Papers

Static-analysis checker synthesis requires agents to interpret a defect specification, inspect a repository, implement analyzer-specific logic, and refine the checker through repeated compilation and analysis feedback. Existing coding-agent benchmarks focus on tasks such as patch generation or vulnerability detection and rarely assess whether an agent can develop a working checker in a repository from start to finish. We introduce CheckerBench, an executable benchmark of 300 tasks derived from 297 CVEs across 167 repositories, 85 CWEs, and five language ecosystems. Each task includes vulnerable and fixed revisions, a pinned analysis environment, and a checker scaffold. We further introduce CheckerLab, a common evaluation framework that independently rebuilds submitted checkers and measures vulnerable-fixed diagnostic contrast, patch localization, false positives, and tool use. Across 21 model-harness configurations and three independent repeats per configuration, mean Pass@1 is 32.30%, while the best reaches 45.33%. These results show that reliable, reusable checker development remains challenging for current coding agents.

Learned orchestration can automatically construct effective language-model multi-agent systems, but existing approaches couple planning to fixed worker pools and train decomposition and collaboration from the same terminal outcome, limiting transfer and obscuring credit assignment. We introduce DeOrch, which separates worker-agnostic planning from concrete worker selection. Its two-stage planner first decomposes the task without worker information, then chooses collaboration operations using compact, worker-identity-free matchability feedback from the pool, enabling conditional credit assignment to decomposition and collaboration decisions. A lightweight matcher estimates worker suitability from behavior on a fixed probe set and adapts online with a contextual bandit, allowing new workers to be incorporated without retraining the planner or matcher. Across diverse in- and out-of-distribution tasks, DeOrch outperforms prior automatic MAS orchestration methods with fewer worker calls than competing learned orchestrators, remains effective when transferred to an entirely unseen worker pool without retraining, and shows consistent gains from both components.

Optimal-transport couplings have been shown to reduce training variance in unconditional flow models, but their role in conditional generation remains unclear. A natural approach constructs separate couplings for each condition, but this is impractical for large or continuous conditioning spaces found in modern image foundation models. We introduce Global Transport (GT), a global class-agnostic optimal-transport coupling, computed without class labels. GT can associate different conditions with different regions of the source noise, and consequently worsens performance without guidance. However, when combined with classifier-free guidance (CFG), GT consistently improves generation across domains, model scales, and sampling budgets. This reversal suggests that couplings for conditional flows should be evaluated both empirically and theoretically under the guided flow used at inference, rather than on unguided generation. We evaluate GT over both discrete class and continuous text conditioned image generation across model scales, and investigate how coupling choice alters guided trajectories. These results identify coupling design in the guided flow setting as a simple training time axis to improve performance without modifying existing architectures, samplers, or guidance mechanisms.

We develop and analyze a B-spline scaling-function Galerkin method for the tempered fractional partial differential equation governing European and American option prices under the CGMY process. The spatial generator couples exponentially tempered left- and right-sided Riemann–Liouville derivatives of order $Y\in(0,2)$, whose nonlocality and asymmetry present discretization challenges. We derive a corrected Fourier-symbol construction of the tempered connection-coefficient matrices whose interior blocks are Toeplitz, maintain exact fractional scaling, and enable $\mathcal{O}(N\log N)$ matrix–vector products via the fast Fourier transform. In the regime $Y\in(1,2)$, we prove full-line and finite-domain Gårding inequalities, coercive whenever the risk-free rate is positive, via zero-extension to the full-line symbol. On this foundation, we establish elliptic-projection, semidiscrete, and fully discrete error estimates with spatial rate $\mathcal{O}(h^{p-Y/2})$, alongside an unconditional shifted-energy stability estimate for the forced Crank–Nicolson scheme with mesh-independent constants. For positive rates the admissible shift vanishes, governing the plain scheme exactly as implemented. A single backward solve prices an entire strike strip at fixed maturity with smooth, analytically differentiated Greeks. The same discretization prices American puts through a projected linear complementarity solve, with early-exercise premiums and exercise boundaries validated against a resolved deterministic reference and a Longstaff–Schwartz Monte Carlo benchmark. Calibration to S&P~500 option surfaces across low-volatility, crisis, and recovery regimes yields implied-volatility errors near two vol-points on well-identified surfaces, while an accuracy-matched comparison with a stabilized Grünwald–Letnikov baseline delineates precisely when the high-order variational structure pays off.

LoRA fine-tuning adapts small language models (SLMs) to heterogeneous instruction data within a low-rank update subspace, making it vulnerable to three structural problems: conflicting gradients that cancel, static data selection that cannot track evolving learning dynamics, and subspace saturation that causes later updates to overwrite useful directions. We argue that effective adaptation therefore requires controlling which data-induced gradients enter the LoRA subspace and when. We propose GRADE (GRadient-Aligned Data-centric rEcipe), a data-centric framework combining two mechanisms: a state-aware selector that continually admits samples aligned with the evolving multi-task gradient field, and a self-calibrating step-level gate that rejects updates likely to cause destructive overwrite near saturation. Across three current-generation backbones and a heterogeneous seven-dataset instruction pool, GRADE outperforms strong data-selection and PEFT-stabilization baselines in accuracy and robustness. It is the only method to improve consistently over standard LoRA on every architecture, while producing more coherent gradient trajectories and less destructive overwrite. These results show that successful SLM adaptation depends not only on which data are selected, but also on which gradients are allowed to enter and persist in the constrained update subspace.

The rapid growth of AI/ML methods for materials discovery has increased the importance of the quality, physical completeness, and information content of computational databases. We examine these issues for magnetic materials, where incomplete sampling of competing magnetic states, chemical disorder, and finite-temperature properties can lead to apparently accurate but physically unreliable predictions. We propose a hierarchical discovery workflow combining independent ML models with explicit electronic-structure, magnetic, dynamical, thermodynamic, kinetic, and experimental validation. To incorporate kinetic accessibility into high-throughput screening, we introduce a dimensionless synthesis-window descriptor based on the relation between the chemical order-disorder and Tammann temperatures. A cluster-expansion analysis of Fe-Co-B illustrates the importance of this criterion: although ordered Fe3CoB2 is energetically favored at T=0, its small ordering energy produces an order-disorder temperature far below room temperature, making long-range Fe/Co ordering kinetically unachievable. These results demonstrate that reliable AI/ML discovery of magnetic materials requires databases and validation procedures that contain physical information relevant to the target functionality, rather than simply larger numbers of calculated structures.

Learning Gaussian mixture models (GMMs) using the Expectation-Maximization (EM) algorithm and its gradient-based variants is a fundamental problem in machine learning. It is known that randomly initialized (gradient) EM fails to learn multi-component GMMs in the exact-parameterized setting, where the number of components matches that of the ground-truth GMM. Recently, global convergence of gradient EM has been established in the over-parameterized setting, where more components are used, provided that the ground-truth components are well separated. In particular, the minimum separation between ground-truth components is required to scale as $\Omega(\sqrt{d})$, where $d$ is the dimension. In this paper, we show that this dimensional dependence is unavoidable in high-dimensional settings. Specifically, we consider a hybrid EM algorithm that uses standard EM updates for the mixing weights and gradient EM updates for the component means. For any $\epsilon > 0$, we prove that when the dimension is sufficiently large, in the worst case a separation of order $\Omega(d^{0.5-\epsilon})$ is insufficient to guarantee global convergence of population gradient EM in sub-exponential time under random initialization, even in the over-parameterized regime. Our result establishes an almost optimal worst-case lower bound on the ground-truth separation required for learning Gaussian mixtures via gradient EM in high dimensions.

Random access (RA) is one of the most foundational medium access control (MAC) layer scheduling schemes for handling unpredictable data traffic from multiple terminals. While multi-agent reinforcement learning (MARL) has been explored to optimize RA-based wireless networks, its reliance on experience-driven, distributed policy learning incurs significant training overhead for each optimization task, limiting its feasibility in real-world applications. In this work, we propose to leverage a foundation model (FM) to improve MARL efficiency across diverse RA network optimization tasks. Specifically, we design an FM-aided actor-critic algorithm within a consensus-based decentralized MARL architecture and provide its convergence analysis under local reward exchanges and nonlinear value function approximations to show that our algorithm achieves the same convergence order as the conventional MARL with critic model exchanges and linear approximations. Our numerical results show that our FM-based approach significantly enhances MARL speed for RA network optimization.

The basic hypergeometric series ${}_{n+1}\phi_{n}$ whose upper and lower parameters are non-negative integer powers of $q$ satisfying a certain constraint can be, after a normalisation, computed as a positive polynomial. We derive three formulas for this polynomial. The first formula is a generalisation of the MacMahon identity in which an infinite binomial series is identified with a normalised generating function of des and maj statistics of multiset permutations. The second formula is a generating function of multiset permutations for the inversion statistic and a generalisation of the statistic mstc. The third formula is a finite sum containing $2n$ $q$-binomial coefficients.

We construct the bulk local states of a massive spin-1 field in AdS$_3$/CFT$_2$ and three-dimensional Carrollian holography. In AdS, covariance under the isometries that leave a bulk point invariant determines the three components of the vector there and splits the solution into two branches of definite helicity, whose inner products reproduce the massive vector two-point function of AdS$_3$. By expanding the massive vector HKLL representation in boundary descendants, one can reproduce the three component states term by term. The two branches together describe the parity-invariant Proca field, and one alone the single propagating mode of Maxwell–Chern–Simons theory. The same conditions are solved in the induced representation of the Carrollian theory, where the solution again splits into two branches. Under the flat limit of the generators the highest weight module becomes the induced module, and the AdS states and Green's functions become their Carrollian counterparts, which supports the induced representation as the home of massive bulk excitations in Carrollian holography. This is the first bulk local state built with spin, complete from the state itself to its bulk-to-bulk propagator and to the limit relating the two descriptions.

Atomic oxygen (AO) erodes polymer surfaces in low Earth orbit, but durability data are limited. We evaluated a regression ensemble trained on 37 MISSE-2 polymers using eight selected bulk properties and five elemental-composition ratios. Its archived leave-one-out R^2(log) was 0.768; this estimate is conditional on feature selection using the complete training table. For nine post-MISSE-2 polymers, the ensemble gave R^2(log) = -0.037 and all predictions were within a factor of two. The training arithmetic-mean baseline also met that tolerance for all nine, highlighting the limited discrimination provided by their narrow target range. Ground-based comparisons included four PLAVIS grades. For an out-of-training mystery polymer with undisclosed composition, the predicted erosion yield was within a factor of 1.92 of the ground-based measurement. Conflicting PLAVIS measurement records prevent a definitive residue-erosion ordering or ground-validation score. Thermal residue, SEM contrast and surface chemistry provide distinct observational evidence; they do not establish a protective-layer mechanism. The results support further evaluation of orbital-data screening models with reconciled measurement records and prospectively fixed inputs..

An $(s,p)$-edge-coloring of a graph $G$ is an edge coloring using $s$ colors such that exactly $p$ colors appear at each vertex. To generalize the notion of proper edge-coloring, these colorings are defined to be equitable: the numbers of edges of each color incident to a vertex are fairly distributed. We find the necessary and sufficient conditions for embedding an equitable $(s_1,p_1)$-edge-coloring of $K_{n_1}$ into an equitable $(s_2,p_2)$-edge-coloring of $K_{n_2}$. We focus on the values of $n_1$, $p_1$, $n_2$, and $p_2$ where $s_1$ is necessarily larger than $p_1$ and $s_2$ is necessarily larger than $p_2$.

We present a fully offline speech-to-speech translation pipeline that runs on a Jetson Nano (4 GB) and corrects its own weak translations without retraining. A Whisper-tiny ASR feeds an Opus-MT translator; multilingual BERT cosine similarity acts as a Quality Estimation (QE) gate, triggering a secondary-pass when confidence falls below a pre-defined threshold $\tau$. We compare three correction methods: QE reranking (M1), Minimum Bayes-Risk decoding (M2), and constrained beam search (M3). On 1,012 FLORES-200 sentences (English-Spanish), M2 at $\tau=0.90$ produces statistically significant improvements over greedy decoding on BLEU (+0.67, p<0.001), ChrF (+0.51, p<0.001), and COMET (+0.0020 at N=3, p=0.002); M1 yields no significant gains, and M3 is significantly worse than baseline (p>0.99). Our central finding is that QE functions effectively as a gate but poorly as a ranker: removing the QE model from candidate selection (M1$\to$M2) does not hurt quality and frees 680 MB from the critical path. Using a gain-to-edit ratio adapted from the post-editing-effort literature, we further show that smaller candidate pools (N=3) yield more surgical corrections with better semantic adequacy, while larger pools (N=10) maximise lexical reward. We release the system and demonstrate live translation across six language pairs.

Large language models (LLMs) are increasingly used to judge how well an advertisement matches a query, but the fairness of these judgments has received limited attention. We conduct a systematic study of fairness in relevance judgments made by LLMs for queries and advertisements. Our counterfactual framework examines the effects of advertiser identity and possible popularity, input language, and demographic wording. We study GPT-4o as a categorical relevance judge and a Qwen-7B model trained specifically for relevance prediction. The advertiser and language experiments use query and advertisement pairs sampled from real advertising logs. Controlled synthetic queries are used to study demographic associations in employment, housing, and credit. For both models, changing the advertiser identity or input language can alter the relevance assessment. Selected demographic comparisons also show patterns consistent with common stereotypes, particularly those involving gender and occupation. We further study mitigation during model inference and training. The results indicate that its effectiveness depends on whether advertiser information is relevant to the query and how advertiser labels are distributed in the training data. These findings can help advertising practitioners identify fairness risks and develop suitable mitigation methods for LLM relevance systems.

This article is the third of a series of three articles whose goal is to generalize the Cartier transform of Ogus and Vologodsky to the logarithmic setting. In the second one, we generalized a topos-theoretic version of this transform, due to Oyama. Let $W$ be the ring of Witt vectors of a perfect field of positive characteristic $p$ and equip $\operatorname{Spf}W$ with the trivial log structure. Let $\mathfrak{S}$ be an fs log $p$-adic formal scheme, locally of finite type and log flat over $\operatorname{Spf}W$ and denote its special fiber by $S.$ For a log smooth morphism of logarithmic schemes $X \rightarrow S,$ we constructed crystalline-like ringed topoi $\mathcal{E}'$ and $\underline{\mathcal{E}}$ and subcategories of crystals of quasi-coherent modules $\mathcal{C}'$ and $\underline{\mathcal{C}},$ equivalent respectively, under some lifting assumption, to modules with Higgs fields and integrable connections, both satisfying certain nilpotence conditions, and a morphism of topoi $\underline{\mathcal{E}} \rightarrow \mathcal{E}'.$ We proved that the pullback functor of this morphism of topoi preserves quasi-coherent crystals and induces a fully faithful functor $\mathcal{C}' \rightarrow \underline{\mathcal{C}}.$ Since the Frobenius morphism is not, in general, flat in the log smooth setting, it is not clear that this functor is essentially surjective. To address this issue, we refine the topoi and crystals mentioned above by endowing them with an indexed structure, inspired by Lorenzon's indexed extension of Cartier descent to smooth logarithmic schemes. Using the Azumaya property of the indexed algebra of logarithmic differential operators, we then obtain an equivalence between the corresponding categories of indexed crystals, thereby generalizing the Cartier transform.

The production of adipic acid generates nitrous oxide as a by-product, which, if emitted, very effectively warms the climate and depletes the stratospheric ozone layer. There exists cost-effective technology to mitigate these nitrous oxide emissions from adipic acid production, but the degree of implementation of this technology, most notably at sites in China, is uncertain. Using satellite observations from summer 2026, we quantify nitrous oxide emissions from three adipic acid production sites in China, finding emissions that total over half a million tonnes per year, a rate that implies limited emissions abatement. With respect to climate change, eliminating the annual emissions from just these three sites would be equivalent to eliminating the annual emissions from 49 coal-fired power plants or 43 million gas-powered vehicles in the United States.

Warp primitives such as tensor core operations, shuffles, reductions, and barriers are critical to high-performance GPU kernels, and every major GPU language supports some set of them. The threads that participate in a primitive, and therefore synchronize, are determined dynamically by how threads diverge and reconverge, and by intra-warp scheduling such as independent thread scheduling. In practice, many high-performance kernels use these primitives and behave as expected, following an intuitive but unwritten data-race-freedom contract that has never been stated precisely or empirically tested. We present WarpDRF, the first abstract warp programming model to make this contract precise, with participation rules parameterized by reconvergence guarantees and per-primitive requirements so that instantiations match different GPU languages. We prove (formalized in Rocq) that a kernel satisfying WarpDRF executes every warp primitive with the participants the reference semantics assigns and produces the same results, so a programmer can reason in the reference semantics alone. We implement the model in MLIR with a reference interpreter and fuzz 10K conformance tests for each of three configurations across CUDA, HIP, HLSL, Metal, and SPIR-V on 16 device and backend pairs; at least one WarpDRF configuration describes each, with CUDA honoring the strictly weakest configuration. Finally, we extend Faial, a static data-race analyzer for CUDA, into the first DRF checker for an empirically validated warp model, and apply it to llama.cpp, where most kernels that use warp primitives already satisfy the contract, but three contain previously unknown data races, showing the need for tools that check it.

Robotic systems often exhibit unstable modes, along which small perturbations and disturbances can cause unbounded growth unless corrected through feedback. Controlling such systems from high-dimensional visual observations requires representations that preserve these modes. Joint-embedding predictive architectures (JEPAs) provide a natural framework for learning such representations and their dynamics from visual data. However, we demonstrate that next step prediction combined with anti-collapse regularization does not guarantee that controllable unstable modes are preserved: the training loss can be minimized while these modes are collapsed, making stabilization from the learned representation impossible. To address this, we augment world-model training with an action reconstruction objective (i.e., an inverse dynamics loss) that encourages control-aware representations, namely, visual representations that preserve crucial features for control. We prove that exact action reconstruction makes the encoder injective on the finite-horizon reachable subspace. Thus, the encoder cannot discard any state direction reachable by an action sequence within $H$ steps. Moreover, we show that, as $H$ grows, the dominant eigenspace of the finite-horizon controllability Gramian converges to the controllable unstable subspace. We establish our theoretical results for linear systems and demonstrate empirically that our findings extend to nonlinear visual control tasks (CartPole, Walker2D, and PointMaze), highlighting the benefits of control-aware representation learning.

Consider a finite planar graph with positive real edge weights and designated boundary vertices, called nodes. Such a graph is called a circular planar electrical network. A grove is a spanning forest in which every component contains at least one node. The connected components of a grove determine a partition of the nodes. We relate weighted grove counts to invariant theory for the symplectic group. To a planar electrical network $G$, we associate an $\mathrm{Sp}(2n)$-invariant tensor $Z_G$. For $\mathrm{Sp}(2)=\mathrm{SL}(2)$, we expand $Z_G$ in the Temperley–Lieb basis indexed by noncrossing matchings and relate its coefficients to the Kenyon–Wilson grove formulas. For $\mathrm{Sp}(4)$, we give reduction rules for superpositions of two groves and prove that the tensors indexed by $3$-noncrossing matchings form a basis of the space of $\mathrm{Sp}(4)$-invariant tensors. The coefficients of $Z_G$ in this basis are weighted counts of reduced double groves, up to normalization.

Consider an arithmetic progression inside the finite field $\mathbb{F}_{q^n}$. We study the existence of progressions whose terms are all primitive and at least one of them is normal. Using a combination of sieve estimates, ramified bounds, and computational arguments, we improve significantly upon previous literature for progressions of length $m=4$ and obtain partial results whenever $5\leq m\leq 10$. Finally, we provide a general bound which outperforms previous results in the range $m\geq 53$.

We answer a long-standing open question of Woodin by proving that if $\kappa$ is a strongly compact cardinal and $GCH$ holds below $\kappa$, then $GCH$ holds.

Artificial viscosity can be used to deal with shocks and other discontinuities such as star surfaces, that can arise in matter. We explain in detail how a local discontinuous Galerkin method can be used to add artificial viscosity terms to hyperbolic evolution equations. These artificial terms are multiplied by a viscosity that is non-zero only in elements that are flagged as non-smooth by a newly developed smoothness indicator. Our artificial viscosity method works well in several simple shock test cases such as with Burgers' equation and blast wave tests. For the first time, we are also able to evolve single neutron star in full general relativistic hydrodynamics. To achieve this, we use different viscosities for the momentum density than for the rest of the matter variables.

Multi-agent simulations with large language models (LLMs) often operate networks of agents with a single base model. This overlooks the inter-model effects which may dominate engagement dynamics in real-world deployments. To show this, we simulate a heterogeneous social network powered by several different base models and show that the amount of engagement an agent receives depends more on its base model than on its assigned persona. The attraction or repulsion effects of a base model strengthen dramatically when more models are added in the mix, suggesting that networks dynamics may converge to base model effects at scale. To help explain this effect, we conduct a series of content-mediating analyses, showing the predictability of base models across contexts as well as the relationship between a model's lexical patterns and an engagement-maximizing style. In light of recent developments in mass multi-agent interaction, this work underscores the relevance of heterogeneous compositions in driving the outcomes of those networks

One practitioner we interviewed said their team writes "must" instead of "should" when instructing a coding agent, because the agent may treat "should" as optional. Small wording choices matter because agents often fill gaps in their instructions with their own assumptions. Spec-driven development (SDD) asks developers to write a specification, plan, and tasks before the agent writes code. SDD frameworks provide templates for these artifacts, but the templates do not help developers judge whether they have written enough or clearly enough. We studied what good SDD specifications contain. We scored the artifacts of 114 open-source SDD repositories against ISO/IEC/IEEE 29148 and derived practices from the highest-scoring ones. The resulting framework, EPIC, has 40 practices in 10 quality dimensions that guide developers in making expectations and decisions explicit in specifications, plans, and tasks for coding agents. The majority of SDD practitioners (N=15) endorsed every practice. Repositories in the top third of specification quality spent 11.8% of their commits on bug fixes, compared with 20.4% in the bottom third, and had a median of 4x as many contributors. Developers can use EPIC to find and fill the gaps in a specification before the agent does.

Audio language models must handle dozens of distinct skills, from pitch comparison and speaker counting to musical tempo estimation and emotion recognition. Joint training on all skills at once causes interference: gains on one skill often come at the cost of another. We propose SkillFormer, which decomposes audio understanding into skill-specific low-rank adapters and composes them at inference time through a learned router. The router examines the question to decide which adapters to activate and how much weight each should carry, so that a pitch query engages different parameters than a genre classification query. An alternating training schedule updates each adapter on its own skill cluster before jointly calibrating the router, preventing the gradient conflicts that arise in standard multi-task optimization. SkillFormer adds fewer than 4% of the base model's parameters and requires no changes to the audio encoder or language backbone. Evaluated on three architecturally distinct models across MMSU, MMAU-Pro, and MMAR, it raises the average accuracy by 2.5 to 4.1 points, with balanced gains across perception, reasoning, and semantic subcategories.

A single camera cannot distinguish between a slowly moving object nearby and a rapidly moving object in the distance without additional information. For a shuttlecock, air drag provides a strong deceleration cue that can be combined with known court geometry. We fit a gravity-and-drag trajectory to smartphone video and compare a free monocular fit with a fit constrained to cross the net plane. Camera calibration uses court lines, and the trajectory is fitted to observed shuttle positions and frame timestamps by robust nonlinear least squares. We evaluated 51 badminton shots from two recording sessions, spanning drop shots, drives, half-smashes and smashes. The net-plane fit reduced the median absolute speed difference from a drag-based stereo reference from 13.3% to 8.6%. A second reference, obtained by differentiating triangulated positions without a drag model, also gave a smaller median difference for the net-plane fit. The results support combining court geometry with flight dynamics for badminton speed estimation from a single camera.

Model-theoretic semantics comes with a computational counterpart: to check validity is to search for a countermodel, and tableaux, resolution, and satisfiability solvers make the semantics something one can run. Proof-theoretic semantics – the younger paradigm in which meaning is given by inferential role rather than by denotation – has no established counterpart. This paper proposes one: logic programming. We show that Sandqvist's base-extension semantics for intuitionistic propositional logic is executable in a precise sense. Bases are logic programs, and to each formula $\varphi$ we assign a goal $[\![\varphi]\!]$ of a second-order hereditary Harrop language such that $\varphi$ is supported in a base $\mathcal{B}$ exactly when $[\![\varphi]\!]$ is operationally derivable from $\mathcal{B}$ read as a program. The encoding is the Russell-Prawitz translation of the connectives with quantification restricted to atoms, and this restriction is not a limitation but the point: it is what the support clauses themselves demand. The correspondence is exact base by base, not merely at the level of validity, so that intuitionistic validity becomes the success of a single query against the empty program. Logic programming stands to proof-theoretic semantics as tableaux and satisfiability solvers stand to model-theoretic semantics.

Fine-grained action recognition often depends on how an action unfolds over time rather than on appearance alone. Visually similar classes may differ in preparation, execution, or completion phases, making uniformly aggregated temporal context suboptimal. We study this problem under a practical objective: i) improving recognition performance with few tunable parameters, ii) being compute efficient, and iii) using only the readily available video modality. Motivated by this missing temporal observation and the accuracy–compute efficiency–unimodality objective, we propose Directional Adapter (DiA), a parameter-efficient adapter on top of CLIP. DiA formulates temporal specialization into causal and anti-causal directions, allowing past-to-present and future-to-present cues to be modeled differently. To keep this directional modeling parameter- and compute-efficient, both directions use depth-wise temporal convolutions in a compact bottleneck space and are combined through a learnable fusion. DiA achieves state-of-the-art performance among methods using a single video input modality across seven public benchmarks, with only $\sim$3.6M tunable parameters, while also delivering higher throughput and lower inference latency than prior methods. Code is available at https://github.com/prakashchhipa/DiA-Directional-Adapter.

Manual annotation remains the principal bottleneck in surgical video understanding. Live intraoperative narration by the operating surgeon is a scalable alternative that requires no manual annotation, but the information it contains has not been characterized. We analyzed 13 endoscopic lumbar spine decompression procedures (13.2 h) by two senior spine surgeons, using time-synchronized close-mic German audio. An annotation-free pipeline (word-level speech recognition, deterministic correction of surgical-term errors, utterance fusion, LLM-based translation to English, and relevance filtering) converted speech into time-stamped procedural cues labelled with the surgical entities they reference and their speech act. Of 3,171 utterances, 1,201 (37.9%) were retained as procedural cues, while 57.6% were non-procedural and removed. Retained cues referenced instruments (60.7%), anatomy (48.0%) and surgical actions (57.1%), and 98.5% carried at least one of these. 39% of cues (472 of 1,201) were directives, mainly instrument requests and commands. All 467 manually verified instrument requests preceded the instrument's appearance in the field, by a median of 8.2 s. Domain-term transcription corruption affected 18.0% of retained cues. Live narration is thus a dense, annotation-free supervisory signal describing instruments, anatomy, actions and team coordination, but it is noisy and temporally displaced from the events it names, and this displacement must be modelled before safe downstream use.

Diffusion transformers have achieved remarkable success in image generation, but accelerating attention while preserving generation quality remains a major challenge at high resolutions. Inspired by statistical physics and the widespread occurrence of scale-free laws in nature, we introduce Scale-Free Attention (SFA), which preserves fine local interactions while summarizing distant regions at progressively coarser scales. For $N$ tokens, mean pooling, token-count correction, and a shared softmax over a scale-free spatial hierarchy give SFA $O(N\log N)$ complexity in both forward and backward passes. Our Triton implementation achieves $357.3\times$ forward and $265.6\times$ backward attention speedups over FlashAttention at 262,144 tokens. After 40 epochs of PixelFlow training from scratch on ImageNet at $256\times256$ resolution, SFA achieves the best generation quality among five attention mechanisms, including full attention. Fréchet inception distance evaluated on 50,000 generated samples (FID-50k) improves from 13.25 for full attention to 12.51, and the advantage over full attention persists through 80 epochs. SFA also achieves the best generation quality among the compared efficient-attention methods on Flickr-Faces-HQ (FFHQ). Beyond training from scratch, Stable Diffusion 3.5 Medium with attention replaced by SFA retains the generation quality of its full-attention counterpart under matched fine-tuning budgets. We obtain similar results when fine-tuning pretrained PixelFlow on ImageNet. SFA establishes scale-free structure as an effective design principle for efficient, high-quality diffusion transformers.