论文库

Within the CG animation production pipeline, the challenges artists tackle in Lighting can be immense. Even minor adjustments require re-rendering massive scenes with slow offline renderers; global illumination has to be sampled and complex shaders have to be evaluated, leading to iteration times of minutes to hours per frame. To accelerate this process, we introduce a novel neural render proxy (NRP) that enables differentiable relighting of static scenes with fixed camera and materials at interactive rates. Our main insight is the decoupling of traditional rendering into path sampling and emission computation. From a single, light-agnostic render pass, we collect light transport data in the form of path samples. This enables rapidly sampling lighting conditions on the fly, and training a scene-specific lightweight neural network that learns how light is transported from any scene location to any image pixel. This approach is compatible with non-differentiable production renderers, induces minimal memory requirements during inference, and scales only with resolution and the number of light sources and parameters, but independently of scene and appearance complexity. Our evaluation demonstrates interactive frame rates for relighting (~30-60 Hz) while closely approximating the visual fidelity of ground-truth path tracing. In addition, the differentiable NRP enables fast, gradient-based inverse workflows, allowing artists to efficiently solve for lighting parameters from intuitive image-space edits or generative targets.

The quadratic complexity of the attention mechanism severely limits the context scalability of Video Diffusion Transformers (DiTs). We find that the highly sparse spatio-temporal attention patterns exhibited in Video DiTs can be naturally represented by the Monarch matrix. It is a class of structured matrices with flexible sparsity, enabling sub-quadratic attention via an alternating minimization algorithm. Accordingly, we propose VMonarch, a novel attention mechanism for Video DiTs that enables efficient computation over the dynamic sparse patterns with structured Monarch matrices. First, we adapt spatio-temporal Monarch factorization to explicitly capture the intra-frame and inter-frame correlations of the video data. Second, we introduce a recomputation strategy to mitigate artifacts arising from instabilities during alternating minimization of Monarch matrices. Third, we propose a novel online entropy algorithm fused into FlashAttention, enabling fast Monarch matrix updates for long sequences. Extensive experiments demonstrate that VMonarch achieves comparable or superior generation quality to full attention on VBench after minimal tuning. It overcomes the attention bottleneck in Video DiTs, reduces attention FLOPs by a factor of 17.5, and achieves a speedup of over 5x in attention computation for long videos, surpassing state-of-the-art sparse attention methods at 90% sparsity. Code is available at https://github.com/EricLina/VMonarch.

Differential morphing attack detection (D-MAD) compares a trusted live capture with a document image that may contain a facial morph. We investigate whether labelled validation data can improve a frozen-parameter vision-language evaluator through prompt selection rather than model fine-tuning. Version~1 generates 50 candidate prompts and selects the top-ranked instruction using a security-oriented loss that prioritises missed morph attacks (MACER) while accounting for bona fide classification errors (BSCER), response failures, parse failures, and decision inconsistency. Version~2 recursively revises prompts using aggregate validation feedback. On a subject-disjoint test set of 160 pairs spanning LMA, MIPGAN-2, and PIPE, the selected Version~1 prompt achieves a (19.17\%) morph attack classification error rate (MACER), with (0.00\%) observed bona fide sample classification error rate (BSCER), compared with (70.00\%) MACER for a controlled manual prompt. The paired (50.83) percentage-point difference has a (95\%) subject-bootstrap interval of ([43.33,58.33]). Recursive revision does not surpass the best Generation~0 validation loss; the reported revised descendant obtains (31.67\%) test MACER. All 23 attacks missed by Version~1 are landmark-based (LMA) morphs. Overall, the results show that prompt formulation can substantially influence frozen-parameter VLM-based D-MAD performance, while also highlighting the challenge of recursive prompt refinement under a compact 10-subject validation setting.

Remote sensing change detection usually relies on large collections of temporally aligned, pixel-level annotations, which limits its scalability across regions, imaging platforms and application domains. Recent zero-shot approaches reduce the need for supervision, but they often localize fine-grained changes poorly because their features are coarse or inconsistent across time. We present SpotDiff, a training-free, plug-and-play framework for zero-shot change detection built by composing complementary vision foundation models. SpotDiff uses DINOv3 for semantically robust features, AnyUp to reconstruct them at pixel resolution, and SAM 3 for text-guided region-of-interest proposals, which an inference-time multi-scale strategy (SWIMS) makes reliable on small objects. Change is then measured by comparing the reconstructed features of the two dates inside the proposed regions, with no learnable components, fine-tuning or change labels. To our knowledge, SpotDiff is the first change detection method to use a universal feature upsampler, which alone adds 13.8 F1 over bilinear upsampling of the same features. On LEVIR-CD, SpotDiff reaches 82.46 F1, 2.06 points above the strongest prior training-free method, and on the multi-class SECOND benchmark it improves building change detection by 10.3 IoU and the six-class average by 2.2 IoU. Qualitative results on PlanetScope (3 m) and Sentinel-2 (10 m) imagery show that the same pipeline transfers across resolutions without modification. Our results indicate that composing modern foundation models is an effective route to scalable, label-free change detection.

The lack of diversity in LM content is widely attributed to the alignment process, but how and where exactly in the pipeline this collapse begins is unknown. We argue that output homogeneity is likely learned during the pretraining phase, and only revealed or magnified during the alignment process. Specifically, we find that semantic convergence is observed from the first alignment stage--the instruction-tuning phase (SFT)--suggesting that homogeneity might already exist in the pre-alignment model. To investigate this, we conduct controlled SFT experiments examining how training data influences output convergence on specific input/output pairs. We find that convergence can be revealed and amplified, but not introduced by the SFT data, supporting its role as a catalyst rather than a cause. To further test whether homogeneity originates before alignment, we measure convergence in base models. We find that instruct-like collapse can be induced through prompting alone, even without alignment. Taken together, our results suggest that semantic convergence may arise naturally from the objectives underlying LM training, making it difficult to mitigate through post-alignment interventions alone.

A small fine-tuned parser that emits a well-formed but wrong tree fails silently: nothing downstream can tell. We ask whether the output format, a free design choice, controls how often this happens. On the PIZZA ordering benchmark we fine-tune two sub-billion models (Qwen3.5-0.8B and LFM2.5-350M) with LoRA on six lossless encodings of the same target tree, from 44.3 down to 10.9 tokens per order, including two matched-length pairs that vary surface form at fixed length. Across 300 runs (5 training pool sizes, 5 seeds) with a hierarchical bootstrap and Holm correction fixed before the results were read, shortening alone never helps, but writing the target in ordinary words does. The word encoding is the most accurate format on both models (87.5% vs. 85.8% exact match for the native s-expression on Qwen) while emitting 4.1x fewer tokens, and it cuts the silent error rate by 2.3x on Qwen and 1.8x on LFM2, separating at every pool size. Error analysis locates much of the gain in negation. All runs also learn to refuse off-menu input (held out: 95.9-99.0% handled) and to apply in-utterance self-corrections. Strict refusal has a measured price: 2.1-10.1% of real, mostly polite, orders are refused.

Decision models answer a question with a typed output, such as a choice among options with a probability for each, so one model can choose an agent's actions, evaluate actions already taken, or score every action to teach a policy with reinforcement learning. Prior work has studied them mainly as actors, judging their usefulness by how well they act. We ask whether a decision model that acts poorly can still teach well. We compare three decision models and three general LLMs as actors and judges on ALFWorld, and use JEV as a teacher on ALFWorld and WebShop. As actors, all six models are pulled away from a subgoal that must come first by an easier one. As judges that see the whole trajectory, the same models credit the steps toward that subgoal, and asking whether each action helped, rather than whether the task will succeed, separates useful from useless actions. As a teacher, JEV raises a Qwen2.5-1.5B policy over GRPO from 72.8% to 82.9% success on ALFWorld and from 56.8% to 71.4% on WebShop. These results suggest that a poor actor can still teach well, because judging an action after the episode ends reveals the consequences that acting must anticipate. Code is available at https://github.com/ZJU-REAL/decision-agents.

This tutorial provides an accessible and implementation-oriented introduction to Gaussian process learning-based model predictive control (GP-MPC), which combines probabilistic residual modeling with receding-horizon control. Its central tutorial contribution is a detailed, step-by-step derivation of multi-step mean and covariance propagation for GP-augmented prediction models. The derivation shows how commonly used propagation formulas follow from the laws of total expectation and total covariance, and clarifies the roles of the GP posterior mean, GP posterior covariance, uncertain inputs, and query-output cross-covariances. Building on this foundation, the article distinguishes mean-only unconstrained GP predictive control from uncertainty-aware constrained GP-MPC, clarifies regulation and output-tracking formulations, and provides concise implementation and computational-complexity guidance, including practical software-tool references and discussion of the scope of closed-loop guarantees. Mobile-robot path-following examples illustrate mean-only unconstrained implementations, whereas mixed-vehicle platooning illustrates uncertainty-aware constrained GP-MPC with uncertainty-dependent safety-distance tightening. The tutorial is intended to help researchers and practitioners understand, implement, and critically evaluate GP-MPC designs for robotic and autonomous systems operating under modeling uncertainty.

Language-guided UAV navigation has largely focused on long-horizon route planning, leaving short-range, object-relative flight maneuvers underexplored. To address this gap, we formulate Flying-on-a-Word (Flow), a task requiring a UAV to convert concise natural-language instruction, egocentric visual observation, and current state dynamics into continuous flight actions. Flow encompasses both primitive motion commands and complex object-interactive behaviors, assessing motion intent comprehension and spatial context grounding. To enable policy learning and closed-loop evaluation, we introduce UAV-Flow, a comprehensive dataset comprising 30,692 expert-piloted real-world trajectories across eight distinct motion categories, alongside 10,109 simulation trajectories in the UAV-Flow-Sim environment equipped with task-specific completion criteria. Furthermore, we develop FlowVLA, which undergoes imitation pretraining on mixed real-sim demonstrations and online post-training via Tail-Aware GRPO (TGRPO). By leveraging group reward distributions, TGRPO effectively reinforces high-reward action sequences. Evaluated on 273 held-out UAV-Flow-Sim episodes, pretrained FlowVLA achieves a mean success rate of 75.58%, outperforming the strongest baseline by 6.72 percentage points, which TGRPO post-training further elevates to 93.16% with superior sample efficiency. Physical UAV experiments demonstrate robust execution of object-relative commands in both indoor and outdoor real-world scenes. Together, Flow establishes a novel benchmark and learning framework for language-conditioned flight maneuvers.

When an LLM agent fails, repair methods choose which action to replace and what to put in its place. They often score candidate repairs with greedy replay, a stronger model, an environment oracle, or a judge that reads the failed log. We ask whether these evaluators pick repairs that help the agent that will actually execute them. Holding failed trajectories and candidate single-step repairs fixed, we label every candidate with independent rollouts of the deployed policy and measure how much deployment value each evaluator's choice loses. On failures not screened for repair value, pre-specified tests show that greedy replay loses 0.08-0.13 in expected net success gain in ALFWorld and 0.06-0.13 in net reward in WebShop, relative to selection with the deployed agent's own rollouts. At an equal rollout budget, selecting repairs for Qwen2.5-7B with Qwen2.5-14B rollouts loses 0.25 in ALFWorld and 0.07 in WebShop, twenty 14B rollouts per candidate are worth no more than two of the 7B agent's own, and fresh rollouts confirm that the two models prefer opposite repairs on selected disagreement cases. The agent's own fine-tuned successor is a worse evaluator still: after one round of training on verified repairs, its rollouts lose 0.55-0.60 when they select repairs for the agent it was trained from, as a pre-specified test confirms. Relative to what selection is worth, substitute models forgo 11-53%; on $τ^2$-bench, where selecting is worth only 0.10, they lose 0.03. Repairs should be verified with the executor that will run them, and repair labels should name their executor, including its training version.

State space models (SSMs) are an emerging alternative to Transformers. SSMs offer linear-time inference and parallel training, but most implementations make the parallel scan efficient by making the recurrence diagonal: each channel of the hidden state is multiplied by an input-dependent real or complex number. These numbers commute, so the state records which inputs occurred but not their order. We formalize this with representation theory, which explains and fixes the state-tracking failures previously reported for SSMs. The state update defines a linear representation of the monoid of input sequences, and we show that an SSM tracks a group at every sequence length exactly when that group is a quotient of the group generated by its transition matrices. Two consequences follow. Diagonal real- and complex-valued SSMs can track only commutative groups, while the transition matrices of any SSM that can track a group must split into independent blocks over the real, complex, and quaternion numbers. We then introduce HyperSSM, which adds quaternion channels while keeping linear-time inference and parallel training. Trained only on final answers, HyperSSM solves the group tasks of recent state-tracking benchmarks exactly with at most five state numbers. This includes $S_5$ with 32 generators, where previous SSMs are at chance, and HyperSSM generalizes to sequences 25 times longer than those seen in training. Because the theory says what the transitions must encode, the learned weights of the linear recurrence can be interpreted directly. We demonstrate this by recovering the calibration of a real gyroscope, the rotations of imperfect qubit gates and the geometry of protein backbones.

Retrieval-augmented generation systems attach citations to attribute claims to their sources. These citations frequently misstate provenance, and standard verification cannot detect this. We measure provenance by counterfactual value substitution: a fact in the retrieved passages is replaced with a false variant and the model re-queried. We find that across five sizes of one model family and a second comparator family, 1.2% to 15.7% of citable claims derive from parametric memory rather than the retrieved context. At the most conservative substitution distance these figures are 4.6% and 8.5% on the two models we tested at it, with the rate rising with model size within a family. We further find that natural language inference verification, the standard approach, gives results indistinguishable from chance across models, even when restricted to the cited passage. We determine that the cause is structural: a model answering from memory selects a citation that agrees with its output. Comparing ten attribution methods on identical claims, the contrast between the model's distribution with and without context best separates provenance. Acting on that signal does not improve provenance honesty at these base rates. The resulting detector is hence still not accurate enough to act on: it measures what the model knew rather than what it used, and separating those remains the open problem.

Multi-teacher on-policy distillation can yield uneven domain capabilities. We propose Domain-Routed Delta Optimization (DRD-MOPD), composing a frozen base with expert-initialized trainable deltas. Domains update their deltas with independent optimizer states and clipping. After joint training, Post-separate exports the base plus one trained delta per domain, selected by known task domain without retraining. Using AdamW, the unweighted full student reaches 32.83% Macro across six benchmarks after 20 updates, versus 32.21% for a matched shared control. Post-separate raises Macro with Open-MOPD weighting.

Embodied interaction requires an agent to integrate multimodal perception with verbal and nonverbal behaviors to effectively communicate its intentions, including speech, body motion, and facial expressions. We present XBot-Omni, a Thinker--Talker--Actor architecture that extends the Omni paradigm to robotic modalities, including body motion and facial expressions, with ``X'' representing the extensibility to robotic-specific modalities. The Thinker processes text, images, audio, and video to formulate contextually appropriate responses, which are then realized by the Talker and Actor through speech and coordinated bodily expressions, respectively. In this report, we innovatively introduce the concept of ``Actor'' as a dedicated component for embodied agents: a coordinated, streaming module that leverages the native representations of the Thinker and Talker within an end-to-end pipeline, introducing only modest latency overhead. This unified architecture enables multimodal generation conditioned on omni-modal inputs directly to prevent information loss in staged designs. We comprehensively evaluate XBot-Omni across seven benchmarks, demonstrating strong multimodal understanding capabilities on both general-domain and robot-centric videos, as well as high-quality, natural, and well-coordinated motion generation. Furthermore, our results validate its capability to generate embodied multimodal responses directly from omni inputs, meeting the diverse requirements of embodied interaction. XBot-Omni enables robots not only to perceive the omni world but also to express themselves through speech and bodily behaviors, extending Omni models from conversational interaction to embodied communication in physical environments.

Robots that learn from demonstrations should extrapolate their policy despite state perturbations such as sensor noise, scene variation, and compounding errors. One approach to this robustness is regularization. Prior regularizers consider each state on its own and limit how sensitive the policy is to any perturbation at that state. By contrast, we explore how the policy's output shifts when given the same perturbation at different states. Our insight is that, locally, the robot should respond to state perturbations in a consistent manner: e.g., if changing an irrelevant object position does not impact the policy at one state, that same change should not shift the policy output at a nearby state. Applying this insight, we develop inner product regularization (InneR). We show that the inner product on action space induces an inner product on state space through the policy Jacobian (the Fisher information for stochastic policies), and that the null space of this inner product governs which state perturbations have a first-order effect on the policy. By clustering states whose inner products are aligned, averaging the inner products within each cluster, and minimizing the smallest eigenvalues of that average, InneR shapes the learned policy so that nearby states share a common null space. InneR enforces consistent responses to perturbations without limiting policy expressiveness, and can be appended to existing imitation learning pipelines for deterministic or stochastic policies. In simulated experiments, InneR makes learned policies more robust to irrelevant state dimensions and shifts between training and test distributions, whereas norm penalties often perform worse than behavior cloning. On a real robot arm performing three manipulation tasks, InneR raises the average success rate of a diffusion policy from 18% to 57% without additional data.

Neural surrogates for computational fluid dynamics (CFD) are trained and ranked on datasets that each vary one design family in one flow domain. Such datasets cannot test the least that an aerodynamic foundation model must do: predict CFD-quality incompressible flow in seconds for new shapes in new enclosures. On DrivAerML, larger models and longer training barely lower Transolver's error; doubling the training cases cuts it by about 12% up to 400 cases, exhausting the dataset. We introduce AeroZoo, a benchmark and training corpus of 3,815 time-mean solutions to the unsteady Reynolds-averaged Navier--Stokes (RANS) equations with a transition model, across seven geometry families, eight enclosure configurations and Reynolds numbers from $10^3$ to $10^8$. Each sample poses a well-posed PDE problem in the form a solver receives it: the boundary meshes grouped by condition type, the prescribed values and the fluid properties. The 2,816 samples with a stable time mean form the canonical benchmark: two tasks, engineering metrics and 22 frozen splits, disjoint in geometry, that hold out geometry families, enclosures, boundary-condition layouts and Reynolds ranges. On the random split, boundary-condition tokens cut GeoTransolver's median velocity error from 76.9% to 8.4% of the inlet speed in single-seed runs (uniform inlet flow: 11.0%). ISLA, a transformer for boundary-driven PDEs, reaches 5.3% over four seeds, about 4 times the change from a finer mesh. On the same test samples, a held-out geometry family barely changes ISLA's error relative to uniform flow, and a held-out enclosure or boundary-condition layout raises it 2.4 to 9.2 times. The data and the scoring code will be released.

The proliferation of user-generated video content has made accurate and robust no-reference video quality assessment (NR-VQA) a highly important challenge. Existing research increasingly relies on architecturally complex multi-branch pipelines or computationally intensive fine-tuning of single-branch backbones. In this paper, we investigate whether a single frozen video foundation model can serve as an effective NR-VQA backbone, and propose HiVE-VQA, the first framework (to our knowledge) to apply a JEPA-based video encoder to this task. HiVE-VQA extracts token grids from multiple depths of a pre-trained V-JEPA 2.1 encoder and stacks them into a hierarchical spatio-temporal volume, which a compact 3D convolutional head fuses across its layer, spatial, and temporal axes, combining low-level distortion cues from shallow layers with the semantic and temporal context captured at greater depths. The encoder remains frozen throughout, so only the head is trained: 5.2M parameters (5.7\% of the total), with no auxiliary multi-modal models and an inference cost that is invariant to video resolution. Trained solely on LSVQ and evaluated without any fine-tuning, HiVE-VQA outperforms 17 state-of-the-art (SOTA) methods in SROCC and PLCC on both intra-dataset (LSVQ-test, LSVQ-1080p) and cross-dataset (KoNViD-1k, LIVE-VQC) benchmarks. These results indicate that hierarchical, depth-wise fusion of features from a single frozen video foundation model provides a more effective and efficient basis for NR-VQA than task-specific multi-network fusion. Code and model checkpoints are available at https://github.com/RobHam99/HiVE-VQA.

Understanding how visual attention influences human preference formation is a fundamental problem in consumer behavior and cognitive science. In this work, we investigate how eye movement patterns can be used to predict preference decisions when subjects compare two alternatives. To address this problem, we analyze an eye-tracking dataset, where subjects view pairs of face images and their gaze behavior is recorded throughout the decision process. We model this problem using multiple machine learning classifiers and apply the Multi-layer Combinatorial Fusion (MCF) framework to iteratively combine diverse models. The MCF framework expands models to a large pool of model fusions and then reduces them based on cognitive diversity, repeating this process across layers until diversity diminishes. Experimental results show that MCF significantly improves performance over individual models. The best single model achieves a Pre@k of 0.90, which is improved to 0.9143 after the first layer of MCF, and further increased to 0.9286 in the second layer. This progression demonstrates that the second layer not only surpasses the performance of the first layer but also outperforms the best individual model, highlighting the effectiveness of leveraging diversity through multi-layer fusion.

We study data poisoning attacks on logistic regression with $\ell_2$-regularization, where an attacker chooses the features of a fixed number of injected observations with prescribed labels to maximize the logistic loss on the clean training data. Strong convexity of the defender's problem yields an exact reformulation through its first-order optimality conditions. Using convex conjugacy of the softplus function, we replace the sigmoid relations by a single convex optimality-gap inequality and express the attacker's objective as a concave function. A change of variables makes the stationarity condition affine, leaving bounded bilinear equalities as the only nonconvex constraints. We derive finite bounds on all variables involved in these equalities, including a global bound on the coefficient norm and tighter bounds obtained by preprocessing based on convex optimization. Experiments on eight binary classification datasets evaluate attack effectiveness and compare the two reformulations when solved with \textsf{SCIP}. Attacks obtained by local optimization degrade predictive performance on most datasets, while the convex-analytic formulation produces substantially stronger reported dual bounds within a fixed time limit.

Humanoid loco-manipulation requires coordinated locomotion, balance regulation, and manipulation under frequent contacts, making teleoperated demonstration collection difficult: a practical controller must simultaneously provide precision, stability, and contact awareness. We propose FORGE, a force-aware multi-granularity control system with a coarse-to-fine command space spanning velocity navigation, planar foot placement, wrist-pose tracking, and optional force-conditioning inputs. Policy training requires no full-body human motion trajectories or joint-angle references; deployment consumes only sparse bilateral wrist poses relative to the pelvis, coarse motion commands, and optional desired wrist-load inputs. For sustained contact, FORGE combines a momentum observer with a learned correction trained using simulator contact-force supervision. At deployment, it estimates low-frequency, wrist-equivalent interaction forces from proprioception and reconstructed low-level arm PD torques, without wrist force/torque sensors. A quasi-static virtual-target loop uses this estimate and an optional desired applied-force input to adapt the wrist target. A frozen locomotion prior and asymmetric residual actions stabilize manipulation-induced whole-body corrections. Experiments evaluate unified-control stability, task-space tracking, quasi-static wrist-equivalent force estimation, squeezing forces during sustained contact with a rigid box, and downstream diffusion-policy execution of a squat-and-lift task across payload conditions. Website: https://bb0928.github.io/forge/.

A foundational principle of physics is that all coordinate systems are equally valid: the behavior of a PDE-governed system does not depend on the origin or orientation of the frame, nor on the units or mesh used to write it down, and neither should a surrogate model. Yet the neural surrogates with the lowest errors on today's aerodynamics benchmarks lack these invariances and fail under simple transformations of the same problem: for models trained on PDEs of cars in one pose (i.e., translation and rotation) and from one mesher, rotating the full PDE raises the surface-pressure error from 4% to more than 100%, and re-sampling its boundary raises it to about 30%. We introduce ISLA (Invariant Slice Attention), a slice-attention transformer for steady, boundary-driven PDEs in which no learned operation receives a coordinate and every sum over the boundary is a quadrature. Positions enter only as distances and relative angles to learned slice anchors, lengths are measured in a unit computed from the boundary, and interior points query the boundary through a tree-accelerated learned kernel without influencing one another. With every boundary cell a token and an exact boundary read, ISLA is equivariant to rigid motions and reflections, invariant to the units of its declared data, and consistent under mesh refinement, while at every setting each query is independent of all other queries. On AeroZoo, 2,816 flows that vary in shape, size, speed, and boundary conditions, its mean reference-scaled velocity and pressure errors are at least 3.1 and 3.3 times lower than any baseline's. On AirfRANS its errors are second only to those of a boundary-integral model (GLOBE) and at least 1.6 times lower than those of Transolver and GeoTransolver.

Large language models (LLMs) show growing promise in specialty medicine, yet existing evaluations largely focus on isolated clinical encounters rather than assessing the distinct clinical responsibilities along longitudinal care pathways. We introduce NPC-RubricBench, a phase-aware benchmark for evaluating multistage clinical reasoning across the cancer care continuum, using nasopharyngeal carcinoma as a defined disease context. Curated from real-world records of 405 patients at two cancer centers, it comprises 941 evaluation units across six sequential care stages, from work-up through follow-up, and 17,688 patient-specific criteria across five dimensions. Each assessment is restricted to clinical evidence available at the corresponding decision point. Across 30 general-purpose and medical LLMs, we find strong diagnostic scores did not ensure reliable downstream care. Decision Quality ranged from 56.8 to 94.7 in diagnosis on a 0-100 scale, but dropped to 8.9-56.0 in staging and spanned 32.4-94.2 in treatment. Crucially, failure modes differed across phases: explicit decision errors predominated in anatomical staging (48.2%), whereas decision omissions were more frequent during initial work-up (33.4%) and post-treatment follow-up (30.4%). Varying query granularity revealed a trade-off between decision quality and other clinical requirements. Broad open-ended requests reduced decision quality across all 30 models, whereas more detailed, checklist-like prompts improved decision quality but were accompanied by lower evidence faithfulness and safety adherence. These findings highlight the need to assess clinical reliability against the distinct clinical requirements of each phase along the longitudinal care pathway. Finally, NPC-RubricBench informs the design of stage-aware clinical guardrails and offers a reusable blueprint for evaluating LLMs across other cancers and chronic diseases.

Managing children on ECMO (Extracorporeal Membrane Oxygenation) requires balancing short-term physiological stability with the long-term goal of preventing neurological injury. Standard reinforcement learning can propose risky, out-of-distribution actions, while imitation learning is limited to reproducing clinician behavior. We introduce Support-Constrained Fitted Q-Iteration (FQI-SC), an offline reinforcement learning framework that cautiously improves upon expert behavior by restricting optimization to empirically supported state-action pairs and defaulting to the clinician baseline in low-data regions. On 78 pediatric trajectories totaling over 6,000 hours of clinical data, we evaluate our framework using Fitted Q-Evaluation (FQE) and a behavioral disagreement. Additionally, we formalize the performance-safety trade-off by introducing a unified Balance Score ($S(π)$). Our empirical evaluation shows that support-constrained policy improves the Balance Score by 38\% relative to FQI, improves estimated value by 1.28 FQE units over imitation learning, and reduces disagreement with clinician actions by 71\% relative to FQI.

LLM agents combine information from multiple principals while acting externally. We formalise privilege escalation (PE) as the system-level security objective: an action may proceed only if every principal that influenced it would allow it. Model-level defences remain probabilistic, so enforcement must sit outside model judgement. We present Conflux, a principal-context defence. Its kernel, ITES, derives authority from provenance and the organisation's existing access-control system and intersects the permissions of every influencing principal. We prove a conditional authority-confinement theorem: under complete mediation, no-laundering provenance, a known non-empty decision context, fresh permission checks, and action binding, the kernel admits no action that escalates privilege with respect to any recorded influencing principal. The theorem covers the represented value-influence fragment; control influence from tool, branch, recipient, or plan-step selection remains a refinement obligation. Conflux extends the kernel with independent policy ports, argument-aware authorisation, disclosure, consent, delegation, restriction, and exposure planning. We describe two system- level evaluators. SLED is a black-box evaluator: it exhaustively explores a declared finite model of a defence up to a bound. SLED-V is a white-box evaluator: it checks a serialisable intermediate representation against security and provenance properties. The evidence is bounded and partial. The appendix reports retained AgentDojo, SLED, and SLED-V results, including the limits of those runs. No result in this paper is an unbounded deployment guarantee.

Hierarchical Reinforcement Learning promises, among other benefits, to efficiently capture and utilize the temporal structure of a decision-making problem and to enhance transfer learning capabilities, but theoretical guarantees lag behind practice. In this paper, we propose a Feudal Q-learning scheme and investigate under which conditions its coupled updates converge and are stable. By leveraging the theory of Stochastic Approximation and the ODE method, we present a theorem stating the convergence and stability properties of Feudal Q-learning. This provides a principled convergence and stability analysis tailored to Feudal RL. Moreover, we show that the updates converge to a point that can be interpreted as an equilibrium of a suitably defined game, opening the door to game-theoretic approaches to Hierarchical RL. Lastly, experiments based on the Feudal Q-learning algorithm support the outcomes anticipated by theory.

Federated learning experiments can produce reasonable-looking results even when the experiment implementation differs from the intended experimental protocol. Small errors in configuration, data partitioning, aggregation, randomness, evaluation, or checkpoint recovery may therefore affect a comparison without causing the program to fail. We present FedBrew, a research codebase for controlled and reproducible federated learning experiments. Its focus is controlled simulation with checks of specified conditions of the intended experimental protocol. FedBrew uses configuration checks, identifiable client data partitions, checkpoint recovery checks, self-describing run records, and development-time checks of scientific conditions. These checks assess, for example, whether client partitions are disjoint and exhaustive and aggregation weights match the prescribed rule; numerical and recovery checks examine non-finite client model states and interrupted runs. The same structure also provides clear extension points and automated checks for new algorithms, datasets, models, and tasks, whether they are added by researchers or AI coding agents. FedBrew therefore supports reproducible and inspectable FL experiments through checks of specified software and scientific conditions and the run records associated with reported results.

Smaller language models (SLMs) make it feasible to deploy clinical natural language processing where large proprietary models are too costly, and parameter-efficient fine-tuning allows them to be specialised on commodity hardware. One obstacle persists after training: on medical tasks these models answer at far greater length than the clinical question requires. In clinical decision support and point-of-care use, where time is short, such verbosity is hard to act on even when the underlying reasoning is sound. Existing length-control methods depend on preference optimisation, reward models, or proprietary teachers that undercut the accessibility making SLMs attractive. We introduce Parameter-Efficient Conciseness Adaptation (PECA), which adds a single differentiable conciseness term to the QLoRA supervised fine-tuning objective. Because the token count is discrete and non-differentiable, PECA instead raises the end-of-sequence probability once generation exceeds a token budget, a smooth signal optimised during ordinary training, with no preference data, no auxiliary model, and a single commodity GPU. On a 1.7-billion-parameter model fine-tuned on the medical-o1-reasoning-SFT corpus, PECA reduces mean response length by 72.8% while improving multiple-choice diagnostic accuracy from 40.8% to 44.0%. A sweep over the conciseness strength traces an interpretable accuracy-to-length frontier for balancing brevity and diagnostic completeness.

Distributed fiber-optic sensing (DFOS) technologies such as Distributed Acoustic Sensing (DAS) generate continuous, high-resolution spatio-temporal data at rates exceeding hundreds of megabytes per second, creating severe challenges for real-time storage and transmission. While neural compression methods can achieve strong rate-distortion performance on sensing data, prior approaches either require per-segment retraining incompatible with streaming, or have only been validated offline. None have been demonstrated end-to-end on a continuously streaming DAS acquisition system at scale. We present STeP-C (Spatio-Temporal Predictive Compression), a neural compression framework that decouples spatial encoding, temporal prediction, and entropy-optimized residual coding, trained end-to-end with a rate-distortion objective motivated by conditional entropy reduction. Critically, we identify entropy coding as the dominant bottleneck for streaming deployment. We address it through GPU-side symbol computation, cached probability tables, and parallel GIL-free arithmetic coding, together with FP16 inference and batched entropy coding, enabling a full real-time pipeline built around a GPU ring buffer. We validate this pipeline on a live DAS testbed, achieving a 98.85% file size reduction at 31.04 dB full-band PSNR and an average throughput of 535 MB/s, exceeding the 324 MB/s requirement for 1 kHz, 81 km DAS streams. We further show that STeP-C substantially outperforms conventional codecs at practical streaming bitrates, can be applied to unseen channel ranges within the same acquisition system without retraining, and largely maintains downstream detector consistency, with 96.62% high-confidence class agreement and 100% Jaccard similarity on detected event types.

When humans hand each other objects, they incorporate both geometric and semantic information into this process. For example, passing a knife with the handle towards the recipient, rather than the blade, is both more ergonomic and safer. Recent state-of-the-art methods for task-oriented robot-human handovers have progressed from modeling object geometry to incorporating object affordances. However, they often forgo predicting the explicit, task-specific hand poses a human selects to utilize an object. Since many objects support multiple interaction modalities, e.g., a claw hammer used to strike or pull nails, this variability must be modeled to achieve robust task-oriented handovers. To tackle this, we propose a novel approach, GENESIS-Handover (GENErative HypotheSIS), which leverages VLM image generation to produce a variety of task-specific hand-object interaction hypotheses. These hypotheses are matched in real time to the observed human hand pose, enabling inference of the most suitable handover configuration. By leveraging VLMs as priors of plausible hand-object interactions, the method produces task-conditioned handover strategies for previously unseen object-task pairs. We evaluate the standalone interaction proposal module before deploying the full system on a mobile manipulator. In a user study with 12 participants across five task-object pairs, 83.3% perceived our method to have better task understanding than the previous state of the art.

DNA-encoded libraries (DELs) are commonly used to find small-molecule binders to a target protein. But the hard part is no longer finding a hit. It is building a machine-learning model good enough to guide the medicinal chemistry that follows, one that can resolve the shift in affinity when a chemist changes a single atom. Such a model needs two things from its training data: dense sampling of the small patch of chemical space around the hit, and an adequate number of genuinely tight binders within it. A diverse naive DEL gives neither. It samples an enormous space, so any one neighborhood is covered far too thinly, and the tightest binders are rare. A model trained on that naive DEL data must extrapolate to binders tighter than any it has seen, and that is where models can fail. One way to get training data with both properties is a second-generation DEL built around the hit. We ask how complex that library must be. We treat binding as a sum of small, independent interactions, called pixels on the target and antipixels on the ligand. From the spread of affinities measured in the naive library, with no structural information and nothing yet synthesized, we estimate the complexity such a library must have to supply that dense sampling and those tight binders. The method does two useful things. It turns the number of tight binders a model needs into the number of compounds a library must contain. And it estimates the expected yield of a randomly designed second-generation DEL, so the gain from designing the library with machine learning or active learning can be measured against a baseline, without the expense of a control library as large as the experiment itself.