arXiv:2606.01843v1 Announce Type: new
Abstract: Deepfake detection suffers from poor generalization across forgery methods, as existing models tend to rely on spurious method-specific shortcuts that fail to transfer to unseen manipulations. While recent approaches attempt to improve generalization, they lack an explicit mechanism to identify and suppress such shortcuts in learned representations. In this work, we propose Shortcut Subspace Suppression (S^3) framework that explicitly characterizes and suppresses method-specific shortcuts via subspace modeling. Our key insight is that variations distinguishing different forgery methods capture method-specific artifacts and thus serve as an effective proxy for method-specific shortcuts. To this end, we train a lightweight linear probe for forgery method classification and perform Singular Value Decomposition (SVD) to extract the dominant shortcut subspace. Building on this formulation, we develop two complementary strategies to reduce shortcut reliance. During training, we softly suppress the shortcut subspace in feature representations, encouraging the model to rely on more generalizable cues for real/fake discrimination. At inference time, we introduce a training-free counterpart that attenuates neurons aligned with the identified shortcut directions, enabling plug-and-play generalization enhancement with improved interpretability. Extensive experiments on multiple benchmarks demonstrate that our method significantly improves cross-method generalization while maintaining strong in-domain performance. The code will be released upon acceptance of the submission.
Science Journals
arXiv:2606.02151v1 Announce Type: new
Abstract: Effective scheduling in the energy sector is essential to ensure the reliable operation of electrical grids and their connected assets by, for instance, optimizing the dispatch of generation units and storage systems. An effective planning strategy must (a) accommodate advanced and potentially non-linear system models -- exploiting the increasing data availability of modern grids, and (b) explicitly handle uncertainties arising, for instance, from the integration of renewable energy sources. While existing approaches can address either non-linearity (e.g., Monte Carlo Tree Search) or uncertainty (e.g., stochastic mathematical optimization), there is a lack of planning techniques capable of addressing both challenges simultaneously. To bridge this gap, we propose a Stochastic Scenario-Structured Tree Search (S3TS) algorithm that explicitly represents uncertainty through scenario trees while enabling the integration of advanced non-linear models. We evaluate S3TS on a simulated demand response signal publication problem, largely mimicking the imbalance settlement mechanism in Belgium. The results demonstrate near-optimal performance in linear, analytically tractable settings, with costs within 14% of the mathematically optimal solution conditioned to the scenario trees. In highly non-linear scenarios, S3TS significantly outperforms baseline methods, achieving cost reductions of up to 51% and 5.4% compared to a myopic algorithm and deterministic MCTS, respectively.
arXiv:2606.02154v1 Announce Type: new
Abstract: We measure picosecond dynamics of labyrinthine stripe domains with chiral N\'eel-type domain walls using time-resolved x-ray resonant magnetic scattering (XRMS). At the stripe-domain wavevector, the helicity-summed signal shows ultrafast demagnetization, recovery, and a weak oscillatory contribution that we identify as a signature of laser-launched coherent surface phonons. In contrast, the dichroic signal, which is sensitive to the in-plane magnetization inside the N\'eel walls, shows a strong oscillatory response whose frequency decreases with increasing pump fluence. We attribute this softening to pump-induced changes in the effective anisotropy and saturation magnetization, which modify the restoring field of an internal domain-wall mode. Time-resolved XRMS thus isolates wall-specific dynamics and provides access to internal domain-wall motion in disordered stripe textures on picosecond time scales.
arXiv:2511.21397v2 Announce Type: replace
Abstract: How does irrelevant information (i.e., distractors) affect test-time scaling in vision-language models (VLMs)? Prior work on text-only language models has shown that textual distractors can intensify inverse scaling, causing models to reason longer but less effective reasoning traces. In this work, we investigate whether similar phenomena arise in multimodal settings. We introduce Idis (Images with distractors), a visual question-answering dataset that systematically varies distractors along semantic and numerical dimensions. Our analyses reveal that visual distractors affect reasoning VLMs in a fundamentally different way from textual distractors: although inverse scaling still emerges, visual distractors reduce accuracy without increasing reasoning length. We further show that attribute counts extracted from reasoning traces provide key insights into how distractors interact with reasoning length and accuracy. As a sanity check, we propose a simple prompting strategy that mitigates distractor-driven predictions in reasoning vision-language models.
arXiv:2606.01416v1 Announce Type: new
Abstract: Tool-augmented large language model (LLM) agents rely on orchestration layers that coordinate planning, retrieval, tool invocation, validation, memory, and recovery. In these systems, failures arise not only from model errors, but also from orchestration-level issues such as tool timeouts, malformed arguments, stale context, contradictory evidence, retry loops, and unverified intermediate outputs. This paper presents a self-healing agentic orchestrator that treats reliability as a bounded runtime control problem. The orchestrator maps observable failure signals to inferred failure classes, selects targeted recovery actions under explicit budgets, verifies recovered trajectories, and records observability traces. We evaluate the approach on a 100-task controlled fault-injection benchmark against static workflow, retry-only, ReAct-style, and full-replanning baselines. Self-healing achieves 98.8\% task success, compared with 94.5\% for retry-only and 93.8\% for full replanning. A matched recovery-budget sweep shows that self-healing outperforms retry-only and full replanning at every tested budget, with the largest gap under a single recovery attempt: 94.0\% versus 85.3\% and 88.2\%, respectively. Under a controlled semantic silent-failure setting, verifier-guided self-healing reduces silent failures to 0.0\%, while non-verifying baselines return wrong-but-plausible outputs more often. A compact model-in-the-loop validation shows that the same recovery mechanism can operate when a live tool-calling model performs tool selection, argument generation, and answer synthesis over local fault-injected tools. These results provide controlled evidence that failure-aware, budgeted, and verification-guided orchestration improves reliability and diagnosability in tool-augmented LLM systems.
arXiv:2606.01419v1 Announce Type: new
Abstract: We propose DENSER, a Depth-guided ENSemble with Staged EFA-GS Reconstruction for soccer novel view synthesis. DENSER extends EFA-GS with three key contributions: (1) camera-height-based loss weighting that prioritises ground-level broadcast views, (2) monocular depth supervision from Depth-Anything-V2 to regularise geometry in textureless regions, and (3) a three-model pixel-average ensemble whose members diverge from a shared base checkpoint by varying training length and Gaussian scale clamping. On five held-out challenge scenes we achieve a mean PSNR of 29.89 dB, SSIM of 0.791, and LPIPS of 0.366.
arXiv:2606.01635v1 Announce Type: new
Abstract: Token selection is pivotal for effective LLM post-training. However, existing methods mostly rely on local heuristics and rarely formulate token selection as a principled valuation of individual response tokens. We introduce $\textbf{AlphaToken}$, a response token valuation framework that decouples valuation into $\textbf{adaptation}$ (promoting target-task learning) and $\textbf{stability}$ (preserving pre-trained capabilities), and makes each objective $\textbf{path-aware}$ by combining the direct-path signal from local token gradients with the downstream causal-path signal in autoregressive generation. Since retention data are typically unavailable, AlphaToken approximates stability via a $\textbf{Fisher-drift proxy}$ anchored at the pre-trained reference model. For efficient computation, we extend Ghost Dot-Product to token-level valuation. AlphaToken masks low-value response tokens during fine-tuning and preference optimization, concentrating training signals on more valuable positions. Experiments show that AlphaToken improves post-training performance and mitigates catastrophic forgetting.
arXiv:2606.01857v1 Announce Type: new
Abstract: Efficient resource allocation is a key challenge in business process management, with direct implications for cost, throughput time, and utilization. While recent Reinforcement Learning (RL) approaches have shown promise in deriving adaptive allocation policies, they typically neglect inter-resource collaboration patterns that can strongly influence real-world task handovers. Recognizing this, this paper introduces the first approach for multi-objective optimization of resource-level decision-making, enabling the recommendation of person-specific handover policies. To achieve this, our work combines an existing Multi-Agent System-based process simulator with a multi-objective evolutionary algorithm. The resulting approach produces Pareto-optimal, resource-specific policies that optimize the process across multiple objectives. Experimental results on synthetic and real-world datasets show that our approach reduces costs by an average of 37% and waiting time by 58%, consistently outperforming heuristic baselines and demonstrating the potential of leveraging collaboration-aware optimization to improve process performance.
arXiv:2606.01862v1 Announce Type: new
Abstract: Translating user intents into physical radio signals represents the critical yet notoriously tedious final step in wireless prototyping, as it requires intricate knowledge of physical layer details and presents immense implementation challenges. Large Language Models (LLMs) and multi-agent systems have revolutionized conventional software engineering, raising the compelling question of whether they can resolve these formidable difficulties. However, our investigations reveal that current models experience significant limitations and fail to accomplish this task when applied to radio signal generation. This performance degradation primarily stems from severe domain ignorance and a fundamental insensitivity to physical hardware constraints. To bridge this gap, we introduce RadioMaster, a fully autonomous multi-agent framework designed to seamlessly translate user input into real-world wireless emissions. RadioMaster operates on three synergistic pillars: RadioWiki for domain-specific knowledge retrieval, RadioAgent for collaborative I/Q sample generation alongside hardware configuration, and RadioEmulator for closed-loop physical layer verification. Furthermore, we construct RadioBench, the first comprehensive benchmark tailored specifically for the radio signal generation domain. Extensive real-world evaluations demonstrate that RadioMaster significantly outperforms state-of-the-art (SOTA) baselines regarding configuration viability and signal fidelity.
arXiv:2606.02347v1 Announce Type: new
Abstract: Algorithm registers are public-facing databases that display basic information about algorithms employed in public administration. While several such registers exist across Europe and globally, their capacity to deliver meaningful transparency remains contested. In Germany, the landscape is notably fragmented: no federal-level register exists, yet at least five state- and federal-level initiatives publish information about AI systems with varying scopes and objectives. A recent conceptual proposal by Alina Lorenz (2025), outlines technical and governance requirements for a national AI transparency register in Germany. We repurpose this proposal as an audit instrument, extracting structured checklists from the transparency goals and subgoals it formulates. The resulting checklists, translated from German into English, is made publicly available to support practitioners auditing existing registers or designing new ones. We apply this framework to conduct an external audit of the two main existing German transparency initiatives, MaKI and Lernende Systeme, evaluating the extent to which they fulfill the proposed goals. Our audit reveals that several adaptations are likely needed for these registers to serve as an useful transparency instrument. We further propose a visualization of register transparency levels and derive concrete action items for improving existing German platforms.
arXiv:2606.02358v1 Announce Type: new
Abstract: We present Chimera, a flexible and scalable Microcontroller Unit (MCU) designed to accelerate real-time inference of rapidly evolving transformer-based models at the ultra-low-power edge (hundred of mW). The chip, implemented in 22 nm FDX technology, integrates a transformer accelerator tightly coupled within a compute cluster featuring nine general-purpose RV32IMA cores. Scalability extends to the memory hierarchy through a novel L2 memory island subsystem, which enables data sharing across multiple clusters while delivering 563 Gb/s aggregate bandwidth. The L2 subsystem enforces quality-of-service guarantees for latency-critical traffic, achieving up to 16x latency reduction. Chimera achieves peak energy and area efficiencies of 3.1 TOPS/W and 281 GOPS/mm2, demonstrating 1.37x higher energy efficiency and up to 100x higher area efficiency compared to State of the Art (SoA) SoCs. Compared to SoA standalone accelerators, Chimera achieves comparable energy efficiency and up to 1.8x higher area efficiency.
arXiv:2606.01967v1 Announce Type: new
Abstract: While prompt engineering is instrumental in maximizing the capabilities of Large Language Models (LLMs) during inference, the role of prompts during training remains critically underexplored. Prevailing fine-tuning paradigms typically treat training prompts as mere surface forms, assuming that semantically equivalent instructions yield identical learning outcomes. However, we reveal that this equivalence is deceptive: while paraphrased prompts often lead to comparable in-task performance, they induce drastically different cross-task impacts regarding catastrophic forgetting and generalization. Crucially, these impacts are positively correlated across tasks, indicating the existence of superior prompts that consistently yield better performance. Furthermore, we discover that these superior prompts can be robustly identified by task loss prior to learning. Leveraging these insights, we introduce State-Adaptive Prompt Optimization (SAPO), a lightweight yet effective training strategy that shifts task formulation from a static input to a dynamic, state-adaptive variable. Comprehensive experiments on diverse benchmarks confirm its effectiveness, which significantly mitigates forgetting while improving generalization, achieving substantial performance gains over state-of-the-art methods. These results provide insights into how training prompts shape learning dynamics and offer a practical recipe for robust fine-tuning. Our code is available at https://github.com/Eric8932/SAPO.
arXiv:2606.02366v1 Announce Type: new
Abstract: We present PRIMA (*PRI*ors for *M*esh *A*daptation), a framework for robust 3D quadruped mesh recovery under severe species and pose imbalance. Existing animal reconstruction methods often regress toward mean shapes and poses due to limited 3D supervision and long-tailed species distributions, resulting in poor generalization to underrepresented animals and rare articulations. PRIMA addresses this challenge through three key contributions. First, we incorporate BioCLIP embeddings as biological priors to inject semantic and morphological knowledge into the reconstruction process, enabling more accurate and generalizable shape prediction across diverse quadrupeds. Second, we introduce a test-time adaptation (TTA) strategy that refines SMAL predictions using 2D reprojection constraints together with auxiliary keypoint guidance, improving pose and shape estimation while enabling the generation of high-quality pseudo-3D annotations from existing 2D datasets. Third, leveraging this TTA framework, we construct Quadruped3D, a large-scale pseudo-3D dataset that covers diverse species and pose variations to systematically improve model performance. Extensive experiments on Animal3D, CtrlAni3D, Quadruped2D, and Animal Kingdom demonstrate that PRIMA achieves state-of-the-art results, with particularly strong improvements on underrepresented species and challenging poses. Our results highlight the importance of biological priors and adaptation-driven data expansion for scalable and generalizable animal mesh recovery. Code is available at https://github.com/AdaptiveMotorControlLab/PRIMA.
Attention-guided Fine-tuning of Multimodal Large Language Models Improves Chain-of-Thought Reasoning
arXiv:2606.01558v1 Announce Type: new
Abstract: The effectiveness of Chain-of-Thought (CoT) prompting in Multimodal Large Language Models (MLLMs) remains uncertain: across several visual reasoning benchmarks, CoT prompting often degrades performance compared to direct prompting. In this paper, we provide a systematic analysis of CoT behavior in three modern MLLM families across model scales on datasets requiring step-wise visual evidence. Our analysis identifies two recurring failure modes: premature answer commitment and limited direct visual-token access during rationale generation. We further find that standard CoT-style Supervised Fine-Tuning (CoT-SFT) can mitigate these issues only partially, while often increasing reliance on textual priors and reducing counterfactual visual dependence. Motivated by these findings, we propose Attentive-CoT (Att-CoT), an attention-guided fine-tuning objective that encourages CoT trajectories to delay answer commitment while maintaining sustained visual-token access. Att-CoT can be plugged into any CoT-SFT training run without architectural changes. Experiments on three visual reasoning benchmarks across six MLLMs show that Att-CoT enhances CoT performance over standard fine-tuning.
arXiv:2606.01560v1 Announce Type: new
Abstract: Graph Neural Networks (GNNs) are vulnerable to adversarial attacks, which inherently invert connectivity patterns by introducing disassortative edges in assortative graphs and assortative edges in disassortative graphs. This structural inversion creates structure-feature mismatches that disrupt neighborhood aggregation across different graph types. However, we find that existing defenses are limited, as they either treat neighborhoods as monolithic under fixed assortativity assumptions or rely on standard softmax classifiers that fail to account for perturbation-induced representation shifts. To further exploit this observation, we adopt a robustness perspective that jointly disentangles node representations and decision spaces, isolating perturbation effects while enforcing well-separated decision regions. Based on this principle, we propose Graph Joint Disentanglement Network (GJDNet), a unified framework for robust node classification across diverse graph assortativity regimes. GJDNet enhances robustness at both representation and decision levels: it employs feature-driven soft structural disentanglement with skewness-aware neighbor filtering to suppress perturbation-induced structure-feature mismatches, and introduces a Spherical Decision Boundary (SDB) to promote intra-class compactness and inter-class separation in the embedding space, thereby stabilizing decision boundaries under perturbations. Theoretical analysis provides insights into the effectiveness of the proposed disentangled representation and decision mechanisms, while extensive experiments demonstrate that GJDNet consistently achieves strong robustness across graphs with different connectivity regimes.
arXiv:2606.02010v1 Announce Type: new
Abstract: PlanarBench tests whether LLMs can draw planar graphs as ASCII art given only an edge list -- a spatial reasoning task that resists memorization because edge order, edge orientation, and node labels are all permutable.
We evaluate 91 models on the 199 simplest non-isomorphic connected planar graphs (2 - 7 vertices).
Edge count is the dominant difficulty predictor ($r = -0.85$) -- a finding not reported in prior LLM graph benchmarks, which use only node count as the difficulty axis.
arXiv:2606.01968v1 Announce Type: new
Abstract: Hamming Quasi-Cyclic (HQC) has been selected by NIST for standardization as an additional code-based key-encapsulation mechanism, providing algorithmic diversity alongside lattice-based post-quantum cryptography. Efficient deployment of HQC on mobile and embedded platforms, however, requires careful optimization of its decoding procedure, whose Reed-Muller and Reed-Solomon components dominate the computational cost. This paper studies HQC decoding on Qualcomm Hexagon processors in NPU-integrated devices, focusing on the Hexagon Vector eXtensions (HVX) backend rather than a tensor-inference engine. We observe that HQC decoding naturally exposes vector-structured computation, including Reed-Muller reliability vectors, Hadamard-transform coefficients, Reed-Solomon syndrome vectors, finite-field products, and packed support-point evaluations. Based on this observation, we redesign the dominant decoding kernels around HVX-friendly data layouts and execution patterns, including a vectorized Reed-Muller Hadamard transform, scalar-equivalent peak selection, HVX-oriented finite-field arithmetic, vectorized syndrome computation, and shortened-support locator-root evaluation. We implement and evaluate the optimized decoder using both Hexagon simulator measurements and real-device experiments on a Snapdragon~8 Gen~2 hardware development kit. The results show that Hexagon/HVX-assisted decoding substantially reduces latency and energy consumption, improving energy efficiency by up to $18.13\times$ while significantly offloading host CPU work. These results indicate that NPU-integrated mobile platforms can serve as effective backends for structured post-quantum cryptographic decoding when the underlying kernels are reformulated around vector execution.
arXiv:2606.01969v1 Announce Type: new
Abstract: Background: Developers increasingly review multi-file code changes generated by LLM-based agents, yet no validated end-to-end workflow or IDE tooling design exists for this scenario.
Aims: We investigate (RQ1) the challenges developers face when reviewing LLM-generated multi-file changes and (RQ2) how developers envision effective workflows for this task.
Method: In collaboration with JetBrains, we conducted a participatory design study structured using the double-diamond design process with Discover, Define, Develop, and Deliver phases. Industry practitioners participated in the Discover phase (N=17); seven of these returned for the Develop phase. The Define phase was an author-led synthesis. The Deliver phase produced a conceptual design and a high-fidelity semi-interactive prototype evaluated through a follow-up survey with N=43 practitioners.
Results: Participants identified trust-calibration as the central challenge. The study yielded a three-level review workflow (overview, file-analysis, code snippet review) supported by seven design constructs (chunk, risk-per-line, risk-per-file, judge, walk-through, zooming in/out, and security cage). In the validation survey, all three workflow levels scored above the neutral midpoint (means 3.50--3.91 on a five-point scale). Of the respondents, 63% expected reduced overall review effort, and 52% reduced trust-assessment effort, relative to their current tools. These findings suggest that the design constructs indicate a positive direction for future tool development.
Conclusions: Reviewing LLM-generated multi-file changes is a trust-calibration problem rather than a diffing problem. The three-level workflow and the seven constructs we report give tool designers a conceptual framework for building AI-ready code review tools that surface risk and confidence signals at the granularity at which developers allocate attention.
arXiv:2606.01581v1 Announce Type: new
Abstract: As the reasoning capabilities of Large Language Models (LLMs) continue to advance, LLM-based agent systems offer advantages in flexibility and interpretability over traditional systems, garnering increasing attention. However, despite the widespread research interest and industrial application of agent systems, these systems, like their traditional counterparts, frequently encounter anomalies. These anomalies lead to instability and insecurity, hindering their further development. Therefore, a comprehensive and systematic approach to the operation and maintenance of agent systems is urgently needed. Unfortunately, current research on the operations of agent systems is sparse. To address this gap, we have undertaken a survey on agent system operations with the aim of establishing a clear framework for the field, defining the challenges, and facilitating further development. Specifically, this paper begins by systematically defining anomalies within agent systems, categorizing them into intra-agent anomalies and inter-agent anomalies. Next, we introduce a novel and comprehensive operational framework for agent systems, dubbed Agent System Operations (AgentOps). We provide detailed definitions and explanations of its four key stages: monitoring, anomaly detection, root cause localization, and resolution.
arXiv:2606.01640v1 Announce Type: new
Abstract: Human mobility generation aims to synthesize realistic trip chains for target populations based on individual features. Existing paradigms, including deep generative models, LLM-based methods, and traditional heuristics, struggle to satisfy the complex demands of this task while simultaneously maintaining interpretability, behavioral plausibility, population-level distributional alignment, and inference efficiency. To bridge this gap, we introduce MobEvolve, the first agentic self-evolving heuristic framework for human mobility generation. MobEvolve initializes a behavior-inspired heuristic system and employs an LLM agent to iteratively evolve its internal logic. By diagnosing empirical misalignments and failure cases on a validation set, the agent proposes targeted updates and accumulates evolution memory for cumulative self-improvement. Extensive evaluations on the Singapore and Montreal benchmarks demonstrate that MobEvolve significantly outperforms state-of-the-art deep generative and LLM-based methods in individual trajectory fidelity, population-level distribution alignment, and behavioral plausibility, while preserving interpretability and high inference efficiency.
arXiv:2606.02381v1 Announce Type: new
Abstract: In this study, a generalized operator-based mathematical conflict framework is presented to explicitly represent structural discrepancies between raw data and contextual data. The proposed structure treats conflict as a local, directional, and context-sensitive quantity, integrating components such as weighting, scale behavior, and output mapping under a unified abstract operator. Without being reduced to a specific learning algorithm or optimization method, the framework is defined as a general structure adaptable to different classes of problems. While existing approaches typically treat conflict merely as an implicit side effect embedded within the optimization process, the proposed framework considers conflict as an independent, operator-based, and component-level mathematical object.
arXiv:2606.02382v1 Announce Type: new
Abstract: Modern command, control, communications, computers, cyber, intelligence, surveillance, and reconnaissance (C5ISR) environments place substantial attentional demands on mission commanders. Failures in attention allocation in these high-risk settings can have severe operational consequences. This study investigates the efficacy of gaze-driven, attention-guided adaptive decision support tools, including visual-only and multimodal designs, in a high-fidelity simulated military command center. To characterize gaze and attentional dynamics during interaction with these tools, recurrence quantification analysis was applied to eye-tracking data. Stepwise regression using the Bayesian information criterion was then used to identify recurrence-based gaze metrics associated with performance. Results showed that the multimodal adaptive decision support tool was associated with significantly higher performance than the visual-only attention-guided tool. Average diagonal line length showed a negative linear association with performance, whereas entropy showed a positive linear association. Recurrence rate, determinism, and entropy also showed nonlinear quadratic relationships with performance. In particular, recurrence rate and determinism followed an inverted-U pattern consistent with the Yerkes-Dodson law. These findings suggest that effective performance in dynamic C5ISR contexts depends on a balance between structured and flexible visual scanning, and that recurrence-based gaze metrics can help characterize attentional dynamics during interaction with adaptive decision support systems.
arXiv:2606.01999v1 Announce Type: new
Abstract: Modern deep learning models for forecasting groups of time series rely on increasingly longer observation windows. However, the benefit of increasing the window size is often simply attributed to capturing long-range dependencies, and broader discussion on how global forecasting models leverage input observations has been limited. In this paper, we show that forecasting groups of time series involves two objectives: (i) generative process identification (GPI), i.e., inferring the specific process generating the input sequence, and (ii) conditional forecasting (CF), i.e., predicting future values given input observations. From this perspective, optimal predictions can be interpreted as an average over plausible data-generating processes, weighted by their likelihood given the input window. This suggests another explanation for the benefits of long context windows: they reduce the uncertainty about which specific process is generating the input time series during operation. We prove that even for processes with memory length $P$, an input window size strictly larger than $P$ is necessary to achieve the minimum attainable error. Finally, we show how decoupling GPI and CF can improve computational scalability without compromising accuracy. Experiments on synthetic and real-world data validate our insights and their relevance for designing forecasting architectures.
arXiv:2606.02388v1 Announce Type: new
Abstract: Reinforcement learning (RL) improves large language model (LLM) agents by teaching them which actions lead to high rewards, but provides little supervision on what those actions do to the environment. World modeling (WM) can fill this gap, yet existing approaches often require separate simulators, extra training stages, or additional inference-time computation. We observe that on-policy RL rollouts already contain the needed signal: each transition pairs an action with its resulting next observation. Based on this observation, we propose PaW, a Policy and World modeling co-training framework that adds auxiliary WM supervision to the same policy during RL, without changing the inference paradigm. To make auxiliary WM supervision informative and stable, PaW introduces three components: action-entropy-based WM data selection, noise-tolerant WM loss, and reward-adaptive loss balancing. Experiments on three agentic task benchmarks show consistent improvements over strong RL baselines across models and RL algorithms. These results suggest that standard RL rollouts are a practical source of WM supervision for language-agent training.
arXiv:2606.01596v1 Announce Type: new
Abstract: Learning chaotic dynamical systems from data requires more than short-term predictive accuracy: the learned model must preserve the attractor geometry and its invariant statistics. Trajectory (zero-order) and Jacobian (first-order) matching supervise the values and tangent structure of the vector field, but neither constrains how the field bends away from its tangent plane. A model can thus match values and tangents at the supervised states yet curve differently from the truth, remaining locally accurate while drifting toward spurious attractors and distorting long-time statistics. We show that enforcing second-order consistency mitigates these failures, but forming the full Hessian is prohibitive in high dimensions. We propose model-constrained randomized Jacobian matching, which compares the Jacobians of the true and learned vector fields at randomly perturbed inputs. A Taylor expansion shows that the expected randomized Jacobian loss decomposes into the nominal Jacobian mismatch plus a Hessian mismatch scaled by the noise variance, implicitly enforcing second-order consistency at $\mathcal{O}(d^2)$ cost without forming the $\mathcal{O}(d^3)$ Hessian tensor. Using only Jacobian evaluations, the method scales to high dimensions where explicit Hessian matching does not. Numerical experiments confirm that second-order methods are robust. For Lorenz~63, first-order methods produce catastrophic Lyapunov-exponent outliers under minimal temporal supervision, which second-order methods eliminate while recovering the correct attractor. For coupled Lorenz~96, an out-of-distribution forcing sweep separates the methods: all agree up to $F=16$, but beyond $F=18$ only second-order methods preserve the invariant measure and Lyapunov spectrum. On both systems, randomized Jacobian matching performs comparably to explicit Hessian matching at much lower cost.