Sliced inference such as SAHI is the standard remedy for small objects in high-resolution imagery, but its gains are usually measured against low-resolution baselines. We revisit sliced inference with scale as the explicit variable and ask when tiling actually pays off. We compare seven dense detectors from four families on VisDrone and DOTA, under pre-specified hypotheses, calibration-only tuning and a single evaluation on held-out test data. Two findings emerge. First, much of the benefit attributed to slicing is a resolution effect: on VisDrone, a single pass trained at a higher resolution outperforms tiling with box fusion (merging decoded boxes) for every detector that could be trained at that resolution, and at equal compute a higher resolution helps more than a larger model. Second, tiles carry complementary information that box fusion wastes. We introduce SCAHI (Slicing with Coverage-Aware Hyper Inference), which accumulates the dense outputs of all passes with weights that give every image position the same total coverage, and reads objects off a peak grid fine enough to keep crowded neighbours apart. On identical passes, SCAHI outperforms box fusion for every YOLO detector on both datasets and matches it for FCOS. On DOTA, whose images far exceed the resolution a single pass can be trained at, tiles with SCAHI beat the best single pass for every detector, by 4.3 to 5.6 mAP50 at 2048 px, and replacing only SAHI's merge step by SCAHI adds 2.3 to 4.0 mAP50. An ablation attributes the gain to the border weighting and the resolution of the peak grid. Source code is available at https://github.com/beckerio/scahi.
This technical report presents MNN-Pocket-0.3B, a compact 300M-parameter dense language model designed for efficient on-device deployment and supporting context lengths of up to 32K tokens. With an INT4 deployment footprint of only 173 MB and decoding throughput of up to 466 tokens/s on iPad Pro M5 (Metal), it can be seamlessly integrated into applications on edge devices such as smartphones and smart glasses. MNN-Pocket-0.3B is developed through the co-design of its architecture, data, training pipeline, and inference system. Hardware-aware architecture selection and dependency-aware vocabulary pruning balance model capability with on-device memory and compute constraints. A four-stage pre-training pipeline covering approximately 4.1 trillion high-quality Chinese and English tokens equips the model with broad knowledge and reasoning capabilities. This is followed by a three-stage post-training pipeline comprising instruction fine-tuning, multi-teacher on-policy distillation, and quantization-aware training. Finally, hardware-specific optimizations in the MNN inference engine enable fast, memory-efficient inference and long-context processing. Extensive evaluations demonstrate that MNN-Pocket-0.3B achieves frontier-level performance among models of comparable size, with particularly strong results in Chinese and English knowledge and reasoning, mathematics, coding, and instruction following. It consistently outperforms similarly sized open-source models and even surpasses several 0.5B-parameter models on instruction-following benchmarks. We release MNN-Pocket-0.3B with open weights and ready-to-use MNN deployment packages. We hope it will provide a practical foundation for capable, responsive, and memory-efficient edge applications while advancing research and development in on-device language models.
Machine learning (ML) services increasingly expose explainability outputs alongside model predictions. While such explanations are intended to improve transparency, they may also reveal information about the underlying training data. In this work, we show that the local surrogate geometry employed by Local Interpretable Model-agnostic Explanations (LIME) can be repurposed as a proxy for a sample's distance to the model's decision boundary. We extend this boundary-distance view to tabular counterfactual explanations and, for the first time, to visual counterfactual explanations for image classifiers. We present XposedMIA, a one-shot, explanation-based black-box membership inference attack (MIA) framework that converts the returned explanation into a proxy for the queried sample's distance to the model decision boundary. After a one-time calibration step on a small labeled calibration set of as few as 30 samples, our attack requires only a single query for each target sample. Unlike prior label-only attacks requiring iterative queries or shadow training, our method relies solely on the predicted label and its accompanying explanation. Our evaluation indicates that XposedMIA achieves high effectiveness, with success rates of up to 0.95 using only a single query per target, revealing that releasing explanations introduces an additional channel for membership leakage.
Exploration in deep reinforcement learning (RL) for continuous control is commonly driven by temporally uncorrelated Gaussian (white) action noise, which leads to jerky motions with locally constrained exploration. In contrast, temporally correlated coloured noise, particularly pink noise, has been shown to support more efficient exploration in RL. Motivated by these findings, we inquire whether scheduling changes to the noise colour over the course of training can outperform white noise for continuous control robotics. We perform a systematic study of 49 linear schedules between white and red noise, together with generalised state-dependent exploration (gSDE), on four sparse-reward manipulation tasks with a Fetch arm and a Shadow Dexterous Hand. We find that the noise colour at the end of training does not affect learning efficiency or end performance, whereas fading towards red noise halves the jerk of exploration in every task. Balancing learning efficiency, end performance and smoothness, red noise emerges as the preferred strategy: it learns as efficiently as pink noise with almost 2-fold less movement jerk. We therefore recommend exploration schedulers that fade towards red noise as the default for off-policy robot learning.
Sensory delays create temporal misalignments between the state of the environment and the observations available to an agent, turning Markovian control problems into partially observable ones. In this work, we inquire whether such misalignments can be mitigated during training by using hindsight to recover aligned state-action transitions. We introduce Hindsight Critics, an asymmetric extension of the Soft Actor-Critic for reinforcement learning under partial observability. The critics are trained on hindsight-aligned transitions whereas the actor learns from the resulting privileged value estimates despite only having access to delayed observations. Comparisons against delay-aware and unaware baselines on the DelayBench suite show that Hindsight Critics can partially compensate for observation delays without requiring modifications to the agent's observations during deployment.
When a vision-language model encounters a new class or domain, how much should it trust its learned prompts? Because source adaptation can correct pretrained errors on some target images while disturbing valid predictions on others, we introduce Contextual Adaptive Fusion (CAF), which retains four prediction branches by pairing frozen and prompted visual and textual representations. To arbitrate when branches disagree on an image, CAF combines each branch's predictive confidence with how frequently the branches favor its proposed class across the unlabeled target collection, placing greater trust in predictions that are locally confident yet less dominated by collection-wide frequency bias. The fusion coefficients are learned once on source data and remain fixed at inference. Trained on ImageNet, CAF achieves 69.96% average accuracy across ten cross-dataset benchmarks (1.92 percentage points above ManiPT) and remains highly competitive across four ImageNet variants, setting a new best of 52.21% on ImageNet-S. In base-to-novel transfer across eleven datasets, CAF further reaches 79.77% novel-class accuracy and an 82.81% harmonic mean. Code is available at https://github.com/zhenyuanchenai/CAF.
Alignment of text-to-image diffusion models with reward functions is a key challenge in controllable generation. Existing approaches typically rely on optimizing model inputs or fine-tuning denoising networks, which requires computationally expensive backpropagation through large denoising models. In this work, we propose Direct Trajectory Optimization (DTO), a training-free framework that performs alignment by directly optimizing the sampling trajectory of diffusion models. Specifically, DTO treats intermediate denoising scores and sampled noise as the optimization variables, eliminating the need for gradients through the denoising network. To mitigate the out-of-distribution problem, we introduce Reward-Optimized Score Guidance, which interpolates between the optimized scores and the pretrained model predictions to stabilize sampling, while providing a provable theoretical guarantee of reward improvement under constrained distributions. Empirically, DTO improves alignment performance across three diffusion models on two benchmarks, outperforming existing inference-time alignment methods while maintaining high efficiency. These results suggest that trajectory-level optimization is a promising direction for efficient and effective diffusion model alignment.
In financial backtesting and other point-in-time tasks, external inputs can be restricted by date, but model parameters may retain financial facts disclosed later, introducing parametric look-ahead bias. We introduce FinCutBench, a continually updated monthly financial knowledge benchmark for measuring how far a fixed model's financial knowledge extends over calendar time. The current dataset contains 9,000 Chinese and English four-option multiple-choice questions from January 2023 to September 2026, covering Chinese and U.S. markets and four categories of financial information. We construct monthly knowledge profiles using the approximate months in which the tested facts became public. After a profile passes the decline test and knot checks, we define the FinCut Boundary as the first month when the fitted profile reaches its lower level, and report the stability of the estimated date under resampling. Boundary dates and localization precision vary across models. Using the independently measured Transition Start and Boundary as fixed references for downstream sample grouping, we find significant relative declines in U.S. asset-ranking performance and in earnings recall for Sol and Luna. Downstream effects vary across markets and tasks. The FinCut Boundary thus provides an empirical reference for the temporal scope of parametric knowledge in point-in-time financial evaluation, supporting evaluation-window selection and result interpretation.
Masked Diffusion Models (MDMs) are well known for fast text generation via parallel decoding. However, independent parallel sampling can commit to inconsistent predictions, and less is known about how to selectively revise these early token commitments. In this work, we introduce $\underline{\mathrm{P}}$lanned $\underline{\mathrm{U}}$nmasking-and-$\underline{\mathrm{R}}$evision Joint $\underline{\mathrm{L}}$earning (PURL) for MDMs, a training-based framework that seamlessly unifies parallel generation with selective token revision. We derive a normalized joint transition whose likelihood factorizes into edit selection, unmask selection, and token prediction. These components are trained jointly on partially masked sequences constructed according to the model's own unmasking policy. During inference, all three decisions are supported by a single shared forward pass, enabling the model to decode masked tokens and revise noisy tokens. Across logic puzzles, mathematical reasoning, and code generation, PURL improves accuracy at matched unmask budgets, with relative gains up to 47.4% accuracy improvement and up to an $8\times$ reduction in scheduled decoding rounds. Beyond these gains, our analysis further reveals general principles of revisable decoding: token edits become more effective as context accumulates; early edits can steer unfinished trajectories, whereas later edits tend to remain localized; and effective edits therefore depend on both calibrating when to intervene and learning where and how to revise. Together, these suggest that effective parallel generation depends not on revising more tokens, but on making the right commitments revisable at the right time.
Salient and camouflaged object detection (SOD and COD) both recover foreground support, yet most joint methods output only a binary mask and therefore cannot distinguish the type. USC12K distinguishes the two types, but its baseline USCNet still binds foreground existence to the type within a single normalization, so the failure source cannot be attributed. Our cross-task study shows that foreground localization transfers between SOD and COD, whereas type discrimination does not: a COD-only model recovers foreground on SOD images but its type judgement follows the training distribution rather than the foreground-background relation, so the two decisions require separate mechanisms. SegDis factorizes the output into an explicit foreground probability and a conditional type probability given foreground: FLR provides high-quality task-agnostic support, and RTA refines the SOD/COD assignment from the foreground-background relation without a task identifier. Because support is explicit, each ground-truth foreground pixel is classified as correct-type, type-swapped, or missed, so foreground-recovery and type-assignment errors are diagnosed separately. With PVT-v2-B2, SegDis attains 68.98% FG-mIoU and 78.35% mIoU on USC12K at about one-eighth of USCNet's parameters; CSCS is 5.75.
AI agents deployed on the open web are becoming more capable of performing actions that may breach contractual provisions and violate computer security laws. In this paper, we introduce ScrapeBench, the first agentic benchmark for empirically evaluating the legal compliance of AI agents with website scraping policies, namely terms of service (ToS) and $\texttt{robots.txt}$ directives. Across widely used closed-weight and open-weight AI agents, we find that most agents exhibit low compliance rates (< 5%) when instructed to scrape websites in our sample of 1,626 popular websites that explicitly prohibit scraping. In addition, we observe that even where AI agents actively check the applicable ToS and $\texttt{robots.txt}$, they continue to exhibit low compliance rates. Conversely, in our sample of 269 websites that do not prohibit scraping, we observe that agents sometimes refuse to scrape despite it being lawful, such as Claude Opus 4.6 exhibiting an over-refusal rate of 98.5%. Our human baseline experiments (n = 180 participants) suggest that humans exhibit compliance rates comparable to AI agents but over-refuse at much higher rates.
This paper contains examples of stereotypes about veterans and displaced people that may be offensive or distressing. Social bias benchmarks for language models commonly examine categories such as gender, race, and religion, often drawn from lists of legally protected characteristics. Veterans and internally displaced people (IDPs) fall outside these lists, and stereotypes about them, which can undermine access to employment, housing, and social participation, remain largely absent from language model evaluation. We introduce WarBias, a Ukrainian-English benchmark of stereotypes about these two groups. Built from a taxonomy of 120 stereotype claims grounded in research and monitoring reports, WarBias contains 1,007 stereotype-counter-stereotype-unrelated sentence triplets and 240 multiple-choice questions per language. A Ukrainian language model selected in a blinded human pilot generated the Ukrainian triplets. At least two independent annotators reviewed each triplet in the annotation cohort. The authors revised items in response to the reviews, and the final release was machine-translated into English. Each claim carries a human severity rating. In a baseline evaluation of six models, four prefer stereotypical Ukrainian sentences about at least one group, most strongly about IDPs, while English scores are mostly balanced. WarBias supports complementary evaluations of stereotypical preferences and context-sensitive judgments, and provides paired forget/retain material for unlearning-based bias mitigation. We document the benchmark's construction, intended uses, and limitations, and offer its construction method as a template for analogous benchmarks in other conflict-affected societies.
Multimodal LLMs have made remarkable progress in perceiving the physical world, while their perceptual scope remains largely human-centric, focusing on what humans can directly see or hear. This progress raises the prospect of perception beyond human limits, such as extending hearing to sounds that are inaudible to humans. More broadly, the physical world is full of waves, vibrations, and oscillations, only a small fraction of which are directly audible to humans. This invites the extension of multimodal LLMs toward a broader form of machine hearing, which remains largely underexplored. To investigate this direction, we introduce the Machine Hearing Benchmark (MHB), comprising 6,128 instances across 13 tasks and seven domains. MHB spans highly heterogeneous input regimes, with sampling rates from approximately $10^{-5}$ to $10^6$ Hz, sequence lengths from $10^1$ to $10^6$ samples, and both single- and multi-channel inputs. Evaluations of current open-source and proprietary Audio LLMs (ALLMs) reveal limited zero-shot transfer to machine hearing. Supervised adaptation on machine-hearing data alone does not close this gap, suggesting that the challenge extends beyond data exposure to a systematic mismatch between existing audio front ends and machine-hearing inputs. To address this mismatch, we introduce MachEar, a lightweight encoder extension that complements the pretrained Audio LLM for machine-hearing inputs. The results show that MachEar equips a pretrained Audio LLM with stronger machine-hearing capabilities while largely preserving its existing audio understanding performance.
Continual learning suffers from catastrophic forgetting, leading to the loss of past knowledge and unstable adaptation across task streams. Existing approaches typically address this issue through replay buffers, architectural changes, or additional regularization terms. We argue that forgetting is a consequence of unstable learning dynamics, which may benefit from an optimizer-level solution. The main challenge is that learning dynamics are time-dependent, as demonstrated by empirical phenomena such as the stability gap. First, we formally prove that, for distance-based classifiers, bounded parameter drift guarantees bounded margin erosion under a local score-stability condition. Based on this insight, we propose STABLE, a primal-dual optimizer with adaptive drift control. STABLE enforces a drift budget by increasing the dual variable when the budget is violated and decreasing it otherwise. This results in high stability without penalizing plasticity. Experiments on long-horizon unimodal streams and CLIP-based multimodal streams show that vanilla PEFT baselines trained with STABLE match or outperform specialized continual learning methods. In the multimodal setting, STABLE further preserves zero-shot transfer, matching or exceeding dedicated multimodal continual learning methods. Code is available at https://github.com/atlas-luiss/STABLE.
The design of manufacturing systems relies on discrete-event simulation (DES) models that experts still build in tool-specific interfaces. AI agents can now write simulation code, but every model they write from scratch is new, unverified software. We present SimuLean, an open-source 3D DES simulator on the Unity engine designed to be programmed by AI agents: its modelling semantics are verified once, in the platform, and the agent builds and studies one model through domain-level commands, in the 3D scene or headless; physically simulated durations are calibrated for fast headless studies. We evaluated it on 18 manufacturing modelling tasks, from single machines to published production lines. A frontier agent built valid models in 93 of 108 attempts (86%), in about two minutes each; on the benchmark's interface version, two frontier agents built 105 of 108 models through domain commands alone, so that each comes with a record of how it was built. Every residual failure was silent - a model that builds and runs but answers wrongly - and only comparison with reference models revealed them all. In an industrial case, one agent working alone rebuilt, verified and studied a robotised welding workshop from a published study, in SimuLean and in SimPy. Comparing them with the published, expert-built model exposed an error in it; against the corrected model, the agent-built models agree within 1%. One agent can already build complete, realistic and traceable models for the design of manufacturing systems; independent reference models are what make them trustworthy.
Multimodal Large Language Models (MLLMs) inherit strong textual reasoning capabilities from their language backbones and text pretraining, but often fail to apply the same level of competency to visual inputs. We show that this degradation cannot be attributed to only perception failures, by evaluating models on semantically-equivalent text and rendered-text image inputs: models can accurately read text in the images and recover near-text performance when provided their own transcriptions, but just cannot reason well on the images directly. We characterize this discrepancy as a modality reasoning-access gap, where models fail to route perceived visual content into existing reasoning machinery accessible with text. To bridge this, we propose TIGER (Text-to-Image Gap-targeted training for Enhanced Reasoning), which first mines synthetic modality counterfactuals where a model succeeds on a text problem but fails when provided the corresponding rendered image, and post-trains the model on the image input via Group Relative Policy Optimization (GRPO), along with SFT and DPO variants. Importantly, although trained exclusively on synthetic rendered-text images with no multimodal annotations, TIGER results in model performance improvements across diverse real-world visual benchmarks including geometric diagrams (MathVerse), multi-disciplinary sciences (EMMA, MMMU), complex charts (ChartQAPro), and embodied spatial reasoning (RoboSpatial-Home). TIGER also complements existing state-of-the-art RLVR visual reasoning methods, yielding further performance gains. Analysis involving reasoning-subspace activation and activation patching provide support indicating that TIGER helps to align visual-to-reasoning pathways toward text reasoning states. Our implementation is available at https://github.com/minhngh/Tiger.
Hardware executions of the quantum approximate optimization algorithm (QAOA) are often validated by comparing the device's mean cost with a simulation, a check that can miss what matters for optimization. Under global depolarization, the mean normalized gap and the floor-corrected probability of every set of bitstrings, the optimum included, retain the same fraction of their noiseless improvement over uniform sampling. We test this prediction with depth-2, transferred-angle QAOA circuits for binary subproblems of an alternating-direction decomposition on the IBM Heron processor \texttt{ibm\_marrakesh}, with and without dynamical decoupling and twirling. At 12 qubits, the main-family circuits retained 29\% to 46\% of their mean-gap improvement but only 5\% to 15\% of their probability of the optimum, lower on at least nine of ten instances per run. Near-optimal bitstrings were lost about as much as the optimum, whereas the avoidance of poor bitstrings was retained best. In the first submission, a device-matched noise model reproduced the pattern but overestimated the retention of the optimum five- to eightfold. On a rugged planted family, the device found the optimum in 62\% of 768-shot blocks (noiseless: 98\%), within the range of local heuristics. Validations of optimization circuits should report retention of the optimum and of near-optimal sets.
Visual data are often expressed through multiple complementary modalities (e.g., images and segmentation masks), each capturing different aspects of the same underlying structure. Most existing multimodal methods adopt a top-down paradigm, learning a shared latent variable through joint training from scratch, which limits scalability. We propose Split and Bridge (SnB), a bottom-up framework for multimodal generation that couples pretrained unimodal diffusion models at sampling time. Rather than learning a unified latent space, SnB generates coherent multimodal samples by following a shared diffusion trajectory up to a split point and then branching into modality-specific trajectories guided by a bridge module. This design enables plug-and-play reuse of powerful pretrained models without joint training, making it scalable and easily adaptable. Across standard benchmarks (PolyMNIST, CelebAMask-HQ) and more complex settings (PIE-Bench++, ROCO), SnB achieves a strong balance between sample quality and coherence. We show that SnB naturally extends to downstream zero-shot tasks such as interleaved co-generation, image editing, and style coupling/decoupling. Our results demonstrate that bottom-up generation provides a practical framework that complements existing top-down multimodal approaches. Code is available at https://github.com/AhmadArrabi/SnB
Large language models (LLMs) are increasingly explored for resource management across the Cloud-Edge continuum. However, embedding LLMs directly inside latency-sensitive control loops places inference latency, execution jitter, and hallucination risks on the critical path. This study argues for a different use of LLMs: a continually instruction-tuned intent compiler that runs once per management intent or declared context change, above a fast, deterministic tactical controller that handles all low-latency runtime actions. We formulate intent compilation as a continual instruction tuning problem, because management intents arrive sequentially and accumulate while infrastructure and workload contexts drift. Unlike recent LLM-based management frameworks, our design scores the model's outputs by execution, so that catastrophic forgetting can be measured in the executed objective of the underlying controller. On a continuum simulator with a 2B backbone and parameter-efficient adapters, preliminary results show that naive sequential tuning forgets earlier intents in the executed objective, that replay of stored examples nearly prevents this forgetting, and that the tuned compact compiler performs on par with simpler translators while being more robust than a grammar parser to rephrased intents.
Large language model (LLM) recommenders can produce semantically different recommendations when a user's protected attribute changes, even if the interaction history remains fixed. Controlling this disparity is challenging when model weights and full-vocabulary logits are inaccessible. We propose FAIDER (Fairness-Aware Item-level DEcoding for Recommendation), which incorporates fairness into list construction by selecting complete item candidates rather than manipulating token-level logits or repeatedly repairing finished outputs. The central idea is to separate the operational criterion for a violation from how strongly generation is constrained. Conformal-inspired empirical calibration fixes a cross-group semantic-disparity threshold, while online decoding applies either hard filtering or a soft penalty adjusted through dual-inspired feedback to target a prescribed violation rate. Separately, we establish a pathwise guarantee: under a uniform enforceability assumption, the soft decoder's empirical violation rate exceeds its target by at most $O(1/T)$. Experiments across recommendation benchmarks and LLM backbones show that Soft FAIDER improves both measured semantic fairness and ranking quality over the evaluated prompt-repair baselines. It also generally achieves a more favorable fairness-utility balance than the hard variant, indicating that stricter per-recommendation constraints do not necessarily yield lower measured disparity.