In briefLatentTriad: Pre-action risk representations for detecting malicious agent skills
SummaryLarge language model agents are increasingly extended with reusable skills, but untrusted skill content can trigger operations beyond the user's task. The authors find that hidden states formed during prefill contain readable information about whether a task and its operations are consistent. They introduce LatentTriad, which combines contextual boundary features, group-adjusted operation directions, and departures from benign state relationships. Detection runs on a single shared prefill without generating output tokens, invoking tools, or executing skill code.
In briefSparse Anchors for Aligning Repeated Robot Recordings in Changing Environments
SummaryThe paper addresses aligning a new robot recording of an environment to a previously built reference map, even when the scene has changed. The method gathers absolute pose estimates via place recognition, feature matching and geometric checks, keeps the consistent ones as anchors, and combines them with within-session relative poses in a robust pose-graph optimization. On 15 cross-session experiments from the HILTI26 construction-site dataset, it reaches 94.4% recall within 1 m and 100.0% within 5 m, outperforming independent image localization and direct registration of reconstructed geometry.
Yunus Talha Erzurumlu, Niklas Trekel, Saurabh Gupta et al.
In briefImage matching method using deep feature manifold embedding and neural adjoint maps
SummaryThis paper addresses finding semantic correspondences between different object instances of similar categories in real-world images. It adapts functional maps, used for 3D shape matching, by deriving a discrete Laplace-Beltrami operator from a manifold embedding of deep image features via Galerkin's method, giving image grids a meaningful geometric structure. A neural adjoint map is added to represent non-linear mappings. The authors report state-of-the-art results among zero-shot image correspondence methods on multiple dense correspondence benchmarks.
In briefCRONOS: A benchmark for reinforcement learning on robot manipulation without episode resets
SummaryThis paper introduces CRONOS, a simulation benchmark for fine-tuning vision-language-action (VLA) robot models with reinforcement learning when scenes are not reset between episodes. The authors find that naively fine-tuning pre-trained policies fails in this setting because of shifts in object configurations and end-effector poses. They report that targeted reset allocation and addressing biases in pre-trained models can mitigate these problems. Non-episodic training is also found to improve long-horizon manipulation and generalization to unseen object configurations and task sequences.
Po-Yi Wu, Djengo Cyun-Jyun Fang, Dian Cheng et al.
In briefRollout-free world model for scoring self-driving trajectories
SummaryThis paper proposes RFWorld, a rollout-free world model for evaluating candidate trajectories in end-to-end autonomous driving. Instead of predicting full future scenes for each candidate, it builds a shared scene memory that candidate trajectories query directly through sparse spatial readout. The authors report that this reduces inference cost and avoids accumulated prediction errors, and that it achieves state-of-the-art results on NAVSIM and Bench2Drive.
In briefVIGOR: Visual Gain Ordering to Reduce Hallucination in Diffusion-Based Multimodal Models
SummaryMultimodal discrete diffusion language models generate answers by iteratively predicting masked tokens, but they can still hallucinate content that fits language patterns yet lacks support in the image. The paper proposes VIGOR, a training-free, plug-and-play inference method that reorders which masked positions to fill next, based on how much each prediction depends on visual evidence. It compares token confidence with the original image against a pass where the visual input is ablated. The authors report that VIGOR reduces hallucination and improves multimodal reasoning on LLaDA-V and Lumina-DiMOO.
In briefFloMo: 3D Scene Flow as an Intermediate for Learning Robot Actions from Human Video
SummaryThis paper presents FloMo, a 5B video-generation model fine-tuned to predict 3D scene flow and action tokens. By rendering scene flow as RGB and encoding it in the same latent space as video, a pretrained video backbone adapts efficiently to motion prediction. Co-trained on action-free human video and robot teleoperation data, the model generalizes to unseen objects, motion primitives, and tasks.
In briefDo Agentic Policies Need to Look at Every Step? A Geometric Abstraction for Reasoning and Action
SummaryThis paper proposes replacing repeated raw-image reasoning in robot control with a 3D geometric abstraction written as text and symbols. Typed, evidence-aware references to object parts are grounded from calibrated RGB-D data, and after each action the system returns compact textual checkpoints with metric residuals. Across offline policy synthesis and online execution, reasoning over this text and symbolic state improved multi-step manipulation completion while substantially reducing repeated multimodal queries.
In briefAblation study of protein ensemble generator design choices across structural types
SummaryThis paper examines design choices for generative models of protein conformational ensembles. Using a minimal Cα flow-matching model, the authors ablate architecture, sequence conditioning, rigid-body alignment and training data source. Variants are evaluated against experimental SAXS and NMR chemical-shift data across ordered, disordered, multi-domain and transitional proteins. The training data, rather than the architecture, is identified as the main constraint.
In briefLearning to Recognize Footsteps Across Unseen Footwear
SummaryThis paper studies identifying people from footstep patterns recorded on pressure- or force-sensitive walkways, where enrollment uses one footwear type and recognition uses unseen footwear. The authors propose BiStepNet, which splits spatial and temporal identity cues into two independently trainable streams and outperforms compact 3D spatiotemporal baselines on the StepUP-P150 dataset. Results show cross-footwear performance improves systematically with the volume and diversity of footwear in training, and auxiliary gait datasets with partial modality overlap can further improve robustness.
In briefHow correlations between emitting partons affect direct-photon yields and elliptic flow in heavy-ion collisions
SummaryThis paper studies how correlations between partons that emit photons affect the yield and elliptic flow of direct photons in heavy-ion collisions. The authors show that the local 2→2 photon rate depends on the joint distribution of incoming partons, so a correlation can change photon emission without altering single-particle distributions. Folded over the evolving medium in Au+Au and Pb+Pb collisions, the correlation raises the photon yield toward fitted PHENIX and ALICE data and increases v2 mainly at intermediate transverse momentum.
In briefPositional encoding to relax equivariance constraints during training, more efficiently and scalably
SummaryThe paper addresses optimization difficulties in strictly equivariant neural networks, which can struggle when data only approximately follow a symmetry. Prior work relaxed equivariance during training with non-equivariant dense layers, but at significant computational cost. The authors propose injecting a positional encoding with learnable scale to break the symmetry, gradually annealing it to zero so the original equivariant model is recovered by the end of training. They report significant gains for translational and rotational equivariant models with only marginal compute and memory overhead.
In briefALBI: LLM-guided inductive logic programming through iterative generate-test-constrain refinement
SummaryThe paper introduces ALBI, a framework that replaces the ASP-based hypothesis generator in Learning from Failures with a large language model while keeping SWI-Prolog as a formal verifier. It runs an iterative loop in which the LLM proposes candidate hypotheses, Prolog tests them, and diagnostic feedback guides refinement. On 30 standard ILP benchmarks, three frontier LLMs each reached 100% success, matching Popper with about 1.4 to 1.7 iterations on average. Under semantic anonymization, all models still achieved 100% success, with a moderate increase in iterations.
Aboubaker Seddiq Benterki, J\"uri Vain, Gert Kanter et al.
In briefToolSpace: Tool-centric environment synthesis for training long-horizon AI agents
SummaryToolSpace proposes a framework that builds training environments around tools rather than data, mining dependencies between adjacent tool calls from real MCP service execution traces using deterministic rules. It clusters the resulting tool-composition graph into environments and generates tasks via state-aware random walks, keeping only those that replay successfully. On seven long-horizon tool-use benchmarks, the method substantially raises the average scores of Qwen3.6-35B-A3B and GLM-5.2.
WhyThe authors propose tool-centric environment synthesis mined from execution traces, reporting gains across seven long-horizon tool-use benchmarks.
In briefFederated Transfer Learning for Medical Image Classification under Mixed Skew
SummaryThis paper addresses federated learning on multicentric medical data, which exhibits mixed skew across feature, label, concept, and quantity shifts between centers. It proposes PAFP, a personalized federated learning framework built on a pre-trained backbone that inserts client-specific Feature Alignment modules before fine-tuned blocks to align hidden features toward a server-aggregated target. On the FedISIC benchmark, PAFP achieves the highest unweighted balanced accuracy and reduces inter-client performance variance roughly six-fold relative to the strongest baseline.
Alexandre G. Leclercq, S\'ebastien Bougleux, Aur\'elien Corroyer-Dulmont et al.
In briefWhose Effect Is It? How Objectives Inflate Learned Subspace Interventions
SummaryThe paper examines activation patching, a method that copies part of a language model's internal state between inputs to locate concepts by measuring output change. The authors prove that, to first order, the effects of orthogonal parts sum to the whole state's effect, so a learned part exceeds the whole only when the remaining components push the output back. Objectives that reward larger output effects select such parts, and changing the objective to target the network's own output removes most of the excess.
In briefComparing representations and fitting methods for learning port-Hamiltonian models
SummaryPort-Hamiltonian models describe physical systems by separating stored energy, dissipation and external inputs. The paper compares neural networks, analytic expressions and combinations of the two for representing these components, and compares gradient-based optimisation with sparse regression for fitting. Tests on simulated mechanical and electrical systems, with and without noise, report prediction accuracy and computational cost for each approach.
In briefDRAP: A Replay-Free Method for Continual Learning in Physical AI
SummaryThis paper studies continual learning for physical AI, where small per-step policy errors can accumulate along a trajectory and cause closed-loop task failure. It proposes DRAP, a replay-free method that estimates each adaptable tensor's historical importance from action gradients weighted by local trajectory-amplification risk, and uses this to reduce plasticity in components that supported earlier behavior. Applied to a fixed subnetwork of OpenVLA and tested on the LIBERO Spatial, Object, and Goal suites, DRAP is reported to retain previously learned behaviors substantially better than standard AdamW fine-tuning and EWC while still acquiring new suites well.
In briefHaptic Shared Control for Multi-Robot Systems: Letting Operators Steer Robot Groups at Multiple Scales Through Touch Feedback
SummaryThe paper modifies a global connectivity maintenance algorithm so robots can use measurements from their local frames, taken with sensors that have limited fields of view. This algorithm is combined with three adaptive supervision strategies that let an operator control a single robot, pin constraints on several robots, or act on the group's centroid. These methods keep inter-agent constraints satisfied while giving the user haptic feedback about the system's connectivity state. A user study compared performance with and without haptic feedback across scenarios designed for each control mode.
In briefDeep learning for on-phone RGB photo enhancement: Mobile AI 2026 challenge report
SummaryThis report describes the Mobile AI 2026 challenge, which asked participants to design fast models that enhance RGB photos while running on smartphone GPUs from Qualcomm and Arm. Participants used the DPED dataset of image pairs captured by an old iPhone 3GS and a professional Canon 70D DSLR. Models were timed on Snapdragon 8 Elite Gen 5 and Dimensity 9500 chipsets, and the top solutions enhanced Full HD images in under 300 milliseconds on both platforms.
Andrey Ignatov, Georgy Perevozchikov, Radu Timofte et al.
In briefHow finite spectral depth affects early rank selection and bypassing the first gradient-flow saddle in matrix factorization training
SummaryThis paper studies training dynamics of matrix factorization, a simple model of neural network training, under small initialization. It analyzes continuous-time dynamics that applies a finite number of Newton-Schulz iterations to the normalized matrix gradient (Muon). The authors show that the number of iterations can change the rank selected early in training and let trajectories bypass the first nonzero stationary point that gradient flow approaches. Under an additional residual-inertia condition, they prove bypass occurs for all sufficiently large finite depths, give an explicit sufficient threshold, and construct an example where the threshold is four.
