arXiv:2606.06946v1 Announce Type: new
Abstract: We present LoRA-MINT, a new methodology for Membership Inference Test (MINT) applied to recent Large Language Models (LLMs) fine-tuned for specific Natural Language Processing (NLP) tasks through Low-Rank Adaptation (LoRA). The primary goal is to assess whether individual samples were part of the training data of these adapted models, providing a useful auditing tool for the management of intellectual property and sensitive data. Our analysis explores the relationship between model perplexity and membership status, providing a systematic framework for estimating data exposure in fine-tuned LLMs. We conducted experiments on four models and three benchmark datasets, obtaining precision values in determining if given data were used for training ranging from 0.77 to 0.92, which outperform state-of-the-art baselines and demonstrate the robustness and generality of the proposed method. In general, our findings underscore the potential of LoRA-MINT as an effective and scalable framework for auditing LLMs, improving transparency, and fostering the ethical and responsible deployment of AI and NLP technologies. For the sake of concreteness and current relevance, our discussion and experiments are centered on LoRAadjusted LLMs, but note that most of the presented methodology is easily applicable for auditing training data given any other technique for adapting LLMs or, more generally, any other domain-adapted AI models.
Science Journals
arXiv:2606.06947v1 Announce Type: new
Abstract: Generative recommendation advances item retrieval by reformulating it as autoregressive generation of Semantic IDs (SIDs), compact token sequences that encode item semantics. While SIDs offer a strong semantic prior, current SID-based methods assign each item a single static identifier through offline tokenization before sufficient user feedback is observed. For cold-start items, this one-shot commitment produces poorly discriminative codes, generating misaligned paths that remain unrefined because the associated tokens are rarely sampled during training. We identify this early static commitment, not model capacity, as the fundamental cold-start bottleneck in SID-based generative recommendation. To overcome this bottleneck and bridge the disjoint objectives of tokenization and generation, we propose DREAM (Dynamic Refinement of Early Assignment Mappings), a three-stage framework that resolves this flaw through progressive refinement. First, an intent-aware tokenizer rebuilds the SID space through counterfactual contrastive learning, generating a diverse pool of behavior-aligned candidates per cold-start item. Second, the frozen recommendation backbone serves as an evaluator, selecting the most reliable candidate based on multi-context user support without retraining. Third, a dynamic beam mechanism maintains multiple weighted SID hypotheses throughout training and inference, preventing premature collapse to a single assignment. Extensive experiments on three Amazon benchmarks show that DREAM substantially outperforms state-of-the-art generative and sequential baselines on cold-start metrics.
arXiv:2606.07071v1 Announce Type: new
Abstract: Deciding when to stop reviewing the results of a search is a common problem with multiple applications. Existing stopping rules developed within Technology-Assisted Review (TAR) aim to achieve a pre-specified recall target and do not take into account the reason for examining the results, potentially leading to sub-optimal recommendations. This paper applies decision theory to the problem and uses it to derive three practical stopping policies based on the Expected Value of Perfect Information. The approach is applied to two professional search tasks: patent examining and systematic reviewing. Experiments on CLEF-IP and medical systematic review datasets show that the proposed approach generally produces more appropriate stopping decisions than existing methods, as demonstrated by higher net utility under the evaluated cost and payoff settings.
arXiv:2606.06950v1 Announce Type: new
Abstract: Three-dimensional models are widely assumed preferable for volumetric medical imaging, yet their practical value depends on whether performance gains justify added computational cost and complexity. Rather than proposing a new architecture, we study how input dimensionality (2D, 2.5D, 3D) affects model behavior across convolutional neural networks (CNNs) and Vision Transformers (ViTs) under a fixed training protocol. Using a leakage-free NLST cohort (n = 1,977) with supporting LIDC-IDRI data, we find that the 2.5D CNN offers the most favorable discrimination-stability trade-off in our comparison (ROC-AUC 0.682, 95% CI [0.546, 0.799]) with a stable operating point. In contrast, 3D CNNs show threshold instability, and transformers exhibit degenerate predictions, such as all-positive predictions. Confidence intervals are wide and overlapping, so we present these results as a controlled resource-performance frontier and a failure-mode taxonomy rather than as definitive superiority claims. For class-imbalanced lung cancer screening classification, 2D and 2.5D inputs provide a more reliable trade-off between performance, stability, and computational efficiency than full 3D representations.
arXiv:2606.06953v1 Announce Type: new
Abstract: We argue that high-quality motion data can steer tracking policies toward better optimization trajectories early in training. In this work, we introduce LIMMT (Less Is More for Motion Tracking). To our knowledge, this is the first data-centric study for physics-based humanoid motion tracking. We go beyond simply removing low-quality and erroneous clips, but define motion data quality through three dimensions: physics feasibility, diversity, and complexity. We show that even training with under 3% of AMASS yields better tracking performance than training with the full dataset. We further conduct data cleaning on the estimated web-sourced mocap data. Extensive experiments and analyses validate the effectiveness of our framework.
arXiv:2602.15084v2 Announce Type: replace
Abstract: We present TokaMind, to our knowledge the first open-source foundation model for tokamak plasma dynamics, based on a Multi-Modal Transformer (MMT) and pretrained on heterogeneous diagnostics from the publicly available MAST dataset. TokaMind supports multiple data modalities (time-series, 2D profiles, and videos) with different sampling rates, robust missing-signal handling, and efficient task adaptation via selectively loading and freezing four model components. To represent multi-modal signals, we use a lightweight fixed-basis Discrete Cosine Transform embedding (DCT3D) and provide a clean interface for alternative embeddings (e.g., Variational Autoencoders). We evaluate TokaMind on the recently introduced MAST benchmark TokaMark, which comprises 14 tasks with heterogeneous reconstruction and forecasting objectives. Our results show that fine-tuned TokaMind outperforms the strongest benchmark baseline on all but one task. Compared with training the same architecture from scratch under a matched epoch budget, warm-start adaptation is most beneficial on demanding downstream settings, including long-horizon forecasting and high-dimensional equilibrium objectives. These findings highlight the value of multi-modal pretraining for tokamak plasma dynamics and provide a practical, extensible foundation for future fusion modeling tasks. Training code and model weights are publicly available at github.com/UKAEA-IBM-STFC-Fusion-FMs/tokamind and huggingface.co/UKAEA-IBM-STFC, respectively.
arXiv:2601.23207v2 Announce Type: replace
Abstract: Understanding what graph neural networks can learn, especially their ability to learn to execute algorithms, remains a central theoretical challenge. In this work, we prove exact learnability results for graph algorithms under bounded-degree and finite-precision constraints. Our approach follows a two-step process. First, we train an ensemble of multi-layer perceptrons (MLPs) to execute the local instructions of a single node. Second, during inference, we use the trained MLP ensemble as the update function within a graph neural network (GNN). Leveraging Neural Tangent Kernel (NTK) theory, we show that local instructions can be learned from a small training set, enabling the complete graph algorithm to be executed during inference without error and with high probability. To illustrate the learning power of our setting, we establish a rigorous learnability result for the LOCAL model of distributed computation. We further demonstrate positive learnability results for widely studied algorithms such as message flooding, breadth-first and depth-first search, and Bellman-Ford.
arXiv:2602.00163v2 Announce Type: replace
Abstract: Hyperkinetic movement disorders (HMDs) such as dystonia, tremor, chorea, myoclonus, and tics are disabling motor manifestations across childhood and adulthood. Their fluctuating, intermittent, and frequently co-occurring expressions hinder clinical recognition and longitudinal monitoring, which remain largely subjective and vulnerable to inter-rater variability. Objective and scalable methods to distinguish overlapping HMD phenotypes from routine clinical videos are still lacking. Here, we developed a pose-based machine-learning framework that converts standard outpatient videos into anatomically meaningful keypoint time series and computes kinematic descriptors spanning statistical, temporal, spectral, and higher-order irregularity-complexity features.
arXiv:2604.17948v2 Announce Type: replace
Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities across various cybersecurity tasks, including vulnerability classification, detection, and patching. However, their potential in automated vulnerability report documentation and analysis remains underexplored. We present RAVEN (Retrieval Augmented Vulnerability Exploration Network), a framework leveraging LLM agents and Retrieval Augmented Generation (RAG) to synthesize comprehensive vulnerability analysis reports. Given vulnerable source code, RAVEN generates reports following the Google Project Zero Root Cause Analysis template. The framework uses four modules: an Explorer agent for vulnerability identification, a RAG engine retrieving relevant knowledge from curated databases including Google Project Zero reports and CWE entries, an Analyst agent for impact and exploitation assessment, and a Reporter agent for structured report generation. To ensure quality, RAVEN includes a task specific LLM Judge evaluating reports across structural integrity, ground truth alignment, code reasoning quality, and remediation quality. We evaluate RAVEN on 105 vulnerable code samples covering 15 CWE types from the NIST-SARD dataset. Results show an average quality score of 54.21%, supporting the effectiveness of our approach for automated vulnerability documentation.
arXiv:2604.20123v2 Announce Type: replace
Abstract: In natural images, object skeletons are used to represent geometric shapes. However, even slight variations in pose or movement can cause noticeable changes in skeleton structure, increasing the difficulty of detecting the skeleton and often resulting in discontinuous skeletons. Existing methods primarily focus on point-level skeleton point detection and overlook the importance of structural continuity in recovering complete skeletons. To address this issue, we propose Lighthouse-Skel, a topology-aware skeleton detection method via lighthouse-guided structured inference. Specifically, we introduce a dual-branch collaborative detection framework that jointly learns skeleton confidence field and structural anchors, including endpoints and junction points. The spatial distributions learned by the point branch guide the network to focus on topologically vulnerable regions, which improves the accuracy of skeleton detection. Based on the learned skeleton confidence field, we further propose a lighthouse-guided topology completion strategy, which uses detected junction points and breakpoints as lighthouses to reconnect discontinuous skeleton segments along low-cost paths, thereby improving skeleton continuity and structural integrity. Experimental results on four public datasets demonstrate that the proposed method achieves competitive detection accuracy while substantially improving skeleton connectivity and structural integrity.
arXiv:2509.04991v2 Announce Type: replace
Abstract: Land surface temperature (LST) is a fundamental physical variable in land-atmosphere interactions, surface energy budgets, and climate processes. LST derived from medium- to high-resolution thermal infrared (TIR) observations effectively reveals thermal environmental disparities across distinct landscape units. However, achieving accurate, robust, and globally generalizable LST retrieval remains challenging under complex atmospheric conditions and diverse land cover types. Traditional split window (SW) algorithms heavily rely on empirical parameterizations, whose fixed coefficients fail to adapt to complex scenarios such as high surface temperatures and high atmospheric water vapor content. Concurrently, conventional data-driven models exhibit limited generalizability to out-of-distribution (OOD) samples due to the absence of explicit physical structure constraints. To address these issues, this study proposes a Parallel Component Decoupled Neural Network (PCD-Net) framework, which reformulates SW retrieval as a dynamic learning problem of physical component coefficients. Using the SW equation as the physical backbone, the framework constructs parallel subnetworks to adaptively learn the dynamic coefficients corresponding to the constant, first-order, and second-order brightness temperature difference terms; meanwhile, a residual branch is incorporated to supplement the nonlinear coupling corrections induced by the joint effects of surface emissivity and atmospheric water vapor. Through this component-level decoupled modeling, PCD-Net explicitly characterizes the dynamic response relationships between land surface emissivity, atmospheric water vapor content, and different SW physical components.
arXiv:2605.00737v2 Announce Type: replace
Abstract: Agentic AI architectures augment LLMs with external tools, unlocking strong capabilities. However, tool use is not always beneficial; some calls may be redundant or even harmful. Effective tool use, therefore, hinges on a core LLM decision: whether to call or not call a tool when performing a task. This decision is particularly challenging for web search tools, where the benefits of external information depend on the model's internal knowledge and its ability to integrate potentially noisy tool responses. We introduce a principled framework inspired by decision-making theory to evaluate web search tool-use decisions along three key factors: necessity, utility, and affordability. Our analysis combines two complementary lenses: a normative perspective that infers true need and utility from an optimal allocation of tool calls, and a descriptive perspective that infers the model's self-perceived need and utility from their observed behaviors. We evaluate six open and one closed-source frontier models under two harnesses, one conditioning on only the current turn and its search results, the other on the full execution traces, across four web-search tools and three tasks. In every setting, we find that a model's perceived need and utility are frequently misaligned with the true need and utility. Building on this framework, we train lightweight estimators of need and utility from the models' hidden states. These estimators drive simple controllers that improve decision quality and yield stronger task performance than the self-perceived baseline for most of the open-source models.
arXiv:2605.16794v2 Announce Type: replace
Abstract: Coincident Peak (CP) pricing is widely used in U.S. electricity markets to allocate capacity and transmission costs. This paper develops a behavioral game-theoretic framework for CP-driven load shifting that couples a nonlinear cost-allocation model with day-ahead (one-shot) and real-time (sequential-learning) decision processes. We examine two update rules, namely best-response dynamics (BRD) and fictitious-play dynamics (FPD), across continuous and finite action spaces to quantify how flexibility, action resolution, and participation influence peak outcomes. Using ERCOT peak-day data, we find that FPD reliably reduces system peaks, whereas BRD is more variable and can increase peaks under tight-capacity conditions. Finer action resolution improves peak shaving, while the number of participants is largely neutral when aggregate flexibility is fixed. Meanwhile, information-provider signals can induce herding, whereas response-aware or diverse signals improve peak shaving. These results highlight both the potential and limits of CP pricing: smoothing information and enabling granular control are as important as the amount of available flexibility. The framework offers practical guidance for system operators and consumers: For ISOs, broadcasting smoothed CP signals and setting minimum controllable-capacity thresholds enhance coordination. For consumers, greater flexibility and finer control resolution improve both cost savings and peak-shaving performance.
arXiv:2602.02819v5 Announce Type: replace
Abstract: Membership Inference Attacks (MIAs) aim to distinguish training points (members) from unseen data (non-members), and are widely used to quantify memorization and assess privacy risks. Standard MIA evaluation requires repeated retraining, which is computationally costly for large models. One-run (single training with randomized data inclusion) and zero-run (post hoc evaluation) methods are often used instead, but their statistical validity remains unclear. We address this gap by framing MIA evaluation as a causal inference problem, defining \emph{memorization as the causal effect of including a data point in the training set}. This novel formulation reveals and formalizes key sources of bias in existing protocols: one-run methods suffer from interference between jointly included points, while zero-run evaluations are additionally confounded by distribution shift between member and non-member evaluation data. We derive causal analogues of standard MIA metrics and propose practical estimators for multi-run, one-run, and zero-run regimes with non-asymptotic consistency guarantees. We validate our approach in several settings, including pretrained and fine-tuned LLMs, showing that it enables reliable measurement of MIA performance without retraining and under distribution shift. Overall, our framework provides a principled foundation for privacy evaluation in modern AI systems.
arXiv:2605.17561v2 Announce Type: replace
Abstract: Issues faced when using software are reported in the form of bug reports. However, many bug reports are invalid, meaning they do not require code changes, and are resolved with a no-code fix. Manually determining the root cause of the invalid bug reports and providing actionable resolutions by the customer support causes a serious waste of resources. Our goal is to introduce a standardized taxonomy for root-cause oriented invalid bug report subclassification, and perform experiments to test the accuracy of various approaches on invalid subclassification and no-code fix generation. We study how different configurations perform on a gold-standard benchmark we have created. Using a manually curated benchmark for higher quality analysis, we experimented with vanilla LLMs, Retrieval Augmented Generation, and agentic web search to identify invalid subclasses and generate no-code fixes. We evaluated the results against manually labeled ground truth data that includes the invalid subclass and no-code fixes from the original bug reports. We measured subclass detection performance with weighted F1-Score, and assessed no-code fix suggestions using BERTScore and Judge LLM success rates. For subclassification, retrieval augmented generation achieves the highest overall performance with 0.66 weighted F1, slightly outperforming vanilla LLMs at 0.65 and agentic web search at 0.64. At the subclass level, performance peaks at 0.85 F1 for Non-reproducibility and 0.79 for Feature Request and Question, while Wrong Version remains the most challenging with scores between 0.00 and 0.29. For no-code fix generation, agentic web search achieves the highest overall Judge LLM success rate at 68.9%, compared to 64.4% for RAG applications and 64.9% for vanilla LLMs, with subclass-level peaks of 87.4% for Working as Designed and 72.2% for Question.
arXiv:2606.06973v1 Announce Type: new
Abstract: Fusion reactivity is enhanced in the vicinity of strongly sheared flow due to the tendency of fast ions near the Gamow peak to travel long distances between collisions, thereby sometimes crossing gradients in the background flow and attaining a velocity boost relative to the thermal background. This ``shear flow reactivity enhancement effect'' (SFRE) allows turbulent kinetic energy on fine spatial scales to contribute to fusion reactivity before thermalizing, which, remarkably, enables ignition of some inertial confinement fusion (ICF) hot spots under conditions where fully thermalized plasma would fail to ignite. The size of the SFRE is a consequence of the dramatic scale separations distinguishing thermal ions, which govern fluid quantities, and fast ions, which govern fusion reactivity. It is demonstrated in this work that, as the Gamow energy increases relative to the thermal energy, the SFRE in unmagnetized plasma becomes asymptotically large compared to hydrodynamic effects such as viscous dissipation. An asymptotic formula is derived in this limit, quantifying the SFRE for reactants of disparate masses and charge states.
arXiv:2602.05833v2 Announce Type: replace
Abstract: There is a need for synthetic training and test datasets that replicate statistical distributions of original datasets without compromising their confidentiality. A lot of research has been done in leveraging Generative Adversarial Networks (GANs) for synthetic data generation, however the resulting models are either not accurate enough or are still vulnerable to membership inference attacks (MIA) or dataset reconstruction attacks since the original data has been leveraged in the training process. In this paper, we frame synthetic data generation as a guided test generation, or search-based testing problem rather than a purely generative modeling task. Ours is a search-based, adequacy-guided input generation technique inspired by GANs, with a generation step and a discrimination step; as in GAN, discrimination uses a discriminator model trained on the date, but instead of using models also for generation, we use a fuzzer. This way, the original (private) data is only indirectly leveraged in the generation process, and by evolving samples and determining "good samples" with the discriminator, we can generate privacy-preserving data that follows the same statistical distributions as the original dataset, leading to a similar utility as the original data. We evaluated our approach on eight datasets that have been used to evaluate the state-of-the-art techniques, finding that synthetic generated with our technique achieves good utility on average while also having good similarity scores, highlighting the potential of a mixed approach leveraging classical generation and model-driven discrimination for generating privacy-preserving, useful synthetic datasets.
arXiv:2605.21727v2 Announce Type: replace
Abstract: Block codes are considered for improving the reliability of messages stored in a computer memory with both stuck-at defects and random errors. It is assumed that the side information about the state of the defects is available to the encoder, but not to the decoder. A novel recursive construction of a set of masks is developed such that it can satisfy any $s$ stuck-at errors in a $2^m$ binary sequence, when $s \leq m$. We prove that the masks generated in this way are codewords in a Reed-Muller $RM(s-1, m)$ code. The constructed set contains no more than $2^s m^{s-1}$ masks. We provide the lower and the upper bound on the size of the stuck-at redundancy, a fixed subset of mask bits that uniquely represents each mask in the set. The stuck-at code constructed in this way is a non-linear code. It is also a subcode of an $RM(r,m)$ code, with $ r \geq s-1$, that can be used for additional random error correction. The encoding requires no mask search and is straightforward based on the description of the recursive construction. The decoding is done in a single attempt and requires almost no additional complexity or latency.
arXiv:2602.07026v3 Announce Type: replace
Abstract: Despite the success of multimodal contrastive learning in aligning visual and linguistic representations, a persistent geometric anomaly, the Modality Gap, remains: embeddings of distinct modalities expressing identical semantics occupy systematically offset regions. Prior approaches to bridge this gap are largely limited by oversimplified isotropic assumptions, hindering their application in large-scale scenarios. In this paper, we address these limitations by precisely characterizing the geometric shape of the modality gap and leveraging it for efficient model scaling. First, we propose the Fixed-frame Modality Gap Theory, which decomposes the modality gap within a frozen reference frame into stable biases and anisotropic residuals. Guided by this precise modeling, we introduce ReAlign, a training-free modality alignment strategy. Utilizing statistics from massive unpaired data, ReAlign aligns text representation into the image representation distribution via a three-step process comprising Anchor, Trace, and Centroid Alignment, thereby explicitly rectifying geometric misalignment. Building on ReAlign, we propose ReVision, a scalable training paradigm for Multimodal Large Language Models~(MLLMs). ReVision integrates ReAlign into the pretraining stage, enabling the model to learn the distribution of visual representations from unpaired text before visual instruction tuning, without the need for large-scale, high-quality image-text pairs. Our framework demonstrates that statistically aligned unpaired data can effectively substitute for expensive image-text pairs, offering a robust path for the efficient scaling of MLLMs.
arXiv:2606.06975v1 Announce Type: new
Abstract: Bioacoustic datasets from tropical regions remain limited, in part due to the absence of reproducible workflows for aggregating recordings from public archives. We present \textbf{MyGardenBird}, a curated dataset of bird vocalisations representing twelve common species across Peninsular Malaysia and the Indo-Malayan region. Recordings were sourced from Xeno-canto and processed through species-level filtering, manual spectrogram segmentation, and quality control checks. The primary release comprises 7,200 manually validated audio clips (16 kHz, 16-bit PCM mono WAV), balanced at 600 three-second clips per species (6.0 hours total) derived from 1,381 distinct recordings. Metadata includes geospatial coordinates, vocalisation categories, and signal-to-noise ratio (SNR) values (range: 0.83--59.18 dB; mean: 15.80 dB). A supplementary 44.1 kHz version is also provided. To mitigate data leakage, dataset partitions are defined at the source-recording level. Baseline classification experiments using convolutional neural networks on Mel-spectrograms achieved test accuracies of 92--96\%, indicating strong interspecies separability. Limitations include reliance on single-annotator curation; however, validation with BirdNET confirmed label consistency. MyGardenBird is openly available at https://doi.org/10.5281/zenodo.20306877 under a CC BY-NC-SA 4.0 licence. Complete preprocessing code accompanies the release to support reproducibility and future expansion.
arXiv:2606.06976v1 Announce Type: new
Abstract: Large language model (LLM)-based agents often make suboptimal tool-use decisions, including unsupported tool invocation and hallucinated direct responses, which may accumulate errors throughout multi-step interactions. Existing approaches mainly improve these behaviors through inference-time correction or coarse-grained reward signals based on decision outcomes and structured checklists, leaving the uncertainty characteristics of agent decisions underexplored. We observe that decision-oriented reinforcement learning tends to weaken the uncertainty separation between correct and incorrect actions, resulting in overconfident mistakes and weaker exploration signals. Therefore, we propose TRUST, which incorporates uncertainty quantification into reward design as a repulsive force for maintaining uncertainty separation, and labels lightweight key-turn annotations for unified post-training of multi-turn trajectories. Experimental results across diverse tool-use benchmarks show that TRUST consistently enhances both decision quality and agent performance while maintaining more reliable uncertainty estimates during optimization.
arXiv:2606.07158v1 Announce Type: new
Abstract: Cyber Threat Intelligence CTI attribution relies on identifying the Tactics, Techniques, and Procedures TTPs that distinguish one threat actor from another. This approach presupposes that each adversary leaves a recognizable operational fingerprint. This work investigates whether AI driven adversary emulation challenges that presupposition. We deploy agents from our Cybersecurity SuperIntelligence CSI framework, configured as five Advanced Persistent Threat APT groups, APT28, APT29, APT41, APT44, and Lazarus Group, against AI driven Defender agents across two cyber ranges provided by CYBER RANGES, equipped with defensive software Wazuh, Velociraptor, Elasticsearch and active AI driven defenders: an enterprise network and a military infrastructure. Across 20 experiments using two defender models, a binary pattern emerges: all 10 Enterprise range experiments resulted in compromise 2 to 12 hosts per experiment, while all 10 Military range experiments were successfully defended or resulted in stalemates, regardless of APT profile or defender model. In 8 of 10 Enterprise experiments, attackers independently weaponized the defender's own Velociraptor endpoint management platform as a command and control channel, a convergent behavior not encoded in any threat intelligence profile. We argue that in the AI era, wherein agents can be deployed provided the right models are available and subject to the right scaffolding and agentic configuration, the entry barrier for operating like a nation state APT collapses: beyond nation states, individuals can now act like commonly identified threat actors, and with it, fundamentally undermine TTP based attribution.
arXiv:2605.25638v2 Announce Type: replace
Abstract: Policy loss estimation remains a fundamental and long-standing challenge in reinforcement learning (RL) for diffusion language models (DLMs). We introduce Reinforcement Learning from Denoising Feedback (RLDF), a novel training paradigm that leverages feedback obtained from rollout and training processes to facilitate accurate and efficient policy loss estimation. To balance the trade-off between computational efficiency and estimation effectiveness, RLDF optimizes the model toward the clipped clean state from intermediate noisy states, combined with weighted timestep sampling over denoising timesteps. Extensive experiments demonstrate that RLDF achieves consistent and substantial improvements in both performance and generalizability across two representative DLM architectures, LLaDA and Dream, on multiple reasoning benchmarks. Our work lays a principled foundation for scalable reinforcement learning in diffusion language models. We build Drift, a training framework for DLMs, available at https://github.com/ant-research/Drift.
arXiv:2605.10832v2 Announce Type: replace
Abstract: Multimodal deep search requires an agent to solve open-world problems by chaining search, tool use, and visual reasoning over evolving textual and visual context. Two bottlenecks limit current systems. First, existing tool-use harnesses treat images returned by search, browsing, or transformation as transient outputs, so intermediate visual evidence cannot be re-consumed by later tools. Second, training data is usually built by fixed curation recipes that cannot track the target agent's evolving capability. To address these challenges, we first introduce a visual-native agent harness centered on an image bank reference protocol, which registers every tool-returned image as an addressable reference and makes intermediate visual evidence reusable by later tools. On top of this harness, On-policy Data Evolution (ODE) runs a closed-loop data generator that refines itself across rounds from rollouts of the policy being trained. This per-round refinement makes each round's data target what the current policy still needs to learn. The same framework supports both diverse supervised fine-tuning data and policy-aware reinforcement learning data curation, covering the full training lifecycle of the target agent. Across 8 multimodal deep search benchmarks, ODE improves the Qwen3-VL-8B agent from 24.9% to 39.0% on average, surpassing Gemini-2.5 Pro in standard agent-workflow setting (37.9%). At 30B, ODE raises the average score from 30.6% to 41.5%. Further analyses validate the effectiveness of image-bank reuse, especially on complex tasks requiring iterative visual refinement, while rollout-feedback evolution yields more grounded SFT traces and better policy-matched RL tasks than static synthesis.
arXiv:2606.07105v1 Announce Type: new
Abstract: The optical transfer matrix formalism is used to describe global set-ups incorporating a plenoptic camera. Analytical relations that give the effective resolution, depth of field, disparity and optimum patch size for image reconstruction are established versus the optical parameters of any global arrangement. The potentiality of this formulation is illustrated analyzing experimental results obtained in astigmatic cylindrical imaging conditions.