Evaluating general-purpose multimodal large language models (MLLMs) for surgical scene understanding remains limited due to scarce surgical data and fragmented evaluation protocols. We address these challenges through a two-stage evaluation framework for surgical scene understanding. In Stage I, we introduce SurgKnowBench, a diagnostic benchmark that harmonizes public surgical datasets across procedures and tasks and evaluates proprietary and open-source MLLMs to map the boundary of pretrained surgical competence. SurgKnowBench reveals that current models retain partial visual priors but remain weak on relational and procedural understanding and workflow assessment. Building on this failure frontier and motivated by the scaffolded nature of surgical training, we introduce SurgAgentGym, an agentic environment where general-purpose MLLM agents can use bounded knowledge search over SurgWiki and general-purpose visual grounding for surgical tasks. SurgAgentGym evaluates whether generalist agents can improve surgical understanding at inference time through tool-mediated evidence use, without additional surgical training. Our results show that agentic inference narrows the gap mainly for perception-heavy tasks, while fine-grained localization and expert-style assessment remain persistent bottlenecks. This opens a path toward future studies of agentic surgical understanding.
Machine unlearning has become increasingly important in face recognition, where regulations like the GDPR empower users to request the removal of their personal data from deployed models. However, most existing unlearning methods are designed for closed-set classification with small-scale models and training data, and therefore do not generalize to real-world face recognition systems, which operate in an open-set setting with large-scale models and data. In this work, we propose Rapid Activation Dampening (RAD), a fast and targeted unlearning method tailored specifically for large-scale open-set face recognition. RAD approximates differences in average neuron activations between forget and retain identities, identifying and selectively suppressing neurons that are most associated with the targeted identity through a sensitivity-controlled dampening mechanism. We further propose the Forget Score (FS), which addresses the mismatch between existing unlearning metrics and open-set problems by providing a probabilistic interpretation of forgetting directly in the representation space. Experiments across several recent large-scale face recognition architectures and datasets show that RAD achieves strong forgetting while preserving overall model utility and substantially improving unlearning speed. These findings establish RAD as a scalable and practical approach to machine unlearning in face recognition systems.
Wireless foundation models learn reusable representations across tasks, configurations, and deployment environments. Their pretraining requires environmental breadth, yet wireless datasets often sample few sites densely. We introduce Channels4World, a scene-centered dataset of 100,000 distinct outdoor RF scenes and a procedural method for generating them. Each scene links a generated environment and RF material assignments to 32 base-station (BS) positions, 16,384 UE positions, and aligned variable-length multipath and ordered interaction records at 3.5 and 7~GHz. The corpus comprises 52.43 billion physical BS-UE pairs, 104.86 billion frequency-specific evaluations, and a 29.14-TB packaged payload. Scene identity aligns geometry, placements, channels, and interactions across frequencies, allowing environmental breadth and within-scene observation density to vary independently. The public release offers predefined subsets at multiple scales and separate access to corpus components. A deterministic 90K/10K scene-level train-test split keeps both frequencies and all scene components together. We characterize morphology, RF materials, and transceiver placement across three generation profiles. Controlled propagation and path-retention analyses show that diffraction materially affects retained-link availability and multipath structure, while the retention policy reduces storage substantially with little loss in represented aggregate power. Under matched training exposure, models trained on broader scene collections perform better on held-out scenes, although an equal-episode control shows that scene count alone cannot explain the gain. By making environmental breadth explicit and controllable, Channels4World supports the development and evaluation of wireless foundation models for generalization across deployment environments.
Recent zero-shot image compression methods exploit the powerful priors of pretrained diffusion models to produce perceptually pleasing reconstructions without compression-specific training. However, image compression serves not only human viewing but also machine analysis, while existing zero-shot codecs have limited flexibility in meeting the diverse reconstruction requirements of these applications. In this paper, we propose Flex-DDCM, a framework for zero-shot task-oriented image compression via diffusion trajectory guidance. Building on Turbo-DDCM, we adapt the coded diffusion trajectory through task-guided noise selection, embedding the effects of task guidance into the compressed representation. Guidance terms can be applied individually or combined, with timestep-dependent strengths controlling their influence throughout denoising. This design enables task adaptation without introducing additional decoding complexity. Experiments demonstrate flexible trade-offs among pixel fidelity, perceptual quality, and downstream task performance, with competitive performance against trained task-oriented codecs and faster decoding than Diff-ICMH.
A legal answer is only as good as the article it rests on. Vietnam's largest open legal corpus pairs 109,543 published questions and answers with 573,851 legal documents, yet its citations are free text: "Khoan 3 Dieu 101 cua Bo luat nay" ("Clause 3, Article 101 of this Code") names neither the code nor its version. Following the five stages of NVIDIA Auto Ontology, Discover, Ground, Validate, Expose and Learn, we turn four open datasets into a knowledge graph of 11.5 M nodes and 16.0 M edges in which every node and edge carries its source record and method, with no duplicate keys, dangling edges or schema violations. Of 67,868 cited provisions, 61,487 resolve to a dated law version, 51,694 to a parsed article, clause or point, and 19,012 are confirmed word for word. The ontology then shows what flat records hide. 36% of the 78,361 answers whose citations resolve now cite a law that is no longer in force; 2,525 verified citations point to an article number that says something else in the current law; 2,520 citations were re-versioned because their words match a different version than their date suggests; and 5,575 document numbers are shared by several local authorities, one of them by 72. Every number in this paper is computed from the graph.
We study offline policy optimization in finite-horizon Markov decision processes where every deterministic Markov policy has an action-value function linear in known features, with coefficient norm at most $B$. From independent complete trajectories of one unknown Markov behavior policy, and under all-policy occupancy coverage, we give an algorithm with $\widetilde O(\varepsilon^{-2})$ sample dependence and computational complexity polynomial in $\log(2+B)$. The algorithm performs exactly one convex regression per layer. Its stopped-return targets are the values of a single consistent Markov witness policy. A shared reference design and bounded prediction constraints convert fast clipped-regression error into pointwise confidence for the raw linear predictions. Truncated leverage controls the performance loss without requiring empirical covariance approximation at scale $B^{-2}$. Rounding confidence radii upward preserves the exact witness and yields a polynomial-time bit implementation for rational numerical inputs. The guarantees allow any fixed initial-state distribution; realizability and coverage need only be checked for deterministic policies. We also establish a covariance-coverage variant with uniform approximate realizability and polynomial dependence on $B$. A two-action lower bound shows that the accuracy exponent $2$ is necessary.
Aligning video generation models with human preference depends critically on the quality of the reward model that scores generated videos. Current video reward models, however, source their reward signal from a model's representational rather than generative capacity: they either repurpose vision-language models as pixel-space scorers, or train auxiliary scoring heads on top of frozen generator features. The generator's actual ability to predict velocities, denoise latents, and produce videos plays no role in scoring its own generations. We argue that this overlooks a structural property of offline preference learning. Any preference optimization objective that approximates an offline preference dataset mathematically targets a strictly defined KL-constrained optimal policy. By isolating the reward term from this analytical solution and mapping the exact generation log-likelihood to its flow-matching surrogate, the empirical reward emerges deterministically as the relative difference in continuous-time velocity-prediction errors. We formalize this equivalence into GenRM-Flow, a framework where a preference-tuned video generator inherently operates as a process-aware reward model directly in the noisy latent space. We further show theoretically that distinct fine-tuning objectives, whether contrastive like DPO and IPO or non-contrastive like RWR, are alternative optimization pathways toward the same optimal policy, each inducing a reward readout from the generator's velocity-prediction errors. Furthermore, the extracted process reward seamlessly substitutes external VLM scorers in FlowAWR. Operating purely on native latent outputs without auxiliary reward networks, this generative reward drives measurable improvements in downstream video generation.
Humans engage in pragmatic reasoning when encountering mentions of highly predictable events. Such mentions may appear redundant, leading comprehenders to infer that the event is atypical, and accommodate this inference with an explanation for why it is atypical. However, these inferences are not drawn uniformly across individuals, and experimental items vary in inference rates due to differences in event typicality. We show that these patterns can be predicted with a resource-rational cognitive model in which comprehenders select interpretation strategies based on individual exploration tendencies and item-level predispositions. These parameters are estimated independently from participants' Raven's Matrices score, event typicality rating and language model based accommodation likelihood. Taken together with recent findings from reference games, our results motivate a new perspective on the mechanisms that support pragmatic inferences.
Earth observation foundation models produce global metre-resolution embeddings that underpin work across agriculture, ecology, hydrology, and beyond. Yet these embeddings are opaque: each dimension entangles unrelated physical concepts, which makes them difficult to interpret and their robustness hard to assess. Here, we propose sparse autoencoders (SAEs) as a principled way to decompose Earth embeddings into sparse, approximately monosemantic features over an overcomplete dictionary. Because LLM judges of interpretations do not transfer to satellite pixels, and spatial autocorrelation matters, we introduce a tailored evaluation protocol. It scores any SAE on reconstruction at matched sparsity, alignment with reference concepts, and spatial scale, each against baselines that include an SAE whose decoder is a frozen random dictionary. On this protocol, SAEs reconstruct AlphaEarth and Tessera embeddings better than every baseline at matched sparsity. Their features align with far fewer reference concepts than raw embedding dimensions (on AlphaEarth, 88% align with at most one of 20 concept groups, most with none, against a median of 12 per raw dimension). They also carry more of their variance at fine spatial scales (23 to 32% within a 640 m patch for BatchTopK and Matryoshka, against 9%). An SAE with a frozen random decoder shares both properties, which therefore follow from sparsity, whereas only a learned dictionary reconstructs the embedding accurately: on AlphaEarth, trained SAEs leave eight times less variance unexplained. We test the sparse code on two applications, discovering land-cover sub-concepts and detecting out-of-distribution pixels by area-of-applicability estimation. Both indicate how sparse features could be used, but neither is yet reliable enough for practical use. Together, these results are a step toward Earth observation models whose features can be inspected.
Dense retrievers are meant to select passages that match a query's semantic specification. We show they suffer from search-by-answer bias: they favor passages that mention answer objects the model knows (scores as likely), over passages that support valid alternative answers. We test this with controlled factual contrasts from LAMA/TREx. For each retriever, we measure how strongly the model scores the canonical object over an alternative object using a token-prediction head, and ask whether this score predicts passage rankings. Across eight retrievers, including DPR, Spider variants, Contriever, E5, and Qwen3-Embedding, larger token-prediction scores consistently predict higher retrieval scores for passages containing the corresponding object. This holds when passages differ only in the answer object, and even when the passage that mentions the high-scoring object expresses a wrong relation. Dense retrieval thus exhibits search-by-answer bias during passage selection, shaping which evidence is retrieved.
What if a face recognition system could learn from every encounter without a single gradient update? We present Gallery Self-Evolution (GSE), a training-free framework where the gallery itself adapts at test time. The core mechanism---dedup-FIFO curation---archives novel appearance samples after successful recognition while permanently locking the enrollment template, bounded at 30 embeddings per identity. On CFP-FP and IJB-C benchmarks, GSE closes the side-view gap by +23 points at 45{\deg} yaw and boosts TAR@FAR=0.001 by +9.4 points, all at CPU-level cost. No retraining. No GPU. Just a gallery that grows smarter every time it meets you.
Single-loop minimax methods are appealing for modern machine learning, but standard stochastic gradient updates can be unstable and oscillatory under tightly coupled primal-dual dynamics, making performance notoriously sensitive to stepsizes. To fundamentally improve stability, we develop an orthogonal update framework tailored to minimax optimization, which applies updates along orthogonalized directions to reduce update anisotropy. However, orthogonalization reshapes the optimization geometry and makes coordinating primal-dual stepsizes even more delicate. To overcome this challenge, we develop AdaSGDA, which equips orthogonal updates with an adaptive stepsize rule driven by accumulated gradient norms. This mechanism automatically balances coupled primal-dual progress under orthogonalization, yielding an effectively parameter-free method for nonconvex-strongly-concave minimax optimization. We prove that AdaSGDA avoids problem-dependent tuning while attaining the state-of-the-art $\mathcal{O}\left(T^{-1/4}\right)$ convergence rate. We further propose AdaSGDA-VR, which incorporates a carefully designed variance reduction scheme and achieves the faster $\mathcal{O}\left(T^{-1/3}\right)$ rate. Extensive experiments demonstrate that our methods deliver competitive performance compared to existing minimax optimization methods.
Cross-lingual concept analysis requires distinguishing the information that a subspace retains from the classifiers and erasers that transfer through it. We study these operations in six frozen encoders using 10,359 Bangla (Bengali) idiom senses, eight input conditions, and figurative-class and sentiment labels. Class-mean directions provide the most consistent evidence of shared coarse meaning: in post hoc analyses, they outperform random-subspace, permuted-label and covariance-matched nulls in all 12 model-label tests under balanced accuracy. A prospective internal re-split supports idiom-alone sharing in 10 of 12 tests, spanning five encoders. The original confirmatory logistic-regression family rejects both planned nulls in 4 of 12 tests, where English directions gain only a few accuracy points above the majority-class floor (3.0-7.3). Classifier transfer depends on the input map and regularization: English-trained class-mean readouts beat both zero-shot nulls in 8 of 12 discovery tests, and stronger source-selected penalties improve transfer through logistic and LEACE directions. These readouts use Bangla-validation-selected layers and unlabeled target preprocessing. Input controls show greater coarse-class sharing for explicit glosses than for idiom counterparts. Reading, however, does not imply effective erasure: English-fitted LEACE removes less than one percentage point of Bangla figurative-class accuracy, compared with 4-11 points for Bangla-fitted LEACE; INLP shows a similar asymmetry. Pattern-filter analysis, covariance-weighted input maps and a binary Gaussian overlap bound explain how these operations can diverge. The analogous multiclass ordering is an empirical diagnostic, observed in 11 of 12 cells. The results support an estimator-sensitive account of cross-lingual concept sharing and motivate evaluating reading, classifier transfer, erasure and overlap separately.
Domain adaptation or DA methods are indispensable when target labels or computational resources are scarce, yet the domain kernel density and subspace alignment approaches completely assume a Hilbertian geometry. They measure distributional discrepancy with symmetric L2 divergences over Euclidean feature spaces or reproducing kernel Hilbert spaces (RKHS). This preprint summarizes the manuscript Non-Hilbertian Domain Adaptation. Total Variation Alignment via L1 Density Regularization on Feature Manifolds, which proposes Banach Manifold Transfer Learning or BMTL. BMTL makes two coupled design decisions. The first one is that features are embedded on a geodesic manifold rather than treated as points of a flat Euclidean space, and the second one is that the density matching regularizer is formulated in an L1 Banach norm, whose minimization is equivalent, by Scheffe's theorem, to contracting the Total Variation or TV distance between domain densities, a bounded metric that weights all regions of the sample space uniformly. The L2 formulation, by contrast, squares deviations and therefore amplifies errors in the tails, over emphasizing density outliers that do not reflect meaningful distributional shift. Experiments on MNIST-USPS, COIL20, and Office+Caltech10 show consistent gains; the L1 variant beats an otherwise identical L2 ablation on every task, for example 92.44% vs. 89.80% on USPS->MNIST and surpasses the strongest baseline DMDA by +3.68% on average over twelve Office+Caltech transfers. This document situates the contribution with an explicit comparison of Euclidean, Hilbert, and Banach geometries.
Tabular foundation models based on prior-data fitted networks (PFNs) achieve strong predictive performance, but their value for causal inference is unclear. We evaluated whether PFNs, which predict through in-context learning, could match or outperform conventional nuisance learners in supporting accurate causal estimation and reliable uncertainty quantification (inference). Learners modeled propensity scores and outcome regressions within doubly robust estimators of average causal effects, with and without cross-fitting, across synthetic and semi-synthetic settings with varying confounding complexity and effect heterogeneity. Under simple parametric data-generating processes, the PFNs were competitive with correctly specified parametric regressions and ensembles containing them. In more complex settings, they outperformed all conventional learners evaluated, including richly specified Super Learner ensembles, for both accuracy and uncertainty quantification. As sample size increased, only the two PFN models (TabPFN and Mitra) consistently approached nominal confidence interval coverage, whereas under-coverage persisted for conventional learners. In some settings, PFNs achieved near-nominal coverage and well-calibrated standard errors even without cross-fitting. At the smallest sample size (n=200) and in the higher-dimensional semi-synthetic settings, PFNs and other flexible learners exhibited poor coverage without cross-fitting. For PFNs, this appeared to be driven by propensity score overfitting consistent with in-sample memorization of treatment labels. These findings highlight the promise of prior-data fitted networks as in-context nuisance learners for causal estimation, while underscoring the need for further evaluation in challenging settings.
Diffusion models have demonstrated the effectiveness of score-based objectives for generative modeling, yet their potential for self-supervised representation learning remains underexplored. We introduce Entropy-Based Regularization (EBReg), a framework that uses latent score estimation to regularize the latent distribution during self-supervised training. EBReg jointly learns an encoder and a latent score model, and utilizes the learned score to provide a diversity-preserving signal that prevents representational collapse. Motivated by a mutual-information lower bound, we formulate the objective as view alignment with smoothed marginal entropy maximization and optimize the marginal entropy without explicit density evaluation. The same score model can additionally be used at test time to refine corrupted representations toward the training latent distribution. Empirically, EBReg learns competitive representations on ImageNet-1K, achieving 75.57% linear-probe accuracy with ViT-L. Across diverse architectures, score-guided test-time adaptation further improves robustness on ImageNet-C, yielding an average relative accuracy gain of approximately 5%.
User plane function (UPF) selection in 5G networks is an active area of research, with various limited approaches proposed to enhance network efficiency and manage resources effectively. In this paper, we address this gap in the related literature by proposing a Deep Reinforcement Learning (DRL) UPF selection method that utilizes a dynamic, multi-objective reward function and validating it on a real-world 5G emulation platform. The proposed method is based on the large language models' (LLMs) translation of network operator intents and the network data analytics function (NWDAF) abnormality alerts. The experimental results validated in the real-world experimental setup demonstrate that our dynamic DRL solution achieves an effective balance between latency, power consumption, and traffic distribution, outperforming the heuristic baselines and fixed DRL by 16.89%. These findings highlight the real-world applicability of our approach for dynamic, intent-driven UPF selection.
Block-diffusion drafters propose a whole block of tokens in one forward pass, and draft trees built from their position-wise marginals increase the expected accepted length of speculative decoding. Every tree node occupies a query row in target verification, so under concurrent serving low-yield branches compete with other requests for verification capacity. Existing tree methods select verified nodes by draft probability under a fixed or cost-selected per-request budget, and thus always retain a prefix of the drafter's ranking. We present D3Tree, which learns which nodes of a block-diffusion draft tree are worth verifying. A tree-size head with 2H+2 parameters predicts each candidate's conditional survival from the draft state, the candidate embedding, and its draft log-probability. Thresholding the product of these scores along parent chains yields an ancestor-closed subtree whose size adapts to each request and decoding round. The head is trained on a frozen drafter with conditional labels on the full candidate tree; an optional joint mode also steers the drafter toward the retained reference path. At serving time, D3Tree verifies whole trees under light load and, under heavy load, allocates one batch-wide node budget by predicted survival, sized to the target's matrix-multiplication tiles. Retained trees share one ancestor-masked target pass, while construction, pruning, and scheduling run on the GPU and overlap with the host. Across six vision-language verifiers and 17 text and multimodal benchmarks, D3Tree is fastest in 70 of 72 load settings, with 3.2-4.0x speedup over autoregressive decoding for one request and 1.3-1.8x at 64 concurrent requests, where fixed-budget trees fall below autoregressive speed. Under a time-varying arrival trace, it has the lowest mean end-to-end latency in every arrival phase.
Large language models (LLMs) have made substantial progress in mathematical reasoning but still produce incorrect solutions to complex problems, indicating the need for more reliable self-correction mechanisms. Mathematical self-correction typically identifies errors through verification, which can proceed backward or forward. The two directions provide complementary verification capabilities. Backward verification checks a candidate answer against the original constraints, bypassing complex intermediate reasoning, but cannot ensure correct reasoning or objective completeness. Forward verification follows the problem conditions and current solution, enabling it to locate reasoning errors and assess objective completeness, but may overlook key constraints in complex solution chains. We propose BiVerify, a dynamically scheduled framework for bidirectional mathematical verification. BiVerify begins with backward verification to screen the answer against the original constraints, followed by forward verification of the reasoning process; subsequent direction switching, state resetting, and re-verification are governed by verification outcomes. Upon failure, the system revises the candidate or verification flow until the same candidate passes both directions. We also introduce shared mathematical tools that separate logical reasoning from precise computation. We systematically evaluate BiVerify on five challenging mathematical reasoning benchmarks using multiple mainstream DeepSeek, GPT, and Gemini models. BiVerify exceeds the strongest representative public baseline evaluated in this work by 13.5 percentage points in average accuracy, with the margin reaching 31.2 percentage points on GPT. Further analysis shows that BiVerify's overall performance exceeds that of either verification direction alone.
Diffusion Transformer (DiT) models, while having become a powerful backbone for generative synthesis, suffer from high inference cost on generating high-resolution images. Diffusion caching has emerged as a promising acceleration technique to reduce inference cost. However, existing caching strategies fail to accurately identify the varying importance of different computational modules and rarely coordinate cache reusing across the timestep and layer granularities, making it difficult to control error accumulation. To address these limitations, we propose multi-granularity caching (MG-Cache) for DiTs, a novel and fully training-free approach to exploit computational redundancy both on timesteps and Transformer layers. Our core innovation contains a two-stage caching strategy: an offline timestep profiling analysis that identifies redundant timesteps for cache reuse, followed by an adaptive layer-wise caching strategy that combines input stability with importance of each Transformer layer to selectively cache stable Attention and MLP submodules. Extensive experiments on mainstream open-source DiT models validate the effectiveness of our approach, achieving significant acceleration while maintaining high fidelity to the uncached output, with up to a 4.18x acceleration on HunyuanVideo.