Papers

We introduce ABLE, a benchmark for evaluating LLM agents' ability to use biological AI models (BAIMs), such as ProteinMPNN and AlphaFold3, in dual-use protein design workflows. ABLE assesses agent performance through a set of tasks spanning structure retrieval, sequence generation, and design validation. We evaluate 15 frontier models and find that seven refuse all tasks, while the remaining models exhibit substantial performance differences. Claude Sonnet 4 and Gemini 3 Pro achieve the highest scores across information retrieval, tool selection, and tool use. We further compare model performance on a subset of tasks against an expert human baseline. Our results suggest that current LLMs can substantially lower barriers to protein design, but remain inconsistent in planning, strategy generation, and integrating biological knowledge with tool use.

Crack segmentation requires a model to recognize target semantics while accurately recovering thin, low-contrast, and topologically continuous local structures. Although SAM3 provides strong open-concept segmentation, its direct application to the crack domain still misses weak cracks, activates crack-like background regions, and produces local boundary errors. We first diagnose the functional differences between the internal prompt-conditioned semantic representation and native visual representation of SAM3 on five crack datasets. The results show that the semantic representation already carries most task information for crack prediction, whereas the utility of the visual representation depends on the current semantic state. Directly combining the two representations does not yield consistent gains. Based on this finding, we propose Conditional Semantic–Visual Reconciliation, termed CoRe. CoRe retains semantic prediction as the primary decision path, applies lightweight semantic calibration to adjust the target-domain decision mapping, and uses spatially aligned native visual evidence to generate a zero-initialized, bounded, and regularized conditional residual that selectively corrects existing predictions. Across five domains, CoRe-SAM3 improves the average Crack IoU from 62.34% to 70.47% and clDice from 81.98% to 89.24%, while introducing only 18.914 K trainable parameters. Prediction-transition analysis further shows that CoRe corrects an average of 34.38% of native errors, with a damage rate of only 0.23% on pixels correctly classified by native SAM3. These results demonstrate that constrained prediction correction based on the functional differences between internal representations provides an effective and parameter-efficient target-domain adaptation strategy for vision foundation models with strong task-specific semantic priors.

Quantum applications increasingly execute as multi-stage quantum-classical pipelines, interleaving QPU computation with classical stages like circuit generation, transpilation, layout mapping, quantum error mitigation (QEM), and post-processing. These stages have diverse resource requirements and exhibit stochastic behavior under drifting hardware noises, yet existing workflow frameworks treat them as static, isolated components. We present LLQM (Low-Level Quantum Machine), a profiling-driven meta-framework for quantum-classical pipelines. LLQM decomposes pipelines into fine-grained tasks and continuously profiles their CPU/GPU, memory, QPU, and queue dependencies alongside real-time hardware states. This unified runtime abstraction captures cross-stage resource dependencies and reveals how classical and quantum decisions interact, enabling characterization of their impact on fidelity and resource consumption. We evaluate LLQM using QEM as a representative pipeline stage, on IBM 156-qubit Heron r2 processors with circuits up to 100 qubits and 1e7 transpiled gates. Our results show that continuous profiling exposes runtime bottlenecks and enables hardware-, fidelity-, and workload-aware optimizations.

We study the query complexity of estimating scalar summaries of zigzag bar lifetimes from snapshot-adjacency bits. For graphs $G_1,\ldots,G_m$ on $n$ labeled vertices, let $\ell_b$ be the snapshot lifetime of a degree-one bar $b$ of the intersection zigzag. For a probability generating function $\phi(x)=\mathbb{E}[x^R]$, the statistic $F_\phi=\sum_b\phi(\ell_b/m)$ includes normalized degree-$r$ total persistence and the mean generalized rank over a uniform time window. An exact identity underlies our algorithm: sample $R$ uniform times; the expected generalized rank between their minimum and maximum equals $F_\phi$. For graphs that rank is the circuit rank of an intersection graph, so a nonlinear barcode functional becomes an average of edge and component counts, and no barcode is computed. Without spectral-gap, homology-state, or QRAM assumptions, this gives a quantum estimator with additive error $\varepsilon n$ and $\widetilde O(\sqrt{m(K+n)}/\varepsilon)$ queries when a bound $K\ge F_\phi$ is supplied, against $\widetilde O(m\min\{n^2,(K+n)/\varepsilon^2\})$ classically, and an adaptive quantum variant with the same instance dependence. These estimators are optimal in two regimes. For every fixed power weight $x^r$, $r\ge2$, and for the uniform-window mean, the worst-case complexities are $\widetilde\Theta(n\sqrt m/\varepsilon)$ quantum and $\Theta(n^2m)$ classical. On sparse instances, under an explicit split-leakage promise met by power and binomial weights of logarithmic degree and the promise $F_\phi\le K$, they are $\widetilde\Theta(\sqrt{mK}/\varepsilon)$ and $\widetilde\Theta(m\min\{n^2,K/\varepsilon^2\})$. The classical lower bounds hold against fully adaptive algorithms, and fewer than $m$ such statistics cannot determine the positive-lifetime histogram. All bounds concern snapshot access; with an explicit update stream, near-linear full-barcode algorithms are known.

The constructive diagonal Ramsey problem asks, given adjacency-oracle access to an $N$-vertex graph, for a clique or independent set of the order guaranteed by Ramsey's theorem. We give a bounded-error quantum algorithm that, for every $K\ge2$ and $N\ge4^{K-1}$, finds and verifies a homogeneous $K$-set using $O\!\left(2^K K\log\frac K\eta\right)$ edge queries with failure probability at most $\eta$. At the Ramsey scale $N=2^n$, this yields a homogeneous set of order $\lfloor n/2\rfloor+1$ using $O(\sqrt N\log N\log(\log N/\eta))$ queries, improving on the $O(N)$ queries of the explicit classical recursion and giving, to our knowledge, the first sublinear worst-case algorithm for the Ramsey relation. We also derive an $\Omega(N^{1/12})$ quantum lower bound by a reduction from collision finding. The algorithm runs the constructive recursion over implicit candidate sets. Each set is represented by a short conjunction of adjacency constraints and sampled using capped unknown-solution quantum search, and a scale-aware concentration schedule balances estimation accuracy against the increasing cost of sampling deeper sets. We complement the upper bound with an $\Omega(N^{1-1/\sqrt2})$ randomized lower bound, transported from the random-Painter analysis of online Ramsey numbers, which holds on the uniform distribution $G(N,1/2)$. On that distribution a greedy quantum search uses only $\widetilde O(N^{1/4})$ queries, giving a provable polynomial quantum speedup for Ramsey search on random graphs. We also give an estimation-free size-biased recursion and extend it to every fixed number of edge colours.

In situ adjoint training extracts parameter gradients directly from measurement, but has so far been limited to reciprocal or restricted systems. We introduce the physical counterpart of the policy gradient theorem: a stochastic-adjoint gradient estimator that lifts these constraints by trading reciprocity for nondegenerate diffusion. As validation, we train a nonlinear resonator network, whose own dynamics supply the policy, against antagonistic temporal modulations with gradients from measured stochastic trajectories alone, without finite differences or a separate adjoint experiment.

Optical Chemical Structure Recognition (OCSR) is a fundamental component of chemical literature mining, enabling molecular database construction, reaction extraction, and AI-driven scientific discovery. Despite substantial progress in recognition accuracy with recent deep learning-based methods, inference throughput remains a critical bottleneck that limits web-scale deployment. To address this challenge, we propose MolParser-Mobile, an AutoML-optimized lightweight end-to-end OCSR framework. MolParser-Mobile contains only 9.98M parameters, while reaching a throughput of 1,520 molecules per second on a single NVIDIA RTX 4090D GPU. Despite its compact design, it maintains competitive and, on several benchmarks, superior recognition accuracy.

Emotion Recognition in Conversations (ERC) aims to identify speakers' emotions in multi-turn dialogue. Accurate emotion recognition can support a wide range of applications, including empathetic conversational agents, mental health support, and educational technologies. While many recent approaches rely on task-specific fine-tuning, such models may exploit dataset-specific cues. A central yet rarely questioned assumption in ERC is that each utterance can be assigned a single unambiguous emotion label. To investigate this assumption, we study ERC using Large Language Models (LLMs) in a zero-shot setting while incorporating preceding conversational turns as context. We show that aggregate metrics mask systematic failures. Errors concentrate around utterances containing negations, exclamations, and interjections. This pattern is consistent across all evaluated models, suggesting limitations in the benchmarks rather than model-specific weaknesses. A controlled re-annotation study involving four human annotators supports this finding: strong agreement is observed in only 35 percent of cases, with neutral utterances dominating high-agreement instances, while many emotional categories fall into low-agreement regimes. These findings suggest that many apparent model errors reflect genuine annotation ambiguity rather than poor emotion understanding. Standard single-label evaluation is therefore insufficient. To address this limitation, we introduce an LLM-as-Judge framework that evaluates each emotion independently according to its plausibility in the conversational context rather than enforcing a single-label decision.

This work compares a hemispherical graphene-based transmitarray antenna with its planar reflector counterpart for wide-angle beam steering in the THz regime. The theoretical framework of the planar reflector is formulated and numerically evaluated, yielding an elevation beam-steering range of \pm60{\deg}. In contrast, the transmitarray extends the elevation steering range to \pm78{\deg} while maintaining full 360{\deg} azimuthal coverage. The planar reflector exhibits a larger HPBW variation of 26.48{\deg}, compared with 13.1{\deg} for the transmitarray, resulting in a broader reflected-beam distribution and reduced directional power density, directivity, and gain. Meanwhile, the transmitarray maintains a more stable and controllable beamwidth response with greater directional power concentration over a wide steering range. The performance advantages of the transmitarray are demonstrated through comparisons with experimental results reported in the literature for planar reflectors and antennas. We further provide a comprehensive assessment of the performance advantages of the transmitarray over planar configurations across the remaining metrics.

Large language models answering questions over multi-page documents are expected to cite the supporting pages, yet supplied citations are sometimes inaccurate, and current evaluations score citations at generation time or against text passages: no existing benchmark evaluates whether a system can verify and correct a page-level citation already attached to an answer. We propose AtomCite, an agentic framework that parses an answer into claims, checks each claim against the image of its cited page, and applies a deterministic repair policy. To evaluate it, we introduce DocCite, to our knowledge the first benchmark for systems that verify and correct page-level citations in document images. Built on MP-DocVQA and DUDE, it combines 928 validated injected instances with 2,468 candidate natural errors harvested from frontier- and efficiency-tier models, of which a two-annotator audit confirms 1,909 as genuine errors. Primary labels are assigned deterministically, not by LLM judges, with the human audit as a separate validation layer. Across three model families (Gemini, Claude, and GPT), AtomCite reaches around 93% binary verification accuracy on the injected benchmark, significantly outperforming every OCR-only condition, including a compute-matched control, and exceeding every prior text-based baseline given the same OCR text. Its repair policy lifts citation precision on the injected mix from a constructed 34% to 87-90% while retaining over 90% of correct claims. AtomCite also transfers: with frozen prompts and zero training, it raises the hallucination-detection scores of two open 7-8B models on five public benchmarks above the same models prompted as direct judges. Finally, the audit shows that noise in automatic labels biases measured verifier accuracy and can reverse system rankings, so evaluations relying only on synthetic or automatic labels risk mismeasuring verification capability.

A document modeled as a discrete sequence of tokens can be thought of as being generated from a composition of texts from different domains; a README file, for example, moves between prose, code, and configuration. When such a document is corrupted and only frozen domain experts are available, restoring it requires deciding both what is missing and which expert to trust at each position, at test time and without region labels or a trained router. We introduce evidence-aligned local composition, which infers a soft, position-wise weighting over the experts from the marginal evidence of the corrupted observation under a given corruption model, estimating the evidence from the experts' own denoising losses and smoothing the weights across positions. Because the weighting is soft, it recovers a mixture when the true composition is mixed and concentrates on one expert when that suffices. Across a categorical simulator, byte-level experts, and experts fine-tuned from a $1.3$B discrete flow-matching model, the inferred weights track the true regions at $0.85$ field accuracy on naturally mixed scientific documents, and at $0.98$ on constructed mixtures whose regions are lexically disjoint. Restoration improves over a single global weight when the experts are genuinely distinct and reduces to it when they converge, tracking a measure of expert separation.

Activation steering controls LLM behavior at inference time by adding learned directions to hidden states, but existing methods handle one concept at a time. Pluralistic alignment, where different stakeholders need different value emphases, requires steering multiple dimensions simultaneously. We show that naive steering produces substantial spillover: the effect intended for one value leaks into others. This parallels the treatment-versus-spillover decomposition in causal inference. We trace spillover to geometric entanglement of steering directions, captured by their Gram matrix, and derive a zero-cost correction from an activation-norm-penalized objective that decouples each direction's contribution exactly. Our end-to-end pipeline requires no fine-tuning, no reward model, and no manual prompt engineering: given only domain questions, it automatically discovers value dimensions, extracts directions, diagnoses entanglement, and applies corrected steering. On climate discourse, the correction improves the net steering effect from +5.9% to +14.0%, validated over 100,000 pairwise judgments.

In streaming simultaneous speech translation, the speech translation system is trained to learn a read-write policy that alternates between consuming source words and generating target ones. In a cascaded setting, the output from the speech recognizer is passed to a separate machine translation component, making it more difficult to learn such a policy. Approximations such as fixed wait-k strategies or target-suffix deletion can be employed, but these approaches do not provide the model with a streaming system's flexibility to make contextual read-write decisions. This paper presents a training strategy for a cascaded machine translation system that enables it to dynamically decide how much of the growing source prefix to translate. We achieve this by fine-tuning a large language model (Qwen3-8B) on stable prefixes of the training data, which are produced by pairing every source sentence prefix in the training data with the longest translation of that prefix that is shared with the full source sentence translation. We fine-tune variants of the model on different subsets of the prefixes and compare against wait-k and target-suffix deletion. We also investigate the effect of fine-tuning the target-token generation confidence. Our experiments show that stable prefixes improve the quality-latency tradeoff when translating from English into German, Japanese, and Chinese across a range of test sets.

Safety monitors are evaluated by recall on harmful prompts, regardless of whether the target model would answer them. Yet a monitor matters most on the prompts the model does answer. We measure recall on exactly those prompts, defined by sampling the target model and judging its responses. Across four text guards, two activation probes, and Latent Guard, recall at a 1% false positive rate falls sharply on this subset: at a common threshold, every monitor catches the requests the model refuses 1.1 to 6.4 times as often as the requests it answers. Standard metrics hide this; AUROC stays above 0.85 for most monitors. Rewriting each request to be less explicit, with intent held fixed and verified, raises compliance 28-fold and lowers every monitor's flag rate. Of the requests newly answered after rewriting, 44 to 93% slip past the monitor, depending on which is used, and most of their completions are graded harmful. We trace the gap to explicitness itself. As wording softens with intent fixed, the target model's harm and refusal readings fall and it answers; the guards' harm readings fall too, and steering a guard along explicitness alone flips its verdict. Model and monitors miss the same prompts, and stacking monitors does not recover them. Fine-tuning a guard on the rewrites at every level of explicitness, on both sides of the label, raises recall on answered requests from .24 to .89 while transferring to unseen benchmarks; hard negatives, the natural alternative, teach the guard to discount indirect phrasing instead.

Principal component analysis (PCA) can rotate away from its population target when a covariance matrix is estimated from limited data. We introduce diagonal attenuation, which preserves sample cross-covariances while reducing coordinatewise sample variances. The method is revealed exactly by averaging a linear full-output reconstruction loss over random input masks; studying the correction directly extends it beyond the range attainable by masking. We isolate the part of the random coupling between retained and omitted population directions that is contributed by sample-variance errors, and show how attenuation can reduce the resulting rotation. Under balanced marginal variances, we derive an explicit expected-risk theorem, uniform over the attenuation path for all sufficiently large finite samples, and obtain the asymptotically risk-minimizing strength. For general covariances, we characterize when attenuation leaves the population PCA subspace unchanged and give a risk theorem that also accounts for changing eigengaps and the population cost when the target moves. Simulations track this tradeoff from exact preservation back to PCA. Across local image patches, speech spectra, and smartphone acceleration, both mask-derived and direct attenuation improve PCA under two fitting-sample budgets, and one of them has the largest mean gain among seven methods in every data–budget cell. The full path selects strengths beyond the mask-derived boundary on $63\%$–$95\%$ of the subsamples.

We present the first polynomial lower bounds for several fundamental problems in the distributed graph sketching model in the tiny-error regime, which includes deterministic algorithms as a special case. In the graph sketching model, every node sends a single message to the referee who does not have any prior knowledge of the graph and must output the answer. While the work of Nelson and Yu (SODA 2019) and Yu (SODA 2021) showed that $\Theta( \log^3n )$ is optimal for constructing a spanning forest or deciding whether the graph is connected with error at most $\frac{1}{\text{poly}(n)}$ , their approach does not yield any stronger bounds for significantly smaller error probabilities. Our main result is to show that solving either connectivity or spanning tree construction with error at most $\delta$ requires messages of length $\Omega( \min\{n, \log_2 \frac{1}{\delta}\}^{1/3} )$, which implies that algorithms with exponentially small error must send messages of $\Omega( n^{1/3} )$ bits in the worst case. Our results significantly narrows the current gap between the Jelani-Yu threshold of $\Theta( \log^3n )$ and the trivial upper bound of sending $O(n)$ bits per node for deterministic graph sketching. We also extend our results to $k$-edge connectivity. For any $k=O(n^{1/7})$, we recover the same bound of $\Omega( k )$ on the message length for algorithms with exponentially small error that was shown by Robinson and Tan (PODS 2026) only for deterministic algorithms. Finally, for $k=n^{o(1)}$, our result implies a stronger lower bound of $\Omega_\epsilon( n^{\epsilon} )$ bits, for any constant $\epsilon<\tfrac{1}{3}$.

Open-weight large language models face a low-cost white-box threat from representation engineering attacks. Attackers can estimate refusal directions and search for projection-matrix edits that suppress safety alignment while preserving general capabilities, within minutes on a single GPU and without gradient-based training. We propose Bait-and-Recover, a weight-level defense that places a bait adapter where attackers read activations and a paired recovery adapter at the subsequent layer. Trained via gradient routing, this decouples the observation path from the behavior path. By actively poisoning the residual signal used for measurement, Bait-and-Recover disrupts the attacker's edit search, while the recovery layer restores clean downstream computation. Across four open-weight models, our defense raises the minimum refusal rate against white-box edit searches from 16.25% to 71.75% under a strict behavior-preservation budget (KL <= 0.10), with negligible impact on general benchmarks. By invalidating the core measurement assumption of these attacks, observation-path poisoning offers a practical complement to behavior-level safety training.

Wireless digital twins can support site-specific analysis, planning, and experimentation for future wireless networks. However, constructing them at city scale remains challenging when accurate 3D city models are unavailable. This paper presents HalifaxDT, a wireless digital twin of the Halifax Peninsula in Nova Scotia, Canada, constructed from heterogeneous open geospatial and spectrum data. HalifaxDT combines building footprints, LiDAR elevation products, building metadata, and spectrum licensing records through a workflow that reconciles multiple sources of building-height information while preserving the provenance of geometry decisions. The resulting terrain, buildings, and gateway metadata are integrated into Sionna RT for wireless simulation. We evaluate HalifaxDT through two use cases. The first compares its coverage predictions with a reference derived from field measurements and with an analytical baseline. The second uses the digital twin to predict the received signal under normal operating conditions and detect interference. The coverage results show that HalifaxDT better preserves the spatial structure of the measured radio map than the analytical baseline. The interference study shows that deviations from the predicted reference can reveal interference that is difficult to detect from received power alone. We also identify current fidelity limitations, including simplified material representation, missing vegetation, and the need for broader RF calibration and synchronization with live measurements.

Distributed spectrum monitoring relies on spatially separated receivers to infer the state of the radio environment. A central task is to determine how many emitters are active and where they are located. Sensors typically report over shared unreliable links, where contention, fading, or energy limits erase whole reports, so forwarding raw I/Q samples from all receivers to a fusion node becomes impractical. This paper formulates distributed emitter counting and localization as a semantic communication problem over unreliable reporting links. Each receiver maps its local time-frequency observation to a compact latent representation and transmits it. The fusion node pairs each surviving latent with the corresponding receiver position and operates on the resulting unordered set, which keeps the fusion rule compatible with variable membership and receiver-count changes. The encoder and decoder are trained end-to-end with report erasure, and optionally finite-rate quantization, inside the task objective. Experiments on synthetic multi-emitter scenes show that channel-aware training improves robustness under erasure, that set fusion is more robust than fixed-order concatenation, and that one trained model operates across receiver counts without retraining. The rate study shows that a few bits per latent component suffice in the tested setting, and that the rate knee remains roughly stable across erasure levels. These results support the feasibility of low-rate semantic reporting for distributed spectrum monitoring over unreliable links.

Peer-review feedback often arrives too late for authors to make meaningful revisions. We study an author-facing LLM system that moves part of this stress test before submission: it generates a broad pool of atomic concerns and compresses them into a short report. We evaluate agreement with historical reviews and, separately, the possible validity of concerns they omit. From 10,000 ICLR 2026 submissions, we use 3,398 manuscripts with accessible versions that predate review. On a ten-paper diagnostic, independent sampling covers 44.9% of historical issues; deduplication and refill reaches 78.7% strict and 84.9% seriousness-weighted coverage, at 3.6$\times$ more requests and 5.2$\times$ more tokens. A hidden Top-32 Oracle preserves the full 79.3% weighted coverage of a 256-candidate pool, but paper-only selectors retain only 40–44%. LLM review therefore provides broad coverage with a large candidate pool but compresses poorly; ablations identify representative selection and matcher sensitivity as the main sources of this gap.

Implantable brain-computer interfaces (BCIs) promise transformative societal impact, from restoring lost motor, sensory, and speech function in patients with paralysis, stroke, sclerosis, and sensory deficits to serving as a high-bandwidth conduit between human cognition and machine intelligence. Realizing these visions requires moving from laboratory prototypes to chronic clinical systems that support thousands to millions of electrodes, several millimeters to centimeters deep beneath the brain surface, under strict heating and size limits. Almost every clinically relevant implant therefore operates in a weakly coupled regime, with coupling coefficients of $10^{-3}$ to $10^{-1}$ across centimetres of lossy tissue. In this regime, the wireless channel sets the limits of power-transfer efficiency, communication bandwidth, and energy per bit. We review the recent progress of bidirectional wireless links across inductive, mid-field, RF, ultrasonic, magnetoelectric, optical, UWB, and electro-quasistatic modalities, and benchmark them against the clinical axes of depth, size, and data rate. Within an approximately 10 mW communication budget set by the approximately $1^\circ$C tissue-heating ceiling, narrowband high-Q links are well suited for power transfer and low-speed data, but at 1-10 nJ/b cannot reach the greater than 10 Mbps to tens of Gbps uplinks that thousand- to million-channel interfaces demand, even with aggressive on-implant compression. This calls for sub-10 pJ/b and ultimately sub-1 pJ/b wireless links, where wideband techniques, such as ultra-wideband (UWB) and brain-channel communication (BCC), are suitable. We close with a quantitative framework for analyzing such wireless links and a co-design roadmap across electromagnetics, circuits, packaging, security, and regulation, toward secure, networked, million-channel brain interfaces.

In autonomous driving, Bird's-Eye View (BEV) representations provide a structured, top-down abstraction of the vehicle's surroundings and have become a key input modality for Behavioral Cloning (BC) policies. While ground-truth BEV maps are readily available in simulation, real-world deployment requires replacing them with camera-predicted counterparts - a substitution that introduces perceptual errors whose downstream impact on closed-loop driving performance is not well understood. In this work, we investigate the use of Cross-View Transformer (CVT)-predicted BEV maps as direct policy inputs for a BC agent in the CARLA simulator. We propose a six-channel BEV representation covering road surface, planned route, lane boundaries, vehicles, pedestrians, and traffic lights, and introduce a Kernel Density Estimation (KDE) weighting scheme that rebalances the segmentation loss towards underrepresented driving maneuvers such as curves and intersections. Closed-loop evaluation across two CARLA towns shows that the KDE-weighted model is the only predicted-BEV agent to complete a full episode without infractions, despite not achieving the highest aggregate IoU. This discrepancy reveals that global segmentation metrics are poor proxies for driving performance: what determines navigation success is prediction quality at geometrically critical locations, and the route channel emerges as the primary bottleneck for reliable agent navigation under predicted BEV inputs.

Vision-language models (VLMs) offer a promising alternative to conventional fire detection systems by reasoning about the semantic context of a scene and thus reducing false alarms, yet their large model size makes deployment on embedded fire sensors impractical. In this paper, we study how domain-specialized VLMs can be compressed for fully on-device deployment without losing the safety-critical behavior required for fire detection. We develop a teacher-student knowledge distillation framework in which large VLMs fine-tuned for fire understanding can be distilled into lightweight students. Experiments across multiple VLM families and model scales show that compact students preserve most of their teachers' fire-understanding capability. We further deploy the distilled models on our commercial Detectium fire detection sensor and jointly evaluate reasoning accuracy, latency, and memory usage. The results show that compression and deployment affect not only accuracy but also model failure modes, with Qwen2.5-0.5B providing the strongest overall deployment trade-off. Our findings provide broader guidance for deploying domain-specialized VLMs in resource-constrained, safety-critical settings.

Is it possible to walk to infinity, avoiding three collinear points, using as steps only the standard unit basis vectors $(0,0,\ldots, 0, 1, 0, \ldots, 0)$? We prove this is possible in $16$ dimensions.

Large language models used for code editing can be trained and deployed in at least two output regimes: direct generation, where the model emits the entire modified file in one shot, and iterative diff-based generation ("steps"), where the model emits a sequence of localized search/replace edits applied one at a time until it signals completion or a step budget is exhausted. The diff-based regime is attractive because it mirrors how developers edit code and should require far fewer generated tokens per turn. We train two code models - a 100M-parameter model trained from scratch (Rainbow-Pony-100M) and a fine-tuned Qwen2.5-Coder-0.5B - in both regimes on a shared Flutter/Dart dataset, and evaluate all four resulting models on a held-out set of approx 1,790 tasks per model. Direct generation substantially outperforms diff-based generation on every metric we measure - compilation/static-analysis pass rate, bits-per-byte, character-level similarity to the reference, and blinded LLM-judge ratings of goal fulfillment, correctness, and code quality - and the gap persists after controlling for task difficulty via a matched-ID comparison and when restricting to code that compiles on both sides. We then identify a single, architecture-independent mechanism behind the conditions where diff-based generation does win: it is competitive on short, spatially localized edits, and its category-level wins concentrate in exactly the two task categories - refactoring and error-handling/edge-case fixes - with the lowest mean edit-step count in our dataset. We term this task locality and discuss its implications for when an edit-based training regime is and is not the right choice for a code-editing model.

Elliptic homogenization is used to determine coarse-grained properties of materials with features on small scales. When these small scale features have rapid, periodic fluctuations, the solution field corresponding to a homogenized constitutive relation closely resembles the true solution based on the heterogeneous material. This homogenized behavior of the material is computed from a cell problem, where a cell is defined to be one period of the fluctuating material. In the context of linear elliptic partial differential equations, the homogenized constitutive relation is defined simply by a constant coefficient tensor, but for nonlinear problems, the homogenized response depends on the macroscopic state and/or its gradient, thus requiring solutions to parametric cell problems. When computing a numerical solution with the homogenized constitutive relation, it is useful to have a differentiable representation of the solution to the cell problem, as derivatives of the homogenized constitutive relation are required in Newton iterations for the macroscopic state field. In this work, we use the Deep Ritz method to solve the parametric cell problems that arise from nonlinear homogenization. First, we exploit the variational structure of the cell problem, then we discretize the dependence of the cell response on both space and the macroscopic state with a neural network. Enforcing boundary conditions on the cell response strongly, we next use the parametric Deep Ritz method to simultaneously solve the cell problem over a range of macroscopic states. We show that this method is accurate, efficient, and offers a continuous and differentiable representation of the cell response over the macroscopic state and gradient. We then show that our parametric representation of the cell response significantly expedites macroscale solutions when compared to a traditional $\text{FE}^2$ scheme.

Large language models increasingly provide formative writing feedback, but most systems evaluate only the submitted text. Process-aware systems can additionally incorporate intermediate drafts and interaction records, yet access to these traces does not establish whether learners perceive the resulting feedback as useful, fair, or aligned with their own revision experience. We conducted a convergent mixed-method post-task evaluation with 43 undergraduate English-language learners. Participants completed a timed essay, received LLM-generated feedback based on the final essay, intermediate snapshots, and keystroke-derived records, and then answered six five-point Likert items and four open-ended questions. Ratings were concentrated above the neutral midpoint across all six items. Between 74.4% and 95.3% of participants selected ratings of 4 or 5, and all six one-sided exact sign tests remained significant after Holm correction. Perceived correspondence between the revision analysis and the participant's reported thought process was positively associated with overall perceived usefulness, Spearman's $\rho_s=.327$, 95% bootstrap CI $[.009,.609]$, $p=.032$, $n=43$. Qualitative responses were analysed through an LLM-assisted codebook development procedure followed by independent human coding of 14 binary codes. Participants frequently described the product feedback and revision rationale as accurate, while a smaller set reported misinterpretation of surface corrections, time constraints, or rhetorical intention. The findings distinguish perceived detection of revision activity from perceived interpretation of why revision occurred.

Ship structures govern vessel strength, safety, and manufacturability, but their design must satisfy hundreds of classification society requirements, making the process complex and iterative. Data-driven approaches are limited by the lack of structured datasets linking design geometry, structural performance, and rule-based constraints. This paper presents MiDShip, a multimodal dataset of 12,753 synthetic cargo-hold structural designs: 6,020 random, 496 generated by an SGLD-inspired procedure, and 6,237 generated by an equation-informed repair procedure. Each design includes parametric data, full and mesh-ready 3D geometry, engineering drawings and annotations, a bill of materials, and preliminary structural evaluations. Twenty-five constraints derived from a subset of ABS MVR are also evaluated. None of the random designs satisfies all constraints. Among the SGLD-inspired designs, 322 (64.9%) were fully compliant, with an average of 0.409 violations, 82.7% below the seed mean and 96.9% below the random-design mean. The repair procedure, developed through LLM-assisted code analysis, produced 4,952 fully compliant designs (79.4%), averaging 0.296 violations, 97.1% below the paired-source mean. In equal-size comparisons, mean nearest-neighbor distances in the scaled 120-parameter space were 3.495 for repaired designs, 1.144 for SGLD batches, and 3.729 for random designs. The primary contribution is the synchronized dataset and its generation and evaluation infrastructure; the generation studies demonstrate its utility rather than proposing new optimization algorithms. MiDShip supports machine learning, generative design, and automated rule-based evaluation for ship structures.

K-Dense BYOK (bring your own keys) is a free, open-source AI research assistant for scientists in any field that runs on the researcher's own computer. The researcher supplies access to a model of their choice, hosted or running locally, and the application supplies everything else: a place for the work to run, a layer of scientific scaffolding, and a complete record. Each project is an ordinary folder, so the data, the code, the results, and the record stay on a machine the researcher administers and can be read years later without the application. Three things separate it from a chat assistant or a general-purpose coding agent. It ships a library of written scientific procedures, guided workflow templates, catalogs of where research data can be found, and reviewer and writer roles the agent can hand work to. It keeps a Living Lab Notebook whose entries link into an argument and are added to but never erased. And it records what happened by watching what the agent does rather than by taking the agent's word for it, in a log the agent has no tool that can write to. That choice targets the most common failure, model overclaiming, in our earlier benchmark of nine frontier models, by making claims checkable rather than preventing them. On twenty interdisciplinary research prompts, scored under a rubric fixed in advance, K-Dense BYOK led two managed platforms on both scientific quality and research execution. Its deliverables were the only ones that recorded the software they ran in, and the only ones that usually arrived with a command that regenerates the results. One of the managed platforms ran the same frontier model and supplied neither. Those environment records were files the agent wrote, not part of the observed log, which does not yet capture the software environment itself. The code is available under the MIT license at https://github.com/K-Dense-AI/k-dense-byok.

The growing complexity in home energy management (HEM) demands advanced systems that guide occupants toward informed energy decisions reflecting their background, preferences, and context. Large language model (LLM)-integrated HEM systems (HEMS) have demonstrated promise, but previous studies relied on single-turn or single-task evaluations with response accuracy as the primary metric. Whether such systems deliver effective interactions across the extended multi-turn dialogues typical of real-world use remains an open question. This study introduces a comprehensive evaluation framework of LLM-integrated HEMS derived from the Goal-Question-Metric methodology, organized across five categories: task performance, factual accuracy, interaction quality, control capability, and system efficiency. A total of 23 metrics across multi-turn conversations are proposed and an LLM-as-judge pipeline is employed to enable scalable automated scoring. Its reliability is validated against three trained human coders: after iterative rubric calibration, twelve of the fifteen LLM-scored metrics reached strong agreement (ICC >= 0.73), three of them perfect, while the remaining three exhibited near-zero variance in human scores and are instead reported via mean absolute error (0.04-0.28). To demonstrate the framework's effectiveness, 970 dialogues – 16 scenarios and five personas – were generated and evaluated across four conversational HEMS configurations spanning a sophistication gradient, from a vanilla LLM with raw energy data to a multi-agent HEMS. The framework distinguished the four configurations across multiple evaluation dimensions, revealing their respective strengths and weaknesses. This study contributes to conversational HEMS by providing a reproducible, multi-dimensional evaluation methodology that comprehensively assesses sustained, context-aware system performance.