Recently, video world models have advanced interactive generation, with large models offering higher quality and smaller models enabling faster inference. Existing speculative methods combine their strengths through fast draft generation and selective target regeneration guided by rewards on decoded frames. However, verification covers only the external scorer's evaluation dimensions and adds decoding and scoring costs. Moreover, these scores provide no correction, requiring costly target regeneration for rejected drafts. To address these limitations, we propose Speculative World Model (SpecWM), a unified, training-free inference framework for streaming world models. It verifies draft chunks using the target's latent velocity field. Verification also yields a reusable correction; two thresholds select Accept, Refine, or Reject. Specifically, one-step refinement reuses it to improve perceptual quality, while three-step refinement strengthens camera following and trajectory consistency. These findings motivate transition-aware refinement. On WBench Navigation, SpecWM improves visual quality at 1.63$\times$ speedup over the target and raises NavScore from the draft's 75.99 to 85.63, achieving the best quality-speed-control trade-off. Additionally, our LongSpecBench shows sustained quality and control over long, multi-turn rollouts, while experiments on Hy-Nomas validate SpecWM across backbones and denoising budgets.
Visual agents that read dense documents must ration a finite observation budget across candidate regions, yet answer accuracy conflates where an agent looks, how it adapts to the reads that remain, and how it an?swers once observation stops. We introduce budget-paired counterfactual audits, an evaluation protocol that holds public evidence and region bindings fixed while perturbing only the remaining budget or de?clared task parameters, and credits every acquired history with exact continuation values. The resulting accounting decomposes lost expected accuracy into stepwise acquisition loss and terminal decision loss, separating planning from fixed-role heuristics. Across forty paired routing anchors, five open-weight
Human vision and video stereo networks both reuse disparity evidence across time, yet a refresh of the input can invalidate a match formed milliseconds earlier. What tells a system whether an old match still counts? We answer this from limited human data in three steps. First, nine temporal mechanisms share a frozen binocular-energy front end and an equal-capacity readout, so held-out behavior isolates the temporal selection rule: on 240 conditions from 10 observers, down-weighting aged evidence that has lost binocular support lowers subject RMSE from 12.6% to 10.3% and reproduces a period-dependent reversed-depth illusion, across all 59 readings of the unpublished protocol and 8 front-end perturbations. Second, four propositions establish what such data can identify: gating is identified at the level of the refresh schedule, and candidate binding is invisible to standard stimuli but exposed by a single-eye geometry flip that separates the readouts by up to 7.1%. Third, the identified rule becomes a two-field stereo memory that separates sample validity from hypothesis support. It removes a 50 ms recovery lag after geometry flips, and its regime analysis explains why plain integration stays preferable on MPI-Sintel, where 71.1% of pixels change appearance but not disparity.
Recommender systems, which retrieve items of interest to users based on their interaction histories, have recently witnessed several key technological advancements, benefiting from the semantic representation space and model-data scaling. Nevertheless, practitioners and researchers have frequently encountered critical inconsistencies between simple offline evaluations and complex online results, and this gap still exists in the era of large models. In this work, we address several key factors that contribute to this inconsistency: 1) the \textit{large-scale continual learning challenge} that assumes a practical demand to maintain long-term model effectiveness under dynamic item pools and distributional shifts; 2) the \textit{combinatorial diversity challenge} that requires the generation strategy to solve the accuracy-diversity trade-off in both the finer token level and the holistic item level; and 3) the \textit{multi-task reward balancing challenge}, which is particularly critical for modern generative models that can accurately model the sequence distribution but are less controllable for the multi-task demands. To address these challenges and further pave the way for the development of more realistic recommendation solutions, we introduce KuaiRecV2, a deliberately constructed benchmark with industrial-level datasets from Kuaishou. Specifically, we provide a large-scale interaction dataset with million-scale multi-modal item information and long-term interaction records across 30 days, with a dynamic video candidate pool. Additionally, we provide a rigorous benchmark that evaluates retrieval accuracy, diversity, and multi-task performance in one rubric system, where an effective advancement occurs if and only if the new method improves all metrics. The code is available at https://github.com/cxli2333/KuaiRecV2
Let $q$ be an odd prime. Let $ζ_q$ be a primitive $q$th root of unity, and put $K=\mathbb{Q}(ζ_q)$, with ring of integers $\mathcal{O}_K=\mathbb{Z}[ζ_q]$. As $q$ ranges over odd primes, we prove that the decision version of the Shortest Vector Problem ($\mathsf{SVP}$) in the canonical $\ell_2$-norm is $\mathsf{NP}$-complete on full-rank free submodules of $\mathcal{O}_K^2$ by giving a deterministic polynomial-time many-one reduction from Exact Cover by 3-Sets (X3C). The module rank is fixed at two. The main obstacle is closure under the action of $\mathcal{O}_K$. A module containing a nonzero vector also contains every scalar multiple of that vector by a nonzero element of $\mathcal{O}_K$, and some of these multiples may be shorter. Three ideas overcome this obstacle. First, we map the Bennett--Peikert Reed--Solomon lattice to a principal cyclotomic ideal and use Wan's point-count estimates to prove that a coset of this ideal contains many binary coefficient representatives. Second, we construct a checker whose canonical squared norm encodes the X3C equations, with a controlled nonnegative error term. Third, the checker and a second module coordinate combine with a separation bound for ideal cosets to rule out short vectors in NO instances. Each constructed instance consists of an odd prime $q$, two integral generators whose $2\times2$ generator matrix has nonzero determinant, and an integer squared threshold. The construction also gives $\mathsf{NP}$-hardness of search-$\mathsf{SVP}$ under deterministic polynomial-time Turing reductions with at most one oracle call.
We prove that exact decision-$\mathsf{SVP}$ in the coefficient $\ell_\infty$-norm is $\mathsf{NP}$-complete for principal ideals and free modules of rank two over the cyclotomic rings $R_n=\mathbb{Z}[X]/(X^n+1)$, where $n$ ranges over powers of two. For each case, we give a deterministic polynomial-time many-one reduction from the Subset Sum problem. For an instance with $s$ positive input integers, the output lattice has dimension $O(s)$. A solution to the original Subset Sum instance can be recovered in polynomial time from any nonzero output lattice vector of norm at most the specified bound. We also transfer these reductions from $R_n$ to the power-of-two cyclic rings $S_{2n}=\mathbb{Z}[X]/(X^{2n}-1)$. This gives $\mathsf{NP}$-completeness of decision-$\mathsf{SVP}$ in the coefficient $\ell_\infty$-norm for principal ideals and free modules of any fixed positive rank over these rings.
Large language model (LLM) coding agents develop software through experimentation, offering an opportunity to design the serving systems supporting their own operation. Effective serving requires coordinating implementation and operating decisions under session dependencies, latency constraints, and correctness requirements. We introduce ServerBench, a benchmark for agentic design of complete, specialized LLM serving systems. Agents experiment against a common objective, dollars per dollar: latency-qualified token value per dollar of GPU cost. Stateful replay preserves session dependencies, and independent verification checks the submitted system. Across 11 coding-agent models and three deployments, Claude Sonnet 5.5 achieves 2.97 dollars per dollar on Gemma E4B, $4.87\times$ the vLLM reference; 22 of 33 runs yield verified systems, and 13 exceed the strongest baseline. Our analysis identifies three findings: high-value designs coordinate capacity improvements with load selection; agents interpret performance feedback through workload-relevant experiments; and reliable delivery requires verification of the execution paths introduced by optimization. Overall, ServerBench supports systematic study of how agents translate service objectives and feedback into effective, verified serving systems for given task specifications.
Inverse kinematics (IK) converts end-effector pose commands into joint motion, and many vision-language-action (VLA) policies rely on it to act on a given arm and to transfer across embodiments. Precise IK remains laborious: redundancy and singularities complicate the inverse mapping, execution depends on hand-tuned controllers, and training data on physical robots are costly to collect. We present PolicyIK, a goal-conditioned reinforcement learning policy for high-precision IK. It updates joint-position references from pose feedback and observes its accumulated reference relative to the measured joints, so it can retain the offset that the position controller needs at rest. Two-scale pose errors with an error-dependent action gain resolve millimeter residuals while preserving large motions. Progress and dwell rewards teach it to both reach the tolerance and stay within it. In simulation, a single attempt by the trained PPO policy keeps the mean errors over the final second within 1 mm and 0.2° on 93.5% of held-out tasks, and a deployment protocol raises success to 98.9%. When trained directly on a physical arm with an off-policy RL learner, PolicyIK achieves higher reaching and holding success than all four classical IK controllers under the same time budgets. In simulated pick-and-place, PolicyIK executes the end-effector commands of the $π_{0.5}$ VLA policy with 95.7% success. These results show that reinforcement learning can provide high-precision inverse kinematics for robot arms. Code is available at https://github.com/Tinaliu0123/PolicyIK.
We present PhysInsert, a framework for physically plausible reference-based object insertion that jointly adapts objects and their surrounding scenes. Existing methods often rely on masked inputs, generation-based pairs, or training pairs with limited viewpoint variation, leaving pose and illumination adaptation and bidirectional object-scene interactions insufficiently constrained. We construct the Insert World dataset with strictly aligned Blender-rendered counterfactual pairs and object references across viewpoints and lighting conditions. Building on these data, Geometry-Guided Adaptation couples a VGGT geometry stream with image and text streams under target-scene feature supervision to guide spatially consistent insertion. Complementary 2D Object-Scene Spatial Supervision jointly supervises object occupancy and environmental changes while grounding spatial instructions in object regions through image-text attention. We introduce PhysInsert-Bench with dedicated cross-environment and cross-view cases requiring illumination and pose adaptation beyond appearance copying. Experiments on PhysInsert-Bench and DreamBooth demonstrate improved image fidelity; a user study further shows more accurate placement and improved lighting integration, perspective, and pose plausibility.
We introduce Gemma 3, a multimodal addition to the Gemma family of lightweight open models, ranging in scale from 1 to 27 billion parameters. This version introduces vision understanding abilities, a wider coverage of languages and longer context - at least 128K tokens. We also change the architecture of the model to reduce the KV-cache memory that tends to explode with long context. This is achieved by increasing the ratio of local to global attention layers, and keeping the span on local attention short. The Gemma 3 models are trained with distillation and achieve superior performance to Gemma 2 for both pre-trained and instruction finetuned versions. In particular, our novel post-training recipe significantly improves the math, chat, instruction-following and multilingual abilities, making Gemma3-4B-IT competitive with Gemma2-27B-IT and Gemma3-27B-IT comparable to Gemini-1.5-Pro across benchmarks. We release all our models to the community.