Papers

Purpose : To develop and evaluate CAPTOR, a deep-learning vision-language foundation model for oncologic chest-abdomen-pelvis (CAP) CT interpretation. Materials & Methods : CAPTOR (Chest-Abdomen-Pelvis foundaTion model for Oncologic Radiology) is a contrastive vision-language model, trained using a pretraining strategy for oncologic CAP CT (in single- or multiple-series acquisitions), on a multicenter retrospective dataset from North America, Europe and Asia, predominantly from tumor lesion examinations. CAPTOR was compared with state-of-the-art CT foundation models on three downstream tasks: disease detection, survival prediction using TumorImagingBench and image retrieval. Classification and retrieval were both evaluated on held-out internal and public test sets, stratified by anatomical coverage and indication. Bootstrap resampling was used to estimate confidence intervals, and pairwise comparisons were performed using permutation tests. Results : CAPTOR was trained and tested on 25,512 and 10,794 CT-report pairs respectively. TumorImagingBench datasets comprised between 134 and 421 studies each. CAPTOR outperformed all baseline models, on the classification task, on the CAP subset (AUC 0.767, 95%CI: 0.734-0.793, vs. 0.695, 95%CI: 0.645-0.736 for the next-best, p<0.001) and oncologic subset (AUC 0.784, 95%CI: 0.759-0.805, vs. 0.734, 95%CI: 0.703-0.767 p<0.001), ranked first on three of the four oncologic prognostic benchmarks, and achieved superior retrieval performance on CAP and oncologic subsets. Conclusion : A foundation model trained on oncology-enriched CAP CT data outperformed relevant state-of-the-art models in its target clinical context, underscoring the importance of clinically aligned pretraining. We release our preprocessing, training and inference code : https://github.com/Guerbet-AI/CAPTOR and model weights : https://huggingface.co/Guerbet-AI/CAPTOR

LLM agents' capabilities depend not only on their underlying models but also on their execution harnesses. Recursive self-improvement (RSI) offers a promising approach to automatically evolve these harnesses. However, deploying RSI in open-world settings still faces two critical challenges. First, open-world tasks are typically unseen during the harness design phase and novel at test time, leaving no ground-truth verifiers to reliably guide online harness edits. Second, the RSI process remains highly task-specific and fails to generalize to novel scenarios. We introduce Harness^2, a plug-and-play, recursive agent harnessing framework that addresses these challenges through two connected levels of recursion. Task-level recursion refines the current task's harness using a contrastive Propose-Probe-Compose step that reliably guides edits without ground-truth verifiers, and can be scaled further via parallel or sequential recursion. Harnessing-level recursion then extracts experience from these edits to update a continual improvement harness, enabling the agent to better adapt to future tasks. Across three professional benchmarks and four model-harness configurations, a single step consistently improves GLM-5 and Gemini 3.5 Flash over their base harnesses, yielding gains of up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench, and 11.8 percentage points on LAB, with further gains as the task-level recursion budget increases.

Softmax attention in Transformers suffers from quadratic complexity in sequence length, making it impractical for long-context applications. Linear attention alleviates this issue by replacing the exponential kernel with alternative functions that enable linear-time computation. Among existing linear attention approaches, recent studies have shown that polynomial kernels are particularly effective, as they exhibit sparse and spiky behavior similar to softmax attention, which emphasizes large dot products while suppressing irrelevant interactions. However, the exact computation of high-degree polynomial kernels is infeasible for high-dimensional representations. As a result, prior work relies on approximate polynomial kernels. This introduces a non-negligible approximation error. In this paper, we show that, from the perspective of polynomial kernel approximation, existing linear attention methods are still suboptimal. We propose Learnable Low-Rank Polynomial Sketch (LLoPS), a principled and flexible framework for approximating polynomial kernels with linear attention. Our method learns a low-rank polynomial sketch that provably achieves a strictly smaller approximation error than existing approaches. Experiment results show that LLoPS achieves the highest performance across extensive benchmarks, comparing with various linear attention baselines.

A digraph $D = (V,E)$ is 3-transitive (respectively, 3-quasi-transitive) if whenever $(u,v,w,z)$ is a directed path of length 3, $(u,z) \in E$ (respectively, $(u,z) \in E$ or $(z,u) \in E$). These digraphs extend the concepts of transitive and quasi-transitive digraphs, respectively. For 3-transitive and 3-quasi-transitive digraphs, only the strong case has been characterized. In this paper, we introduce the notion of semi-acyclic digraphs and use it to describe completely the structure of 3-transitive and 3-quasi-transitive digraphs. As an application of this type of structure description, we prove that the path partition conjecture holds for the class of extended semi-acyclic digraphs, which generalizes two previously known results on the conjecture.

Coding agents such as Claude Code also follow text files that developers write for them: a project file, skills the agent may choose to open, and rules that load automatically when the agent reads matching files. We call these files the agent's harness. Developers add to it whenever they correct the agent, but one success proves little: the same agent can pass a task once and fail it the next time. We present Cortex, a tool that tests changes to the harness instead of trusting them. It saves fixes from the developer's sessions as replayable tasks, finds the corrections that recur, writes one skill or rule at a time, and keeps it only if repeated runs with and without it show a net gain and no task made noticeably worse. Existing text is tested the same way before it is deleted, and Cortex neither adds nor deletes text the agent never used. We tested Cortex with Claude Haiku 4.5 on a repository built for the purpose, with conventions the model tends to break, one area with nothing to learn, and a script playing the developer. On held-out tasks, the harnesses it built made the agent follow the conventions far more often than no skills or rules did, left the comparison area unchanged, and added no text read on every turn. Rules, which load automatically, did most of the work; skills helped only when offered with the files a task touches, and those offered on every turn were rarely opened. Cortex kept none of the useless or harmful suggestions we planted, but the agent never opened the useless ones or the harmful skills, and it reached no verdict on the harmful rules, so the test shows that Cortex keeps no unused text, not that it detects damage. Under the stronger Claude Sonnet 5 a single run's harness added little, and on a real open-source project, where the agent was rarely corrected, Cortex found nothing to learn. We release the tool, the lab and its data.

The deployment of large language models (LLMs) on edge devices is largely motivated by privacy concerns over sharing sensitive user data with cloud servers. In this paper, we demonstrate that keeping inference local does not eliminate privacy risks: an adversary with physical access to the device may exploit hardware leakage to recover sensitive information without compromising or modifying the software system. In particular, we present an electromagnetic (EM) side-channel attack targeting LLM inference on edge devices. Targeting LLaMA-3.2-1B-Instruct running on a Raspberry Pi 5, we demonstrate that EM traces collected during inference can leak both generated tokens and user prompts. We propose three attacks under different threat assumptions. First, a profiling attack that identifies generated tokens among 50 medical and privacy-related keywords with 97.2% accuracy on profiled words. Second, a reconstruction attack recovers the model's hidden representations and uses them to predict tokens from the full 128k-token vocabulary, achieving 94.3% accuracy on profiled classes. For unseen words, the correct token is ranked within the first 100 positions of the 128k-token vocabulary in 67% of cases, with a median rank of 54. Finally, we present a hidden-state inversion attack that uses the recovered hidden state to reconstruct user prompts semantically. Our proposed attacks require only passive EM side-channel measurements and a limited number of traces during model training. These results demonstrate that local execution alone may not guarantee the confidentiality of generated tokens and user inputs.

Hallucination detection is critical for ensuring the reliability of Large Language Models (LLMs) in real-world deployments. While recent works, such as Semantic Entropy (SE) and Kernel Language Entropy (KLE), identify hallucinations by measuring semantic disagreement across multiple generated responses using Natural Language Inference (NLI) models, these methods depend on separately loaded NLI classifiers to estimate pairwise semantic similarity, introducing significant computational overhead. We propose Embedding Graph Kernel Entropy (EGKE), a variant of Kernel Language Entropy (KLE) that replaces its NLI-derived semantic graph with a Radial Basis Function (RBF) kernel applied to dense sentence embeddings generated by a BGE embedding model, while retaining KLE's graph Laplacian, heat kernel, and von Neumann entropy formulation unchanged. By computing similarities directly in a continuous embedding space, the proposed method provides a symmetric and computationally efficient alternative to NLI-based semantic comparison. We evaluate our approach across 30 experimental setups (covering multiple question-answer datasets and open-weight LLM families) using both AUROC and AUARC, and demonstrate that this RBF-based method achieves competitive hallucination detection performance against state-of-the-art methods.

LeetCode Problem 856, Score of Parentheses, admits a one-pass algorithm using $O(n)$ time and $O(1)$ auxiliary registers. Its correctness follows from a local interpretation: every occurrence of () contributes a power of two determined by its nesting depth. The space bound is immediately optimal in the usual programming-challenge model, but the linear-time bound is less obvious. In particular, a direct adversary that forces an algorithm to inspect every character is difficult to maintain under the promise that the input is balanced. We instead prove a counting lower bound. For every $k$, we construct $2^k$ balanced parenthesis strings of the same length $4k+2$, all having distinct scores. The construction encodes each bit using one of two operations, changing a current score $v$ to either $2v+1$ or $2v+2$. Consequently, any deterministic character-query decision tree computing the score must have depth at least $k$. A two-character padding argument extends the result to every even input length, proving an $\Omega(n)$ worst-case time lower bound. The one-pass algorithm is therefore asymptotically optimal in the unit-cost RAM model.

Understanding how information is distributed over time is fundamental to modeling natural language. While existing empirical studies primarily rely on post-hoc analysis of information-theoretic metrics, such as surprisal, little is known about the generation processes under the surface. We propose a novel approach for addressing this gap using the tool of neural temporal point processes (TPPs) by modeling the key information-bearing events in text sequences with learnable event intensity functions. Our model takes the history of event intervals and the contextual semantic representations from LLMs as input, and predicts the future arrival times for two types of events: new-word and high-surprisal-token. The experimental results reveal an interesting distinction in event intensities: LLMs tend to produce new words at significantly higher rates than humans, while high-surprisal tokens do not exhibit the difference. A decaying pattern of intensity over time is found in new-word events for both humans and LLMs, but not in high-surprisal-token events. Our findings revisit and extend previous studies that use simpler methods; provide supplementary evidence for theories about constrained context representation inherent in human language production; and find a new angle to view human and LLM generation differences.

ScanRefer Multiple queries require spatial disambiguation: several instances of the target class are present, and the referencer must pick one. When per-query logits were never logged, a common substitute is cross-epoch agreement. On the full ScanRefer validation set (9,508 queries; GeoClassifier with ground-truth proposals), object-level softmax is similarly miscalibrated on Unique and Multiple (ECE 0.068 vs. 0.102; the scene-bootstrap gap interval includes zero). Fixed-final agreement on the same checkpoints reports ECE 0.041 vs. 0.272 (gap +0.231). Modal-vote agreement is 0.045 vs. 0.318. Most Multiple errors are same-class wrong-instance (45.6%); geometric miss (H2) is 59.6%, and wrong-category (H1) is 16.7%. A non-monotone H1 pattern does not appear under the default fixed-final protocol, and we retract it. Logit-free thresholds rank spatially hard queries poorly. If only agreement was logged, calibrate abstention on Multiple and report wrong-instance rates with H1.

The most accurate agents executing user-provided tasks are driven by large, cloud-hosted models, which solve them by reading the private data of the user. Agents powered by models small enough to run on edge devices keep that data local, but yield far worse results. A third option is to have cloud models write a program that devices execute: the data stays local, but as the program is written before any data is seen, it fails whenever a step depends on interpreting content that appears only at run time. We explore an intermediate architecture, in which the cloud-written program defers bounded questions over user data to a small local model through an ordinary function call. On AppWorld this improves the success rate of a code-only agent in all six combinations of backbone model and test partition we evaluate, by up to 14.9pp, recovering a substantial part of the gap to a cloud-hosted ReAct loop while keeping user data on the device.

Transmitters in adaptive communication systems may switch their modulation scheme at an unknown time, and a receiver must detect this switch quickly without raising frequent false alarms. Existing modulation classification methods label a finite block of received data and do not provide stopping rules with controlled false-alarm and delay performance. We study the sequential detection of a switch from one phase shift keying (PSK) scheme to another over an additive white Gaussian noise channel. We build cumulative-sum (CUSUM) and moving-sum (MOSUM) detectors from the log-likelihood ratio and its low-SNR approximation. For a switch from quadrature PSK to binary PSK, we derive the moment-generating function and probability density function of the approximate statistic. We use them to relate the CUSUM threshold to the average run length to false alarm and the worst-case detection delay. For a switch between any two PSK orders, we approximate the vector of overlapping MOSUM statistics as multivariate Gaussian and obtain bounds on both measures. Monte Carlo simulations show that our derived expressions for ARL2FA and WADD closely match the empirical results across thresholds, validating the accuracy of the proposed analytical approximations and theoretical bounds.

Under a fixed context budget, which objects and relations should a 3D scene-grounded VLM see first? On ScanNet indoor referring benchmarks (Dai et al., 2017), matching the query to object category names outperforms pairwise geometry scoring at every reported K. The gaps are largest at high recall. Analytic pairwise cues beat per-object cue scoring and query-agnostic kNN graphs, including on paraphrases. MiniLM and BGE-small still score higher on those splits. We evaluate token-budgeted candidate retrieval with axis-aligned boxes, a token-ratio metric, training-free baselines, two small encoders that match q to each object's category name, and QCRR (Query-Conditioned Relation Retrieval), a deterministic pairwise-geometry scorer. Across ScanRefer gt_val plus a custom 20% scan-id split of Nr3D and Sr3D (35,504 queries; not official ReferIt3D val), additive noun matching reaches R@10 0.864 on ScanRefer against QCRR's 0.626; MiniLM and BGE-small reach 0.911 to 0.913 on ScanRefer and 0.941 to 0.947 on Nr3D. A corrected 3DGraphLLM harness (native Acc@0.25 0.67 on the full graph, n=300 ScanRefer queries) shows that context selection changes downstream accuracy under a matched object budget (Random 0.44, QCRR 0.56, MiniLM 0.67, oracle 0.70). Semantic retrieval beats hand-designed geometry on that harness. We will release the protocol, retriever implementations, and per-query outputs so new selectors can be swapped in without rewriting the analysis.

Implicit neural representations (INRs) offer a powerful framework for modeling high-dimensional physical fields, but their training remains inefficient because evaluating dense coordinate grids is computationally expensive. Existing coordinate-pruning strategies rely on local error functions that can oversample correlated points and incur substantial overhead from full-grid error evaluation. We introduce Refined Grid Sampling (RGS), which combines stratified grid sampling, error-based refinement, and dynamic grid-resolution adjustment. This design reduces spatial redundancy without evaluating errors over the full grid. On high-resolution Navier–Stokes fields, RGS achieves higher reconstruction accuracy within a fixed wall-clock budget and reaches the accuracy of full-grid training up to $6\times$ faster than the evaluated baselines.

Contemporary AI development operates on foundational assumptions about intelligence, rationality, and human values that reflect narrow cultural perspectives, limiting alternative conceptions of how AI can be globally beneficial. This paper introduces seven anthropological concepts – origin story, monoculture, thick description, personhood and community, cargo cult, structuralism, and liminality – to expose hidden assumptions and offer alternative frameworks for understanding AI alignment. The paper examines how AI systems embody Western-centric notions of rationality as utility optimization, personhood as bounded individuals, and values as quantifiable preferences extracted from behavior. The concepts surface three assumptions that current alignment debates rarely examine: that values can be inferred from behavior, quantified, and formalized; that such quantified values capture what actually matters; and that values, once defined with sufficient precision, are appropriate optimization targets. Against "thin alignment" – optimization for behavioral proxies stripped of context – the paper proposes directions toward "thick alignment": recognizing that values emerge through social practice rather than existing as extractable properties; defining more appropriate metrics that capture not only end goals but the conditions that lead to long-term flourishing; and moving from models of fixed identity toward relational ontologies where personhood emerges through social participation. The paper concludes that if AI systems are to create conditions conducive to human flourishing rather than optimize metrics that flatten social experiences, three shifts are needed: redistributing interpretive sovereignty to affected communities, protecting what resists optimization, and extending accountability across generations.

We introduce ClothCode, a unified generative framework for garment state estimation and action-conditioned dynamics prediction in a shared discrete configuration space. A material-aligned tokenizer encodes garment configuration as residual vector-quantized patch tokens, while task-specific Masked Generative Transformers predict these tokens through parallel confidence-guided refinement. Across three garment templates, ClothCode reduces mean displacement error by 26.6% for state estimation and 37.0% for dynamics prediction relative to the strongest baselines. Dynamics prediction requires only 16.7 ms per window, compared with 556.2 ms for diffusion. Integrated into the same model-predictive controller, ClothCode improves real-world folding and flattening success from 56.7% to 78.9%. Further results and videos are available at: https://anonymous.4open.science/r/ClothCode-DE32

Late-interaction (multi-vector) models represent documents using a set of token-level vectors, providing a middle ground between cross-encoders and dense encoders in terms of quality and cost of embedding. However, compared to the latter, the indices are costly to store and score. Pooling and pruning based compression techniques are deployed widely but deteriorate in quality noticeably with aggressive compression. We investigate whether a lightweight change to the scoring function can recover retrieval quality after document-vector compression. For each combination of encoder, compressor, and vector budget, we learn a global linear correction by distilling the uncompressed model's score distributions over candidate documents. The encoder and compressor remain frozen, and the correction can be applied to each query at search time or once, offline, to the stored vectors. We evaluate the method with ColBERTv2 and Jina v4 on MS MARCO and BEIR, and with ColPali2 on the ViDoRe v3 visual document retrieval benchmark. Experiments cover four training-free compression methods, multiple vector budgets, in-domain adaptation, and transfer from MS MARCO to BEIR. In the in-domain setting, the correction improves over baseline compression in 154 of 160 BEIR settings, while transfer from MS MARCO is more variable across models and datasets. We also find that matrices trained with synthetic queries can improve over baseline compression on ViDoRe v3. These results show that a small learned correction can recover some of the quality lost to compression without retraining the encoder or recompressing the documents.

Vertical federated learning lets organizations train a shared model over features they hold separately, while one party keeps the labels private. The active label inference attack of Fu et al. [12] breaks this: a feature-holding party runs a modified optimizer, MaliciousSGD, that forces class structure into its embeddings and then recovers the labels with semi-supervised learning. Defenses that perturb the shared gradient fail, because the attack acts on the party's internal optimizer, not on the message it sends. We show the attack leaves a signal the server can read from the embeddings it already receives. The adversary's Fisher class-separability pulls away from the honest party's, and the gap is visible within a few epochs. A defense triggered by this signal acts on the adversary alone. Scaling its gradient down handles few-class tasks, and a norm non-expansive projection handles many-class tasks, with the Discriminative Coverage Ratio explaining why the projection rank must grow with the class count. The defense breaks the alignment between embedding clusters and labels rather than destroying the embedding. Across four seeds on CIFAR-10, CIFAR-100, and CINIC-10, label inference falls below what honest training already leaks, at under two points of task-accuracy cost.

The modified Anderson-Bjorck (modAB) method was first proposed in 2023 by Ganchovski and Traykov for solving nonlinear equations in structural mechanics. It was improved further in 2026 by introducing better switching criteria with account to function symmetry, and bisection fallback based on measuring the bracket reduction rate. Although the algorithm demonstrated superior performance across a great variety of functions, it lacked formal mathematical analysis and proof of convergence. In addition, some discrepancies were noted by L. Tomov in the latest version that need to be addressed: premature application of the Anderson-Bjorck (AB) correction on the first AB step, caused by retaining the endpoint-update record from the preceding bisection step, and using the corrected endpoint ordinates in symmetry factor calculation instead of the true residuals. The purpose of the present paper is to provide rigorous mathematical proof of global convergence in exact arithmetic, remove the above discrepancies, and provide floating-point safeguards for interpolation, switching and stopping, with explicit treatment of infinite and NaN values, underflow, overflow and round-off errors. These safeguards do not certify the accuracy of the supplied function values. Additional performance improvement was introduced by allowing for early bisection fallback when AB fails to make progress on reducing either the bracket width or the residual, while retaining a finite bound on consecutive exceptions to the prescribed width condition.

The codeword stabilized (CWS) framework provides a systematic method for constructing quantum error correcting codes from a simple undirected graph and a classical code. When the underlying classical code is linear, the resulting quantum code is a stabilizer code. In this work, we investigate the conditions under which an arbitrary $[n,k,d]$ classical binary linear code can be lifted to an $[[n,k,d]]$ quantum stabilizer code using the CWS framework. We analyze the random CWS code constructed from a given $[n,k,d]$ binary linear code $\mathcal{C}$ and an Erdős-Rényi random graph $\mathrm{G}(n, \frac12)$. We establish a lower bound on the probability that the resulting quantum code preserves the minimum distance $d$ of the classical code. This probabilistic analysis yields a Gilbert-Varshamov-type condition on the $n,k,d$ parameters of the classical code that guarantees the existence of a graph resulting in a quantum code with the same $n,k,d$ parameters.

Forecasting tissue deformation from future instrument motion could support predictive planning and increasingly autonomous surgical systems. We study this problem using a scene-specific differentiable surface model fitted to an observed tool-tissue interaction window. Rather than forecasting absolute simulated positions, we introduce an observation-referenced formulation that anchors each tissue point to its last valid observation and adds only the deformation induced by the known, preplanned future tool trajectory. This reduces sensitivity to reconstruction, initialization, and persistent simulation offsets while preserving the predicted deformation response. The physical model uses differentiable, mesh-free position-based dynamics with scene-specific rigidity, attachment, region scale, and grasp coupling. We further introduce an optional observation-referenced residual model combining graph-spectral spatial modes with periodic harmonic dynamics for structured motion not explained by the tool input. We evaluate the framework retrospectively on 12 surgical manipulation clips from three datasets. Observation referencing provides the largest gain, with a median within-clip error ratio of $0.598$ relative to the corresponding absolute simulation forecast. The optional periodic residual further improves forecasting when the remaining motion exhibits periodic structure. These results support observation-referenced deformation prediction as a practical building block for future surgical world models, digital twins, and generative simulators that must predict how tissue evolves under instrument actions.

Explaining a vision–language model (VLM) prediction requires tracing training-example influence to the multimodal evidence that mediates it, yet existing attribution methods largely operate only at the sample level. Recent latent-mediated influence methods provide feature-level attribution in language models, but lack the multimodal and spatial structure needed to ground influence in VLMs. We introduce MOSAIC (Multimodal Orthogonalized Sparse Autoencoders for Influence Characterization). It learns a shared sparse image–text latent space whose concepts are 1) sparse within each modality, 2) supported on contiguous image regions, 3) aligned to the phrases that describe them, and 4) approximately decorrelated. It then applies EK-FAC to decompose training-example influence through these concepts and localize their contributions to specific text phrases and image regions. We evaluate MOSAIC on Qwen and Gemma across captioning, chest X-ray understanding, and visual question answering. The learned features are more localized and coherent without degrading task behavior, and ablating influence-ranked latents changes model outputs more than ablating latents ranked by activation or frequency. We find that visual evidence can substantially reorganize training-data attribution while preserving more of the shared latent structure. MOSAIC thus connects where model behavior comes from with what multimodal evidence carries that influence.

Urban traffic systems are becoming increasingly complex, demanding real-time traffic perception, proactive collision prediction, and intelligent management to improve safety and reduce congestion. Although AI-driven traffic management has shown strong promise, its success relies on diverse, high-quality data spanning critical scenarios such as collisions, near-collisions, and traffic rule violations. Digital twins offer a powerful solution by enabling synchronized virtual replicas of physical traffic environments for real-time monitoring, collision prediction, traffic-flow optimization, and diverse realistic scenario generation. However, continuously synchronizing a high-fidelity urban traffic digital twin through edge surveillance cameras requires substantial edge-cloud communication, and transmitting raw video streams leads to severe bandwidth and latency constraints. To overcome this challenge, this paper proposes a semantic communication-based edge-cloud framework for real-time urban traffic digital twin applications. By leveraging edge intelligence to extract and transmit only task-relevant semantic traffic information instead of full video feeds, the proposed framework significantly reduces communication overhead while preserving the essential dynamics needed for accurate real-time digital twin reconstruction. Experimental results demonstrate more than 99% transmission savings and an average trajectory reconstruction error of only 0.12 m, confirming the proposed framework as an effective, scalable, and low-latency solution for AI-enabled urban traffic digital twin applications.

Could model-free reinforcement learning algorithms actually be learning a model of the environment? In this work, we apply counterfactual interventions which hold everything about a state fixed, besides a targeted change, to explore if the value function of a Proximal Policy Optimization agent forms a world model of the Crafter environment. Despite only being trained to optimize reward, we discover that the value function learns concepts such as the rules governing when a diamond is collectible, water and food having an impact on health, heuristics regarding inventory, and more. Finally, we discuss the safety implications of our visualizations and some shortcomings of current reinforcement learning agents that will prevent us from building truly advanced machine intelligence.

Studying model training requires checkpoint trajectories, whose storage cost limits temporal resolution. Although adjacent checkpoints are numerically close, existing lossless codecs either operate on XOR bit patterns or ignore trajectory-specific conditional structure. We show that a reference weight's exponent predicts the bit length of its displacement in ordered floating-point space without side information. We introduce ULPack, a lossless codec which exploits this dependence by conditioning entropy models of ordered-integer residuals on the reference exponent. Across bfloat16, float16, and float32 trajectories, ULPack reduces storage by up to 27% against ZipNN, 21% against zstd-19, and 11.5% against FM-Delta. Across two dense trajectories, compression improves as updates shrink and reference conditioning strengthens, while near uniform low-order bits set the remaining compression floor and earlier checkpoints add little information.

Organisations that want language-model agents for security work face a choice between open-weight models, which they can run and audit themselves, and far more expensive closed frontier models. We study that choice on CTI-REALM (arXiv:2603.13517), a benchmark in which an agent turns a threat-intelligence report into a working detection by querying a live telemetry store. We evaluate twenty-six models, nineteen open-weight and seven closed, under a common harness, scorer and panel of LLM judges. Our main finding is that an open-weight model can match the strongest closed models at less than a tenth of their cost, provided it is given enough turns. The benchmark's published limit on agent messages is a confound: a task that exhausts the limit scores zero on every checkpoint, including those it had already passed, so results at that limit partly measure which models finish in time. At the published limit the leading open-weight and closed models cannot be distinguished. With the limit relaxed, the models it was truncating recover, and the highest rewards belong to open-weight models. Among the leading models, price is a poor guide to capability: models with similar rewards differ in cost by more than an order of magnitude. Because the benchmark's checkpoint scores largely record what an agent did or mentioned, we also introduce stage-level metrics that score what the agent submits, by executing its detection queries in the telemetry store and judging its outputs against the answer key. Rather than combining them into one more overall reward, we show that different models lead different stages, which suggests routing each stage of the workflow to the model that performs it best rather than choosing a single model by its overall score. We also report how many tokens each model generates, and find that the share of output spent on thinking does not track the reward.

Digital libraries increasingly host scientific publications whose figures encode primary research findings, yet making this visual content searchable, accessible, and reusable remains an open challenge. Large language and multimodal models (LLMs/LMMs) provide the possibility to automate many figure parsing tasks such as determining data trends or extracting data from plots. In this work, we consider if these models can reliably support task-oriented visual data analysis when presented with synthetic, but realistic, scientific figures that contain multiple charts. Our Visual Question Answering (VQA) process scaffolds our tasks along a four-level description model from basic conceptual answers to interpretive and analytical understanding. Additionally, we outline our process for synthetically generating our single and multi-panel (charts in which there are multiple panels present) plots and testing the data with various LMMs using quantitative analysis. In general, we find LMMs become less accurate with increasing figure complexity as the number of panels, number of plot elements, and difficulty of questions increase. Finally, because LMM accuracy is expected to change rapidly, our analysis also considers cost relative to performance as a metric to inform future user decisions. We provide the detailed question-answer pairs used in this analysis, along with the code used to generate random plots, question-answer pairs and detailed bounding-box metadata for chart elements. The code and data are available at https://github.com/ReadingTimeMachine/LLM_VQA_MultiPanel and https://huggingface.co/datasets/ReadingTimeMachine/visual\_qa\_multipanel

Deterministic flows can contract Wasserstein distances between probability laws without being contractive in point-state space. We develop a differential contraction theory for pushforward flows on forward-invariant subsets of the 2-Wasserstein space that quotients out the gauge freedom of the continuity equation. Weighted Helmholtz-Hodge decomposition and the Wasserstein contingent cone define the weighted horizontal tangent cone (whtc). Requiring the contraction quadratic form to be nonpositive only on the whtc yields the Intrinsic Wasserstein Contraction Condition (IWCC), which generalizes the classical contraction LMI and recovers it on Dirac measures. An almost-everywhere Euler approximation based on the superposition principle and Banach-valued Lebesgue differentiation theorem proves that IWCC is sufficient for intrinsic Wasserstein contraction. Law contraction without point-state contraction is illustrated by a rotationally symmetric shear example. Finally, a new comparison lemma yields an incremental Wasserstein ISS theorem; for Brockett's integrator, we prove exponential stabilization of a class of axisymmetric laws to the uniform law on a circle via a smooth feedback.

Terminal agents solve tasks in software engineering and scientific research by executing commands in a terminal. Training them via reinforcement learning (RL) requires diverse and verifiable environments, but these environments are still mostly designed by human researchers, so their supply grows only as fast as researchers can work. We introduce AutoEnvScaling, which treats environment design as a terminal task and uses agents to automate the data flywheel. A proposer agent builds new RL environments from the rollouts of the solver being trained, and the solver's new rollouts guide the next round. By turning environment generation into a terminal task, AutoEnvScaling increases the valid-environment rate by up to 25 times compared to prompt-only generation, at up to 75% lower cost per valid environment. We show that training Qwen3.5-35B-A3B on AutoEnvScaling environments from GPT-5.6-sol improves its Terminal-Bench 2.1 score by 4.3 points over the matched Tmax baseline. The broader goal of AutoEnvScaling is to enable recursive self-improvement (RSI): by making both environment generation and solving terminal tasks, it lets one model co-evolve in both roles. We show that Qwen3.6-35B-A3B, acting in both roles, gains 12.2 points on Terminal-Bench 2.1 and raises its Terminal-Bench 4.0 score from 1.6 to 3.2. The pipeline also generalizes across agentic domains, generating environments for software engineering, computer use, professional work, and GPU kernel programming. Overall, these results show a path toward agents building their own training environments, reducing human reliance and taking a core step toward RSI.

Diffusion language models (DLMs) offer parallel sampling for faster generation than autoregressive models, an increasingly important advantage as inference becomes a dominant computational workload. However, DLMs degrade quickly with fewer sampling steps, process the full sequence at every step, and require many passes over the dataset to reach good quality. We introduce Next-Concept Modeling (NCM), a two-stage, coarse-to-fine framework: a multiscale residual VQ-VAE encodes text into a hierarchy of discrete codes, and a transformer generates one level at a time, predicting all codes within a level in parallel. At matched parameter scales, NCM achieves better generative perplexity than the strongest DLM baseline on LM1B while running $21\times$ faster. On OpenWebText, NCM is competitive with the strongest DLM while running $10\times$ faster; on XSum, it matches the strongest DLM's ROUGE-L while being $31\times$ faster. On the reasoning task of Sudoku, NCM outperforms autoregressive and diffusion baselines across all difficulty levels, using fewer function evaluations. On the reasoning task of Sudoku, NCM outperforms autoregressive and diffusion baselines across all difficulty levels, using fewer function evaluations.