Science & Mathematics

This paper describes our submission to the WMT 2026 Multilingual Instruction Shared Task. We adapt Qwen3-8B with LoRA using organiser-provided data, EuroAlpaca, a multilingual instruction dataset created through task-preserving localisation, and CrossEuroAlpaca, a cross-lingual augmentation that assigns different source and target languages to instructions, contexts and responses. Our system targets open-ended question answering (QA-OEG) in ten shared-task languages represented in the auxiliary data. Under matched decoding, adaptation improves the mean Gemma-4 judge score on target-language QA-OEG by 7.8% relative to the base model. On an additional evaluation using the Aya Evaluation Suite across all shared-task languages, it yields relative gains of 41.5% in macro-averaged ROUGE-L and 2.9% in BERTScore-F1, with larger improvements in target languages. However, these gains are task- and language-specific. Gemma-4 scores improve for QA-OEG and summarisation in the target languages but decline for context-grounded QA there and across all three tasks in the non-target languages.

A benchmark is not merely a collection of test questions; it naturally induces an empirical measure over difficulty. In this work, we develop this measure-based view of benchmark evaluation. This perspective allows us to identify and formalize two central failure modes: measure mismatch, where the benchmark difficulty distribution differs from the intended difficulty distribution, and coverage gaps, where uncovered regions along the difficulty axis prevent benchmarks from detecting meaningful localized training progress. We propose two diagnostic metrics that capture these two failure modes and provide rigorous theoretical guarantees. Empirically, we analyze 25 popular benchmarks spanning 7 domains using carefully estimated item difficulties. Our results show that many benchmarks implicitly overweight certain difficulty regions, so their leaderboard conclusions should be interpreted with caution; others leave substantial portions of the difficulty axis uncovered, limiting their ability to monitor the model improvement process. Finally, we turn diagnosis into action by introducing BenchFill, a simple difficulty-targeted item rewriting algorithm that effectively reduces local coverage gaps.

End-to-end piano audio-to-score (A2S) transcription is the task of reconstructing a notated score from an audio recording in a single system. Existing A2S approaches usually rely on pre-segmented excerpts, and few handle whole pieces without external segmentation. This raises the question of whether a musical-time coordinate can guide score transcription while also helping the system segment whole pieces. Previous work on jazz saxophone A2S uses an external beat tracker to obtain such a coordinate for score prediction. We introduce downbeat phase, a musical-time coordinate learned jointly with score transcription that represents each audio frame's relative position within a notated bar. With this coordinate, our system can not only condition its score decoder on downbeat phase but also self-segment at bar boundaries and stitch the segments into a whole piece. Evaluation suggests that the coordinate improves metrical structure and supports whole-piece transcription with comparable notation quality.

In response to vast and widespread online tracking, regulators and technologists alike have worked to develop laws and tools that allow people to express their privacy preferences. These measures seek to offer users greater control over the sale and sharing of personal information, putting agency in their hands to protect their privacy. However, the web is undergoing a seismic shift, away from strictly user-driven behavior towards agentic behavior, where AI agents take autonomous actions on behalf of users. In this study, we ask: What should privacy preference controls look like in the agentic web and agentic systems more generally? To answer this question, we develop a taxonomy of enforceable privacy preferences for agentic systems. Our taxonomy is informed both by laws and empirical studies of existing privacy tools as well as an audit of 30 leading AI assistants and agent platforms. To explore how our taxonomy can be implemented, we design and develop eight open-source prototypes that mirror real-world agentic applications and demonstrate how major aspects of our taxonomy can be structurally integrated in agentic protocols, such as the Model Control Protocol and Agent2Agent protocol. Ultimately, our study seeks to provide the groundwork for ensuring end-user privacy and autonomy is not lost in the newly arriving agentic world.

Can LLM agents make contributions to frontier computer science R&D? We introduce FrontierCS 2, a benchmark of 48 open-ended tasks. We take a paper and the code its authors released and delete the paper's contribution. The deleted code becomes the human reference that sets the bar. Earlier open-ended benchmarks give each agent a problem statement and grade one program after the run. FrontierCS 2 places the agent inside a research environment. The agent inherits a working repository and may submit a patch at any point. For each patch, a judge returns a development-suite score but keeps the evaluator, the held-out final suite and the reference private, and grades the final submission on that suite against the reference. The tasks span 13 domains, from machine-learning systems and high-performance computing to databases, security and robotics. Beyond these 48 tasks, FrontierCS 2 keeps a private set of more than 100 tasks, to be announced in the future. We evaluate 12 agents. The best one is Claude Opus 5.5, passing 58% of its runs. No run passes on 16 tasks. A higher score is not the same as a pass: most runs that beat the reference on the combined final score still fall short on some metric or workload. On a shared capability scale for agents and tasks, the median task is roughly as difficult as what the strongest agent can handle.

Modern Android apps extensively rely on native libraries, bridging Java/Kotlin and C/C++ through the Java Native Interface (JNI). This cross-language interoperation, however, exposes the memory-safe Android Runtime to the unsafe C/C++ world, creating an attack surface that has been repeatedly exploited for full device compromise. While fuzzing is a proven technique for vetting native libraries, state-of-the-art harness generation tools struggle to faithfully capture closed-source cross-language API usage: (i) precise program analysis across the JNI boundary does not scale, whereas lightweight approximations miss critical invocation patterns; and (ii) none tackles the construction of complex composite objects consumed by JNI methods. In this paper, we propose JANUS, the first agentic approach to synthesizing cross-language JNI-aware fuzzing harnesses for Android native libraries. JANUS iteratively resolves both explicit and implicit cross-language call dependencies to recover app-specific runtime context, and recursively assembles cross-language argument objects in a bottom-up manner. JANUS leverages an autonomous agentic workflow to reduce reliance on hand-crafted heuristics, capturing the vast landscape of hidden API usage knowledge embedded in closed-source native libraries. Evaluation on 44 popular real-world apps (such as Google and YouTube Music, covering 21,386 JNI methods total) shows that JANUS produces over 5.9 times as many valid harnesses (6,094) as state-of-the-art automatic harnessing tools. On pre-installed apps from five mainstream Android vendors (e.g., Xiaomi, Vivo, Samsung), JANUS further uncovers 6 previously unknown vulnerabilities in closed-source native libraries.

Retrieval-augmented generation (RAG) frequently integrates knowledge graphs (KGs) for complex multi-hop question answering. However, their performance is increasingly bottlenecked by the knowledge structuring and reasoning processes. (i) Severe semantic loss: Distilling continuous texts into discrete symbols strips away contextual nuances and narrative coherence. (ii) Pronounced modality mismatch: Fragmented graph structures clash with the continuous texts LLMs are pre-trained on, while rewriting them into natural language introduces generative noise and omits information. To address these limitations, we propose ThreadRAG, a semantic-topology-based RAG framework for context-preserving knowledge structuring and modality-native evidence reasoning. ThreadRAG introduces a Semantic Topology Mapper (STM), which organizes retrieved documents into context-intact Semantic Nodes and constructs a Semantic Topology Graph (STG) through narrative, semantic, and lexical edges. It further incorporates Energy-Driven Pathfinding (EDP), which identifies query-relevant subgraphs via energy diffusion and deterministically linearizes them into coherent textual evidence paths. Experiments on three multi-hop QA benchmarks demonstrate that ThreadRAG consistently outperforms strong baselines. Notably, it achieves a 14.4% EM improvement over text-centric methods on HotpotQA and surpasses the best baseline by 5.8% EM on 2WikiMultihop, while reducing token cost to 23.3% of the strongest baseline.

Forming an accurate theory-of-mind (ToM) is key to successful multi-agent collaboration and requires reasoning under uncertainty to capture multiple possible partner states. This is hard to do in the text representation space native to large language model (LLM) agents, which more easily describe a single partner state. To address this challenge, we introduce ToM2Vec, which learns to form abstract theories-of-mind in a rich latent space via a fixed budget of continuous reasoning tokens. During training, supervision for these tokens comes from the partner's actual thoughts, which the ToM tokens are trained to abstract and match. At test-time, ToM tokens are generated by each agent and act as conditioners for that agent's thoughts and actions. Across several environments covering collaborative, competitive, and mixed-motive settings (Overcooked-AI, multiparty negotiation, Decrypto, Hanabi, and Collab Escape/Capture), we evaluate ToM2Vec's ability to support multi-agent interactions and collaboration, finding consistent improvements over the base model and a variety of competitive trained and prompted baselines. We show that latent reasoning is critical, with performance dropping when reasoning is expressed in text. Moreover, we show that ToM2Vec training generalizes to partner model families not seen at train-time, including a proprietary model (GPT-5.4). In our analysis, we provide evidence that the latent tokens encode a ToM, rather than task-specific information: ToM2Vec's tokens improve the agent's ability to directly answer questions about partner beliefs, desires, and next actions.

Recovering metric world-space hand motion from egocentric video supports understanding human interactions and transferring them to embodied agents. Three challenges complicate this task: multi-stage pipelines propagate errors from independently estimated hand and camera motion; monocular observations leave physical scale ambiguous; and occlusion and out-of-view hands admit multiple plausible motion trajectories. We present DewHand, a unified model for direct joint reconstruction of world-space hand and camera motion. To avoid compounding errors from multi-stage estimation, DewHand jointly predicts hand and camera motion in a shared world coordinate system within a single-stage pipeline. To resolve metric-scale ambiguity, it combines scene-level geometric features from VGGT-$Ω$ with metric-space supervision, while hand-specific features preserve fine-grained articulation. To handle occlusion and out-of-view hands, we formulate reconstruction as conditional motion generation with flow matching. Under the same architecture, this generative formulation outperforms direct regression in both accuracy and temporal smoothness with only two sampling steps. Experiments on HOT3D and ARCTIC demonstrate improved world-space accuracy and temporal consistency over state-of-the-art methods. DewHand reduces rigid-aligned world-space error relative to StableHand from 32.9 to 28.4 mm on HOT3D and from 76.2 to 36.2 mm on ARCTIC. It also recovers camera scale within 10% on two thirds of HOT3D clips, versus about a quarter for SLAM. Moreover, in zero-shot evaluations on HoloAssist, H2O, and TACO, captured with different devices, DewHand achieves the lowest rigid-aligned world-space error (WA-MPJPE) on each dataset.

CLIP's 77-token text encoder truncates detailed image descriptions. We present ERCLIP, a small-data adaptation of CLIP for long-caption image--text retrieval. During training, a long caption is split into sentences that the CLIP text encoder encodes independently; the sentence embeddings are fused by mean pooling or a gated residual Transformer scorer and receive their own contrastive loss. At inference ERCLIP is an unmodified CLIP dual encoder: test captions are still read through the 77-token window, and only training sees the full caption. We find that this fine-tuning erodes CLIP's zero-shot classification, most when the fusion output is not a combination of CLIP embeddings, and add a frozen-CLIP preservation loss on the image and short-caption embeddings. With 20k pairs and one epoch, ERCLIP improves long-caption R@1 over CLIP by 6.5--11.2 points, closing 62--77% of the gap to LongCLIP with 2% of its long captions. Preservation recovers 1.5--2.1 ImageNet zero-shot points (30--42% of the loss) and adds up to 2.4 points of COCO image-to-text R@1, at no inference cost.

POEMMA-Balloon with Radio (PBR) is a hybrid fluorescence and optical Cherenkov telescope that is a pathfinder instrument for a future Probe Of Extreme Multi-Messenger Astrophysics (POEMMA), together with a radio instrument, to be flown on a super pressure balloon in Spring 2028. One of its science objectives is follow up of astrophysical event alerts via signals of tau decays in the atmosphere, sourced by Earth-skimming tau neutrinos. The recent report of the KM3NeT Collaboration's muon event with $ E_μ=120^{+110}_{-60}$ PeV sparked searches for beyond the standard model (BSM) physics explanations. Using a specific, boosted dark matter scenario which together with the KM3NeT observation, sets a scale for the fluence of boosted dark matter in this energy range, we assess the capabilities of PBR to constrain or discover such a source for a boosted dark matter energy spectrum above $\sim 10^8$ GeV. PBR has discovery potential for dark matter energies above $10^9$ GeV, and depending on BSM input parameters, to dark matter energies as low as $1.5\times 10^8$ GeV. While the analysis is focused on one BSM example, the methodology is applicable to any BSM physics that affects tau production in the Earth.

Indian court proceedings are held in English, so speakers of languages with limited translation technology have limited access to them. We present the first English Mizo parallel dataset of Indian court judgments. It contains 13,168 sentence pairs from the Gauhati High Court and the Supreme Court of India. Professional native Mizo translators translated complete documents; sentence pairs were manually aligned, and two independent reviewers checked the results. Our study finds that English words are borrowed in a very selective way: official and legal terms are almost always kept in English, but words for everyday things like actions, situations, and results are usually expressed in native Mizo. Mizo diacritics are also used inconsistently across the corpus. When we fine tune the NLLB 200 1.3B model on our training data, it scores 27.47 BLEU on a balanced legal test set of 1,000 sentence pairs, an improvement of 7.16 BLEU over the base model. However, it still uses English words about three times more often than a human translator does.

Power grids use a constantly changing mix of fossil fuels and clean renewables like solar and wind. Because of this, the {\em carbon footprint} fluctuates drastically based on time and geographic location. Time-varying carbon intensity creates opportunities to reduce carbon emissions by shifting workloads to cleaner periods, but deadlines, task dependencies, and limited edge capacity constrain this flexibility. We study carbon-aware scheduling of jobs with release times, deadlines, and arbitrary precedence constraints under piecewise-constant carbon costs. Jobs execute preemptively on a single edge server or are outsourced at job-dependent costs to an unlimited-capacity cloud, modeled as providing instantaneous execution. The objective is to minimize the sum of edge carbon costs and cloud outsourcing costs while satisfying all timing and precedence constraints. For edge-only execution, we give an exact polynomial-time algorithm. Allowing cloud outsourcing makes the problem NP-hard, but we obtain a fully polynomial-time approximation scheme through a reduction to preemptive scheduling with rejection penalties (weighted tardiness). In contrast, even two edge servers without job migration admit no polynomially bounded approximation when carbon-cost variation is unrestricted, unless \(P=NP\). We complement these theoretical results with experiments on controlled synthetic workloads and production-derived Alibaba traces. On synthetic instances, our incremental max-flow implementation is $4.96\times$ faster on average than a direct min-cost flow, while experiements on production-derived workloads evaluate its performance under realistic job structures.

The Gaussian Mumford-Shah model segments an image into regions by minimizing a continuous multivariate Gaussian negative log-likelihood term plus a perimeter penalty. It is used for scalar- and vector-valued images, in particular hyperspectral images, but the existence of a minimizer in general is not guaranteed. The energy is unbounded below whenever the covariance of a single region degenerates, regardless of the perimeter of the regions. A quantitative covariance floor on the covariances of all regions, or equivalently uniform Hölder continuity with exponent at most one of all one-dimensional marginals of the histogram, guarantees a lower bound on the energy and the existence of a minimizer. Our approach therefore shows that existence follows from a property of the histogram of the image, and we furthermore argue that scale-independent covariance floors cannot exist. Admitting a quantitative covariance floor is generic: it holds almost surely for all images with suitable noise. As applications of our theory, we show the convergence of regularized models used in practice and of numerical methods.

Control structures and operability indices for plantwide processes are developed and compared on fixed simulated flowsheets, and the configuration of a study is rarely reported in a form that another group can reproduce. In this contribution, plantbench, a library of plantwide process simulations in which the design, the controller tuning and the instrumentation of a plant are declared as hashed plain-data configurations, is presented. On a heat-integrated reactor--separator--recycle plant with an open-loop unstable reactor, a heuristic plantwide control design prices four heat-exchanger networks (HENs) by the least damping ratio of the linearized closed loop, and 2016 closed-loop runs over networks, tuning, analyzer deadtime and upsets test that ranking. Results show that the damping ratio orders the networks as the runs do with the network that recovers the most heat reaching 0.013 against 0.157 for the selected one and cycling under a throughput increase, but that it is a weak detector of which operation will oscillate. At the recall of a random forest on the first 100 min after an upset, 60 of 82 limit cycles, a damping-ratio threshold raises 173 false alarms against 5. The integral time of the temperature loop, to which the damping ratio is nearly insensitive, decides in part whether the plant holds an operating point.