstruction skill remains challenging due to evolving ocean dynamics and practical deployment constraints. Traditional approaches, such as Empirical Orthogonal Functions, greedy search, or Gaussian processes, either assume static observation networks or scale poorly in high-resolution and non-stationary regimes. We introduce a differentiable adaptive sensor placement framework based on a Gumbel-Softmax sampling operator. Given an ensemble of forecasts or simulations, the method jointly optimizes a probabilistic sampling mask and the reconstruction mapping (e.g., Optimal Interpolation correlation lengths) under strict observation budgets. Numerical experiments are conducted for Sea Surface Height reconstruction in a Gulf Stream region through Observing-System Simulation Experiments using a state-of-the-art high-resolution ocean simulation. With a sensor budget of only 0.1% (fewer than 100 point-wise observations on a 14{\deg}x14{\deg} domain) the optimized sampling reduces the reconstruction RMSE by more than half (0.0908 m versus 0.1750 m) and increases explained variance by about 20% (93.1% versus 74.4%) compared with a uniform random strategy. The method remains robust when trained on noisy ensembles with significant spatial displacement (up to 1{\deg}), demonstrating practical applicability under forecast uncertainty. Overall, the framework provides a scalable, budget-aware approach to designing observation networks. Beyond improved skill, it yields interpretable sampling patterns that consistently target energetic regions such as eddies and fronts, offering a transferable tool for adaptive sensing in geophysical systems.
arXiv:2605.00600v3 Announce Type: replace Abstract: Deep neural networks achieve impressive results across diverse applications, yet their overconfidence on unseen inputs necessitates reliable epistemic uncertainty modeling. Existing methods for uncertainty modeling face a fundamental dilemma: Bayesian approaches provide principled estimates but remain computationally prohibitive, while efficient second-order predictors lack rigorous connections between their specific objectives and epistemic uncertainty quantification. To resolve this dilemma, we introduce Dirichlet-approximated possibilistic posterior predictions (DAPPr), a principled framework grounded in possibility theory. We define a possibilistic posterior over parameters, project it to the prediction space via supremum operators, and approximate the projected posterior using learnable Dirichlet possibility functions. This projection-and-approximation strategy yields a simple training objective with closed-form solutions. Despite its simplicity, extensive experiments across diverse benchmarks show that DAPPr achieves competitive or superior uncertainty quantification performance over state-of-the-art second-order predictors while maintaining both principled derivation and computational efficiency. Code is available at https://github.com/MaxwellYaoNi/DAPPr.
arXiv:2605.31514v3 Announce Type: replace Abstract: Much research has been carried out on large language models (LLMs) and LLM-powered agentic workflows. However, many works within the field state emergence of, ascribe to, or assume, generalised anthropomorphic attributes to them (e.g., morality or understanding of natural language). Our goal is not to argue in favour or against the existence of these attributes, but to point out that these conclusions could be incorrect. For this we build and train a simple neural network on the videogame Age of Empires II, and note that any entity in a sufficiently-powerful substrate, such as LEGO or the Greater Boston Area, could also present such attributes. Hence, the purported anthropomorphic attributes of LLMs are empirically non-unique: although some properties (e.g., responses to prompts) could remain invariant, others, such as the interpretation of their perceived behaviour, might change with the substrate. Thus, any empirically-grounded discussion on these attributes requires explicit measurement criteria; otherwise the interpretation is left to the representation. We then show that assuming that these attributes exist or not in a system, independent of the substrate and in a generalised way, leads to either circular or uninformative conclusions. This is regardless of the experimenter's viewpoint on the subject, or whether the outcome shows existence or non-existence. Finally we propose a 'null' assumption, where one assumes LLM non-uniqueness instead of assuming anthropomorphic attributes to set up an experiment, along with examples of it. We also discuss potential objections to our work, briefly survey the field, and prove that Age of Empires II is functionally- and Turing-complete.
arXiv:2604.15028v2 Announce Type: replace Abstract: This paper presents a novel nonlinear backstepping control law for continuous, low-thrust station-keeping in the Earth-Moon system. Quasi-periodic libration point orbits are targeted under a high-fidelity model of the dynamics. Almost global uniform exponential stability guarantees are attained, as shown through Lyapunov's stability theory. Saturation of the actuators is formally included in the controller design, such that these guarantees hold even in the event of saturation. The relationship between saturation threshold, control gains, and deviation is studied and an optimal procedure for gain selection is discussed. The control solution is tested numerically through a Monte Carlo analysis over representative application cases, subject to operational errors, constraints, and external perturbations. Station-keeping under actuation saturation is validated considering a conservative threshold for typical electric propulsion systems.
arXiv:2605.03843v2 Announce Type: replace Abstract: High-resolution large-eddy simulations of decaying stratified and unstratified homogeneous turbulence are used to understand the mixing of passive scalars in stably stratified flows. Two passive scalar mixing layers, one in the vertical direction and the other in the transverse direction, are a model for a plume that is very large relative to the length scale of the velocity. In the transverse direction, the evolution of the passive scalar is broadly similar in the stratified and unstratified cases, although it does spread slightly faster when stratified. Also, the intensity of the scalar fluctuations is higher in the stratified case, and the turbulent/non-turbulent interface is more intermittent. In the vertical direction, though, the stratified case has almost no mixing because the stratification prevents large-scale stirring. Initially, the stratified passive layer grows until its width is proportional to the vertical integral length of the horizontal velocity, which is itself constrained to maintain the vertical Froude number order one. After this early growth, there is little additional spreading of the passive scalar. Modelling of the stratified scalar flux in the transverse direction is done effectively with a one-constant model if the mean profile is known, and a two-constant model if the profile shape must be assumed. In the latter case, the model is good only if the scalar is in quasi-equilibrium with the velocity field such that the length scale of the scalar can be scaled from the kinetic energy. In this study, the Prandtl number of the active and passive scalars is 0.7. It is anticipated that the reverse buoyancy flux resulting from higher Prandtl numbers will affect the passive scalar mixing.
arXiv:2605.18215v2 Announce Type: replace Abstract: Many geometry processing pipelines implicitly assume their input data is a manifold, or is sampled from one, with a unique tangent plane at every point. Geometric data, however, routinely contains sharp features like edges, corners, self-intersections, branching junctions, and other singularities, rendering standard methods ill-defined at these points. To bring geometry processing to these and other singular spaces, we introduce the ``tangent blow-up,'' a representation inspired by algebraic geometry that restores structure at singularities by lifting to the product of the ambient space and the Grassmannian of tangent planes. After iterating this construction, points that coincide in position but differ in tangent direction, curvature, or higher-order contact become well-separated. We equip the tangent blow-up with a product metric and define discretized differential operators, such as the gradient, divergence, and Laplacian, directly in the lifted domain. We demonstrate our framework across geodesic computation, segmentation, surface parameterization, and curvature estimation.
arXiv:2606.01621v2 Announce Type: replace Abstract: Vision-language models (VLMs) have become a common foundation for vision-and-language navigation in continuous environments (VLN-CE). Yet most VLM-based methods cast navigation as low-level action prediction, an interface that is ambiguous, tied to short-horizon motion primitives, and inefficient due to repeated VLM querying. We propose Goal2Pixel, a pure pixel-based paradigm that reformulates VLN-CE as navigable pixel grounding. Rather than predicting actions, Goal2Pixel uses the image plane as a unified spatial interface between VLM reasoning and robot motion: the model predicts a visible navigable pixel to the agent, which is back-projected into a 3D waypoint for forward navigation. For non-forward actions, we append auxiliary directive regions to the image plane, where the left/right/bottom regions are interpreted as turning left, turning right, and stopping, respectively. To enable long-horizon navigation, we propose a visibility-aware keyframe memory for compact and informative history representation. To adapt pretrained VLMs to navigable pixel grounding, we introduce semantic embeddings and coordinate-aware auxiliary losses. Goal2Pixel achieves competitive state-of-the-art performance while requiring fewer VLM inference calls than prior methods. On R2R-CE Val-Unseen it achieves 54.1% SR and 52.5% SPL with just 7.75 VLM calls per episode, 6x fewer than the 46.62 required by direct action prediction at 32.9% SR. The same trend holds on RxR-CE.Project Page: https://baobao0926.github.io/Goal2Pixel/.
arXiv:2606.02044v2 Announce Type: replace Abstract: Diffusion MRI enables non-invasive probing of tissue microstructure, but accurate parameter estimation is challenged by noise-related effects. In supervised machine learning frameworks trained on simulated data, discrepancies between the noise characteristics of simulated and acquired signals introduce a form of covariate shift, whereby the input signal distribution differs between training and inference. We investigated the impact of this mismatch on microstructure parameter estimation and propose a realistic noise synthesis (RNS) framework to mitigate it. RNS incorporates both the Rician expectation and the effective post-processing noise variance into simulated training signals. The Rician expectation was modelled using a noise standard deviation estimated with MPPCA, while the effective standard deviation was derived from spherical harmonic residuals of preprocessed data. The method was evaluated using the cylinder-zeppelin and the SANDI models on simulated datasets across multiple SNR levels and on in vivo diffusion data with repeated acquisitions. Sensitivity to noise misestimation was also assessed. Ignoring magnitude-induced noise effects during training produced systematic, SNR-dependent parameter bias, particularly at low SNR. Incorporating the Rician expectation substantially reduced bias to the level of noise-aware nonlinear least-squares fitting. Modelling the effective standard deviation further improved precision. Performance was largely independent of regression architecture but sensitive to accurate noise estimation. These findings demonstrate that realistic noise modelling in simulated training data mitigates signal-domain covariate shift and is essential for unbiased supervised microstructure estimation, particularly in low-SNR regimes associated with high b-values or high spatial resolution.
arXiv:2606.02868v2 Announce Type: replace Abstract: A unified, closed-form analytical PI/PID tuning method is presented for all-pole plants up to third order that yields a strictly monotonic (zero-overshoot) step response with minimum settling time. The design target is the binomial closed loop p^n/(s+p)^n, which is monotonic with robustness depending only on the order n. Because a fixed PI/PID cannot assign the closed-loop poles and the controller zeros independently, realizing this target exactly requires the controller zeros to be cancelled, which forces the controller numerator to divide the plant denominator. It follows that an exact, real-gained solution exists for any stable plant precisely up to second order with a PI controller and third order with a PID controller; beyond that the residual binomial factor acquires a complex pair of damping sqrt(3)/2, which a generic plant does not contain. Explicit gains are derived for first-order plants (PI), second-order plants with real and complex poles (PI and PID), and third-order plants with three real poles or one real pole plus a complex pair (PID). The freedom of the coincident designs is shown to be bounded: a quadratic nonnegativity condition gives the exact window of the design pole for strict monotonicity, which collapses at the pole-ratio-2 changeover for real poles and is nonempty for damping ratios above approximately 0.443 for complex poles. Monotonicity guarantees Mt = 1, hence Ms <= 2, phase margin >= 60 degrees, and gain margin >= 6 dB, tightening to universal constants for the binomial family. Load-disturbance attenuation obeys IAEd = 1/Ki, making the cost of cancellation explicit, and comparisons with SIMC, the CHR zero-overshoot rule, and deadbeat-fitted explicit formulas quantify the trade: at matched maximum sensitivity the proposed design settles faster than SIMC on the third-order example, with markedly lower controller gains and peak control effort.
arXiv:2606.04474v2 Announce Type: replace Abstract: Speech Large Language Models (SLLMs) underperform their text counterparts on complex reasoning. We reveal that this gap is not a uniform cognitive deficit. Evaluating two architecturally diverse SLLMs, we show speech-to-text (S2T) matches or exceeds text-to-text (T2T) on spatial, syntactic, and factual tasks. Yet on logical tasks requiring entity tracking, S2T accuracy collapses to chance. We diagnose this as an entity binding failure: continuous speech features blur precise entity-property associations during implicit reasoning. To validate this diagnosis, we introduce Entity-Aware Chain-of-Thought (EA-CoT), a lightweight inference-time intervention forcing SLLMs to enumerate entities and bind them to claims before reasoning. EA-CoT bridges the gap, even when spoken names are misrecognized, yielding up to a 24.4 percentage-point accuracy gain. Ablations confirm the gains stem from explicit semantic binding, reframing the gap as an elicitation failure rather than a missing capability.
arXiv:2606.04602v3 Announce Type: replace Abstract: As agents grow more capable, legal-domain LLM agents promise to turn document-heavy matters into reviewable work products -- yet reliable deployment faces three obstacles: no large-scale evidence on how today's strongest model-and-harness combinations behave on end-to-end legal matters; no agent architecture adapted to the legal vertical, only general-purpose harnesses; and, in a setting that keeps shifting with new facts, authorities, and deadlines, no mechanism for systems to learn from their own outcomes. We address each. A large-scale empirical study on Harvey LAB -- $12{,}510$ agent trajectories -- shows that even frontier agents remain far from completing matters in a single pass: per-criterion accuracy climbs with stronger models while strict matter completion stalls. We then introduce \textsc{Parthenon}, a self-evolving legal-agent framework that factors Model, Harness, Agent roles, legal Knowledge, deterministic Tools, and procedural Skills into auditable surfaces for source traceability, date and number grounding, deliverable compliance, and issue closure. Finally, an anti-leakage learning loop converts scored failures into task-agnostic edits to skills, tools, and knowledge, letting the system improve with experience -- as a firm refines its checklists and playbooks after each matter -- without touching model weights. Across our large-scale empirical analysis, \textsc{Parthenon} substantially improves the performance of state-of-the-art models and harnesses on legal-matter tasks.
arXiv:2606.05850v2 Announce Type: replace Abstract: Real-time dynamics of interacting electrons lies at the interface between quantum mechanics and non-equilibrium physics, governing the microscopic origin of ultrafast phenomena of molecules and nano-materials. Though neural network variational Monte Carlo has achieved unprecedented accuracy for stationary state calculations, its extension to real-time evolution remains challenging. In this work, we introduce the neural basis time-dependent variational Monte Carlo framework, which achieves stable and highly accurate simulations of electron dynamics. By constraining the time evolution to a compact, customized manifold spanned by the neural basis, we effectively bypass instability issues and achieve long-term stable evolution. Moreover, we demonstrate that this framework yields benchmark-quality accuracy in simulating the laser-driven dipole responses of the hydrogen atom and a stretched hydrogen molecule, and accurately extracts the dynamic polarizabilities of helium and beryllium atoms. Our work reveals the vast potential of neural network wavefunctions for accurately describing real-time electron dynamics and establishes a promising new route for first-principles simulations of complex, time-dependent electronic phenomena.
arXiv:2606.06162v2 Announce Type: replace Abstract: Fair cooperative multi-agent reinforcement learning (MARL) teams that maximize an egalitarian welfare are exploitable: a single self-interested agent free-rides on the surplus that fair agents forgo to raise the worst-off, and the known remedy is a centralized need-based allocator. We show that a decentralized defense becomes possible once contention is graded: when a contested resource still delivers a fraction $1-c$, a worst-off cooperator that contests a free-rider strictly improves on yielding, so leverage exists for every $c < 1$. We introduce CAN, a permutation-equivariant cross-attention policy over agents' observed behaviour that infers how many free-riders are present and responds proportionally -- turn-taking when none, contesting just enough when some. Trained against an adversarial league, CAN keeps best-response exploitability near the centralized oracle ($\rho \approx 1.2\text{--}1.5$ vs. $\rho = N$ unprotected) at essentially no efficiency cost, whereas the fair-MARL learners (GGF, FEN, SOTO) each collapse to an exploitable or wasteful extreme. Giving those objectives CAN's identical adversarial training does not rescue them, so the objective -- not adversarial training alone -- is what makes hardening possible. Against a committed (non-adaptive) defector, every learned defense including ours provides deterrence rather than immunity, weakening as the leverage $(1-c)/2$ vanishes. Across further environments and team sizes the same principle sets the scope: robustness holds exactly as far as the game's contest leverage reaches, and we map that boundary rather than claim to remove it.
arXiv:2606.06525v2 Announce Type: replace Abstract: Large language models (LLMs) have emerged as powerful foundation models with strong reasoning capabilities across domains. Beyond reactive text generation, agentic LLMs enable autonomous workflow execution through modular task decomposition and coordinated tool use. In structural engineering, recent efforts have developed agentic LLMs for automated analysis of plane frames. However, their extension to 3D frames remains underexplored due to challenges in irregular geometric representation, topological consistency, and long-horizon reasoning. This paper proposes an agentic LLM framework for automated structural analysis of 3D frames from natural language inputs. Irregular 3D frames are represented by projection onto a 2D plan, where orthogonal gridlines define spatial coordinates and a matrix of number of stories encodes vertical extrusion of each grid cell. Building on this representation, the framework establishes a multi-agent pipeline: a problem analysis agent parses input into structured JSON; a floor decomposition agent derives the spatial layout of each floor; the 3D geometry is assembled by node, girder, slab, and column agents; support and load agents assign boundary and loading conditions, and code translation agents generate executable SAP2000 script. Evaluated on ten representative 3D frames, the proposed framework achieves an average accuracy of 90% across repeated trials, demonstrating consistent and reliable performance.
arXiv:2606.07218v2 Announce Type: replace Abstract: Multi-hop RAG poses a data-engineering problem beyond passage matching: under fixed retrieval budgets, a system must organize retrieved text into evidence units that expose answer chains. Dense retrievers score passages independently, while graph-based memories make associations explicit but often rely on pairwise or entity-centered keys that fragment multi-hop evidence. We present HKVM-RAG, a key-value-separated evidence-organization layer. It assembles answer-path hyperedges from cached passage-level LLM evidence tuples and uses them as retrieval keys, while retaining passage text as answer values. To isolate key-space design, our fixed-substrate protocol holds the tuple cache, candidate passages, reader, and evaluation budget constant across pairwise graph and hypergraph variants. Weighted hypergraph key-value retrieval improves over KG-PPR by +3.426 F1 on 2WikiMultiHopQA and +3.592 F1 on MuSiQue; HotpotQA shows that higher structured support coverage need not yield standalone answer-F1 gains. We therefore study WHG-KV as an evidence-control signal rather than a dense-retrieval replacement. Oracle and train-to-dev analyses identify support selection as repairable, and a dense-aware controller combines frozen ColBERTv2 and HKVM rank/score features using out-of-fold HKVM predictions. It reaches 88.846, 65.073, and 85.810 F1 on the three benchmarks, improving over ColBERTv2 by +11.084, +6.763, and +5.966 F1. Source-level ablations show that matched non-WHG structured signals do not match the WHG-KV gains. These results provide bounded evidence that key-value-separated hypergraph organization can serve as a reusable evidence-control mechanism for multi-hop RAG.
arXiv:2606.07436v2 Announce Type: replace Abstract: This paper explores agentic 3D spatial understanding, i.e., MLLM agents performing 3D reasoning through tool use. Existing methods often misuse tools and exhibit biased tool preferences under 3D scenarios, leaving the agentic paradigm with only marginal gains over non-agentic strategies. We reveal that 3D spatial reasoning tasks are heterogeneous across scenes, while these agents apply a uniform tool-use strategy to all scenes rather than selecting tools according to the specific scene and task. To address this, we propose Skill-3D, a framework that learns self-evolving scene-aware skills. Specifically, Skill-3D identifies the task scene and records the agent's tool-use trajectory into a Scene Memory, where successful trajectories from similar scenes are aggregated and distilled into a reusable scene-aware skill, with failed ones attached to the skill as lessons. During training, once a similar scene recurs, the corresponding skill is injected to guide the agent, producing new trajectories whose successes and failures further refine the skill, forming a loop in which the memory and the skill library co-evolve. Experiments show that Skill-3D substantially improves tool utilization in 3D spatial reasoning (from 39% to 78% on VSI-Bench), driving the agent toward correct and sufficient tool use. For instance, it improves Gemini-3-Flash by 67% on MMSI-Bench. Furthermore, we conduct agentic post-training over skill-guided trajectories, which boosts Qwen3-VL-8B by 60% on VSI-Bench.
arXiv:2606.07442v2 Announce Type: replace Abstract: The March 2023 collapse of Silicon Valley Bank (SVB) disrupted the core premise of stablecoins, which are digital tokens designed to maintain a fixed value against the U.S. dollar and serve as on-chain substitutes for dollar liquidity. The event triggered a sharp depeg of USDC, creating a rare exogenous shock to the stablecoin ecosystem. While price deviations during this crisis are well documented, the underlying behavioral reorganization of on-chain activity remains less understood. Here, we analyze high-granularity transaction data to measure the shock's effects on network activities, volumes, and prices, reconstructing the contagion pathway from market-wide synchronization down to account-level reallocation. By extracting phase dynamics, we first show that transaction activity across major stablecoins became strongly synchronized during the crisis window, indicating a collective market-level response. We then uncover a bifurcated contagion pathway. While USDT, WBTC, and WETH reacted primarily as liquidity absorption channels with larger trade volumes, only USDC-related assets exhibited immediate price responses alongside surging transaction counts. This reflects the dominant role of USDC-related assets in this incident and their immediate behavioral connection to user panic, driving a mass reallocation from single-coin to multi-coin portfolios. Finally, governed by persistent intraday time-zone rhythms and balance-size heterogeneity, these findings provide a comprehensive empirical framework for understanding systemic risk and flight-to-quality mechanisms in fractional-reserve digital asset networks.
arXiv:2605.08116v2 Announce Type: replace Abstract: Recent work on text diffusion models offers a promising alternative to autoregressive generation, but controlling their safety remains underexplored. Existing safety approaches are geared toward autoregressive models and typically rely on post-hoc filtering or inference-time interventions. These are inadequate for effectively addressing safety risks in text diffusion models. We propose the Safety-Aware Denoiser (SAD), a safety-guidance framework in text diffusion models. The SAD modifies the iterative denoising process such that the text sample at the final denoising step is steered toward provably safe regions of the text space. This inference-time method can integrate safety constraints into the denoiser, avoiding computationally expensive retraining of the underlying diffusion model and enabling flexible, lightweight safety guidance. We evaluate the safety of the generated text using the SAD, with respect to hazard taxonomy, memorization, and jailbreak. Experimental results show that SAD substantially reduces unsafe generations while preserving generation quality, diversity, and fluency, outperforming existing methods. These results demonstrate that our safety guidance during denoising provides an effective and scalable mechanism for enforcing safety in text diffusion models.
arXiv:2605.16430v3 Announce Type: replace Abstract: Scaling LLMs requires tremendous computational resources, and recent advances in AI have gone hand in hand with massive amounts of capital expenditure. While it is established that scaling up LLMs reliably increases model quality (quantified in terms of loss or downstream evaluations), it is unclear how these quality improvements translate to potential revenue, and whether revenue increases would offset costs of larger-scale training and inference. In this work, we develop an economic model for characterizing the rational behavior of an LLM training firm by combining scaling laws with microeconomic theory. Under our model of firm behavior, LLM quality can be increased with more parameters and training tokens, leading to more potential adoption by consumers, who each have a quality threshold for using the LLM. On the other hand, additional parameters and training tokens both incur additional costs. We analyze the profit maximization problem for this model under compute-bound and data-bound regimes. In the compute-bound regime, optimal model size and token budget track hardware efficiency $E$ (FLOPs/\$) at a near-linear rate; total training cost then scales sub-quadratically in $E$. Data efficiency improvements incentivize larger models and training expenditure. When we are limited to $D$ data, profit-optimal training expenditure scales as $D^2/E$, i.e, increase with data and decreases with hardware efficiency (as well as data efficiency). Finally, we analyze practical trends in training expenditure: current trends are consistent with our most permissive model variants in the compute-bound regime, but are not profit-optimal in the data-bound regime or assuming hardware advances will stall. Overall, our results provide a theory of profit-optimal LLM training, providing a foundation for engaging critically with industry statements and supporting long-term economic decision making.
arXiv:2606.07489v2 Announce Type: replace Abstract: Frontier AI systems are bridging the gap between intelligence and utility by shifting from conversational assistants to autonomous agents that execute tasks end to end. Using production data from Perplexity's Search and Computer products, we study this transition by examining how AI agents accelerate and reshape knowledge work. Three key empirical findings emerge. First, using sessions with near-identical initial query pairs as natural experiments for the same underlying task attempted with both products, Computer performs 26 minutes of autonomous work per user session, versus 33 seconds for Search. Computer automates task decomposition and execution that Search users might otherwise manually orchestrate and implement. As a result, Computer shifts follow-up query distribution toward higher-order work such as verification and extension. Autonomy also increases execution quality, with per-query dissatisfaction rates 55% lower on Computer than on Search. Second, due to its autonomy advantage, Computer reduces completion time from 269 to 36 minutes on matched tasks, lowering estimated time and cost by 87% and 94%, respectively, compared to humans equipped with Search alone. Third, Computer changes the scope of work that users attempt: Computer queries more often cross occupational boundaries, require higher-order cognition, draw on broader expertise, take the form of composite tasks that bundle interdependent subtasks into a single query, and unlock work activities that are essentially absent from Search usage among the same users. Together, the evidence indicates that AI agents accelerate workflows, enhance output quality, reduce costs, and expand the breadth and depth of automated work.