Deep learning models for dynamic survival analysis (DSA) achieve strong predictive performance by incorporating longitudinal patient data, but their black box nature limits clinical trust and adoption. Existing explainability methods cannot handle longitudinal, irregular inputs and functional survival outputs simultaneously, which limits their usability in DSA. We propose DynSHAP, a SHAP framework suited specifically for dynamic survival analysis. It extends common marginal SHAP estimators to this setting by treating time--feature pairs as players in the Shapley game. We further introduce Temporal DynSHAP, which learns linear dependencies in features over time and uses conditional sampling to address them in explanations. When applied to synthetic data with known ground-truth attributions, Temporal DynSHAP recovers temporally dependent features more accurately than marginal estimators for a given state-of-the-art model. Applied to two real-world clinical datasets and two DSA architectures, DynSHAP produces attributions faithful to model learning, allowing medical experts to see which patient information drove the prediction and when.
Nicolás Segura-Kunsagi, Juan Luis Suárez-Díaz, Iván Sevillano-García, Francisco Herrera
Machine unlearning aims to remove the influence of specific training data from a trained model without retraining from scratch. In real deployments, however, deletion requests arrive over time, calling for incremental class unlearning (ICU): classes are removed sequentially, previously forgotten knowledge must remain forgotten, and the remaining classes must be preserved. We present DIstributional SynthetiC rEplay for Retain-free iNcremental unlearning (DISCERN), an intervention-based class unlearning method that replaces access to the retain set with synthetic embeddings sampled from per-class parametric distributions. Unlike prototype-based replay, which assumes an isotropic Gaussian around each class centroid, DISCERN fits Euclidean (Gaussian, Laplace, Student's $t$) and directional (von Mises--Fisher, Fisher--Bingham) families to each class and selects among them through statistical tests and, when none of them accepts, cross-validated log-likelihood. This design is motivated by the architecture-dependent latent geometry: transformer embeddings lie approximately on a hypersphere and concentrate around class directions, unlike those of convolutional networks. We evaluate DISCERN against seven retain-free baselines on MNIST, CIFAR-10, and CIFAR-100 with two convolutional and two transformer backbones, under single-class, multi-class, and incremental unlearning. DISCERN is competitive when classes are removed at once and is the most stable method in ICU on transformers: on ViT-B/16, after 20 sequential deletions it retains a harmonic mean of $0.797$, whereas the strongest method in the simultaneous setting drops from $0.918$ to $0.048$. A paired ablation shows that per-class distribution selection substantially reduces the resurfacing of previously forgotten classes, acting as a stabilizer of accumulated forgetting rather than a uniform improvement.
Gautam Gupta, Dikshant Kukreja, Unique Kumar, Avni Mittal, Deepanshu Dabas, Kshitij Sah, Erik Cambria, Rajiv Ratn Shah 等 10 位作者
Quantization and pruning reduce the cost of deploying language models, but their effects are usually reported as benchmark scores. Such scores record only the loss that remains at the output. Inside the model, compression damages some components more than others, and the rest of the network can partly compensate. The observed loss therefore reflects both the damage and this compensation, which benchmark scores cannot separate. How the two combine, and whether their balance changes with model size, remains unclear. We study this compensation for in-context retrieval, a capability carried by a small, identifiable set of attention heads. To separate the two, the expected loss from ablating these heads is decomposed exactly into the part restored by the rest of the network, attributed to the components responsible, and the part that reaches the output. This decomposition is then compared across model sizes. For damage of comparable size, smaller models restore most of the loss, whereas larger models retain most of it, because the compensating responses offset less of it as models grow. Compressing the heads produces the same pattern. Compression that is nearly harmless in small models can therefore be costly in large ones.
Huanran Chen, Huaqing Zhang, Xiao Li, Yinpeng Dong, Ke Shen, Jun Zhu
The foundational capabilities of large language models are acquired during pretraining on internet-scale, highly heterogeneous data mixtures. In this work, we investigate an interesting geometric question regarding the converged state of pretraining: Does the model converge to a common minimizer across all data sources (e.g., Fig. 2b), or merely a minimizer of the summed loss (e.g., Fig. 2a)? We hypothesize that the geometric "closeness" of task-specific minima is intrinsically linked to downstream generalization. We reveal that standard optimizers (e.g., AdamW) often converge to points where task-specific minima are distant from each other. To address this, we propose the Nexus optimizer, which encourages the closeness of these minima by maximizing gradient similarity during optimization. Experiments across models ranging from 130M to 3B parameters, various data mixtures and hyperparameter schedules, show that Nexus significantly boosts downstream performance, despite achieving the same pretraining loss (see Fig. 1). Notably, on the 3B model, Nexus reduces the out-of-distribution loss by 0.012 and yields up to a 15.0% accuracy improvement on complex reasoning tasks (e.g., GSM8k). This finding challenges the reliance on pretraining loss as the sole proxy for model evaluation and demonstrates the importance of implicit biases in unlocking downstream generalization.
Hallucination remains a persistent challenge for Multimodal Large Language Models (MLLMs), severely limiting their reliability in high-stakes applications. Existing evaluations, predominantly based on static benchmarks, suffer from narrow taxonomical coverage and rapid performance saturation, failing to reflect model robustness in evolving real-world scenarios. To bridge this gap, we present a systematic evaluation framework integrating a comprehensive benchmark with self-evolving stress testing. First, we introduce UniHall, a fine-grained dataset grounded in a unified taxonomy spanning Object, Instruction, and Knowledge dimensions. Second, to address benchmark saturation, we propose Self-Adaptive Multimodal Fuzzing (SAMF), a self-adaptive framework that employs evolutionary mutation strategies to explore the boundaries of model hallucinations. Crucially, to ensure reliable assessment of dynamic inputs, SAMF incorporates a structured metric suite driven by an ensemble of multi-modal oracles. Our extensive experiments reveal that state-of-the-art MLLMs exhibit significant performance degradation under fuzzing compared to conventional settings, exposing a dissociation between reasoning capabilities and factual grounding. Furthermore, we identify a helpfulness-hallucination trade-off, where reinforcement learning alignment inadvertently exacerbates sycophancy in instruction-following tasks. The framework, code and benchmark are available at https://github.com/LanceZPF/EvalHall.
Micheline Bénédicte Moumoula, Serge Lionel Nikiema, Tailia Malloy, Abdoul Kader Kaboré, Jacques Klein, Tegawendé F. Bissyande
Maintainers sometimes cite a low token-similarity score as evidence that an LLM-assisted rewrite of a copyleft project is independent of the original. We test what such a score can support under \emph{source-conditioned} rewriting: a coding agent keeps the original in its working copy, rewrites the repository, and iterates until a fixed behavioural harness passes, with no human edit to the code. On 69 rewrites of six copyleft Python projects we find that: \ding{182}~Agent-only rewriting is inexpensive and often yields low scores. Every controlled rewrite passes the harness after repair, usually within half an hour for the initial rewrite. Under a ground-up rearchitecture prompt, most rewrites fall below 25\% JPlag similarity and still pass the harness. \ding{183}~A single low score is an unreliable witness. Identical project--agent--prompt configurations produce wide spreads in similarity; run-to-run variation dominates project and agent effects, and in more than a third of configurations the three repeats straddle the threshold. Rewrites retain more of the original's imports and fingerprint facts than unrelated projects do, yet hand-written independent implementations of the same task often retain as much, and neither signal separates the two groups better than chance. Thus a low score on one rewrite is weak evidence of independence: another run of the same agent may score high. Claims based on similarity should report repeated generations, and candidate provenance signals should be checked against same task independent implementations, not only against unrelated projects.
We develop a state discrimination approach to unassisted quantum communication through a noisy quantum channel. Starting from the operational duality between entanglement recovery and state discrimination, we construct entanglement-generation codes with explicit decoders constructed from Petz recovery maps. The resulting error bounds hold for every number of channel uses and are expressed in terms of Petz-Rényi coherent information. For product sources, they yield exponential error decay at rates below coherent information.
Tabular Foundation Models (TFMs) adapt to a new task only through the labeled rows in their context. However, labels are expensive to acquire in real-world scenarios due to the need for highly specific domain knowledge to ensure high-quality data. Active in-context learning transfers classical acquisition rules to this setting, in which the model that guides acquisition changes with every label added. We ask how the value of diversity, predictive uncertainty and target-aware information gain changes as a frozen TFM acquires more context. For that, we study Target-Aware Nearest Neighbor (TANN) acquisition and the Expected Predictive Information Gain in Context (EPIC) and evaluate them against classical rules. Our study across twenty benchmarks with two well-established TFMs and classical GBDTs highlights three findings: (i) When: diversity-based acquisition is best during the first acquisition iterations; (ii) Where: labeling pool rows that serve the target inputs via a simple target-aware nearest-neighbor rule matters substantially; (iii) How: EPIC yields the best models at the label budget, retains on average 80--87% of the chance-corrected full-data performance with only 5-10% of the pool labeled on the eight benchmarks with the largest pools, and selects labels that also improve the predictive performance of gradient-boosted decision trees.
Kun Yuan, Hao Ding, Yiqing Shen, Gasper Podobnik, Benjamin Killeen, Tito Porras, Axel Krieger, Mathias Unberath
Vision foundation models are the backbone of modern surgical image analysis, yet most surgical systems still rely on encoders pretrained on natural images or on specialist models trained from scratch on comparatively small datasets. Pretraining encoders on large amounts of surgery-specific data has the potential to generally enhance the performance of surgical image analysis methods simply by providing stronger backbones. Here we present InnerLogic-Surg, an open family of surgical vision encoders (21M--300M parameters) obtained by continually pretraining a transformer-based image encoder on millions of surgical images curated from 11,965 internet-sourced surgical videos spanning ~2,900 hours and more than fifteen procedure types. Our training recipe inherits DINOv3's pretraining stage to preserve representation ability after surgical adaptation, and additionally applies a learned feature-upsampling module for high-resolution dense prediction at low token cost. Across a benchmark covering workflow, instrument and anatomy recognition, segmentation, detection, and monocular depth, under the same architecture, the largest InnerLogic-Surg model achieves the best linear-probe performance in all 20 surgical tasks. We demonstrate that surgical continual pretraining yields transferable representations that generalize across procedures, anatomical regions, and task types, showing our surgical image encoders as a promising and scalable alternative to the procedure-specific representation learning approaches that have traditionally dominated surgical computer vision.
Loop invariants are essential for deductive program verification, yet automatically inferring them for {multiphase programs}, whose loops contain conditional branches that induce multiple execution phases with distinct behaviors, remains an open problem. Existing LLM-based methods either reason about loops as a whole or use explicit branch information to generate atomic assertions jointly across branches, without maintaining complete candidate sub-invariants for independent generation and targeted repair. We present CLOVER, a structured neurosymbolic reasoning framework for multiphase invariant inference built on two interlinked mechanisms. First, compositional inference: CLOVER statically decomposes a loop into symbolic {loop branch sequences}, conjunctions of the loop guard and branch conditions that characterize distinct execution scenarios, and guides the LLM to generate candidate sub-invariants for each one independently, so that each LLM query focuses on a single execution scenario rather than the entire loop. Second, counterexample-guided localized refinement: when SMT verification rejects a composed disjunctive invariant, CLOVER traces the counterexample back through the branch structure to identify which specific loop branch sequence caused the failure and repairs only that sub-invariant, preserving all correct components. Both mechanisms are unified by a shared symbolic structure derived from the loop's branch semantics, which guides what the LLM generates, how candidates are composed, and where refinement is directed. We evaluate CLOVER on 316 linear arithmetic, 50 nonlinear arithmetic, and 54 multiphase benchmarks. CLOVER solves 35 of 54 multiphase benchmarks, outperforming the best LLM-based baseline (29) by 20.7\% and the best traditional method (22) by 59.1\%. On linear/nonlinear benchmarks, CLOVER remains competitive, solving 355 of 366 programs overall.
Yijiong Lin, Wenhao He, Toby Jack, Max Yang, Haoran Li, Nathan F. Lepora
Robotic simulators that represent shear within tactile sensing are becoming increasingly important for contact-rich manipulation. However, simulating contact shear with high-fidelity deformable or soft-body simulation is computationally demanding, limiting scalability to large-scale reinforcement learning. Here we introduce \textit{TactileLab}, a scalable multimodal tactile simulation framework built on IsaacLab, with a particular focus on a new optical-flow-based tactile representation termed \textit{simulated tactile flow}. Instead of relying on soft-body physics to model shear deformation, TactileLab uses rigid-body simulation to efficiently estimate dense tangential motion at the contact interface, producing tactile flow fields that encode shear-relevant contact dynamics. This flow field complements geometric features from a contact depth field. We further introduce a novel policy architecture for jointly processing long-horizon proprioceptive signals and rich tactile observations. Across multiple contact-rich tasks, policies trained using simulated tactile flow learn efficiently at scale and exhibit emergent contact-aware behaviours. To facilitate sim-to-real policy deployment, we then propose a novel real-to-sim tactile translation method that maps real tactile image streams to simulated tactile flow and depth fields, together with a new tactile data collection method, and validate it across diverse tactile sensor designs. Using this pipeline, we demonstrate real-world deployment on tasks encompassing dual-arm and dexterous in-hand manipulation. Overall, TactileLab provides an end-to-end sim-to-real framework for scalable tactile policy learning, enabling diverse shear-sensitive manipulation across heterogeneous tactile robotic systems.
We give deterministic polynomial-time constructions of vertex multicut-mimicking networks of quasi-polynomial size. Given a graph G, a terminal set T, and a positive integer c, our algorithm constructs a graph with |T|^7 c^{O( log c ) } vertices that preserves every minimum vertex multiway cut value of at most c, for every partition of the terminals. Setting c=|T| gives an exact vertex multicut-mimicking network with |T|^{O( log |T| ) } vertices, resolving an open question of Wahlström [ACM Transactions on Algorithms 2022], who showed a bound of κ^{O( log^3 κ)} for the edge multicut-mimicking variant, where κ:= \sum_{t \in T} |N_G(t)| is the capacity of the terminal set, which was later improved to κ^{O( log^2 κ)} by Anand, Lee, Li and Saranurak [STOC 2024]. As applications of our vertex multicut-mimicking network, we obtain deterministic kernels of size k^{O( log k )} for Vertex Multiway Cut, p^{O(1)} k^{O(log k)} for Vertex Multicut with p requests and |Γ|^{O(1)} k^{O( log k)} for Group Feedback Vertex Set, where k is the deletion budget and Γis a finite group. Our first kernelization result, in particular, resolves an open question by Wahlström [ACM Transactions on Algorithms 2022], which was reiterated in a recent survey on kernelization by Jansen [Computer Science Reviews 2026]. Also, to the best of our knowledge, our kernel for Vertex Multicut is the first one whose size is completely polynomial in p and (quasi-polynomial) super-polynomial only in k. Finally, again using our vertex multicut-mimicking network for graphs, we obtain a hyperedge multicut-mimicking network for hypergraphs, which preserves hypergraph edge-multicuts of size up to c, of size |T|^{O(1)} c^{O(log c)}. Our result removes the dependence on the rank of the hypergraph from the exponent of the size bound, that existed in prior work.
Estimating causal effects from observational network data requires accounting for both individual treatment and neighborhood exposure. Existing outcome models commonly incorporate these intervention variables into a joint predictor and estimate causal effects through prediction contrasts, leaving the corresponding response components implicit. We propose the $\underline{N}$etwork $\underline{E}$ffect $\underline{D}$ecomposition $\underline{Net}$work (NEDNet), which parameterizes the conditional outcome response as a query-invariant pre-treatment baseline plus separate treatment- and exposure-response pathways. Path-specific neighborhood contextualization supplies local pre-treatment information to the two pathways. We characterize the fitted response through a unique reference-normalized decomposition, which is equivalent to a zero mixed treatment--exposure contrast. Experiments on semi-synthetic benchmarks derived from BlogCatalog and Flickr show that NEDNet achieves the lowest PEHE among the compared methods for individual, peer, and total effects across homogeneous and covariate-heterogeneous response settings.
Large language models increasingly judge code, yet without external feedback such judgments are unreliable, and verifying them by executing tests or by extended reasoning is costly. Instead, execution-free correctness predictors use signals from the model's output probabilities, sampled solutions, or hidden states to identify judgments that warrant further verification. However, these predictors may degrade under distribution shifts when generalizing across projects and languages. We compare six execution-free predictors from four families on Qwen3-Coder-30B and gpt-oss-20b under benchmark shift between LiveCodeBench and BigCodeBench and language shift across Python, C++ and Java. We evaluate two targets: whether the model's own verdict on a program can be trusted, and whether the program itself is correct. Besides established approaches, we adapt two novel causal localization methods to code, which have been shown to be more robust out-of-distribution on natural language tasks. For verdict correctness, causal localization improves robustness to language shift. For Qwen3-Coder a probe on the causally localized subspace stays 0.09 AUROC ahead of output confidence under language shift (0.11 in distribution), significantly outperforming the best non-causal predictor by 0.04 AUROC on average. This advantage, however, does not generalize consistently to gpt-oss or to benchmark shift. For program correctness, predictors transfer more successfully between benchmarks, largely by identifying problems on which the model tends to fail. However, their ability to rank candidates within the same problem is limited.
Armed conflict can increase the value of Generative AI (GenAI) while undermining the conditions required for its responsible use. Yet the perspectives of community-based humanitarian responders working under these conditions remain underrepresented in responsible AI research. We examine GenAI use in Sudan's Emergency Response Rooms (ERRs), a community-led humanitarian network, through an exploratory survey of 89 respondents and 15 semi-structured interviews with GenAI users and non-users. War reconfigures GenAI's role, making it most useful where its outputs are hardest to verify. ERR responders turned to GenAI when conflict made human expertise difficult or dangerous to reach. However, participants cautioned that GenAI could not independently establish rapidly changing local conditions, and that relying on its outputs could lead to harmful decisions when human verification was difficult. Risks arose before GenAI was used. For some volunteers, carrying a phone or accessing the internet could expose them to danger. Others were concerned that sharing sensitive beneficiary or operational information with GenAI could put people at risk. Participants feared that AI-assisted reports could undermine their credibility with donors, whose funding sustains the food and medical aid ERRs provide. Access to GenAI was uneven, and volunteers using it sometimes relied on colleagues with relevant expertise or local knowledge to verify its results. Our findings show that responsible use of GenAI during conflict depends not only on model performance or individual adoption, but also on whether responders can safely access these tools, verify their outputs against local conditions, and recognize when limiting or avoiding GenAI is the more responsible choice.
Reinforcement learning (RL) promises to unlock capabilities beyond imitation learning for Vision--Language--Action (VLA) models, but its requirement for massive real-world interaction prevents direct deployment on physical robots. Recent work attempts to use learned world models as simulators for policy optimization, yet closed-loop imagined rollouts inevitably suffer from hallucination and long-horizon error accumulation. Such errors not only degrade visual fidelity, but also mislead policy optimization by providing unreliable learning signals. We propose WoVR, a reliable world-model-based RL framework for post-training VLA policies. Instead of assuming a faithful world model, WoVR explicitly regulates how RL interacts with imperfect imagined dynamics. It improves rollout stability through a controllable action-conditioned video world model, reshapes imagined interaction to reduce effective error depth via Keyframe-Initialized Rollouts, and maintains policy--simulator alignment through World Model-Policy co-evolution. Extensive experiments demonstrate that WoVR enables stable long-horizon imagined rollouts and effective policy optimization, achieving superior LIBERO performance and consistent real-world gains across multiple robotic platforms. These results show that world models can serve as practical simulators for RL when hallucination is explicitly controlled. Additional visualization results are available at https://wovr-rlinf.github.io.
Panoramic video depth estimation is pivotal for applications such as Virtual Reality and World Models. However, advancements in this field are impeded by two primary obstacles: the scarcity of large-scale training data and the unique spatiotemporal challenges of Equirectangular Projection (ERP), which hinder the direct transfer of perspective models. In this paper, we first present PanoCARLA, a large-scale synthetic RGB-D panoramic video dataset, featuring natural motion trajectories and drone-like roaming perspectives. Building on this foundation, we propose PVDepth, an end-to-end framework adapted from perspective video depth models. To tackle ERP-specific geometric distortions and consequent non-linear temporal dynamics, we introduce two core mechanisms: (1) A Progressive Sphere-aware Noise Initialization strategy that anneals the noise distribution from planar to spherical, guiding the model to adapt to non-uniform information density; and (2) A Cube-rectified Temporal Modeling module that incorporates an auxiliary cubemap temporal branch to rectify non-linear temporal dynamics in the ERP domain. Extensive experiments demonstrate that PVDepth achieves superior performance, generating geometrically accurate and temporally consistent depth sequences. Code and data will be released at https:// github.com/ChuanxinSong/PVDepth.
Hao Hong, Yanhui Zhong, Bei Zhang, Shuo Xu, Minchuan Tan, Quansheng Zang, Hussein Rappel, Stéphane P. A. Bordas
Moving loads in semi-infinite media generate extended wave fields while complex local structures require fine geometric and spatial resolution. This paper develops a projection-coupled two-and-a-half-dimensional (2.5D) THB-GIFT and quadtree-SBIGA computational formulation that assigns independent discretizations to these regions. The near-field THB-GIFT representation separates the geometry and analysis spaces and permits hierarchical local refinement, while the bounded propagation domain is discretized by quadtree-SBIGA with unified Bézier trace extraction and connected to the homogeneous unbounded exterior through a scaled-boundary impedance. On a common physical integration partition, an $L^2$ projection maps the nonmatching traces; Hermitian-adjoint force transfer then gives a virtual-work-consistent interface reduction. The regional dynamic operators and moving-load response reconstruction follow a single Fourier convention. Homogeneous half-space comparisons and a cylindrical-cavity analytical solution verify the components and the complete coupled system. In the cavity study, hierarchical THB refinement reduces the displacement error in the target inner band by a factor of approximately $41$. At a specified frequency--wavenumber point in the cavity benchmark, a comparison with matched outer-boundary discretization gives approximately $15\%$ lower sampled full-domain error for the coupled model, using $34\%$ fewer system degrees of freedom and $48\%$ fewer S-elements; the warm per-sample times are similar. A layered pavement with a buried conduit and twin modified-D tunnels demonstrate open, closed and multiple disconnected nonmatching interfaces and the analysis of local structural demand and dynamic interaction under moving loads.
Yiyang Zhang, Yuli Zheng, Bin Li, Meimei Zhang, Qiang Sun
Large language models use auxiliary conditioning to incorporate additional information into time-series forecasts. However, auxiliary conditioning does not consistently improve accuracy and can increase forecasting error when its influence is poorly matched to the observed temporal context. We propose Forecast-Aware Residual Refinement (FARR), which learns to correct forecasts using endpoint evidence. FARR obtains a valid forecast under the standard condition and a null forecast under a predefined null condition. Their difference reveals how conditioning changes the prediction. A shared residual refiner combines this endpoint evidence with historical features to estimate the remaining error of the valid forecast. Horizon-wise gating then controls the contribution of the predicted residual to the final output. Experiments on four long-horizon forecasting benchmarks show that FARR achieves the best dataset-average MSE and MAE among the compared methods. Code will be released at https://github.com/blossomrichzyy-star/farr-forecast-aware-residual-refinement.
Building a playable world requires more than generating plausible video: it demands responsive control, persistent spatial structure, and coherent interactions with dynamic entities. Using generated frames as the sole recurrent world representation ties these requirements to view-dependent appearance synthesis, making it difficult to maintain long-term spatial and temporal consistency in generation and interaction. We present PlayGen2, a real-time interactive world model that separates learned 3D state evolution from neural rendering. PlayGen2 maintains a persistent world memory comprising static geometry represented by a volumetric occupancy grid and identity-indexed 3D states for the player and other entities. Within a fixed-size local region, its world-state model jointly learns action-conditioned entity dynamics, geometry completion, and entity discovery, committing updates to a persistent world memory. A separate neural renderer synthesizes first-person observations from depth and semantic projections of the world state, without feeding generated images back into state prediction. Using Doom as a testbed, we demonstrate that the model successfully learns world dynamics including geometry-aware player motion and player--NPC interactions. The framework is computationally efficient, enabling the complete simulation and rendering pipeline to achieve approximately 160ms backend latency on a MacBook Pro with an M4 Pro chip. Beyond single-player interaction, we demonstrate zero-shot deployment on unseen Doom maps, multiplayer interaction through the composition of single-player models, and programmable gameplay through rule-based state interventions. These capabilities demonstrate the transferability and extensibility of the learned game-world model. Code and pretrained model weights are publicly available.
Hongyang Su, Beibei Kong, Lei Cheng, Chengxiang Zhuo, Zang Li
LLM-based recommender systems increasingly generate chain-of-thought (CoT) text before each item prediction to activate the LLM's reasoning capacity, costing tens to hundreds of extra tokens per recommendation at inference. We ask whether inference-time reasoning artifacts (textual or latent) can be replaced by training-time supervision alone, without loss of accuracy. We propose LatentRec, a framework for LLM-based generative recommendation that replaces inference-time reasoning artifacts with training-time per-step head supervision. LatentRec has three components. A 2.5M-parameter Continuous Thought Module iterates K latent reasoning steps over the LLM's hidden state during training. Soft Token Selection bridges the ~14x norm gap between hidden states and input embeddings. A Per-step Reasoning Loss (PRL) supervises every intermediate state through the LLM's prediction head against the first token of the target Semantic ID. The trained LLM retains this accuracy without the module at inference, so the module can be discarded at deployment. On three Amazon benchmarks, LatentRec improves HR@10 by 4.0-5.0% (relative) over a CoT-distilled baseline at 18.3x lower latency (batch=1). Per-step training supervision consistently improves accuracy across three datasets, paid once at training rather than at every recommendation.
Can Cebeci, Nikolaj Bjørner, George Candea, Clément Pit-Claudel
SMT solvers make automated verification convenient. At the same time, solvers suffer from instability, whereby seemingly inconsequential changes to the input may cause a previously quickly produced proof to fail or time out. This paper addresses a common cause of outcome instability (i.e., unsat/unknown fluctuations) in the context of program verification. We demonstrate that the generation (i.e., depth) accounting used to make quantifier instantiation practical and implemented in multiple state-of-the-art SMT solvers is non-confluent (i.e., prone to divergence), and that this leads to instability. We then address this deficiency and develop a new accounting method that is stable. The new method is justified using a sequence of refinements from an abstract, confluent solver model all the way to our implementation in Z3. Our empirical evaluation demonstrates that our implementation reduces outcome instability by 94% in the unstable core of the Mariposa benchmark without leading to performance regressions.
Kaixiang Mai, Jiaqi Yu, Yuzhu Ji, Yiqun Zhang, Zhijie Wang, Haijun Zhang, Yiu-ming Cheung
Accurate segmentation of 3D anatomical structures is crucial to medical image analysis. Recently, Mamba-based approaches have achieved promising segmentation performance in capturing long-range dependencies with linear-time computational complexity. However, heuristic selective scanning strategies inherently disrupt the intrinsic spatial hierarchy and increase computational redundancy, ignoring the inherent sparsity of 3D anatomical structures. To this end, we propose OcTreeMamba, a hierarchical framework that incorporates Octree structures into SSM, achieving 3D topology-aware long-range context feature propagation and integration. Specifically, a Structure Learner performs anatomy-adaptive structured sparsification through hierarchical gating and Gumbel-Softmax sampling, enabling differentiable selection of anatomical regions. The selected regions are subsequently organized by a Dynamic Octree Builder into a multi-level topology that preserves their hierarchical connectivity. Based on the constructed Octree, a bidirectional Tree-SSM propagates and integrates contextual features along connectivity-aware leaf--parent--root and root--parent--leaf trajectories. Extensive experiments on four benchmark datasets show that our proposed OcTreeMamba achieves competitive performance \wrt topological continuity and computational efficiency, particularly for recovering the complex structure on vascular datasets. The code is available at https://github.com/MikeMak123/OcTreeMamba.
Statically scheduled accelerators such as Groq's Tensor Streaming Processor (TSP) achieve low-latency large language model inference by stripping dynamic synchronization hardware from the chip. With no caches, arbiters, or hardware schedulers, the compiler must fix the exact slice and issue cycle of every operation, so when an operation runs is a condition for correctness. Existing kernel languages cannot target such machines because they rely on hardware to resolve cycle timing and instead encode schedules implicitly in code structure. We present TauLang, a kernel language and compiler that makes time a first-class citizen of the program. The programmer schedules operations, loops, and composed kernels with parameterized explicit times, so exploring schedules means changing parameter values rather than rewriting code. TauLang validates explicit times algebraically in TauIR, its timed intermediate representation. It then solves tensor placement and cycle scheduling jointly in an integer linear program, using explicit times to group connected operations into timing clusters that prune the search space. On a cycleaccurate model of a TSP-like machine, joint search achieves a $1.40\times$ latency speedup across ten kernels over a space-first flow. A few explicit time pins keep latency within one cycle of the untimed schedules and scale joint search to 1,636 GEMM blocks, whereas untimed search finds no schedule at 32.
Hoang Tran Vuong, Ngoc Quan Pham, Hung Bui, Fatih Porikli, Dat Quoc Nguyen
Deep research agents answer complex questions through repeated search. They rely on a harness that runs their tool calls and returns evidence to their context. This creates a trade-off: Search works best over long passages, which give enough context for reliable semantic matching. But the returned evidence should be short, because every observation stays in the agent's context for the rest of the trajectory. We introduce CORE, a training-free harness that combines coarse-to-fine search with context management. CORE retrieves passages with a dense retriever. It then reuses the token-level similarities from a late-interaction reranker to return only the most relevant snippet from each passage. Context management lets the agent expand the context around sources it has already retrieved and suppresses repeated snippets. In controlled closed-search comparisons with the Tongyi DeepResearch and OpenResearcher harnesses, CORE achieves state-of-the-art answer accuracy and evidence recall across three agent models and three benchmarks (BrowseComp-Plus, BrowseComp-Wikipedia and BrowseComp-Plus-CM), on corpora containing up to 553M documents.
Dephasing is a prominent source of decoherence in quantum communication, degrading phase coherence and transmission reliability. Therefore, knowledge of the physical channel parameters for reliable transmission is imperative, motivating the integration of quantum communication and sensing at the physical layer. Consequently, we study the transmission of coherent states over a von Mises dephasing channel affected by thermal noise. We consider a heterodyne receiver, for which, we first derive the symbol-error probability (SEP) and an accurate closed-form approximation, through which an irreducible asymptotic error floor due to the dephasing channel is identified. We note that the derived SEP is affected by the mismatch between the receiver's mean phase estimation and the true mean phase, thereby formalizing the interplay between sensing accuracy and communication reliability. In addition, we derive the Fisher information (FI) for the mean phase acquisition of the channel and provide closed-form lower and upper FI bounds. Moreover, the quantum FI (QFI) of the channel is studied, and lower and upper bounds on the QFI are given, which become asymptotically tight as the photon number increases. Despite the initial complexity of the FI and the QFI, the bounds given are relatively simple and effective. In addition, by comparing the FI with the QFI, we show that heterodyne detection approaches the quantum sensing limit as the number of photons increase, while an unavoidable sensing floor exists due to dephasing. Then, using the SEP and the FI analysis, we discuss the conditions under which communication and phase estimation requirements are jointly feasible. Finally, numerical results validate the derived SEP, estimation bounds, and the sensing-communication feasibility tradeoff.
Nischay Dhankhar, Alisher Amantay, Zhanibek Datbayev, Darkhan Akhmetov, Yash Shah, Nurmukhamed Ubaidullayev, Raman Saparkhan, Dos Baha 等 9 位作者
Autoregressive approaches with large language models, such as long chain-of-thought reasoning, have become the default for new AI applications because they reach higher accuracy on a wide range of tasks. In production, however, they give inconsistent answers, require parsing of outputs, and add cost and latency. Decision models address this: given a question and a set of options supplied by the user, they return a probability distribution over the options without autoregressive decoding. We present Drex, a family of four open decision models with 8B (DLM), 9B (v1.5), 35B with 3B active (v2.0), and 27B (Max) parameters. Each request consists of a state (a document, record, or agent context, optionally with images) and questions about it, each asking for a choice among options, a yes/no judgment, or a score on an ordered scale. Training combines supervised fine-tuning with reinforcement learning from adjusted feedback (RLAF), plus vision tuning for Max. Ablations show that RLAF improves accuracy and that consistency losses improve robustness to option order. On Decision Index 0.4.1, Max sets a new state of the art among the decision models we compare (67.52 on the public suite vs. 60.70 for Jev 1.14.0). On Decision Bench v4, it is statistically tied with d3-turbo and GPT-6 Sol and trails Claude Opus 5.5 by 1.6 points, and it has the highest accuracy among decision models on DecideBench. Max answers a typical request in 183 ms, against about 2.0 s for GPT-6 Sol and 3.5 s for Claude Opus 5.5, and its outputs are stable under option reordering. All four models are open source: https://github.com/nace-ai/drex-decision-models
Zhiquan Liu, Haolin Gao, Bo Wu, Jingjing Guo, Heng Sun, Jianfeng Ma
The Controller Area Network (CAN) bus lacks built-in security mechanisms, making real-time intrusion detection necessary. However, resource and latency constraints limit vehicle-side intrusion detection systems (IDSs) to lightweight models, whose detection performance and generalization can degrade under non-independent and identically distributed (non-IID) data. Meanwhile, uploading original CAN messages to the cloud for more powerful model inference risks exposing sensitive vehicle information. We propose CANMentor, a large-small model collaboration framework in which a cloud-side large language model (LLM) teacher supervises vehicle-side lightweight IDS models for CAN bus intrusion detection. Each vehicle first performs local inference, and a privacy-aware generator then produces proxies from the original CAN messages for cloud-side teacher inference to protect data privacy. The vehicle uses the teacher prediction only when the response arrives within a prescribed latency bound and otherwise retains the local prediction; all returned teacher logits are converted into soft targets for periodic knowledge distillation. On non-IID partitions of the ROAD dataset and the HCRL Car-Hacking dataset, knowledge distillation yields maximum relative Macro F1 improvements of 25.01% and 20.79%, respectively. Teacher predictions also improve detection accuracy for CAN messages near the local decision boundary by 38.71-42.00 percentage points.
Most recommender systems treat user preferences as a single signal to infer and optimize, yet human preferences are a mix of a stable preference and situational preferences that turn on and off depending on context. Through Tasteply, we externalize this preference structure as manipulable spatial objects. Tasteply represents each situational preference as an attractor and the stable preference as an underlying global field. Users can adjust their feed by moving around the space and adjusting the strength of the global field, turning the preference structure that remains latent within the model into something they can see and act on directly to steer the recommendation algorithm. Through an exploratory study (N=5), we found that this spatial control was effective, along with its limits and future direction as a means of manipulating preference-based recommendations.
Hao Wu, Yuqi Li, Yizhou Zhao, Kun Wang, Xiao Luo, Yu Cheng, Bo An, Yingli Tian 等 10 位作者
Robot world models predict action-conditioned future visual states to support robotic planning and decision making. However, visual reconstruction does not fully capture task progress, semantic evaluation alone provides limited constraints on fine-grained visual errors, and homogeneous predictions limit comparative feedback. We present RoboImagine, a framework for post-training autoregressive robotic world models through candidate-future comparisons with semantic and visual feedback. Diverse search promotes diversity among candidate trajectories under shared visual observations, action sequences, and task instructions. A frozen multimodal evaluator provides task-relevant semantic preferences through pairwise video comparisons in both presentation orders, complemented by pixel- and perceptual-level visual consistency feedback that constrains local structural and appearance errors. These signals drive policy optimization through group-relative advantages. Fixed initial context and branch-specific rolling histories support long-sequence generation with bounded context and consistent conditioning between candidate generation and policy optimization. Experiments with separately trained models on RT-1, BridgeV2, and RoboCasa cover real-world and simulated manipulation. On RT-1, under matched initialization, update counts, and evaluation protocols, RoboImagine improves PSNR by 1.26 dB and reduces LPIPS by 11.93% relative to continued supervised fine-tuning. Ablations show that diverse search improves prediction quality and that combined semantic and visual feedback yields higher external multimodal evaluation scores than visual feedback alone while maintaining comparable visual fidelity. Rolling histories further improve long-horizon prediction. These results support improving visual fidelity, task relevance, and long-horizon prediction through candidate comparison and complementary feedback.