Science & Mathematics

Inference-time harnesses, i.e., external procedures that call a model many times and keep state across calls, can substantially improve model capabilities. In this paper, we investigate the benefits of using such harnesses and the extent to which these benefits can be internalized through training. More precisely, we conceptualize a harness $\mathcal{A}$ as an arbitrary wrapper around a base model $M$ and attempt to distill a student model from the harness's traces. Empirically we find that, even with simple scaffolds and small models, the inference-time harness significantly outperforms a distilled student. We then prove that such a gap is due to a fundamental representational limitation: inference-time harnesses can compute a closure of their per-call primitive that is $\NL$-complete even when the primitive is a trivial $\AC^0$ lookup, and that no fixed-depth forward pass can represent under the widely believed conjecture that $\NL\not\subseteq\TC^0$. Our theory suggests that expanding the representational capacity of the student relative to the task at hand, e.g., by enlarging the model, making the task easier, or lengthening its chain-of-thought, is necessary to close the gap, which we verify empirically. Finally, we scale to an agentic harness on competition math, where the same gap reappears. Overall, our results show that a harness confers on the base model a profound gain in representational power, one that training a same-class student on its traces cannot transfer back into the weights.

Public charging of electric vehicles (EVs) adds a large and synchronized load to the distribution feeders. In most networks, it is rare to overload a feeder at the same time. When it does happen, the consequences can be severe: an overloaded feeder stresses its transformers and can trigger cascading outages across a city. Planners who design for high reliability therefore need to estimate very small outage probabilities. To that end, plain Monte Carlo estimation with an expensive grid simulator can require a prohibitively large number of runs for this task. To reduce the number of runs, we propose a control variate (CV) built from a learned classifier of the outage, trained on cheap observable features such as temperature and load forecasts. We show that for any binary learner, the variance reduction is an explicit function of its recall and false-positive rate. Furthermore, we show that control variate based on a Gaussian proxy of the load loses its effect as outages become rarer, while a learner with non-vanishing precision keeps removing a fixed fraction of the variance even in the extreme rare-event regime. We verify these results in numerical experiments using a feeder-overload simulator across three orders of magnitude of outage probability and obtain that even weak learner with a single feature can remove a quarter of the variance, while a boosted-tree learner is able to remove about 60%. We then apply the estimator to the feeder network of Surabaya, Indonesia, in which the Gaussian load model leaves every feeder three to seven times the one-day-in-ten-years blackout target. At the same simulation budget, planning the reserve with the learned CV and a 95% confidence limit reduces the city-wide blackout days by a third compared with Monte Carlo. These results show that learned control variates can help planners design distribution grids that stay reliable against rare outages.

In robot-assisted education, a robot must understand how to complete a task in order to effectively guide a student through it. We present a novel method that enables a semi-autonomous, camera-equipped robot to recognise the task goal and plan actions to achieve it. Our work focuses on Shoebox Tasks, an activity commonly used in the education of individuals with Autism Spectrum Disorder (ASD). These tasks are designed so that the goal can be inferred directly from their visual structure. We have developed a solver that uses a vision-based Large Language Model to convert photos of a Shoebox Task into a Planning Domain Definition Language (PDDL) representation and generate action plans for solving it. We have also implemented a robotic interface that enables a Nao robot to autonomously execute various Shoebox Tasks and assist a student by suggesting next steps or signalling detected failures through action demonstrations, voice commands, or gestures. With an adapted robotic interface, the solver can be integrated into a variety of educational robots. No programming expertise is required from teachers, and the solver operates exclusively on captured images as input. Research in special educational needs indicates that students with ASD can benefit substantially from robot-assisted teaching, and our system aims to strengthen these learning opportunities while also fostering interest in robotics among individuals with ASD.

Tripos theory constructs categorical models of higher-order logic from indexed logical data. We develop higher tripos theory and higher tripos-to-topos construction recovering the classical construction on zero-types through a covering criterion. The higher tripos-to-topos construction yields elementary $\infty$-toposes with closed bounded hierarchies of univalent universes, thus establishing a logical adjunction between $\infty$-triposes and elementary $\infty$-toposes with univalent universes. We also construct a logically conservative higher completion of regular $1$-categories with weak section families and a finite-limit-closed projective cover. This completion preserves original maps, predicates, and existing dependent products, and gives criteria for extending interpretations while preserving logical structure. With strongly inaccessible size bounds, Kleene assemblies instantiate the construction, producing a cumulative hierarchy of univalent universes that classifies every family in the completion and recovering the effective topos as the category of zero-types. We then establish an arithmetical completeness result: for a uniformly computably presented family, the numerical index set for identity with a fixed classifier code is $Π^0_2$-complete. We also construct two inequivalent univalent classifiers with the same global homotopy type: every internal map from one to the natural-number object is constant, whereas the other admits a nonconstant such map. These reconstruction, preservation, and computability results contribute to the higher effective-topos program of Awodey-Emmenegger and Anel-Awodey-Barton.

Despite substantial progress in Deep Search agents, their ability to conduct end-to-end research in real-world open Web environments remains insufficiently evaluated. Unlike static retrieval settings, information on the real-world Web is distributed across diverse sources, modalities, and heterogeneous formats, often involving complex interaction mechanisms that require agents to autonomously search for, access, and integrate relevant evidence during execution. To address this gap, we introduce the OpenDS task, designed to evaluate agents' ability to perform multi-step deep search in the live, open Web. This task requires agents to autonomously explore the real Web, discover and connect relevant evidence from multiple independent sources through Web search, browser interaction, and evidence processing, and ultimately determine a unique factual answer. Building on this task formulation, we construct OpenDS-World, a benchmark comprising 200 challenging tasks spanning diverse information modalities, heterogeneous file formats, and interactive access requirements, with independent human verification to ensure evidence validity and answer uniqueness. We evaluate 12 frontier models under a unified tool environment and execution protocol. The best-performing model achieves 39.0% accuracy, while human evaluation across 1,000 independent attempts yields 21.1% accuracy. Further analyses of tool usage, search cost, termination behavior, and failure recovery reveal substantial differences in research strategies across models, as well as persistent challenges in tool coordination and long-horizon search management. OpenDS-World provides a unified evaluation platform for systematically assessing and advancing agents' autonomous deep search capabilities in real-world open Web environments.

A growing research community spanning quantum batteries, molecular and excitonic energy transport, polaritonics, quantum materials and nuclear control has begun to describe its work in terms of quantum energy or quantum energy science. What connects these otherwise disparate efforts is the deliberate use of quantum states, couplings, collective effects and engineered environments to control energetic states and pathways. A key difference from quantum information science is that the complete quantum state does not always need to be preserved; engineered dissipation or dephasing can even improve the desired outcome. We organize this emerging domain through a framework that connects mechanisms, functions, capabilities, and applications. We identify five elementary functions (Excite, Hold, Release, Transfer, and Transform) and examine the mechanisms that implement them. The same functions recur from the molecular to the nuclear scale, although their relative implementation difficulty can differ sharply. Moving toward technology, we highlight three engineering capabilities (Storage, Routing, and Conversion) through the performance metrics that matter for their applications. Stretch scenarios illustrate what performance targets could make these capabilities technologically consequential.

We study the shift equivalence testing problem for high-degree sparse multivariate polynomials. Given explicit spar se representations of two polynomials $f,g\in\mathbb{Q}[x_1,\ldots,x_n]$, with exponents encoded in binary, sparsities $s$ and $t$, and total degree at most $d$, the task is to decide whether there is a vector $\mathbf{a}\in\mathbb{Q}^n$ such that $f(\mathbf{x}+\mathbf{a})\equiv g(\mathbf{x})$. Furthermore, one needs to output such a vector if it exists. Dvir, Oliveira, and Shpilka (ICALP 2014) studied shift equivalence for low-degree arithmetic circuits. Their framew ork does not apply directly to sparse polynomials, which are not closed under shifts, and they left open whether shift equivalence of sparse polynomials admits an efficient deterministic algorithm. We answer this question positively over $\mathbb{Q}$. If the input coefficients have bit-size at most $τ$, we sh ow that a rational shift of polynomial bit-size exists whenever the polynomials are shift equivalent. Moreover, our alg orithm outputs such a shift and has a running time of $\operatorname{poly}(n,s,t,τ,\log d)$ in the bit model. The ar ithmetic and structural parts of our result also hold over any field of characteristic zero or characteristic larger th an $d$.

Agentic systems increasingly carry out complex work, where a single turn routinely runs for minutes. This produces a micro-wait that is too long to watch and too short to leave. In a formative study with 12 daily agent users, we found that staying felt like wasted time, while switching away fragmented attention and forfeited the context needed to re-engage. We introduce Sidecar, a parallel assistant situated in the same interface as the agent, which reads the run and surfaces implicit decisions the agent made, along with task directions worth exploring. In a within-subjects study (N=16), Sidecar helped participants stay focused on their main task, freely workshop unfinished ideas without fear of unwanted action, and recognize under-specification in their own prompts. Together, these findings contribute to designs that help users stay context-aware and engaged during agent micro-waits.

Academic research in wireless communications has no tradition of benchmarks so far. Remarkable progress in benchmark-oriented fields like machine learning suggests that the adoption of benchmarks could benefit our field. In the absence of benchmarks, however, a scientific debt has accumulated: since the academic community has not canonized specific problem setups, methods that address essentially the same problem were often developed for slightly different setups and lack a common framework for comparison. Any attempt to establish a new benchmark is faced with the challenge of resolving this debt. Recent advances in agentic artificial intelligence (AI) create an opportunity to do so by leveraging AI-assisted coding. This makes it possible to implement the state-of-the art at scale and to adapt it to a common framework for comparison. We present Sionna Bench to initiate benchmarks in academic wireless communications research. As a pilot study, we use the problem of MIMO-OFDM channel estimation. At the time of writing, Sionna Bench's channel estimation benchmark features twenty channel estimation methods that were implemented from ten different papers. The ability to compare methods against the state-of-the-art will be crucial for navigating the research landscape in an era where AI-generated methods coexist with human-developed algorithms.

Editing agents on 2D infinite canvases must reason about application state that pixels may not reveal. Targets may be offscreen or covered, while visual containment and touching arrows may disagree with recorded frame membership or connector bindings. We ask how a structured observation layer can support accurate object selection during agentic editing on 2D infinite canvases. We present CanvasObs, which augments visual context with a structured payload and guidance for reading four components: Identity & State, Visibility & Structure, Relations, and Offscreen Object bounds in action coordinates. A Remote Team Handoff case study illustrates retrieving an editable offscreen copy, choosing a covered chart by color, separating cards by frame membership, and preserving an evidence connection. We turn these cases into ten controlled configurations and compare CanvasObs with the baseline agent in 1,200 browser trials across three models. CanvasObs improves Correct target for every model. Combined across models, it raises Correct target from 16.5% to 48.3%, an increase of 31.8 percentage points (95% confidence interval [29.0, 34.7]), and Full task success from 4.3% to 25.0%. These results show that compact structured observation can improve object choice on 2D infinite canvases without redesigning the editing pipeline.

Platforms that host young users screen messages with commercial text-safety APIs, but independent evidence on how often such systems miss harm or flag harmless text is scarce, especially outside English, and the documentation of the one studied here leaves its false-positive rate unquantified. We evaluated one such API, Tuteliq Detect as deployed under the release identifier tuteliq-detect-v5.20261011 (not the separately published v6 model), on six harm categories: authorised push-payment fraud, money-mule recruitment, social engineering, romance scams, gambling harm and bullying. We built SITO (Synthetic Independent Text-safety benchmark with Overflagging tests; sito is Ukrainian for 'sieve'), 2,332 synthetic cases in English, Ukrainian and mixed Ukrainian, with each harmful case matched by hard harmless ones: same-topic messages, minimal-edit contrasts, warnings and retrospective accounts. On English text the system flagged 84.9-94.1% of harmful cases and 0-4.3% of harmless ones per category. It missed third-party reports of ongoing harm far more often than perpetrators' messages (63.6% against 95.1% recall, all languages), when sent without the settings the API offers for reports of harm. Ukrainian and mixed-language input was detected at least as often as English, recall was lower on text from a held-out generator family, so headline recall may be optimistic, and 3.2% of cases received a different decision in repeated runs. A platform should measure false positives on its own traffic and check how bystander reports are handled. The datasets, generation prompts and code are released.

Learning latent dynamics is central to assessing the current states and forecasting the future trajectories of high-dimensional time series. For locally-stationary latent dynamics $\mathcal{F}(\mathbf{z}_{

From 3 October, every non-bot merge into our main branch passed a non-author review gate, yet 71 of those 263 merges (27%) rested only on an approval that a coding agent had submitted through the technical lead's GitHub account, and the platform recorded each one as an ordinary approval by the lead. We report twelve weeks of agent-driven development at Bashi AI, a fintech startup preparing for a SOC 2 audit: 786 merged pull requests, 657 of them associated with agents by explicit markers or by inference, analysed from platform records and saved agent-session transcripts. In the gated phase, an agent session submitted every approval the lead's account gave on a colleague's pull request, and issued nearly all of its merges; the records cannot show who decided any of them. We recompute, for every merge, a core gate property: a non-author approval and a passing required check, both before the merge and both on the exact revision merged. The recomputation confirmed the gate wherever branch protection applied and found where it did not: agents switching enforcement off around individual merges, a merge script that recorded failed checks as passing, and an agent collapsing a stack of pull requests through unprotected branches. The engineers' survey answers show that one engineer's approvals also rested on review agents, with the person validating findings and deciding, and both engineers say review mattered when it tested a change against something outside the diff. We draw lessons on recording decision authority separately from account identity, keeping agents' credentials away from the controls they operate under, and checking evidence at the branch and revision where integration happens.

Modern language models are increasingly asked to handle several tasks within a single prompt. Doing so requires solving each task correctly while integrating the resulting answers into a coherent response. In this paper, we identify a fundamental limitation in this process, where models make substantially more errors when tasks are composed, even when they can solve every constituent task reliably in isolation. This degradation cannot be fully explained by the accumulation of independent errors. Across both open-weight models and the frontier model Claude Opus 5, our findings instead point to an implicit thinking budget underlying this effect. In particular, under composition, models use substantially fewer tokens per task than in isolation, even when the full sequence remains well within the context window. Tasks that require longer solution traces in isolation are consequently more $\textit{fragile}$ under composition, i.e., models make more errors on prompts containing these tasks. In isolation, non-fragile tasks receive more reasoning than they need and waste compute. However, by grouping tasks according to their fragility, the implicit budgeting can be manipulated to reduce token usage by $43$\% across a range of math and linguistic tasks while also preserving accuracy.

Multimodal large language models (LLMs) are increasingly used for structured extraction from financial documents, but their behavior under visual ambiguity is poorly understood. We study CARLAR (Courtesy Amount Recognition and Legal Amount Recognition) influence: a model returns the same amount for both cheque fields even though their annotations differ, including cases with a blank annotation encoded as zero. Across ten visual prompting conditions, we evaluate GPT-4.1 and Claude 4.5 Sonnet on 243 primary cheques (124 with two observed unequal amounts) and 598 strict mismatches from two additional datasets. Under courtesy redaction, GPT-4.1 CARLAR reaches 73.7 percent (95 percent bootstrap interval: 68.3 to 79.0; p below 0.001 versus baseline). Model-reported confidence can remain high when predictions agree; expected calibration error reaches 0.93 among CARLAR cases. We also test a compressed-image few-shot intervention that presents examples as one visual table without model fine-tuning. On the 243-case cohort, Claude's Control condition increases paired joint extraction accuracy by 11.5 percentage points (95 percent interval: 6.2 to 16.9; adjusted p = 0.00054), but its strict-124 gain is smaller and not significant after correction. The intervention's effect on predicted-field agreement is model-dependent: it decreases for Claude and increases for GPT-4.1. These results motivate validation methods that account for visual context and dependencies between extracted fields.