Mobile agents are increasingly expected to operate everyday applications from screenshots and language goals, where reliable control requires reasoning over screen affordances, multi-step navigation, and future state changes. However, many agents externalize this computation as long textual chains of thought, which slows interaction, increases supervision cost, and complicates deployment. We introduce MIRAGE, a framework that learns continuous latent reasoning representations from visible textual reasoning traces. MIRAGE transfers explicit reasoning into compact hidden states, enabling the agent to reason internally without decoding long rationales. It also incorporates a generative world-model objective: latent reasoning vectors are aligned with future screenshots, encouraging the agent to anticipate upcoming interface states before acting. This turns hidden computation into both a compressed thought representation and a forward-looking model of environment dynamics. At inference time, MIRAGE reasons in continuous latent space, reducing token generation while improving execution efficiency. On AndroidWorld, MIRAGE matches explicit chain-of-thought supervised fine-tuning in the 4B ablation with a 3-5x lower decoded-token budget and improves a comparable instruction-tuned baseline by 10.2 points; on AndroidControl, it improves action grounding while generating over 75% fewer tokens.

Many current agentic systems and LLM pipelines correct mistakes by optimizing outcome reward. This addresses only the what of failure: when an outcome diverges from prediction, the why and when of the mismatch are not systematically logged, reviewed, or corrected, so the same error can recur episode after episode. We argue that this is a structural problem, not merely a model-capacity one. We propose long-horizon temporal regret as a first-class objective alongside outcome regret and epistemic regret over the working causal model. Temporal regret captures when failure persists: how long a miscalibrated causal model is tolerated before correction. Epistemic regret captures why failure persists: residual uncertainty or error in the working causal model. Together, the three regrets give a falsifiable account of what, why, and when a long-lived agent can fail. Modeling the agent as a stream of E episodes, we prove three conditional results under explicit causal-probing, persistence, and detectability assumptions. First, under observationally equivalent confounding, outcome-only learning cannot distinguish causal from spurious structure without an intervention channel, so temporal miscalibration can persist linearly even after outcome regret is driven to zero. Second, with a persistent causal log and budgeted probes, total probe complexity is logarithmic in the episode horizon, inducing O(log E) temporal regret. Third, under K detectable change-points, the rate extends to O(K log E). We instantiate Trivium and pre-register five falsifiable predictions. On CausalBench-Seq, Trivium follows the predicted logarithmic envelope while outcome-only baselines grow linearly. A pilot real-LLM stream provides preliminary external-validity evidence across one full E = 500 run and three E = 100 frontier-model pilots. Self-learning here means revising an external causal model, not retraining LLM weights.

Modern deep neural networks usually have large parameter scales and nonlinear hierarchical structures, and they have achieved strong performance in computer vision. However, the source of their generalization performance remains difficult to explain using traditional statistical learning theory. Among the factors that may affect visual generalization, data scale, model complexity, and input modalities are fundamental and controllable variables. This study empirically analyzes how these three factors influence model generalization performance. Specifically, in a preliminary experiment, we construct a one-dimensional nonlinear function and vary the number of training samples and the polynomial degree to observe the effects of data scale and model complexity on model performance. In the main experiments, we compare model performance on CIFAR-10 and CIFAR-100 under different training data scales, model architectures, and input modalities. The experimental results show that increasing the training data scale consistently improves generalization performance, whereas changes in model complexity do not provide stable gains. In addition, removing color information degrades model performance, while explicit prior features such as gradients, edges, and wavelets have inconsistent effects across different model architectures. Overall, this study provides an empirical analysis of the relationships among data scale, model complexity, input modalities, and visual generalization performance. Code and experimental logs are available at: https://github.com/YidiZhouluo/DeepLearning-Empirical-Studies/tree/main/Exp_01.

Context-augmented language model systems often wrap supplied content with labels such as Reference:, Evidence:, Instruction:, Note:, or Example:, but the effect of these labels on reader-model behavior remains underexplored. We introduce a paired fixed-content probe over 500 MMLU-Pro items: each item receives the same misleading answer-bearing assertion under different discourse-role labels, and adoption is measured by whether the model outputs the injected wrong option. Across GPT-5.5, DeepSeek V4 Pro, Llama-3-8B-Instruct, and Qwen2.5-7B-Instruct, Misleading Adoption Rate shifts by 56-84 percentage points. Binding or source-like labels such as Instruction: and Reference: produce high adoption, whereas Example: consistently suppresses it. Paired tests, bootstrap intervals, final-instruction ablations, and Qwen final-step log-probability probes support a label-conditioned candidate preference. Boundary probes show where the effect weakens or persists: arithmetic tasks reduce adoption, passage-shaped external context preserves smaller label gaps, short-answer evaluation rules out option-letter copying, and nested-label conflicts suggest that illustrative framing can delimit adoption scope. A 200-case single-author manual audit confirms that the short-answer contrasts are stable under conservative adjudication. The resulting claim is bounded but practical: context-utilization and reader-side RAG benchmarks should report and control wrapper labels, because presentation choices can change measured reliance on supplied context.

Reinforcement Learning (RL) has received increasing attention and adoption in real-world use cases. Most of these systems follow a train-then-fix paradigm, where trained agents do not learn while interacting with the world until performance degrades and retraining becomes necessary. In this position paper, we argue that deploying an agent that is incapable of optimality, but receives an evaluative reward signal, is inherently a continual RL problem. We identify four sources of non-stationarity after deployment that necessitate never-ending learning, and highlight why the best deployed agents never stop adapting. We analyze successful examples of continual RL in the real world, and present the community with the advantages and measures to move away from the current train-then-fix paradigm.

Robots solving generalist tasks need to be able to ground instructions in their past experience, since humans may refer to notable past events when giving a task (e.g., ``Take me to where the chemical spill happened yesterday''). Since memory limits make storing all past events infeasible, long-term robot memory must be selective, ideally retaining only those episodes with high utility for future tasks. However, future tasks are not typically given a priori for generalist robots. To select generically useful memories, we propose Bayesian surprise as a gating mechanism for memory formation. We present an approach to compute surprise in a semantically rich deployment-agnostic latent space provided by V-JEPA-2. Using our gated episodic memory to augment 4D scene graph-based spatial memory, we show a consistent improvement over state-of-the-art benchmarks in robot question answering, outperforming prior robot memory methods by $\geq12\%$ for temporal, spatial, and binary questions, and surpassing the performance of supervised and non-causal methods with an unsupervised causal method in event segmentation tasks.

Scholarly text classification supports literature organization, subject indexing, and research intelligence, but Chinese scholarly corpora often contain imbalanced and semantically adjacent disciplinary labels. We propose AutoTail-BSFGM, a class-balance-aware fine-tuning method that combines an automatically gated tail-prior adjustment, a weak Balanced Softmax auxiliary loss, and Fast Gradient Method adversarial regularization. The method changes only the training objective and procedure; inference uses the same single base-size encoder and linear classifier as the corresponding label-smoothed baseline. We evaluate the method on two CSL-based tasks: an abstract-to-discipline task with 67 labels and a title-to-category task with 13 categories. On the primary abstract task, AutoTail-BSFGM improves validation and lockbox accuracy under both Chinese RoBERTa-WWM and MacBERT-base. With MacBERT-base, validation accuracy increases by 0.83 percentage points and lockbox accuracy by 0.49 points, with a pooled paired McNemar signal on validation (p = 0.023). On the title task, the method improves validation accuracy by 0.70 points and validation balanced accuracy by 2.64 points; lockbox accuracy is approximately neutral while lockbox balanced accuracy improves by 1.22 points. The results support a bounded contribution: AutoTail-BSFGM improves class-balance-sensitive behavior and yields consistent gains for abstract-based scholarly classification, without uniformly improving every metric on every split.

Multimodal retail agents should not only recognize what a customer is doing, but also decide whether and how to assist before an explicit request is made. We study this setting through the See--Infer--Intervene (SII) framework, where a device must see pre-interaction behavior, infer latent customer intent, and act by selecting an appropriate service intervention or choosing to wait. We instantiate SII with the Proactive Intent World Model (PIWM), which represents customer state with AIDA (Attention, Interest, Desire, Action) purchasing phases and BDI (belief, desire, intention) psychological fields, predicts action-conditioned intent transitions, and selects from five response classes: Greet, Elicit, Inform, Recommend, and Hold. We further construct GuidanceSalesBench, a smart-retail benchmark containing state manifests, pre-interaction videos, candidate responses, action-conditioned outcomes, and best-action labels. When conditioned on ground-truth customer state to isolate action selection, PIWM achieves 0.641 macro F1 on 30 held-out target videos, outperforming a zero-shot Qwen2.5-VL-7B baseline and training variants without balanced action supervision; end-to-end video-only selection drops to 0.295, below the 5-class balanced random baseline of 0.414, identifying video-to-state grounding as the dominant deployment-time bottleneck. A preliminary staged real-store pilot (recorded with paid participants performing scripted customer behaviors) reaches 0.579 action macro F1 on 20 fully annotated videos, with 10 additional accessible videos released with index-level labels.

Post-training pruning compresses large language models to high sparsity using a small unlabelled calibration set, and recent work has concluded that the choice of calibration source has only modest impact on averaged post-pruning accuracy. We ask whether this conclusion survives once calibration impact is evaluated separately across distinct capability dimensions rather than aggregated. Decomposing post-pruning capability into General, Commonsense, Code, and Math, and analysing $n{=}15$ calibration sources via Spearman correlations between OIT information metrics and per-dimension retention, we uncover an opposite-sign trade-off: calibration perplexity correlates positively with General retention ($ρ{=}{+}0.71$) but negatively with Math and Code retention ($ρ{=}{-}0.53,\,{-}0.59$; $p{<}0.05$), so no single source can preserve all capabilities. We respond with multi-source calibration mixing, and propose IGSP, an information-guided self-calibration protocol that automates multi-source construction without capability-aligned corpora by minimising 4-gram aggregation and balancing perplexity across dimensions. On LLaMA-3.1-8B at SparseGPT 60% sparsity, a uniform multi-source mix reaches 58.8% total retention, outperforming the best single source (MetaMath, 50.0%) by $+8.8$ and the C4 default (40.0%) by $+18.8$; IGSP improves over Self-Cal by $+2.4$ and SGS by $+4.8$.

Inference-time scaling has emerged as a critical avenue for enhancing Large Language Models' performance, yet real-world deployment is constrained by strict computational budgets. In this work, we formulate inference budget allocation as a global constrained optimization problem governed by economic principles. By modeling per-query reasoning utility with a shifted-surge function, we derive an optimal allocation policy based on a global shadow price that equilibrates marginal utility under resource scarcity. Based on this theory, we propose Constrained Latent-utility Equilibrium Allocation for Reasoning (CLEAR). It performs rational abandonment and reallocates resources from insolvent queries to solvable queries near their emergence thresholds. Extensive experiments on several reasoning tasks with different traffic streams demonstrate that CLEAR significantly improves the Pareto frontier of total token cost versus mean accuracy. In resource-scarce regimes, CLEAR achieves up to a 3x improvement in global accuracy compared to uniform allocation.

Large language models (LLMs) exhibit strong natural-language reasoning abilities for clinical decision support, but struggle to effectively model structured longitudinal electronic health records (EHRs). In contrast, EHR foundation models can learn predictive patient representations, yet lack interpretable language-based reasoning. To bridge this gap, we propose ChatHealthAI, a multimodal reasoning framework that aligns structured EHR representations from a pretrained EHR foundation model with the semantic space of a frozen LLM through a task-aware resampler. By integrating longitudinal patient representations with refined clinical event descriptions, ChatHealthAI enables clinically grounded natural-language reasoning while maintaining accurate patient prediction. We evaluated ChatHealthAI on three clinical predictive tasks from the EHRSHOT benchmark. Results show that ChatHealthAI improves reasoning quality and interpretability while preserving competitive predictive performance. These findings highlight the potential of integrating EHR foundation models with pretrained LLMs for interpretable clinical prediction.

The success of the transformer architecture as the backbone of modern LLMs is in large part due to its use of attention layers. An attention layer follows the standard neural network paradigm: it takes the residual stream as input and thereby produces context-dependent query, key, and value vectors. However, we find that model performance meaningfully improves when deeper layers learn only a context-free value vector to preserve the original token information, without drawing on any context from the residual stream. When the model has access to this context-free value vector, adding back the context-dependent component provides little additional benefit for aggregate benchmark performance. Such context-free value vectors can be stored as sparse model parameters, eliminating the need to recompute or persistently cache these values. Through systematic ablations on the key design choices for such context-free value vectors, we propose Bank of Values (BoV), a new way of computing value vectors in attention by learning a lookup table of token-specific value vectors for each of the last third of layers. Across 135M and 780M models, BoV improves validation loss over standard attention and, at 780M, the average score across 21 benchmarks, matching the previous best method that adds token information to the value vector with less compute and memory.

Generalization remains a central bottleneck for vision-language-action (VLA) models: under distractors, appearance shifts, and semantically similar tasks, the policy must often infer local execution details from coarse instructions while also deciding which parts of the image matter for control. We present S2 (See Less, Specify More), a framework for improving VLA generalization by training the executor under a cleaner interface. Specify More preserves the original instruction as a stable high-level goal while relabeling each trajectory into refined trajectory- and subtask-level language that disambiguates the current execution mode. Unlike native attention, See Less imposes an explicit visual evidence budget, training the executor to act from task-sufficient evidence rather than unconstrained visual context, without any region or mask annotation. This interface lets the executor follow detailed guidance without relying on distracting visual patches or resolving avoidable ambiguity on its own, and it remains compatible with off-the-shelf VLM planners through in-context learning. Across our main evaluation settings, S2 improves overall generalization metrics by changing the executor's learning problem: coarse instructions induce avoidable supervision aliasing, goal-preserving local guidance outperforms instruction replacement in our main ablations, and explicit evidence budgeting reduces dependence on broad visual context beyond efficiency considerations. Across eight real-robot tasks on TX-G2 (an AgiBot G2-compatible variant) and HSR, S2 raises mean subtask success from 54.2% to 79.0% over pi0.5. Together, these results suggest that VLA generalization improves when the executor is trained to act from informative local guidance and task-sufficient visual evidence, rather than recovering both from weak supervision.

Agricultural robots are serving as powerful assistants across a wide range of agricultural tasks, nevertheless, still heavily relying on manual operations or railway systems for movement. The AgriVLN method and the A2A benchmark pioneeringly extended Vision-and-Language Navigation (VLN) to the agricultural domain, enabling a robot to navigate to a target position following a natural language instruction. However, almost all the prior methods adopt an ideal assumption that the given instructions themselves are correct, which does not align with the realistic scenarios, because anybody may say an instruction with mistakes. To bridge this gap, we propose the A2A-MI benchmark, in which we build a semi-automatic data annotator to insert three mistake classifications into each original instruction in a more diversified and efficient way. We test several state-of-the-art agricultural VLN agents on it and observe a sufficient drop with -57% on SR and -9% on NE, from which we suggest that an agricultural VLN agent tends to assume that the given instruction is correct, so does not have the awareness to doubt it when the scenes it sees do not align with the instruction it receives. To build the awareness on instruction mistake, we propose the IMAC module analyzing the instruction and the current front-facing image, to judge whether the instruction has mistakes and attempt to correct it when needed. We integrate IMAC into the baseline model, and observe a noteworthy improvement, sufficiently narrowing the gap to the performance on instructions without mistakes. Project: https://github.com/AlexTraveling/IMAC-AgriVLN.

Bayesian optimization (BO) is a popular and effective approach for tuning expensive, noisy experiments, but requires the formulation of an explicit objective function. Preferential BO (PBO) removes this requirement by learning from pairwise human feedback, yet existing methods struggle to efficiently optimize beyond low- and medium-dimensional problems due to their global search approaches. We address this limitation by developing a family of local PBO methods that transfer key ideas from high-dimensional BO to the preferential setting. In particular, we introduce local PBO methods which adapt trust-region and derivative-informed local search to pairwise preference feedback, where the latter exploits first- and second-order derivatives of the Laplace-approximated GP posterior. Our benchmark on GP sample paths, standard optimization benchmark functions, and policy-search tasks shows that local PBO methods are especially effective in high-dimensional and complex landscapes with steep optima. Compared with global preference-based baselines, they can substantially reduce cumulative regret, making them particularly useful for real-world preference-based optimization tasks such as policy search.

Underwater acoustic classification has a wide array of oceanic applications, but faces challenges due to an increasingly complex acoustic environment. Waveform and spectrogram representations have been primarily used as acoustic data features for classification tasks in this domain. Spectrograms model harmonic dependencies, but these reduced representations can filter out acoustic features relevant for discrimination. While phase information from the waveform allows full characterization of the signal, the original waveform can be noisy and complex, rendering this representation difficult for models to process directly. This paper proposes a dual-encoder neural architecture to simultaneously process acoustic waveforms and spectrograms, leveraging pre-trained backbones and parameter-efficient fine-tuning modules, enabling a domain adaptation. To combine these adapted branches, a novel differentiable fuzzy aggregation mechanism based on the Choquet integral is introduced to balance the temporal and spectral representations. This fusion strategy not only yields higher classification accuracy but also provides interpretability. Specifically, by analyzing the learned fuzzy measures, insights are revealed about class-specific shifts in the network's representation reliance. By dynamically shifting attention to the representation least corrupted by potential asymmetric channel distortions, the proposed gating mechanism mitigates the non-stationary challenges of the underwater environment. Evaluations on the DeepShip and ShipsEar datasets demonstrate that the proposed architecture achieves classification improvements over independent single-encoder baselines, while simultaneously restricting the trainable parameter space. This mitigates the risk of overfitting on limited acoustic datasets while alleviating the computational costs associated with fully fine-tuning foundation models.

End-to-end manipulation policies, combined with web-scale pretrained Vision-Language Models (VLMs), show the promise for generalizable and dexterous robotic manipulation. However, they inherit two key limitations from 2D foundation models: 1) the reliance on 2D RGB inputs that ignores the intrinsically 3D nature of manipulation; and 2) the lack of spatial 3D alignment between input-output spaces as well as across diverse robot embodiments, camera setups, and trajectory datasets. In this paper, we present a series of contributions to address these issues. First, we introduce aligned vertex map and vertex spectrum -- a pixel-wise 3D representation that elevates 2D visual inputs to 3D, using camera calibration and optional depth. This novel input representation marries 3D awareness with the generalization of 2D large VLMs. Then, we propose to align the inputs and outputs of manipulation policies by expressing per-pixel 3D information of each camera view and robot actions to a shared coordinate. Based on this, we designate a canonical Bird's-Eye-View (BEV) alignment frame and innovatively propose to construct BEV images, producing a view-invariant representation robust to camera pose variations. To enable training and evaluation at scale, we develop a comprehensive data processing pipeline to perform such alignments; we also introduce a novel temporal alignment scheme for trajectories across diverse robots, human operators, and datasets. These contributions collectively mitigate input and output spatial-temporal misalignments, improving the consistency and generalization for real-world manipulation. Pretrained checkpoint, source code and data processing pipeline are available in https://hnuzhy.github.io/projects/Dex-BEV.

Large language models (LLMs) are increasingly asked not only to write static interfaces, but to construct executable interactive worlds from natural language. Browser-native 3D, commonly built with Three.js, is a natural next frontier: generated programs must integrate assets, obey spatial and physical constraints, and keep user-facing controls synchronized with hidden runtime state. Existing web-generation benchmarks and evaluators, however, largely observe only pixels or DOM nodes, while the mechanics of a Three.js world unfold inside an opaque <canvas>. We introduce WorldCoder-Bench, a benchmark for autonomous, physically grounded 3D world synthesis. WorldCoder-Bench contains 2,026 expert-curated tasks across Simulation, Rendering, and Application scenarios, with optional .glb assets and hidden behavioral contracts. We further propose StateProbe, an execution-based protocol that probes generated programs in a sandboxed browser and verifies hidden, mutation-hardened contracts over runtime states and transitions. Beyond verification coverage, we report Return on Automation and Time Efficiency Multiplier to measure correctness-adjusted cost and time savings. Across nine frontier models, the best system reaches only 27.8% verification coverage on WorldCoder-Core and 19.9% on WorldCoder-Robust, with failures dominated by state-schema drift and broken interaction chains rather than missing scene elements. Utility metrics further show that cheap or fast models can still provide substantial value on easier domains. WorldCoder-Bench is available at https://anonymous.4open.science/r/WorldCoder-Bench/.

As LLMs are increasingly used to draft public-facing arguments, they may flatten public debate by repeatedly introducing the same polished, plausible arguments. We study argument collapse, the tendency of essays generated by different LLMs to converge to a smaller set of main arguments, sub-arguments, and paragraph-level structures. We compare 1,039 human responses from 195 New York Times (NYT) debates, 448 human responses from 61 longer-form Boston Review (BR) forums, and 23,384 LLM-generated essays. In the NYT corpus, 65.3% of human main arguments are unique within a debate, compared to 3.4% of LLM main arguments. Asking LLMs to generate diverse answers adds variation, but a typical model recovers only about half of the distinct human main arguments, with much of the added variation falling outside the observed human argument space. Collapse also appears in sub-arguments, where among essays with the same main argument, 41.0% of human sub-arguments are unique versus 9.1% from LLM responses. Qualitatively, LLMs often reuse generalized and hedged sub-arguments, while humans prefer more concrete and topic-specific ones. Structure-wise, LLM-generated essays tend to follow a more fixed arc, often opening with a direct claim and moving quickly toward proposals. The same patterns hold in longer BR essays, suggesting that argument collapse extends beyond short-form responses.

Large language models are increasingly used in multi-agent systems, where they see and respond to other agents' answers. A key risk is conformity: a model may abandon its own answer simply because others agree on a different one. Prior studies show that LLMs often revise toward a majority answer, but it remains unclear whether these revisions help correct mistakes as often as they introduce new errors. In this paper, we conduct a controlled study in which an LLM first answers a question, then sees simulated peer responses before making a final decision. We manipulate two social cues: consensus structure and authority labels assigned to peers, and measure how they influence beneficial and harmful revisions. Across four open-weight LLMs and seven QA datasets, we find that peer agreement makes it much easier to mislead initially correct models than to correct initially wrong ones. Authority labels make models more likely to choose the endorsed answer, regardless of whether it is correct. More concerningly, generic reasoning interventions such as chain-of-thought and reflection do not reliably reduce harmful revision while preserving beneficial revision. These findings suggest that multi-agent LLM systems should verify peer answers rather than simply aggregate them.

Agentic reinforcement learning (RL) enables LLM agents to improve continuously from environment rewards, yet the resulting policies do not systematically accumulate reusable strategies that generalize across tasks. Modular skills can provide such reusable strategies, yet existing skill-augmented RL methods decouple skill creation from policy optimization, risking adopting skills that conflict with the evolving policy. Inspired by Anthropic's Skill Creator, we introduce ReSkill, an RL-in-the-loop skill creation framework that reconciles skill evolution with policy learning. ReSkill exploits the group-wise structure of GRPO to naturally embed three mechanisms with only marginal additional overhead: (1) an assertion-driven skill creator that diagnoses failures from past experience and proposes conditional, trigger-based skill revisions; (2) within-group rollout sampling that enables controlled comparison of skill versions, capturing which version best supports the policy's ongoing learning; and (3) Thompson Sampling with adaptive discounting to balance exploration and exploitation in skill version selection as the policy evolves. Across several domains, ReSkill consistently outperforms existing memory and skill-based RL methods, with the largest gains on unseen tasks. Analysis of the skill lifecycle shows skills being automatically created, tested, refined, and pruned as the policy improves, demonstrating reconciled skill-policy co-evolution.

Sparse high-dimensional representations are conducive to uncovering nontrivial structures in unsupervised exploration of data. Such a representation can deal with the dense connectivity in graphs relevant to community detection problems. However, sparse high-dimensional representations are capable of doing more, including manifold tiling and feature learning. Conventional algorithms optimize in the space of computationally intractable completely positive matrices or relax the problem to the space of doubly nonnegative matrices that scale with sample size in a way rendering them impractical for large data sets. Some of these methods also impose a row sum constraint, such as double stochasticity. Row sum constraints have the added advantage of being shift-invariant, in the context of manifold tiling. Constraints on the row sum of output similarity matrices require nontrivial online learning rules. Addressing these needs, we propose a versatile online biologically plausible learning algorithm capable of learning sparse shift-invariant representations, useful for clustering, manifold tiling, or sparse coding, depending on the data structure.

High-quality, large-scale synthetic data from simulations is becoming a cornerstone for pushing the capabilities of robot algorithms. While aerial robotics simulators have evolved to support specialized needs such as fidelity, differentiability, and swarms independently, a unified platform that can synthesize data across all these domains is missing. In this work, we propose Crazyflow, a simulator designed to push the limits of aerial-robotics algorithm development, from model-based to data-driven methods, gradient-based to sampling-based approaches, and single-agent to multi-agent systems. Compared to existing state-of-the-art drone simulators, it achieves speeds more than an order of magnitude faster for a single drone and can simulate thousands of swarms of 4000 drones each. Real-world experiments show Crazyflow supports both analytical-gradient-based policy learning, achieving sub-centimeter trajectory tracking accuracy without domain randomization, and sampling-based obstacle avoidance at speeds exceeding half a billion steps per second. Breaking the traditional train-then-deploy paradigm, we show that its unprecedented speed even enables in-flight reinforcement learning; we demonstrate this by throwing a physical drone into the air and training a recovery policy from scratch in 0.38 seconds, successfully stabilizing the drone. Crazyflow supports multiple levels of simulation abstraction, is directly compatible with all open-source Crazyflie models, and enables rapid reconfiguration across custom drone platforms and applications by providing a light-weight system identification pipeline. By pushing accuracy, speed, and differentiability simultaneously, Crazyflow serves as an open-source resource for synthetic data generation, with emerging capabilities for large-scale parallelization for online, in-execution learning and optimization, opening the door to novel algorithm development.

While prompt-based parameter-efficient continual learning mitigates catastrophic forgetting by isolating task-specific prompts, this isolation also limits later tasks from improving earlier ones, leaving backward knowledge transfer underexplored. We address this limitation by proposing Selective bAckward refinement for positive Backward knowledge transfER (SABER), a replay-free framework that enables controlled backward transfer in prompt-based continual learning. SABER determines when backward refinement is beneficial using complementary task-correlation criteria based on prompt-gradient geometry and loss-distribution similarity, and how to perform refinement safely by restricting updates to non-interfering directions in the prompt parameter space. Extensive experiments across multiple continual learning benchmarks and diverse pretrained backbones, including T5-Large, LLaMA, and Qwen, demonstrate that SABER consistently achieves positive backward transfer while maintaining strong overall average performance. Code is available at https://github.com/OptMN-Lab/SABER-ICML-2026/.

Hard negative mining has become the dominant strategy for training retrievers, yet it faces intrinsic limitations: negatives are bounded by corpus availability, selected by retriever score rather than diagnostic value, and increasingly contaminated by false positives as the retriever improves. LLM-based synthesis offers a principled alternative, where negatives that are unconstrained, targeted, and free from false positive risk. But we show that naively incorporating generated negatives into contrastive learning often degrades retrieval performance. We identify and formalize the root cause as a generative-discriminative gap: LLM generation optimizes for fluent, plausible text, while contrastive learning demands strategic violations of relevance at the decision boundary. Our analysis reveals two compounding failure modes: discriminative-agnostic generation, where the LLM lacks an explicit model of query information needs and defaults to generic or topic-drifted text that provides no contrastive signal; and source-dependent shortcuts, where distributional artifacts enable the model to distinguish negatives by origin rather than relevance, causing gradient drift that actively corrupts optimization. To close this gap, we propose CausalNeg consisting of two main modules: (1) CoT-guided counterfactual perturbation for data construction: decomposes why a document satisfies a query into explicit information requirements, then surgically violates individual requirements to construct negatives with controlled, interpretable hardness. (2) Query-view entropy maximization during training: disperses generated negatives across the similarity spectrum, minimizing the mutual information between source identity and similarity scores to suppress shortcut exploitation. We make our code publicly available at https://github.com/mzhangzhicheng/CausalNeg.

Vision-language navigation (VLN) for UAVs demands grounding free-form instructions into 6-DoF flight under partial observability. While Vision-Language-Action (VLA) models excel at semantic reasoning, they suffer from brittleness due to geometric inconsistency and dynamics mismatch. To address this, we propose ImagineUAV, an imagination-driven framework leveraging cascaded world-action modeling. Instead of direct regression, ImagineUAV employs a latent video diffusion model to generate instruction-conditioned future observations, explicitly imagining environmental evolution, from which 6-DoF motions are inferred via an action extractor. A kinodynamic planner then refines these estimates into collision-free trajectories. Additionally, a step-distilled inference pipeline ensures real-time execution. With only 1.3B parameters, ImagineUAV outperforms prior VLN and VLA baselines on benchmarks and real-world flights, validating the practicality of imagination-driven aerial navigation.

Preference alignment has substantially improved the observable behavior of large language models, yet it remains unclear what alignment changes internally. Aligned systems still fail under jailbreaks, prompt injection, and retrieval-time corruption, suggesting behavior-level evaluation alone is incomplete. Post-training should leave measurable traces in internal computation. We ask: when an instruction-tuned (IT) model becomes a preference-aligned (PA) model, what geometric structure changes, where do those changes concentrate, and how selectively do they vary across concepts, prompts, and model families? We introduce MENTIS, a geometry-first framework for measuring alignment-induced internal reorganization in paired checkpoints. MENTIS compares IT and PA models using a primary layerwise covariance-based torsion norm (T1), a secondary spectral torsion diagnostic (T2), and an Energy-Radiance-Activation measure (ERA) for depth localization. Across four 7-8B model pairs on LITMUS, our study reveals that alignment-induced change is selective rather than uniform: normative concepts exhibit larger torsion shifts than factual concepts on average; torsion is negatively correlated with contextual entropy; and peak effects localize to architecture-specific mid-to-late layers. The same pattern appears across word-level, prompt-level, and model-level analyses. These results suggest preference alignment leaves structured, depth-localized geometric signatures in internal computation beyond what behavior-level evaluation alone can reveal.

Generative models for volumetric medical images have found many applications in medical imaging, ranging from data augmentation to serving as priors for inverse problems. For these applications, generating high-resolution 3D images with strong controllability is essential but remains highly challenging. Existing approaches typically control generation either through radiology reports used as text prompts or through full image segmentation. While text-based prompting is flexible, it provides limited spatial control over the location, shape, and boundary of abnormalities. In contrast, segmentation-based methods receive precise spatial guidance but are restrictive in requiring full-organ annotations. In this work, we propose a flexible multimodal framework for controllable volumetric image generation that supports input from radiology reports and segmentation prompts (both optional). Our approach allows users to provide segmentation of a specific anatomy or abnormality without requiring full-organ annotations. The semantic meaning of the segmentation mask is specified through an accompanying text description, resulting in a highly flexible and scalable conditioning mechanism. We develop a memory-efficient architecture based on a modified diffusion transformer that jointly processes image and segmentation tokens. The model further incorporates gated attention to effectively attend to long radiology reports. Experiments demonstrate that our method achieves state-of-the-art perceptual and semantic scores (e.g., 24% relative improvement in mean FID), generates high-resolution anatomically consistent CT volumes, and improves data efficiency when used for data augmentation. Radiologists' evaluation further confirms strong alignment between generated and real medical images.

Strategic classification (SC) investigates scenarios where agents manipulate their features to obtain favorable decisions from predictive models. Existing fairness-aware SC approaches primarily focus on group fairness and typically assume that agents respond independently. However, when individual fairness is required, ensuring similar individuals receive similar outcomes, agents' manipulation becomes interdependent: an agent's preferred manipulation depends on the neighborhoods' outcomes. This induces a mismatch between classical SC formulations and fairness-aware decision settings, where independent models no longer accurately characterize strategic manipulations. To address this issue, we introduce individual fairness-aware strategic classification (IFSC), a framework that models peer-driven manipulation arising from individual fairness, where agents imitate nearby positively decided peers to obtain favorable outcomes. IFSC characterizes strategic manipulation as similarity-based imitation toward visible accepted peers and learns classifiers under the resulting post-manipulation distributions. To account for uncertainty in peer observability, IFSC employs a robust learning process that introduces stochastic perturbations during manipulation simulation. Experiments on synthetic and real-world datasets demonstrate that IFSC improves individual-fairness consistency and mitigates imitation-induced distortions.

Recent advancements in video-based world models have demonstrated an unprecedented ability to synthesize high-fidelity visual sequences. However, a fundamental gap persists between visually plausible video generation and the functional requirements of a world model, particularly in maintaining a stable and reasonable internal state over extended temporal horizons. While existing benchmarks primarily emphasize visual quality, motion coherence, and text-video alignment, they largely overlook memory, the core capability of a world model to preserve consistency across long-term horizons and complex interactions. To address this gap, we present \textbf{MBench}, a comprehensive benchmark dedicated to quantifying and evaluating the memory capability of video world models. We systematically decompose the memory capability of video world models into three hierarchical and complementary core dimensions: entity consistency, environment consistency, and causal consistency, which are further refined into 12 quantifiable sub-dimensions for comprehensive characterization of long-term memory. Our benchmark is built upon rigorously curated real-captured long videos, and evaluated by rule-based quantitative matrices and VLM to enable objective and comprehensive consistency assessment. Extensive evaluations of mainstream state-of-the-art video world models reveal critical systemic limitations of existing methods in long-term state retention, providing a standardized benchmark and clear research direction to advance the field.