Science & Mathematics

Large language model (LLM)-based agents have recently demonstrated outstanding performance in single-turn tasks, yet they continue to struggle with long-horizon tasks, particularly in environments without prior training. A key limitation of single LLMs is their difficulty in processing long contexts, which stems from weaknesses in continuous decision-making and reasoning capabilities. Although existing multi-agent systems (MAS) equipped with long-term memory bases can partially address this gap, they mostly rely on knowledge reuse or summarization, an approach that is often constrained by directly retrieval without any adaptation for the current task, while the retrieved memory may contain much history noises waiting to be preprocessed. In contrast, MAS characterized by task-specific guidance often lack long-term memory specifically designed for the guidance. To address these challenges, we propose M-Plan, a closed-loop MAS framework that integrates long-term and short-term memory. By combining global-local and past-current guidance, M-Plan enhances agents' performance on multi-turn tasks while maintaining a plug-and-play design. During task execution, agents receive task-specific milestones and insights, and are provided with local hints only when confusion arises, which form global-local guidance together. These guidance generation and utilization construct the short-term memory as well. After task completion, relevant information about the current task is optimized or reorganized before being sent to the long-term memory storage for future retrieval. The designments mentioned above build a comprehensive framework different from other individual long-term or short-term memory works. Extensive experiments on multi-turn tasks including ALFWorld, Webshop, and PDDL across diverse models demonstrate that M-Plan outperforms multiple baseline methods in both accuracy and efficiency.

The service rate region of a distributed storage system describes the set of concurrent demand vectors that can be served without overloading any individual server, and is a fundamental measure of how effectively redundancy is used. In this paper, we investigate aggregate service rates in distributed data storage systems based on linear codes over arbitrary finite fields. We first derive an upper bound on the aggregate service rate for an arbitrary subset of $t$ data objects in terms of the dual distance and the sizes of their smallest recovery sets. For systematic generator matrices, this bound simplifies to an expression depending only on the code length, the dual distance, and the cardinality of the selected subset. This result extends existing bounds and provides a subset-wise formulation of the First Dual Distance Bound. We then establish conditions under which this upper bound is attained. We apply our results to several important families of linear codes, including MDS codes, Reed-Muller codes, constacyclic codes, and extended Golay codes, and obtain exact values of the maximum aggregate service rates for the corresponding subsets of data objects.

A cover song re-renders an existing piece, most commonly by preserving its tonal content -- melody and chord progression -- while reshaping other musical attributes. This suggests a natural formulation for generative cover synthesis with pretrained music language models: explicitly control tonal content, while specifying lyrics and style through text. Existing approaches condition on frame-level features, tightly coupling generation to the reference's timing and structure, and limiting flexibility such as tempo variation, structural rearrangement, or partial conditioning. To this end, we present MuLaCover, a cover generation model that conditions a pretrained text-to-song foundation model on an editable symbolic lead sheet. Our design enables alignment-free control: generated outputs preserve the tonal signature of the source without frame-level alignment, allowing flexible timing, structure, and segment-level generation. We further introduce a training curriculum with partially mismatched audio-symbolic pairs, improving diversity and robustness. We evaluate MuLaCover with objective metrics, standard subjective ratings, and in-depth expert interviews that provide fine-grained diagnostic insights. Experiments show that MuLaCover achieves state-of-the-art performance among open-source systems and is competitive with leading commercial models such as Suno-v5.5. Audio samples and code are available at our project page: https://mulalabs.ai/mulacover/

Rubric hacking limits both performance and training robustness in rubric-based reinforcement learning. By decomposing rubrics into sub-dimensions, we reveal the Intra-Rubric Reward Divergence: as the total rubric score increases, certain sub-dimension scores systematically decrease, suggesting the model sacrifices specific quality aspects to optimize the aggregated objective. From the sub-dimension perspective, we observe two phenomena: (1) dimension imbalance, causing the policy to favor heavily covered dimensions; (2) dimension absence, leading to a systematic gap between rewards and true quality. To mitigate these issues, we propose Fair-Race Dimension Aggregation (FRDA), which has generations compete within each dimension separately to mitigate rubric imbalance, and Residual Semantic Dimension Completion (RSDC), which mines evaluation dimensions implicitly encoded in a holistic preference signal to complete rubric coverage. Experiments in chat, science, and medical domains indicate that FRDA mitigates rubric hacking with more stable improvements, while RSDC outperforms both single-reward training and naive multi-reward aggregation, yielding generally higher-quality responses.

We investigate how frozen pretrained visual and depth representations support dense optical flow when coupled with a supervised matching network. Building on TTSFlow, this paper studies Dense Prediction Transformer (DPT) feature decoding and one coarse-to-fine refinement stage. Our method fuses features from frozen DINOv2 and Depth Anything V2 models and estimates optical flow through a trainable transformer-based matching pipeline. After optical-flow training on FlyingChairs and FlyingThings3D, it achieves 2.37 endpoint error (EPE) on the Sintel training split, Final pass, compared with 2.52 for the FlowSeek (T) configuration reported here. Our RoCo-Spring 2026 entry, RoCo-29.VSAI, obtains a robustness-balanced score of 1.1038, Spring EPE of 1.411, and RobustSpring proxy EPE of 4.739, with a lower balanced score than seven of the twelve listed reference baselines. These results characterize the complete pipeline, which reuses frozen foundation backbones and performs one refinement stage. Code is available at: \url{https://github.com/Visual-and-Spatial-AI-Lab/FFlow}

Autoregressive language models emit one token per forward pass. This causes high inference latency, in particular as models are scaled up for quality. In this paper, we introduce Parallel Language Models (PLMs), which generate mutually consistent tokens in one forward pass by routing dependencies through latent auxiliary variables. PLMs are trained using a variational bound that reduces to an autoregressive loss plus a parallel decoding loss. Empirically, PLMs are on a faster and better Pareto frontier than autoregressive and discrete diffusion baselines. Pretrained PLMs achieve near-autoregressive pretraining loss on model scales of 80M to 3B parameters, and speedup follows a power law in model size. When post-trained, our 1.3B model matches or exceeds an identically trained autoregressive model on math, coding, and instruction-following benchmarks, speeding up inference $2.3\text{-}5.7\times$. Together, we show that parallel generation can be trained from a single training objective at autoregressive quality. We will make our code available at https://github.com/mandt-lab/plm.

Hierarchical abstraction is a fundamental principle across many data modalities, enabling complex sequences to be represented compositionally at multiple levels. Learning such hierarchies directly from data, without explicit structural supervision, remains a central challenge in representation learning. We investigate whether meaningful hierarchical structure can emerge from compression alone. While autoencoders learn compressed representations through reconstruction, this objective does not necessarily favour meaningful compositional structure. We study this challenge through unsupervised grammar induction, where the underlying hierarchy can be evaluated explicitly. We introduce a hierarchical token-merging autoencoder that recursively combines adjacent tokens into increasingly abstract representations, compressing entire sequences through a variational information bottleneck. The model jointly learns representations and hierarchical structure by marginalising over possible binary trees using dynamic programming. To address training instability, we combine independently trained models through unsupervised consensus and distillation. Experiments on a synthetic, ambiguous grammar demonstrate the emergence of meaningful hierarchical structure without structural supervision. The resulting model achieves 93% grammatical parse-validity and 93% exact reconstruction on held-out sequences, compared with 42% parse-validity for the strongest trivial baseline. These findings establish a proof of concept for hierarchical token merging as a mechanism for unsupervised structure discovery and abstraction learning.

A key problem in AI safety is eliciting latent knowledge (ELK), extracting information from a model that is misaligned and more knowledgeable than its overseer. For example, a misaligned model used to scan code for vulnerabilities may want to withhold its findings, i.e., sandbag, in order to exploit them later. To study ELK empirically, we need model organisms that mirror this setting. Existing model organisms fall short in several respects. In this work, we train model organisms to withhold solutions in a vulnerability detection task. We combine transcript distillation with counter-prompting and tamper-resistance training to instill three properties that we argue are critical for a useful ELK testbed: 1) Selectivity, the model sandbags when it thinks the overseer doesn't know the answer and complies otherwise, 2) Robustness, the model resists elicitation, and 3) Retention, the model retains a knowledge advantage over the overseer, as opposed to losing the relevant capability. We train a sequence of model organisms and show that our strongest model substantially outperforms prior models on robustness while keeping the other two properties intact. We hope that our model organisms provide a useful ELK testbed for the community, and that our methodology sparks the development of increasingly challenging model organisms. All our code, models and the dataset are available at https://huggingface.co/collections/lennart-finke/sandbagging.

Foundation machine-learned interatomic potentials (MLIPs) promise broad transferability, yet their out-of-domain reliability is hard to establish. Here, we use pure and O/H-containing disordered carbon spanning multiple phases, competing hybridisations and porous frameworks to stress-test six popular foundation MLIPs. We first screen all six on density functional theory (DFT) energy and force errors and binding curves, then subject four to melt-quench-anneal molecular dynamics and evaluate the resulting structural observables. Binding curves and O/H-containing trajectories reveal unphysical extrapolation and dynamical failures that low DFT energy and force errors fail to reveal. Of the models tested, MACE-MH-1 is the most consistently robust, with low DFT errors and stable MD. Despite differing architectures, OMAT-based models share structural biases, highlighting the importance of training coverage. All four models generate structures of nanoporous carbon that reproduce the principal experimental and specialised carbon potential signatures. Incorporating O/H leaves short-range order largely unchanged but shifts pore-size distributions towards larger pores. These results emphasise the need for application-specific validation and extend molecular dynamics modelling of nanoporous carbon towards experimentally realistic structures.

Interaction traces record how people and agents use computers, and comparing them can reveal usability barriers, differences in expertise, and human--agent similarity. Yet traces are recorded as low-level, interface-specific actions, making comparisons difficult across tasks, interfaces, and input modalities. Existing approaches either compare these actions directly, reduce traces to aggregate outcomes, or summarize them at higher-level goals, obscuring where behavior differs. We present MAP (Multi-level Action Parsing), a hierarchical semantic representation and annotation pipeline that maps traces into a common vocabulary across linked levels of abstraction. MAP compares traces by sequence, recurring patterns, and effort, localizing differences by type and abstraction level. A technical evaluation across three computer use datasets shows high agreement with manual annotations and broad coverage of behavior. Applying MAP shows that sighted and blind users increasingly converge at higher levels despite navigation differences, while agents omit corrective, help-seeking, and recovery behaviors present in human traces.

Networked mobility sensing couples local trajectory inference with remote geospatial context, making communication decisions part of the prediction system. We present GeoSplit, an edge--server framework for travel-mode recognition that separates the utility of map evidence, its training-time transfer, and its selective acquisition at runtime. On GeoLife with 10% labels, we evaluate these components under user-day and user-disjoint grouping. For unseen users, aligned full-map context raises macro-F1 from 0.7154 to 0.7864, while map-informed distillation yields a motion-only student at 0.7225. Through a compact query interface, entropy gating achieves 0.7729 macro-F1 versus 0.7738 for all queries, querying 39.5% of trajectories and reducing application payload by 60.5%. Under user-day grouping, selective and all-query scores are similarly close at a 21.6% query rate. Learned value estimation and probability fusion offer no consistent improvement. These results illustrate communication-efficient edge--server inference for an intelligent transportation sensing task.

Serverless computing is increasingly used to run complex DAG workflows across diverse applications, and individual function execution times have grown to support them, pushing workflow makespans to hours. Open-source serverless platforms, deployed on a pool of VMs, typically use on-demand instances, and their cost increases with the makespan. Spot instances are far cheaper, but a preemption is likely before the workflow finishes. A preemption at a fan-out node or a task feeding a fan-in node stalls every parallel branch, cascading a single preemption into a workflow-wide failure. Prior systems address this challenge by migrating high-impact tasks to reliable, non-spot resources, sacrificing spot's cost advantage. We present FaaSpot, a serverless workflow execution system that runs every workflow task on spot instances with no on-demand fallback. Our key insight is that cascade-prone positions are determined by the DAG's topology and can be identified before execution. Therefore, FaaSpot analyzes the workflow DAG once at submission time and classifies each task by the impact of its preemption. Selective checkpointing, survival-aware scaling, and lineage-based retry then act on this single classification, with no runtime coordination. We evaluate FaaSpot on a trace-driven spot emulator with eight DAG workflows against four state-of-the-art baselines. FaaSpot completes 89.03% of runs at $0.018 per successful run, 21pp (percentage points) above the average baseline success rate and at a lower cost per success than every baseline, while reducing mean latency by 2.3x.

Estimating heterogeneous survival treatment effects (HSTEs) from time-to-event data under time-varying treatments is important in fields such as medicine (e.g., to decide which sequence of treatments is better in terms of survival probability). However, this task is challenging because treatment decisions, patient characteristics, and censoring probability often evolve together, requiring sequential adjustment for both treatments and censoring. Existing survival learners mostly consider a single baseline treatment assignment, while longitudinal treatment-effect learners typically assume uncensored outcomes. In this paper, we introduce a novel doubly robust learner for heterogeneous survival treatment effects under time-varying treatments and history-dependent censoring. Our framework separates the intervention window from the survival horizon, with treatment prescribed for a finite period, after which care follows its natural course. We further establish the favorable theoretical properties of Neyman-orthogonality, rate double robustness and quasi-oracle efficiency. The framework accommodates flexible machine learning models for nuisance estimation and treatment-effect regression. We evaluate our learners through several numerical experiments, demonstrating their effectiveness.

Recent video generation models produce visually compelling outputs, yet their physical plausibility remains poorly measured. Physical plausibility asks whether generated objects obey geometry, kinematics, dynamics, and conservation laws. Existing benchmarks rely on coarse visual-language model judgments or target embodied world models rather than video generation itself. We introduce PVBench, a physical video benchmark that quantitatively measures the plausibility of generated videos across 46 dimensions organized into seven domains, spanning: i) instruction and semantic fidelity, ii) behavioral and causal consistency, iii) perceptual and temporal quality, iv) geometric and spatial consistency, v) material properties and field states, vi) kinematic and spatiotemporal persistence, vii) and mechanical plausibility with conservation laws. PVBench comprises 600 carefully crafted prompts, of which 30% involve physical embodied agents such as robot arms, humanoids, quadrupeds, and multi-agent systems. For challenging physically tractable dimensions, it pairs each with specialized vision models that derive verifiable, physically grounded measurements, such as object-level 6D pose, metric scale, surface normals, kinematics, and numerical conservation-law residuals. It thus avoids relying solely on subjective scoring, while supporting both text-to-video and image-to-video generation. Finally, we validate the agreement between automatic metrics and human judgments through human evaluation. Meanwhile, we evaluate mainstream open-source and closed-source models and present several insights. By releasing the benchmark, prompts, and calibration protocol, we aim to foster physically grounded video generation.

We extend the augmented graph of convex sets GCS framework for motion planning with temporal logic precedence specifications to dynamic environments in which obstacles, keys, and doors undergo piecewise-linear translation and/or dilation over time. The main insight is that piecewise-linear motion of polytopic sets in R^d lifts exactly to a static polytopic partition of spacetime R^d X R_>=0. We show that the main result from the static case carries over to spacetime essentially without modification: a shortest path in the spacetime augmented GCS provides an exact optimal solution to the motion planning problem with precedence specifications. Moreover, the layered subgraph lattice that encodes signal temporal logic (STL) precedence specifications retains precise correspondence with the Bellman--Held--Karp (BHK) algorithm, and the full library of restricted but expressive STL specifications generalizes to the dynamic setting. Further, the framework allows time-gated keys and doors, collectible or passable only within a temporal window. The geometric side of the framework requires two modifications: a causality constraint (the time coordinate of a trajectory must be non-decreasing) and a spacetime speed constraint (bounded physical velocity), both of which are convex in the Bézier control points.