In briefLatentIndex: cross-layer sharing with layer-specific selection for sparse attention
SummarySparse attention cuts core attention computation, but its indexers still repeat selection work and store keys for each layer. LatentIndex extends Multi-head Latent Attention's latent-sharing idea across indexer layers: each layer group builds one shared latent cache from its first layer, while each layer scores tokens independently. With four-layer sharing, it reduces logical indexer-cache storage by 61.1% on DeepSeek-V3.2 and gives decode indexer speedups of 2.30-2.72 times over DSA in the hierarchical variant.
In briefDistScene: Object-to-Scene Distillation for 3D Scene Generation
SummaryDistScene is a framework that generates compositional 3D scenes from a single image by jointly modeling the environment and individual objects. It generates scene and object components in a shared coordinate frame, then refines each object using scene context. It also transfers pretrained object-generation priors to scene generation through automatically composed and rendered synthetic scenes. Evaluations on indoor and outdoor benchmarks report improved scene-level spatial coherence over the compared baselines.
In briefThe Numerical Linear Algebra of Large Language Models: a survey explaining LLM core concepts for numerical specialists, the linear algebra methods they use, and recent NLA contributions.
SummaryThis survey article focuses on machine learning techniques, with particular attention to large language models. It has two stated goals: to clarify the core concepts behind LLMs in a way accessible to specialists in numerical methods, and to examine the key numerical linear algebra concepts LLM techniques employ. It also highlights several significant recent contributions of NLA to the field.
In briefTask-sensitive geometry of representation transfer for object detection under image blur
SummaryThis paper studies how transferring knowledge from clean images affects object detection on Gaussian-degraded images. Using a gradient-derived geometry, the authors examine how local responses to representation directions relate to intervention effects. The local relationship holds for a specific direction family but does not clearly yield better global training, which the authors leave as an open problem.
In briefA new diffusion sampling method that better reconstructs PDE coefficients from sparse data
SummaryThis paper addresses recovering PDE coefficient fields from extremely sparse observations, an ill-posed inverse problem. The authors prove that output-space diffusion methods can fail to minimize unobserved error in discontinuous PDE settings. They propose FunBCS, an input-space optimization approach that seeks inputs whose denoiser reconstructions are physics-consistent, and prove it minimizes unobserved error. Across four PDE inverse problems, including discontinuous Darcy flow, FunBCS reduces reconstruction error by 27-64% and runs 1.4-2.1 times faster than the current state of the art.
In briefStance Drift: How AI Tools Can Distort the Positions in Our Messages
SummaryThis paper examines whether large language models keep a speaker's position intact when they draft or summarize text. The authors model the process as a two-step pipeline, where one model writes an argument from a given stance and another extracts the stance back, measured by a stance preservation rate. Across 112 debate propositions, none of nine tested models exceeded 0.7 under default settings. The most common drift patterns were polarization, deviation from neutrality, and flipping.
In briefPerturBot: Breaking Shortcut Priors in Vision-Language-Action Models via Perturbative Training
SummaryThe paper argues that vision-language-action (VLA) policies can rely on modality shortcuts, such as visual, lexical, or motor cues, instead of the task evidence needed for decisions. The authors propose PerturBot, a training method that applies task-preserving wrist-view perturbations, adds decision-relevant captions to instructions, and includes relabeled random and failed trajectory segments. They also introduce GroundingFscore, an offline score that measures how strongly a policy depends on these shortcuts.
In briefJoint tuning of layered controllers for humanoid robots doing walking and manipulation tasks
SummaryThis paper studies how to tune parameters across several levels of a humanoid robot's hierarchical controller together, rather than tuning each level separately. The authors propose a contextual Bayesian optimization framework that models task performance, motion realization quality, and control effort as separate outputs using a correlated multi-output Gaussian process. In a box-pushing task, they jointly tuned a centroidal predictive controller and a whole-body controller under varying box masses. The method achieved the lowest mean empirical regret during both training and adaptation among the baselines considered.
In briefForeAct3D: Policy-Grounded Future World Modeling for VLA Policies
SummaryThis paper presents ForeAct3D, a framework that trains Vision-Language-Action robot policies to predict future semantic 3D scene states, with future predictions conditioned on the policy's planned action chunk. Physical-consistency constraints based on background staticity, instance-level rigidity, and end-effector kinematics shape the shared representation during training, and no future prediction is needed at inference. The authors report 98.3% average success on LIBERO, an average task length of 3.73 on CALVIN, and real-world average success rising from 6.7% to 37.8% over the base policy.
WhyAction-conditioned 3D future forecasting with physical consistency improves VLA policies, backed by benchmark gains, ablations, and real-robot experiments.
In briefSparse-view 4D scene reconstruction with spatiotemporal priors and generative help
SummaryThis paper presents a 4D Gaussian Splatting framework for the Sparse-View Track of the SIGGRAPH Asia 2026 Volumetric Video Challenge, which reconstructs dynamic scenes from only six wide-baseline cameras. It combines region-adaptive spatial priors, motion-consistent temporal priors from frame interpolation and optical flow, and diffusion-restored images from virtual cameras used as pseudo-supervision. On the official test benchmark it reaches 30.04 dB full-frame and 27.88 dB foreground PSNR, ranking first in the track.
In briefConEx: Concept-based saliency maps that show which visual ideas drive a model's predictions
SummaryThis paper introduces ConEx, a framework that links pixel-level saliency maps to human-understandable concepts. It automatically discovers class-specific concepts, represents them as concept activation vectors, and shows where each concept appears in an image and how it affects the prediction. The authors also propose two metrics, Vector-Concept Match and Concept-Class Match, and report state-of-the-art results on faithfulness, segmentation, and concept-quality benchmarks.
Yehonatan Elisha, Oren Barkan, Ziv Weiss Haddad et al.
In briefUsing Large Language Models to Forecast the Effects of Curriculum Decisions
SummaryThis paper treats automatic curriculum learning as a sequential decision-making problem and points out a gap between what determines the value of curriculum choices and what local learning signals capture. The authors propose combining online learning-progress estimates with LLM-based estimates of each task's potential downstream benefit and whether direct training is likely to yield progress. The method is evaluated on a custom benchmark of 256 textual goals in Craftax, with the strongest gains reported when optimizing for individual target tasks.
Octavio Pappalardo, Nathan Herr, Tim Rockt\"aschel
In briefGrouping 3D shape primitives into semantic parts with reinforcement reasoning (RePart)
SummaryThe paper introduces RePart, a method that organizes geometric primitives (superquadrics) of 3D shapes into semantic parts. It formulates this as a finite-horizon Markov decision process and learns merge-and-stop decisions through trajectory-level reinforcement reasoning. Boundary-Aware Surface Labeling maps the inferred part labels back to the original mesh. Experiments on PartNet and 3DCoMPaT++ report strong partition results.
In briefWASP: Weakly Aligned Paired Learning for Linking Fetal Brain MRI and Ultrasound
SummaryThe paper introduces WASP, a framework that links fetal brain MRI and ultrasound without publicly available paired scans. It uses clinical metadata such as gestational age and diagnostic planes to align the two modalities through entropic optimal transport, then trains a lightweight module that maps MRI representations into the ultrasound latent space without fine-tuning the backbone. On USFM, gestational age estimation error falls from 21.9 to 17.4 days and standard plane accuracy rises from 61.9% to 69.0%. Gains are smaller for backbones already pretrained on both modalities.
Francesco Correnti, Gabriele Magrini, Marco Mistretta et al.
In briefAsymptotically optimal fixed-budget best arm identification under differential privacy
SummaryThis paper studies fixed-budget best arm identification in bandits under pure ε-differential privacy, where the learner must recommend an arm while protecting the full interaction transcript. The authors derive an instance-dependent, privacy-aware exponent that upper bounds the optimal error-probability decay rate. They propose AO-Pri-BAI, an adaptive algorithm that is proven to be ε-differentially private and whose failure-probability exponent matches this benchmark, and numerical studies show it also performs well in non-asymptotic settings.
In briefDiffGate: Difficulty-Gated Teacher Guidance for On-Policy Distillation
SummaryThis paper proposes DiffGate, a post-training method for large language models that improves on-policy distillation. It combines GRPO with selective, bounded teacher guidance applied only to failed trajectories and scaled by group difficulty. Experiments on Qwen3-0.6B and Qwen3-1.7B across code and mathematics report pass@8 gains over matched GRPO in all four model-domain settings.
In briefSIFT: Checking Whether Reasoning-Faithfulness Detectors Stay Reliable Under Distribution Shift
SummaryThis paper examines whether detectors that audit chain-of-thought reasoning for faithfulness give consistent verdicts when inputs shift. The authors propose SIFT, a hidden-state trajectory detector trained with cross-environment invariance objectives and certified abstention, and test it on 14,996 traces across four domains and eight models. They report that existing detectors transfer poorly, and that random sampling variation accounts for most detector instability, more than distribution shift does.
Noor Islam S. Mohammad, Md. Basim Al Zabir Shammo, Hasan Siddiki et al.
In briefMathematical bounds and null behaviour of KRATOS, a bibliometric recognition-comparability diagnostic
SummaryThis paper gives a complete mathematical specification of KRATOS, a group-based bibliometric diagnostic that compares how documents are distributed across groups with how citation weight is distributed. It derives bounds, equality conditions, an exact logarithmic decomposition of the composite index, and closed-form moments of recognition ratios under global and stratified permutation nulls. The results are verified numerically with an accompanying script. The specification covers measurement structure only and does not define a measure of epistemic change or justice.
In brief"Frozen in a Frame: The Velocity Blind Spot in JEPA World Models" explains why single-frame world-model encodings miss motion and proposes a fix
SummaryThe paper shows that JEPA world models trained on single rendered frames cannot encode velocity, including the official LeWM weights, as linear velocity probes fall to chance while position probes reach R² near 0.95 across four benchmarks. The authors propose TI-JEPA, which splits the latent into a pose code and an explicit finite-difference motion code predicted jointly. It gives significant gains on a stop-at-goal planning task across three physically grounded environments.
WhyIdentifies a structural velocity blind spot in single-frame JEPAs with a checkable argument, and a lightweight decomposition yields significant, seed-robust planning gains across environments.
In briefStrong Helps Weak: Transferring Cross-Modal Alignment Between Multimodal AI Models
SummaryThe paper finds that merging a well-aligned, data-rich multimodal LLM from one modality into a data-scarce model for another modality substantially improves the target model on its own benchmarks. The authors attribute this to stronger alignment between modality-specific tokens and text tokens, supported by theoretical and empirical analysis. They propose DCAT, which transfers this textual alignment without further fine-tuning, using a closed-form weight-space solution computed from a small calibration set.
WhyProposes a novel one-directional transfer of text alignment from strong to weak modality, with a mutual-information bound and closed-form solution; experimental breadth remains to be verified.
In briefFrom Probe Scores to Alarm Policies: Checking Whether Activation Monitors for Language-Model Agents Work in Practice
SummaryThe paper argues that activation probes can predict safety-relevant properties of language models with high AUROC, but deployed monitors make thresholded alarm decisions under tight false-alarm limits, which is a different task. The authors propose an Operational Validity Contract covering target, timing, calibration and cost, and test it on several benchmarks. They find that high AUROC does not guarantee good detection rates at fixed false-alarm levels, and that thresholds can realize false-positive rates different from their targets. When support is insufficient, the policy abstains rather than reporting success.
WhyProposes an operational validity contract separating AUROC from alarm risk, validated across benchmarks with preregistered protocols showing low detection at tight false-alarm budgets.
In briefProject Only Once: A Loss That Enforces Exact Constraints in Neural Predictions
SummaryThis paper addresses how to make neural network predictions satisfy constraints exactly. It proposes Projection-Adaptive Loss (PAL), which needs only a single detached projection step during training and uses the remaining constraint residual to adaptively weigh penalties. The authors report near-perfect feasibility on highly nonlinear constraints and training 2.5x faster than the DC3 repair-based baseline on an ACOPF benchmark.
In briefFrom Transformers to Weighted Automata: A Formal Bridge for Verifying Large Language Models
SummaryLarge language models are increasingly used in safety-critical settings, but their black-box nature makes formal guarantees about their behavior difficult. This paper links transformers to weighted automata, a classical model from formal language theory, so that verification tools from automata theory can be applied to LLMs. It establishes a formal correspondence between transformer architectures and weighted automata over the reals, and introduces an identity testing algorithm that statistically decides whether two stochastic models define the same distribution up to a tolerance threshold.
Smayan Agarwal, Aslah Ahmad Faizi, Shobhit Singh et al.
In briefGradient-Free Sampling: Drawing Samples from Generative Models via Stochastic Bounded Extremum Seeking
SummaryThis paper proposes a sampling method for energy- and score-based generative models that requires no gradient evaluations of the model. It replaces the score-containing drift term with a high-frequency dithered cosine of the model's value, which in the high-frequency averaging limit yields Langevin MCMC or the reverse-time SDE of diffusion models. The authors prove convergence of the dithered trajectories to the target SDE trajectories and give an O(ω^-1/2) mean-square rate under global bounds. Experiments on CelebA-HQ and CIFAR-10 demonstrate latent-space tracking and latent-space EBM sampling.
