Academic Intelligence · Curated Daily

Explore the Frontier of Global Academia

AcademicHub aggregates real-time literature from top journals and preprint platforms. Build your personal research radar and let large language models compile cross-disciplinary analysis briefings automatically.

01.
arXiv (quant-ph) 2026-06-25

Software compensation of trigger-synchronous control-frame errors in qubits and qudits

arXiv:2606.00358v2 Announce Type: replace Abstract: Quantum control experiments are often subject to coherent, time-dependent disturbances that vary over timescales comparable to the experiment duration. We show that when such disturbances are reproducible with respect to a trigger signal, their effect can be measured and compensated through software-defined updates to the control frequency and phase. We experimentally verify the performance of our protocol using a trapped $^{137}$Ba$^+$ ion experiencing magnetic-field-induced energy shifts synchronous with the laboratory AC mains power. Using this compensation technique, the calibrated AC line contribution to the instantaneous oscillator detuning is reduced by $21(9)\times$, and the fitted AC-induced phase amplitude is reduced below the measurement uncertainty. We use randomized benchmarking to validate the compensation performance in quantum gate sequences, recovering an average single-qubit gate fidelity of 99.93(1)\% with a magnetic-field-sensitive qubit. Finally, we extend the compensation framework to multi-level qudit control. Using the Bernstein-Vazirani algorithm as a benchmark, we increase the algorithm's success probability from 10(7)\% to 70(9)\% in a 16 level system when compensation is enabled. Our results demonstrate that trigger-synchronized coherent errors can be reframed as deterministic control-frame errors and corrected in software.

02.
arXiv (CS.AI) 2026-06-25

Geometry-Aware Online Scheduling for LLM Serving: From Theoretical Bound to System Practice

arXiv:2606.22327v2 Announce Type: replace Abstract: The explosive demand for interactive Large Language Model serving has highlighted the management of the Key-Value cache's dynamic memory footprint as a critical area for performance optimization in inference engines. Modern inference systems overwhelmingly rely on time-centric scheduling heuristics, such as Shortest Job First. However, their theoretical optimality is rooted in traditional schedule modeling, failing to capture the highly dynamic, 2D spatio-temporal geometric growth specific to LLM inference mechanisms. To resolve this, we propose the geometry-aware online scheduling by introducing the Smallest Volume First (SVF) algorithm and its highly efficient variant, 1-bit SVF. Theoretically, we provide a rigorous mathematical foundation for our approach. Via a novel volume-certificate proof, we sharpen SVF's worst-case competitive ratio from the prior best of 48 towards 3 in the high-concurrency regime of LLM serving. Building upon this core breakthrough, we complete a comprehensive theoretical taxonomy analyzing our algorithms across different traffic scenarios and information availability. Practically, we seamlessly integrate our approach as a plug-and-play layer in vLLM. Extensive evaluations on Llama-3.1 models demonstrate comprehensive performance gains: SVF delivers strong reductions in both average and tail latency, while 1-bit SVF, with merely a single bit information, achieves competitive throughput and latency. This work establishes a theoretically sound and empirically proven approach for resolving memory-constrained scheduling in modern LLM deployments. To facilitate future research, our code is available at https://github.com/Aurora-Kl/Geometry-Aware-Online-Scheduling.git.

03.
arXiv (CS.AI) 2026-06-19

Variable-Length Tokenization via Learnable Global Merging for Diffusion Transformers

arXiv:2606.20076v1 Announce Type: cross Abstract: Latent Diffusion Models (LDMs) have become dominant in visual synthesis, but their quality-compute trade-off is largely constrained by the tokenizer's fixed compression ratio. Variable-length tokenizers (VLTs) promise adaptive compression by varying token counts, allowing diffusion models to flexibly balance quality and compute. However, conventional VLTs modulate length by truncating ordered token sequences, which makes token semantics depend on token position and breaks representational alignment across lengths. This leads to a cross-length shift in the latent distribution that hinders a single variable-length diffusion model from operating effectively. To address this, we propose a novel variable-length tokenizer that modulates length by merging tokens. We show that encouraging similar tokens to merge enables direct cross-length representation alignment when the diffusion transformer operates according to the merging pattern. Since conventional merging methods are data-dependent, making the merging pattern inaccessible during generation, we introduce learnable global merging, which is data-independent, to ensure compatibility with diffusion transformers. On ImageNet 256$\times$256 generation, our merging-based variable-length tokenizer integrated with a diffusion transformer achieves a superior gFID-compute trade-off compared to prior VLT methods. Code is available at [this https URL](https://github.com/movinghoon/lgm)

04.
medRxiv (Medicine) 2026-06-15

Filum Terminale Diameter on Routine Pediatric MRI: A Large-Cohort Clinical Reference in 3,406 Children and the Age-Dependent Meaning of the 2-mm Thickened-Filum Threshold

Background. A filum diameter >2 mm is the conventional MRI threshold for a thickened filum, but it derives from small, mostly adult series showing no age dependence; whether one cutoff suits all of childhood is untested. Objective. To build an age-specific filum-diameter reference on routine pediatric MRI and test, adjusting for image resolution, whether the 2-mm threshold is age-stationary. Materials and methods. In this retrospective study an nnU-Net tracer measured the maximal filum diameter on consecutive lumbosacral MRI; versus manual tracing it showed negligible bias but moderate single-measure agreement. After excluding report-confirmed fatty filum, lipoma, or tethered cord, the proportion >2 mm was analysed within one acquisition protocol and by logistic regression adjusting for voxel size and slice thickness. Results. Of 7,245 examinations, 3,869 (53%) were traceable; untraced ones were younger (median 0.75 vs 2.0 years). The presumed-normal cohort had median diameter 1.48 mm. At matched resolution, 2 mm marked the 94th percentile in infants (5.6% exceeded it) but the 83rd by 3-6 years (17.4%); the age effect persisted after adjusting for voxel size and slice thickness (3-6 years vs infants, adjusted OR 4.7; P < .001). Conclusion. Filum diameter clusters near 1.5 mm, and the fixed 2-mm cutoff flags ~5% of infants but ~17% of preschoolers. Caliber should be judged against an age-specific clinical reference, not one fixed cutoff; a thick filum is not itself a diagnosis of tethered cord.

05.
arXiv (CS.CV) 2026-06-18

SpectralDiT: Timestep-Conditioned Spectral Residual Correction for Flow-Matching DiTs

Authors:

We propose SpectralDiT, a lightweight modification to flow-matching Diffusion Transformers that adds timestep-conditioned spectral correction to the MLP residual branch. The module decomposes each residual update into low- and high-frequency components on the patch-token grid, then learns a zero-initialized additive gate so the model initially matches the baseline DiT. On CIFAR-10 pixel-space generation, SpectralDiT improves FID from 20.78 to 19.71 at patch size 1 and reduces the radial Fourier spectrum gap. Furthermore, we scale our method to latent diffusion on ImageNet-100. With 0.6% additional theoretical FLOPs and 1.36% additional parameters, SpectralDiT improves latent flow-matching, achieving an 8.7% relative FID reduction under classifier-free guidance (CFG 2.0). All reported results are averaged over five seeds. Ablations and gate visualizations on CIFAR-10 reveal stable block-specific spectral correction patterns.

06.
arXiv (CS.CV) 2026-06-11

Motion Reinforces Appearance: RGB-Skeleton Gated Residual Fusion for Micro-Gesture Online Recognition

Micro-gesture analysis attracts increasing attention for inferring spontaneous emotion from subtle body movements. Micro-gesture online recognition, which localizes and classifies each gesture instance in untrimmed videos, is a core task in the 4th EI-MiGA-IJCAI Challenge. Compared with typical temporal action detection, MGR emphasizes the localization and classification of actions, requiring the model to output the start time, end time, and category of each micro-gesture. Moreover, since micro-gestures are highly spontaneous, relying solely on a single modality makes it difficult to capture the complete and accurate multi-modal cues. In this work, we propose DyFADet+, which extends DyFADet into a dual-stream RGB-skeleton framework. In our model, both modalities are projected into shared multi-scale temporal embeddings and fused through a gated residual module, which adaptively injects skeleton motion into the RGB representation rather than using naive concatenation. Finally, these fused features are decoded by a Dynamic TAD head for online classification and boundary regression. On the SMG dataset, our method achieves an F1 score of 40.88, ranking 2nd in the Micro-gesture Online Recognition track.

07.
arXiv (CS.CL) 2026-06-16

Learn-to-learn on Arbitrary Textual Conditioning: A Hypernetwork-Driven Meta-Gated LLM

Conventional LLMs may suffer from corpus heterogeneity and subtle condition changes. While finetuning can create the catastrophe forgetting issue, application of meta-learning on LLMs is also limited due to its complexity and scalability. In this paper, we activate the meta-signal of $\beta$ within the SwiGLU blocks, resulting in a meta-gating mechanism that adaptively adjusts the nonlinearity of FFN. A hypernetwork is employed which dynamically produces $\beta$ on textual conditions, providing meta-controllability on LLMs. By testing on different condition types such as task, domain, persona, and style, our method outperforms finetuning and meta-learning baselines, and can generalize reasonably on unseen tasks, condition types, or instructions. Our code can be found in https://github.com/AaronJi/MeGan.

08.
arXiv (CS.AI) 2026-06-18

Surrogate Benchmarks for Model Merging Optimization

arXiv:2509.02555v2 Announce Type: replace-cross Abstract: Model merging techniques aim to integrate the abilities of multiple models into a single model. Most model merging techniques have hyperparameters, and their setting affects the performance of the merged model. Because several existing works show that tuning hyperparameters in model merging can enhance the merging outcome, developing hyperparameter optimization algorithms for model merging is a promising direction. However, its optimization process is computationally expensive, particularly in merging LLMs. In this work, we develop surrogate benchmarks for optimization of the merging hyperparameters to realize algorithm development and performance comparison at low cost. We define two search spaces and collect data samples to construct surrogate models to predict the performance of a merged model from a hyperparameter. We demonstrate that our benchmarks can predict the performance of merged models well and simulate optimization algorithm behaviors.

09.
arXiv (quant-ph) 2026-06-12

Stable, bidirectional electro-optic transduction in thin film lithium tantalate

arXiv:2606.12726v1 Announce Type: new Abstract: Efficient and stable microwave-optical transduction is a key enabling technology for distributed superconducting quantum computing and heterogeneous quantum networks. Electro-optic transducers based on thin-film lithium niobate (TFLN) have shown strong promise, but demonstrations to date have been limited by various factors such as low frequency bias drift, low efficiency, fabrication complexity, and scalability. Here we demonstrate the first integrated electro-optic microwave-optical transducers realized in thin-film lithium tantalate (TFLT), a material platform offering Pockels nonlinearity comparable to TFLN together with improved bias stability and high-power handling. We fabricate superconducting microwave resonators coupled to tunable photonic-molecule optical resonators using wafer-scale deep ultraviolet lithography, offering high-throughput production of hundreds of devices per wafer. Across six devices we observe coherent bidirectional conversion between C-band optical photons and 4.9-5.5 GHz microwave photons, with measured on-chip efficiencies and inferred single-photon coupling rates g_0/2{\pi} ~ 1 kHz consistent with theory. Continuous operation over multiple days is achieved using a static bias field with minimal feedback, demonstrating a major operational advantage. We further characterize optical loss statistics, microwave resonator performance, and optically induced added noise under pulsed pumping, finding less than one added photon for 100 microsecond pulses at the highest measured efficiencies. These results establish TFLT as a scalable and robust electro-optic platform for future quantum interconnects and modular quantum processors.

10.
arXiv (CS.AI) 2026-06-11

"That's AI Slop, You Bot!" Studying Accusations, Evidence, and Credibility in Online Discourse Towards LLM-Generated Comments

arXiv:2606.12073v1 Announce Type: cross Abstract: Generative AI has made fluent prose cheap to produce, breaking the old promise to readers that good writing meant real thinking. How have readers responded, and what can this tell us about changing anti-AI attitudes? We analyzed 25 million comments from Hacker News and Reddit (2023-2026), combining LLM judgment on 7,500 sampled accusations of AI use, sentiment trajectories, speech-act coding of 300 confirmed accusations of AI use, and a matched-control test of accused versus non-accused parent comments. We found that the pejorative-label share of accusations rose more than tenfold on both platforms while a placebo vocabulary of pre-2022 inauthenticity terms (shill, astroturf) did not. This shift reflected a fast-growing trend of branding any suspicious or seemingly inauthentic prose as "AI slop". The slop frame now constitutes 94 percent of pejorative mentions, with the dominant comments shifting in tone from mockery toward gatekeeping and structural protest. The key surprise comes from a matched-control test which found that prose features that statistically distinguish AI from human text do not predict which human text gets accused as AI. The new accusations work as social gatekeeping of perceived authenticity without actually screening for AI. This research extends signaling theory by showing that substitute signals used socially can grow even when inaccurate if the underlying detection problem cannot be solved at the non-expert level. It shows that AI's effects on writing from the reader side are distinct from those on the production (writer) side. Detection technology cannot resolve this dynamic because the social function of accusations is increasingly to perform social gatekeeping and in-group signaling as opposed to identifying AI-generated writing.

11.
arXiv (CS.LG) 2026-06-19

Characterization of Gaussian Universality Breakdown in High-Dimensional Empirical Risk Minimization

arXiv:2604.03146v3 Announce Type: replace-cross Abstract: We study high-dimensional convex empirical risk minimization (ERM) under general non-Gaussian data designs. By heuristically extending the Convex Gaussian Min-Max Theorem (CGMT) to non-Gaussian settings, we derive an asymptotic min-max characterization of key statistics, enabling approximation of the mean $\mu_{\hat{\theta}}$ and covariance $C_{\hat{\theta}}$ of the ERM estimator $\hat{\theta}$. Specifically, under a concentration assumption on the data matrix and standard regularity conditions on the loss and regularizer, we show that for a test covariate $x$ independent of the training data, the projection $\hat{\theta}^\top x$ approximately follows the convolution of the generally non-Gaussian distribution of $\mu_{\hat{\theta}}^\top x$ with an independent centered Gaussian variable of variance $\mathrm{tr}(C_{\hat{\theta}} \mathbb{E}[xx^\top])$. This result clarifies the scope and limits of Gaussian universality for ERMs. Additionally, we prove that any $\mathcal{C}^2$ regularizer is asymptotically equivalent to a quadratic form determined solely by its Hessian at zero and gradient at $\mu_{\hat{\theta}}$. Numerical simulations across diverse losses and models are provided to validate our theoretical predictions and qualitative insights.

12.
arXiv (CS.LG) 2026-06-24

Natural Identifiers for Privacy and Data Audits in Large Language Models

arXiv:2606.24408v1 Announce Type: new Abstract: Assessing the privacy of large language models (LLMs) presents significant challenges. In particular, most existing methods for auditing differential privacy require the insertion of specially crafted canary data during training, making them impractical for auditing already-trained models without costly retraining. Additionally, dataset inference, which audits whether a suspect dataset was used to train a model, is infeasible without access to a private non-member held-out dataset. Yet, such held-out datasets are often unavailable or difficult to construct for real-world cases since they have to be from the same distribution (IID) as the suspect data. These limitations severely hinder the ability to conduct scalable, post-hoc audits. To enable such audits, this work introduces natural identifiers (NIDs) as a novel solution to the above-mentioned challenges. NIDs are structured random strings, such as cryptographic hashes and shortened URLs, naturally occurring in common LLM training datasets. Their format enables the generation of unlimited additional random strings from the same distribution, which can act as alternative canaries for audits and as same-distribution held-out data for dataset inference. Our evaluation highlights that indeed, using NIDs, we can facilitate post-hoc differential privacy auditing without any retraining and enable dataset inference for any suspect dataset containing NIDs without the need for a private non-member held-out dataset.

13.
arXiv (CS.AI) 2026-06-12

An Embodied Simulation Platform, Benchmark, and Data-Efficient Augmentation Framework for Wet-Lab Robotics

arXiv:2606.12936v1 Announce Type: cross Abstract: Wet-lab robots can improve the reproducibility, throughput, and safety of biomedical experiments, but scaling their learning requires customizable simulators for safe and reproducible task generation, open editable laboratory assets, and efficient pipelines that turn limited demonstrations into usable training data. We present Pipette, an embodied simulation platform, benchmark, and data-efficient augmentation framework for wet-lab robot learning. Pipette releases over 43 open-source and re-editable wet-lab assets, together with an extensible asset-building pipeline. A key component of Pipette is its simulation-based data augmentation pipeline, replaying human demonstrations in simulation, applies lighting, camera, speed, and action perturbations, and filters generated episodes with automatic task success checks, rapidly expanding usable training data from limited manual demonstrations. We further introduce an 11-task wet-lab embodied benchmark covering sample handling, culture-ware manipulation, device operation, and precision placement. With only 30 demonstrations per task, ACT achieves 65.5% average success rate, while simulation augmentation improves SmolVLA from 44.1% to 74.7% and {\pi}0 from 40.4% to 46.5%, validating the effectiveness of Pipette for data-efficient VLA training and evaluation. Pipette also supports natural-language-driven scene construction and task registration, lowering the barrier for non-expert users to define new wet-lab robotic tasks.

14.
arXiv (CS.AI) 2026-06-25

Governing Technical Debt in Agentic AI Systems

arXiv:2605.29129v2 Announce Type: replace Abstract: Agentic AI systems are increasingly being explored as production infrastructure: they reason over multiple steps, call tools, act through workflows, and adapt through memory and feedback. These systems create governance challenges that are not fully captured by traditional software or predictive ML technical debt. We define Agentic Technical Debt as the accumulated liability created when prompts, memory, tool schemas, orchestration graphs, control policies, and observability routines are patched together faster than they can be validated, standardized, and governed. We define Stochastic Tax as the recurring operating burden of keeping probabilistic agent behavior within acceptable bounds. The distinction matters: debt is a stock of design and governance liability, while the tax is a flow of operating cost that arises because stochastic agents act through tools and workflows. We outline how managers can make both visible through lightweight dashboards and governance controls.

15.
arXiv (CS.CV) 2026-06-25

Gastroendoscopy View Synthesis: A New Real Dataset and Evaluation

Novel view synthesis (NVS) is an active research topic in computer vision, owing to the success of neural radiance field (NeRF) and 3D Gaussian splatting (3DGS) methods. While NVS opens the door to potential applications in gastroendoscopy, such as extending the field of view of endoscopic images and enabling digital twins for 3D archiving and endoscopist manipulation training, the dataset is insufficient to evaluate NVS for gastroendoscopy. In this paper, we present the first real gastroscopy dataset for NVS, namely the GastroNVS dataset, which contains a set of gastroscopic images, camera poses, and a point cloud for real gastroendoscopy inspection. To assess the suitability of the GastroNVS dataset, we evaluate several 3DGS methods and discuss the challenges for future development. The dataset is available on request from our project page.

16.
arXiv (CS.CL) 2026-06-15

Deja Vu at Scale: Paraphrase-Robust Detection of Duplicate Gherkin Steps in Behaviour-Driven Software Testing with Sentence-Transformer Embeddings and a 1.1M-Step Open Benchmark

Context. Behaviour-Driven Development (BDD) suites in Gherkin accumulate step-text duplication with documented maintenance cost. Prior detectors either require runnable tests or are single-organisation, leaving a gap: a static, paraphrase-robust, step-level detector and a public benchmark to calibrate it. Objective. We release (i) the largest cross-organisational BDD step corpus to date, (ii) a labelled pair-level calibration benchmark, and (iii) a four-strategy detector with a consolidation-savings model linking clusters to ISO/IEC 25010 maintainability sub-characteristics. Method. The corpus contains 347 public GitHub repositories, 23,667 .feature files, and 1,113,616 Gherkin steps, SPDX-tagged. The detector layers exact hashing, normalised Levenshtein, sentence-transformer cosine, and a Levenshtein-banded hybrid. Calibration uses 1,020 manually labelled step pairs under a released rubric (60-pair overlap, Fleiss kappa = 0.84). We report precision, recall, and F1 with bootstrap 95% CIs under the primary rubric and a score-free relabelling, and benchmark against SourcererCC-style and NiCad-style lexical baselines. Results. Step-weighted exact-duplicate rate is 80.2%; median-repository rate is 58.6% (Spearman rho = 0.51). The top hybrid cluster has 20,737 occurrences across 2,245 files. Near-exact reaches F1 = 0.822 on score-free labels; semantic F1 = 0.906 under the primary rubric reflects a disclosed stratification artefact. Lexical baselines reach F1 = 0.761 and 0.799. The savings model estimates 893,357 corpus-wide eliminable step occurrences; on the median repository 62.5% of step lines are eliminable.

17.
arXiv (CS.AI) 2026-06-11

DecompSR: A dataset for decomposed analyses of compositional multihop spatial reasoning

arXiv:2511.02627v3 Announce Type: replace Abstract: We introduce DecompSR, decomposed spatial reasoning, a large benchmark dataset (over 5m datapoints) and generation framework designed to analyse compositional spatial reasoning ability. The generation of DecompSR allows users to independently vary several aspects of compositionality, namely: productivity (reasoning depth), substitutivity (entity and linguistic variability), overgeneralisation (input order, distractors) and systematicity (novel linguistic elements). DecompSR is built procedurally in a manner which makes it is correct by construction, which is independently verified using a symbolic solver to guarantee the correctness of the dataset. DecompSR is comprehensively benchmarked across a host of Large Language Models (LLMs) where we show that LLMs struggle with productive and systematic generalisation in spatial reasoning tasks whereas they are more robust to linguistic variation. DecompSR provides a provably correct and rigorous benchmarking dataset with a novel ability to independently vary the degrees of several key aspects of compositionality, allowing for robust and fine-grained probing of the compositional reasoning abilities of LLMs.

18.
arXiv (quant-ph) 2026-06-24

Syndrome aware mitigation of logical errors

arXiv:2512.23810v2 Announce Type: replace Abstract: Broad applications of quantum computers will require error correction (EC). However, hardware roadmaps indicate that physical qubit numbers will remain limited in the foreseeable future, leading to residual logical errors that constrain the size and accuracy of achievable computations. Recent work suggested logical error mitigation (LEM), which applies known error mitigation (EM) methods to logical errors, eliminating their effect at the cost of a runtime overhead. We introduce syndrome-aware logical error mitigation (SALEM), which mitigates logical errors conditioned on the error syndromes measured during error correction. The runtime overhead of SALEM is exponentially lower than that of LEM schemes which do not make use of syndrome data, enabling substantially larger circuit volumes that can be executed accurately. Compared to the routinely used combination of error correction and syndrome rejection (post-selection), SALEM increases the size of reliably executable computations by orders of magnitude. In the practical setting where space and time overheads are fixed and error reduction methods are compared by their resulting estimation errors, we observe a surprising phenomenon: SALEM, which tightly combines EC with EM, can outperform physical EM even above the standard fault-tolerance (pseudo) threshold. Thus, SALEM can make use of EC in regimes of physical error rates where EC is commonly deemed useless.

19.
arXiv (CS.LG) 2026-06-18

On the Stability of Nonlinear Dynamics in GD and SGD: Beyond Quadratic Potentials

arXiv:2602.14789v2 Announce Type: replace Abstract: The dynamical stability of the iterates during training plays a key role in determining the minima obtained by optimization algorithms. For example, stable solutions of gradient descent (GD) correspond to flat minima, which have been associated with favorable features. While prior work often relies on linearization to determine stability, it remains unclear whether linearized dynamics faithfully capture the full nonlinear behavior. Recent work has shown that GD may stably oscillate near a linearly unstable minimum and still converge once the step size decays, indicating that linear analysis can be misleading. In this work, we explicitly study the effect of nonlinear terms. Specifically, we derive an exact criterion for stable oscillations of GD near minima in the multivariate setting. Our condition depends on high-order derivatives, generalizing existing results. Extending the analysis to stochastic gradient descent (SGD), we show that nonlinear dynamics can diverge in expectation even if a single batch is unstable. This implies that stability can be dictated by a single batch that oscillates unstably, rather than an average effect, as linear analysis suggests. Finally, we prove that if all batches are linearly stable, the nonlinear dynamics of SGD are stable in expectation.

20.
arXiv (CS.AI) 2026-06-17

MathVis-Fine: Aligning Visual Supervision with Necessity via Progressive Dependency-Guided Training for Multimodal Mathematical Reasoning

arXiv:2606.17888v1 Announce Type: new Abstract: Chain-of-Thought (CoT) reasoning has extended from purely linguistic domains to multimodal scenarios; however, existing approaches often treat visual inputs as homogeneous or auxiliary signals, failing to capture the intricate and sample-specific dependencies between text and images in mathematical problem-solving. This gives rise to two core issues: first, the supervisory signals for visual content are generalized and coarse-grained, lacking adaptation to the actual necessity of visual information in each sample; second, training feedback becomes inaccurate when visual rewards are uniformly applied without distinguishing the complementary relationships among inputs. These limitations hinder models from achieving precise multimodal reasoning. In this work, we propose a framework for modeling fine-grained visual dependencies in mathematical reasoning. We first construct the MathVis-Fine dataset, augmenting fine-grained visual annotations with visual dependency ratings. Building upon this dataset, we introduce a two-stage progressive visual enhancement training paradigm that balances answer correctness rewards and visual grounding rewards according to the intrinsic visual dependency level of each sample, thereby mitigating reward bias and improving supervision accuracy. Extensive experiments demonstrate that the MathVis-Fine framework effectively enhances visual perception progressively based on visual dependency, offering a more precise training framework for multimodal mathematical reasoning. We will release the dataset upon acceptance.

21.
arXiv (CS.AI) 2026-06-25

ACT-JEPA: Novel Joint-Embedding Predictive Architecture for Efficient Policy Representation Learning

arXiv:2501.14622v5 Announce Type: replace-cross Abstract: Learning efficient representations for decision-making policies is a challenge in imitation learning (IL). Current IL methods require expert demonstrations, which are expensive to collect. Additionally, they are not explicitly trained to understand the environment. Consequently, they have underdeveloped world models. Self-supervised learning (SSL) offers an alternative, as it can learn a world model from diverse, unlabeled data. However, most SSL methods are inefficient because they operate in raw input space. In this work, we propose ACT-JEPA, a novel architecture that unifies IL and SSL to enhance policy representations. It is trained end-to-end to jointly predict 1) action sequences and 2) latent observation sequences. To learn in latent space, we utilize Joint-Embedding Predictive Architecture, which allows the model to filter out irrelevant details and learn a robust world model. We evaluate ACT-JEPA in different environments and across multiple tasks. Our results show that it outperforms the strongest baseline in all environments. ACT-JEPA achieves up to 40% improvement in world model understanding and up to 10% higher task success rate. Finally, we show that predicting latent observation sequences effectively generalizes to predicting action sequences. This work demonstrates how integrating IL and SSL leads to efficient policy representation learning, an improved world model, and a higher task success rate.

22.
arXiv (CS.AI) 2026-06-16

CoffeeBench: Benchmarking Long-Horizon LLM Agents in Heterogeneous Multi-Agent Economies

arXiv:2606.16613v1 Announce Type: new Abstract: As LLM agents become capable of increasingly long-horizon tasks, evaluating their performance in economic systems is becoming increasingly important. Unlike existing benchmarks that primarily evaluate a single agent interacting with a passive environment, economic systems are inherently multi-agent, requiring autonomous agents to communicate, negotiate, and transact while pursuing their own objectives over extended periods. We introduce CoffeeBench, a benchmark for evaluating LLM agents in a long-horizon multi-agent economy composed of heterogeneous firms. In CoffeeBench, two farmers, two roasters, and two retailers autonomously operate their businesses over a 90-day simulation, each seeking to maximize cumulative net income through communication and transactions while managing cash, inventory, and pricing. The evaluated model controls one coffee roaster, while the remaining firms are controlled by fixed reference agents. Across several recent open-weight and proprietary LLMs, all models outperform a passive baseline that takes no actions, with most achieving positive net income. Analysis of agent behavior reveals substantial differences in long-horizon economic interaction: higher-performing models communicate more actively with other firms, whereas Claude~Haiku~4.5 exhibits an idle-drift failure mode, repeatedly choosing inaction despite producing coherent assessments and plans. We release our code and agent trajectories to support future research.

23.
arXiv (CS.AI) 2026-06-18

Explaining Attention with Program Synthesis

arXiv:2606.19317v1 Announce Type: cross Abstract: A longstanding goal of research on interpretable deep learning is to replace opaque neural computations with human-meaningful symbolic descriptions. In this paper, we propose an approach for approximating the behavior of components of deep networks with executable programs. We focus on attention heads in transformer language models. For a given head, we first compute its associated attention matrices on a collection of randomly selected training examples. Next, we prompt a pre-trained language model with a summary of these matrices, and instruct it to generate a set of Python programs that can reproduce the associated attention patterns given only text from the input sentence. Finally, we re-rank programs according to how well our final set of programs predict behavior on held-out inputs. We demonstrate that a set of fewer than 1,000 such generated programs can reproduce the attention patterns of heads in GPT-2, TinyLlama-1.1B, and Llama-3B, achieving an average Intersection-over-Union similarity above 75% on TinyStories. Moreover, the best-fit programs can replace neural attention heads without substantially affecting model behavior: replacing 25% of attention heads with programmatic surrogates across the three models incurs only a 16% average perplexity increase, while maintaining performance on a variety of downstream question answering benchmarks. This work contributes a scalable pipeline for reverse-engineering attention heads in transformer models using human-readable, executable code, advancing a path toward symbolic transparency in neural models.

24.
arXiv (CS.CV) 2026-06-11

Frames2LoRA: Parametric Video Internalization for Vision-Language Models

Processing video in vision-language models is expensive: each frame occupies hundreds of tokens, and inference cost scales with every frame and every repeated query. We introduce Frames2LoRA, a method for parametric video internalization. A perceiver hypernetwork reads the intermediate representations produced layer-by-layer as a frozen VLM encodes a video, and generates a Low-Rank Adaptation (LoRA) adapter in a single forward pass. Unlike standard LoRA fine-tuning, which requires iterative gradient updates, Frames2LoRA predicts these weights directly from the video. Trained for SmolVLM2 500M and 2.2B on video summarization and captioning, Frames2LoRA enables the same frozen VLM to answer queries from the adapter alone, with zero visual tokens in its context at query time. Frames2LoRA is statistically non-inferior and equivalent to direct video-in-context inference across all five captioning benchmarks at both model scales, and across seven of eight video question answering benchmark-scale pairings. Although trained only on 12 frames at 384px, it remains stable up to 1,024 frames and 1024px, where direct video-in-context inference often degenerates. Across this sweep, it reduces answer-time visual-token load by up to 1,500x and query TTFT by 6-80x, while preserving video-faithful outputs. We also find that independently generated adapters for non-overlapping video segments can compose in rank space, suggesting a path toward chunked long-video internalization.

25.
arXiv (CS.AI) 2026-06-16

Guiding Federated Graph Recommendation with LLM-encoded knowledge

arXiv:2606.15277v1 Announce Type: cross Abstract: Graph-based recommender systems are highly effective at extracting collaborative signals from user–item interactions, and federated learning (FL) allows these models to be trained while preserving user privacy. However, aggregating graph representations across distributed, non-IID clients remains a challenge; structural embeddings learned locally often misalign, and naive averaging fails to capture meaningful cross-client relationships. Most existing federated graph methods rely exclusively on structural aggregation, neglecting the rich, global semantic context available in large language models (LLMs). In this paper, we propose a novel framework that uses LLM-encoded knowledge to guide federated graph recommendation. Specifically, clients learn structural representations from local graphs while simultaneously summarizing their typical interaction patterns into compact semantic vectors via a frozen LLM. The central server then uses these LLM-encoded semantic signals to discover related preference patterns across clients, guiding the selective aggregation of their structural representations. This enables semantically informed cross-client collaboration without exposing raw data. Extensive experiments on standard benchmarks show that guiding structural alignment with LLM-encoded knowledge consistently improves recommendation accuracy over existing federated graph baselines.