Papers

When post-training large language models on tasks with semi-verifiable rewards, there are many factors (training steps, base model size, training order, data quality, verifier accuracy, etc.) that practitioners must contend with to maximize model performance. Yet, it remains unclear how well verifier agreement predicts post-training performance on such tasks. In this paper, we explore this question with over 11k H100 GPU-hours, across HealthBench and PRBench tasks in medical, legal, and finance domains. Across the tested domains, Qwen3 trainees (1.7B-8B on HealthBench; 8B on PRBench), evaluation splits, and frontier LLM reference judges (which we call golden verifiers), higher verifier agreement does not consistently identify the best training verifier. Expensive verifiers need not outperform inexpensive ones, and open-weight Gemma verifiers produce strong training outcomes. We compare two low-cost choices retrospectively – a cost-reducing choice and a balanced choice – with estimated grading cost reductions of 98.8%-99.7% relative to the golden grading protocols and average post-training score gaps of 1-3 points from the best evaluated training verifier. These averages include larger losses in individual settings; they do not establish that verifier choices are interchangeable.

Improving language model efficiency requires not only cheaper operations but also eliminating redundant computation. We argue that effective modeling requires information accessibility and computational depth, but not repeated historical retrieval across layers. Guided by this principle, we introduce the Global State Model (GSM), a causal encoder–decoder architecture that separates historical information acquisition from deep computation. The encoder progressively aggregates long-range information through multi-stage sparse retrieval into a fixed-size shared state, while decoder layers perform successive transformations over this state without repeated historical access. This design makes decoder attention and cache costs independent of history length, while dependency-aware pruning further reduces prefill computation. Experiments across multiple model scales show that GSM achieves downstream performance comparable to matched DeepSeek-V4.1-Flash-based baselines, with improvements at some scales, while preserving long-context retrieval capabilities. GSM also improves training and inference throughput and reduces cache overhead. These results demonstrate that eliminating architectural redundancy offers an effective path toward efficient language modeling without sacrificing information accessibility or computational depth.

World models learn internal representations of environment dynamics to predict future states, enabling agents to optimize action plans without physical interactions. However, developing world models that genuinely internalize underlying causal physical laws to explicitly reason about action preconditions and subsequent state transitions remains an open challenge. In this paper, we propose VIDEAS, a data distillation framework that transforms continuous physical dynamics from operational videos into explicit action semantics for foundation models. Specifically, it deconstructs visual demonstrations into discrete action trajectories and utilizes advanced vision-language models (VLMs) to extract structured knowledge encapsulating action preconditions and effects. To ensure physical consistency, we introduce a prior-guided trajectory simulation mechanism grounded within a text-based environment to rigorously validate the extracted knowledge. Notably, we incorporate negative trajectories to enrich knowledge completeness and enhance data diversity to mitigate cognitive bias. Furthermore, we present VIDEAS-WM, an 8B/9B-parameter suite of language-based world models trained on 34K high-quality samples derived from AgiBot-World dataset. Extensive experiments demonstrate that VIDEAS-WM establishes state-of-the-art performance in high-level embodied action semantic reasoning, exhibiting profound physical understanding and robust generalization across unseen scenarios.

Supervised fine-tuning (SFT) learns most aggressively from tokens that the model deems least likely. This helps acquire new behaviors, but also amplifies noisy or conflicting supervision and can overwrite useful pretrained knowledge. Through a unified policy-loss view, we revisit existing token-reweighting methods and show that they assign nonnegative coefficients to demonstrated tokens. Consequently, they can suppress or amplify supervised updates, but cannot reverse harmful features once learned. Moreover, larger training weights do not amount to feature extrapolation, since they change the optimization trajectory rather than scale a fixed SFT direction. We argue that reversal and extrapolation require a stable reference frame defined by a fixed SFT delta. Motivated by this, we propose SCALE (Selective Control of Adaptation via Local Entropy), an entropy-guided adaptation-strength-control method that freezes the pretrained model and the SFT delta and learns bounded token- and module-specific gates by minimizing predictive entropy alone. These gates suppress, reverse, or extrapolate frozen SFT features according to their alignment with entropy reduction. Across Qwen2.5-Math-1.5B, Qwen2.5-Math-7B, and Qwen3-4B-Base, SCALE achieves mathematical-reasoning averages of 37.84, 43.60, and 36.57, exceeding the strongest corresponding baselines while remaining competitive on general-retention benchmarks. It also attains the best average code-generation performance across HumanEval, HumanEval+, and MBPP for all three models. These results suggest that effective SFT correction can benefit from controlling how already learned residuals are used, rather than only modifying how they are learned.

Omnidirectional or 360 cameras provide embodied AI agents with a holistic, wide field-of-view (FoV) view of their surroundings, motivating the use of Multi-modal Large Language Models (MLLMs) for omnidirectional spatial reasoning. However, most MLLMs are trained on conventional 2D perspective images and struggle with the severe distortions and wrap-around discontinuities induced by spherical geometry. Enabling them to generalize to non-Euclidean 3D spaces without retraining therefore remains challenging. We propose SphMind, a training-free, plug-and-play framework that decouples semantic perception from geometric reasoning. Rather than requiring MLLMs to learn spherical geometry internally, SphMind preserves their semantic capabilities while handling geometry externally. We introduce a Spherical Harmonics-based Spatial Graph (SHSG) that models spatial relationships through equivariant transformations on the sphere, together with Inference-Time Geometric Grounding (IGG), a model-agnostic closed-loop optimization process that aligns MLLM representations with spherical geometric constraints during inference. Experiments on three benchmarks show that SphMind achieves over 21.4% average improvement in directional reasoning on MP3D and Stanford2D-3D, outperforms prompt-engineering baselines by 8.7% on the real-world ODI-Bench, and improves rotational invariance by 5.9% under panorama rotations, without additional training or dataset-specific tuning. In-the-wild evaluations further show that SphMind resolves directional reasoning queries that baseline vision-language models fail to answer correctly.

Runtime monitoring traditionally evaluates a specification that is fixed before execution or externally modified when requirements change. In learning-enabled and data-intensive systems, however, the temporal relationships represented by a specification may themselves evolve. Allowing an AI component to directly replace a formal specification is unsafe: it may overfit transient behavior, weaken protected requirements, or activate statistically unsupported revisions. We introduce an intersymbolic architecture in which an untrusted AI proposer suggests temporal specification revisions and a symbolic governor controls their activation. Two results organize the framework. First, origin-version semantics makes the outcome of each obligation invariant to later revisions. Second, aggregate certification can conceal systematic failures on protected triggers; simultaneous aggregate and core-conditional post-selection certification controls both targets. A structural invariant preserves designer-protected components, and a proposer-independent lifetime error bound supports repeated activation decisions. The statistical bound concerns the predictable means of completed certification samples; interpreting it as future operational validity requires an additional stability assumption. Controlled synthetic experiments use a frozen supervised AI proposer to illustrate the masked-core failure at one decision and across repeated governed revisions. The proposer is a supervised regressor trained offline on synthetic tasks and frozen before use; it ranks candidates by predicted aggregate margin and never observes the protected-trigger success rate, so the masked-core failure arises from optimising the aggregate rather than from an adversary constructed by hand.

Multimodal large language models answer audio questions, but how they represent auditory semantics and use them in decisions remains unclear, limiting our understanding of response formation. We study dog barking in Qwen2.5-Omni-7B using Jacobian lens (J-lens) readout and directional interventions. We define the dog direction as a J-lens-derived hidden-state vector associated with dog; adding or removing its component modulates dog-related information. We find this information decodable without dog/bark prompt cues or animal-identification requirements. Directional interventions change response tendencies and some final answers, with effects concentrated in late-layer states immediately before generation across species classification, vocalization classification, and sound description. The dog direction shows no comparable advantage over controls in animal/other classification. These results provide causal-intervention evidence that the dog direction affects output scores in a task-dependent manner, most consistently at L22 and L24 immediately before generation.

Many dense-prediction benchmarks evaluate rare events by pooling prediction and target over spatial blocks, thresholding each, and scoring the contingency table. At a fixed rare operating point, the max-pooled Critical Success Index (CSI) confounds spatial discrimination with amplitude calibration: sharp observations promote many blocks above threshold, while attenuated predictions from squared-error regression leave the same blocks below it. We repurpose classical monotone calibration as a symmetric audit: a post-hoc transform fitted on held-out data and applied separately to each system. The transform cannot reverse pixel ordering, so any contrast it reproduces cannot establish improved spatial ranking. On SEVIR, two released checkpoints of one architecture differ by -29.5% in extreme-threshold CSI before the control and by +5.3% after it. Across 450 pairwise contrasts among 6 systems, the difference in pooled frequency-bias deviation is associated with how far the CSI contrast moves under the control (r = +0.796), and 51 contrasts reverse sign. At CasCast's published extreme-event operating point, the cascade-over-backbone CSI gap falls from 0.1601 to 0.0339, a 78.8% reduction; the remaining gap stays positive. The effect persists when the transform is fitted on a window before the test period, and calibration also reveals advantages hidden by a better-calibrated baseline. On geostationary infrared imagery the relative gain grows as events become rarer, crowd counting reproduces the bias-gain relationship under patch-sum pooling, and semantic segmentation, where frequency bias is already near one, shows little average change. The confound therefore requires both a fixed operating point and a training regime that leaves the output miscalibrated there. We recommend reporting pooled frequency bias and a symmetric held-out FreeKnob Audit alongside rare-event pool-and-threshold scores.

Weighted split conformal prediction reweights calibration scores by the likelihood ratio between the test and training covariate distributions and guarantees marginal coverage under covariate shift. We study its coverage conditional on the calibration data. An elementary argument, based on a single concentration inequality at a fixed population quantile, gives explicit training-conditional bounds without unspecified constants, and shows that the relevant scale is not the supremum of the likelihood ratio but a variance proxy built from the chi-squared divergence of the shift and from the average of the ratio over the part of the test population, of probability equal to the miscoverage level, where it is largest. A two-point lower bound shows that the root-m rate and the chi-squared contribution are intrinsic to the shift. Run at an explicitly inflated level, the weighted quantile becomes a deterministic PAC prediction set. We compare it with randomized rejection sampling and with importance-weighted learn-then-test and, through a certified choice of a clipping level for the likelihood ratio, map the regime in which each gives the narrower valid set. The analysis extends to estimated likelihood ratios and to tail functionals estimated from an unlabeled source sample, which yields a fully finite-sample certificate.

On-policy distillation (OPD) trains a student model on its own trajectories using dense token-level feedback from a stronger teacher model. Since each update is conditioned on the reasoning prefix already generated by the student, the prefix also shapes how effectively teacher feedback is converted into learning. We find that shared prefixes can lead to weak token-level updates, a phenomenon we call Prefix-Induced Supervision Attenuation (PISA). This attenuation arises in two common cases. (i) High student confidence can weaken corrective gradients even when the teacher disagrees. (ii) Tokens that rely on earlier reasoning can receive learning signals as weak as those for simple local continuations. To solve this problem, we propose Trajectory Dropout, a simple training-time intervention that exposes these weakened signals. The student first performs a standard full-context rollout to generate a complete trajectory. During training, we randomly drop a certain proportion of the student's reasoning trajectory, while the teacher continues to observe the complete trajectory for token-level supervision. This intervention strengthens corrections for overconfident predictions and introduces additional supervision at prefix-sensitive positions. Trajectory Dropout consistently improves average performance across teacher–student model pairs of different scales and six mathematical reasoning benchmarks, while also yielding gains on two out-of-domain benchmarks. It can also be flexibly integrated into existing OPD variants with negligible computational overhead, further improving their performance. These results demonstrate that Trajectory Dropout provides a simple mechanism for strengthening token-level supervision across model scales and OPD objectives.

We present DAOCP, a dual active set solver for linear quadratic optimal control problems with stage-wise equality and inequality constraints. Active set methods are leading Model Predictive Control benchmarks for full-body robotics, but existing solvers operate on dense QPs, while typical problem dimensions favor methods that exploit the optimal control structure. DAOCP combines the warm-starting capabilities of active set algorithms with the better computational scaling of structure exploiting solvers, by relying on a generalization of the relationship between the Riccati recursion and the Cholesky factorization of the condensed Hessian. This result enables dual active set iterations to operate directly on the original optimal control problem through recursive computations, without explicitly forming the condensed quadratic program. On robotics benchmarks, DAOCP is the fastest of four solvers in four of five scenarios, cutting average solve time on a 58-state Atlas model by 9$\times$ relative to state of the art solvers DAQP and HPIPM.

Extracting who is where, on which team, wearing which number from a broadcast frame is typically done by stitching a detector, an OCR engine, and classifiers together – and the stitching step swaps identities under occlusion. We make association native instead: a 0.77B vision-language model (Florence-2) is fine-tuned to emit all per-person attributes as one grammar-constrained sequence, with each attribute generated inside its owner's block. Output is therefore schema-valid on every frame by construction, and no post-hoc binding step exists to attach a correctly read number to the wrong player: residual misassociation is pure perception error, $\approx4\times$ rarer than zero-shot-prompted frontier APIs' (0.057 vs. 0.21-0.24). On a frozen multi-sport test set, this single pass reaches 0.95 detection F1 (APIs: 0.65-0.75). A single extra forward pass yields a per-field confidence that supports a reject option (jersey precision $0.71\rightarrow0.96$ at half coverage) and routes a training-free zoom-and-re-read for small players. Surprisingly, once the grammar is learned, further parameter-efficient tuning yields no measurable gain under the adaptation configurations we test; the identical recipe on WIDER-Attribute reaches 93.1 mAP given-box, yields the first detection-coupled end-to-end results under its standard test protocol (84.5 mAP), and reproduces the same tuning result. In this regime, the gains live in the structure, not in added weights.

Historical manuscript illustrations preserve rich visual evidence of past cultures. They depict people, animals, plants, diagrams, music notations, and decorative forms. Although large digitization projects have made many manuscripts available online, the material itself remains difficult to explore at scale. Extraction systems can find illustrations on manuscript pages, but without meaningful categories, large collections remain hard to search and explore. We address this gap by introducing a manually labeled dataset of 15,000 illustrations from manuscripts dating back hundreds of years across 22 categories, and evaluating modern vision models for image classification on this task. The problem is challenging due to stylistic diversity, degradation, and semantic ambiguity, with many images that fit more than one category. We compare fine-tuned CNN and Transformer-based classifiers, zero-shot CLIP, embedding-based classifiers, and direct vision-language models. Results show that fine-tuned image classifiers perform best overall, with ConvNeXt reaching 88.9% accuracy and 81.3% macro-F1. Using CLIP embeddings with XGBoost provides a strong alternative. In contrast, zero-shot CLIP and direct vision-language classification perform substantially worse, highlighting the limits of general-purpose models in this domain. Beyond overall performance, the analysis reveals which categories are visually separable and where errors reflect genuine semantic overlap, suggesting that some limitations arise from the taxonomy itself.

Guideline documents are published as unstructured PDFs whose evidence is locked in visual structures—tables, flowcharts, and graded recommendations—that standard retrieval pipelines flatten into fixed-size text chunks. We cast evidence access as a document image analysis problem: parse each page image into typed structural elements, then retrieve structure-aware evidence units that follow the document's own layout (sections, table rows, flowchart paths, graded recommendations), each keeping its structural context so a result points to a specific element rather than a page. On 26 clinical practice guidelines from 9 sources (3,619 pages, Chinese and English) with 199 evidence queries, structure-aware units rank the gold element first under BM25, dense, and hybrid retrieval (hybrid Element Hit@1 of 0.382), with a significant element-level ranking gain over per-element OCR text (MRR_e +0.107, p=0.002; the Hit@5 gain is directional, p=0.17), while matching page-level recall (Page Hit@5 0.879 vs. 0.889, p=0.75) at 3.8x less context and clearly outperforming a ColPali visual-RAG baseline (PH@5 0.497).

Security operations centers receive far more alerts than analysts can investigate, and organizations that cannot send their telemetry to hosted models must automate triage with small open-weight LLMs on their own hardware. Current LLM agents leave the investigation procedure to the model, and small local models fail at it: they probe without converging, never commit to a verdict, or dismiss real attacks. In this paper, we present HESP, a controller that holds the investigation procedure outside the model. HESP keeps a ledger of competing explanations, selects read-only probes by expected information gain per cost, accepts only verdicts backed by current evidence, can end an investigation itself, and journals every prediction before its observation. We evaluated HESP in four pre-registered studies with five open-weight models from two families (7B to 72B), totalling 7,272 audited episodes in a controlled triage environment. With likelihood tables counted from LLM-free runs, HESP lifts Qwen2.5-7B from 0.125 to 1.000 verified completion, matching oracle tables. The information-gain ranking adds +0.26 to +0.35 on every model that concludes, and a controller-side stop lifts Llama-3.1-8B, which never concludes on its own, from 0 to 0.917. What to probe and when to stop are therefore separate failures, and different small models exhibit different ones. Because HESP and its planner run entirely on local hardware, it suits environments where telemetry cannot leave the premises. We release all code, protocols, and episode journals at https://github.com/lzwhehe/HESP.

Concept erasure aims to suppress undesirable knowledge in text-to-image generative models. However, existing robustness evaluations typically rely on relearning attacks tailored to specific model architectures. We study concept reactivation across two substantially different generative paradigms: noise-prediction U-Nets and flow-matching Transformers. We introduce Concept Score Relearning (CSR), a unified parameter-level framework that reactivates erased concepts by optimizing each model within its native prediction space. CSR requires no external target-concept image dataset and applies the same concept-directed objective to both U-Net-based Stable Diffusion and Transformer-based FLUX. Experiments across diverse concepts and multiple erasure methods demonstrate consistent concept reactivation across both architectures, highlighting the cross-architecture applicability of CSR and the persistent recoverability of apparently erased concepts. For strict nudity, CSR reaches average ASRs of 50.47% on FLUX and 40.29% on Stable Diffusion, consistently ranking first across all evaluated safety settings.

A nascent family of methods that forgoes the policy gradient and reweights a supervised regression instead has garnered momentum in reinforcement learning for diffusion and flow models. DiffusionNFT, FlowAWR, and RAM are representative regimes with contrasting motivations. It is yet opaque what, if anything, they share. We substantiate that each is the solution of one divergence-constrained reward-maximization problem, and they are differentiated only by the convex generator that defines the constraint. Under the unified modeling framework, we unravel the relaxations that prior art made during building the advantage-embedded regression target: approximating the KKT condition and posterior normalizer for the linear and exponential tilt shapes DiffusionNFT and FlowAWR respectively, while preserving the exact sparsemax projection onto the probability simplex for linear tilt leads to another superior model type in this work. Beyond the theoretical underpinnings, we further empirically investigate the design space and shed light on the training recipe for regression-style diffusion RL. Retaining the merits discovered during our exploration gives rise to DiffusionRFT, our paradigm that converges faster, trains more stably, and attains the top performance.

Full-duplex spoken dialogue models can listen and speak simultaneously like the real-time dynamics of human conversation. For natural dialogue, the ability to search for external information in real-time is also an important capability. Many models remain trapped in parametric knowledge, leaving them unable to access real-time information and tool execution. Furthermore, even when Large Language Models (LLM) retrieve information, many duplex speech models process it within a compressed latent space rather than in its raw text form, which can lead to information loss from compression. To address this issue, we propose Context Spanning, a framework for information injection between a full-duplex speech model and an external LLM backend via real-time chunked prefill. The injected frame is encoded in a single forward pass inside the real-time frame budget. It feeds the retrieved information to the speech model as-is, enabling it to reason over the information independently and generate responses. With this approach, our model achieves high performance on Full-Duplex benchmarks and strong results on Question Answering tasks, demonstrating its conversation potential. Context Spanning shows that external information can be injected directly into a duplex speech model, introducing a new simple and powerful mechanism for duplex systems.

Claims paired with AI-generated images are a rapidly growing form of misinformation. Existing automated fact-checking (AFC) methods mainly treat this as a provenance problem, detecting low-level synthesis artifacts to decide whether an image is AI-generated. However, such methods do not verify what human fact-checkers often check: whether an image's content is consistent with the context implied by its accompanying claim. To address this gap, we introduce MIC (Multimodal Inconsistency Checking), an AFC framework that assists human fact-checkers by detecting AI-generated multimodal misinformation and explaining inconsistencies using world knowledge. MIC first uses supervised fine-tuning (SFT) for task adaptation and then applies Group Relative Policy Optimization (GRPO) to directly optimize component-level verifiable rewards for verdict prediction, inconsistency type classification, visual evidence description, and world-knowledge explanation. We further introduce MIC-Bench, a benchmark comprising 8,812 image-claim instances derived from 4,406 claims, where each claim is paired with an authentic image and an AI-generated counterpart that introduces a controlled contextual inconsistency. Compared with SFT alone, GRPO further improves Macro-F1 by 4.67 and 4.11 points in the in-distribution and out-of-distribution settings, respectively, while also improving the semantic similarity of visual evidence descriptions and world-knowledge explanations to reference annotations. Our code and data are available at https://github.com/UKPLab/arxiv2026-mic.

Objective cognitive assessment from neural signals supports neurorehabilitation, but individual-level prediction from task-based fMRI (tfMRI) remains difficult because neural features coexist with substantial demographic and scanner-related variation. We present the Multi-task Activation and Contrast Network (MAC-Net), a covariate-aware deep learning framework for modeling individual cognitive function from regional tfMRI. By isolating tfMRI features into a dedicated neural pathway and restricting participant variables to a terminal late-fusion pathway, MAC-Net prevents dominant covariates from suppressing high-dimensional clinical representations during feature learning. Evaluating baseline data from 6,500 Adolescent Brain Cognitive Development Study participants under family-aware cross-validation, MAC-Net was benchmarked against linear models, random forests, and alternative deep architectures. The N-back plus Monetary Incentive Delay configuration achieved $R^{2}$ values of 0.174, 0.238, and 0.277 for fluid, crystallized, and total cognition, outperforming covariate-only baselines (0.178) and alternative deep models (0.217). N-back was the most informative paradigm, whereas incorporating the Stop Signal Task marginally degraded performance. Feature attributions via Integrated Gradients, DeepLIFT, and Input Gradient were highly concordant, localizing working-memory-related frontal, parietal, and cingulate regions. These findings demonstrate that covariate-aware multi-task modeling yields reproducible cognitive-function estimations, establishing a robust neural engineering framework for clinical translation.

As large language models advance, AI agents are moving beyond isolated, domain-specific tasks toward long-horizon, cross-domain workflows. This transition exposes two challenges: increasing harness complexity makes manual design difficult to scale, while tighter coupling to specific domains limits the generality of a single harness. The central question thus shifts from how to engineer a stronger harness for one domain to how to autonomously construct specialized harnesses, improve them through experience, and orchestrate them across domains. We introduce Raven, The Harness of Harnesses, an open-source multi-agent ecosystem that automatically constructs and evolves modular harnesses for specific models and domains, treating each executable model–harness pair as a composable unit of intelligence. To support an All-Domain Collaboration Network, its Host Agent decomposes goals, matches subtasks to specialized agents, coordinates execution dependencies, and integrates results, while a host archive and EverOS preserve experience across tasks and Skill Forge makes that experience available as reusable procedures. Our theory establishes sufficient conditions for such composition to expand reliable task coverage beyond that of the available individual agents under a shared resource budget. On complex and long-horizon tasks, Raven significantly outperforms the state-of-the-art agent systems, pushing the frontier of composable agentic intelligence.

Formal specifications are critical for program verification, comprehension, and maintenance. However, manually writing them is costly and difficult to scale. Recent studies have shown that Large Language Models (LLMs) are promising for automated specification generation, but existing methods suffer from quality issues. We analyze a state-of-the-art approach and find that at least 34.6% of successfully verified specifications actually fail to meaningfully capture the program's distinct behavior, which is a quality issue not captured by metrics that only measure verification success. We further found that a major factor contributing to such hidden quality issues stems from the design of existing methods: these methods treat a program as a single unit, resulting in overly general, coarse-grained constraints. To this end, we introduce Path2Spec, a divide-and-conquer framework that addresses these limitations through systematic path-based reasoning. Path2Spec leverages LLMs to extract all execution paths from an input program, generates path-specific specifications for each, and merges them into a comprehensive overall specification. For complex programs where path-based generation struggles, Path2Spec employs a decompose-then-retry strategy that recursively breaks a program into smaller subprograms based on logical branches, generates specifications for each, and merges them back. We evaluate Path2Spec on two public benchmarks: SG-Bench (120 programs) and SV-COMP (265 programs). Results show that Path2Spec can outperform the state-of-the-art baseline SpecGen: 87.5% versus 66.7% on SG-Bench, and 83.0% versus 44.2% on SV-COMP. Human evaluation further validates that Path2Spec generates higher-quality specifications with precise semantic alignment to the code.

Model merging integrates the capabilities of multiple experts without joint retraining, but standard expert training optimizes task loss alone and does not guarantee good performance after merging. Merge-aware training (MAT) aims to improve merged performance, but existing methods do not fully account for common merging operations and add training cost. We observe that, from an expert's perspective, common merging methods can be described by three operations: Scale reweights its own update, Mask removes selected coordinates, and Perturb adds updates from other experts. Based on this view, we introduce SMAT (Simple MAT), which jointly optimizes expert loss and expected loss at simulated merged parameters generated by sampling scaling coefficients, masks, and additive noise. We further introduce periodic scheduling, kernel fusion, and parameter storage switching to make SMAT efficient, with one forward and one backward pass per step. Across four language and vision-language backbones, SMAT improves the mean score across five merging methods by 1.07-2.16 points over the strongest baseline for each backbone, with less than 2% training-time overhead over standard fine-tuning.

Attention provides direct access to past representations, but retaining an ever-growing history is costly. Recurrent models bound persistent state, yet must preserve selected information while processing subsequent inputs. We introduce SchemaMem, an attention-based recurrent memory architecture combining chunk-local attention with a persistent, schema-indexed phase state. Learned schema embeddings provide a shared representational reference for reading and writing. Reads use the current state, whereas writes use the layer input and static schema embeddings, excluding direct feedback from that layer's own state. Chunk-boundary commits aggregate bounded phase increments through forward computation. The same parameters also support full-history attention training before and during recurrent training. We studied selective updates, preservation, and delayed retrieval in a controlled address–value task, comparing three-layer models with approximately matched parameter counts and persistent-state dimensions. Across nine address/value settings and three training seeds, SchemaMem has higher mean written-value retention at four times the maximum training delay than both baselines, which are trained toward a higher in-range accuracy target. Updated-value recovery favors SchemaMem in all nine settings against Mamba-3 and seven against Gated DeltaNet. Defaults consistently favor Gated DeltaNet over SchemaMem at that delay, and SchemaMem requires substantially more optimization steps. These results identify a promising retention–optimization trade-off in schema-indexed recurrence.

Text-to-Audio (T2A) retrievers are typically evaluated with caption style queries, but the same user intent can be expressed in many forms. We introduce CORA (Caption-Offset Retrieval for Audio), a caption anchored diagnostic protocol that rewrites each source caption into five intent preserving forms (Command, Question, Indirect, Key phrase, and Statement) while fixing the target audio. By tracking the same target across query forms, CORA defines RankDrop, a metric revealing failures hidden by Recall@k. Using Pearson's correlation coefficient r, we find that RankDrop is weakly associated with raw text space movement (r=0.084), but strongly associated with Target Alignment Loss and Target Boundary Margin Degradation (r=0.508 and r=0.615). The same pattern appears in OEA retrievers, where RankDrop is better explained by boundary degradation (r=0.472/0.478) than by query movement (r=0.084/0.046). Overall, these results suggest that robust T2A retrieval requires preserving the target's boundary advantage over competing audio under reformulation.

Manual dense annotation remains a major obstacle to deploying semantic segmentation models in new driving environments. Active domain adaptation (ADA) seeks label-efficient transfer by annotating only a selected portion of the target domain. Existing ADA methods commonly implement this process through multiple rounds of acquisition, annotation, and retraining. We study a practical one-shot image-level setting that selects and densely annotates a fixed target subset in a single round, followed by uninterrupted adaptation. Within this setting, we develop Target-Calibrated Active Domain Adaptation (TC-ADA) as a joint design of complete-image acquisition and target-calibrated adaptation. Stage~1 uses visual representations from a vision foundation model (VFM) together with semantic predictions from a fixed unsupervised domain adaptation model to select representative and informative target images without target annotations. Stage~2 jointly uses labeled source data, labeled target data, and the remaining unlabeled target data, while calibrating source and target supervision under limited target labels. Extensive experiments across five synthetic-to-real and real-to-real driving transfers show consistent improvements over representative ADA baselines. With only 23 to 46 labeled target images on four transfers and 140 on Mapillary, TC-ADA stays within 1.9 mean intersection over union (mIoU) points of target-only full supervision. Code will be available at https://github.com/ywher/TC-ADA.

We release MoGround, a vision-language dataset spanning four visual domains in which the answer to every question is guaranteed to be available from exactly one modality. This guarantee enables us to measure modality distraction, the failure in which a model answers a question correctly from one modality alone and then flips to a wrong answer once irrelevant content from the other modality is added. Existing probes rarely establish single-modality answerability this way, making it hard to isolate distraction in the first place. Across seven open-source VLMs, we find that modality distraction is not universal but model-dependent. The weaker-grounded modality is the more distracted one (r = +0.86), and distraction scales inversely with grounding strength (r = -0.90). The single-modality guarantee also enables a mitigation method that needs to distinguish between relevant and irrelevant context. Trained on one split of MoGround alone, a weight-space robustness vector reduces distraction on all seven models by 9% to 51%, at a cost of only 0.1 average points of accuracy on standard multimodal tasks.

Production scheduling requires realistic models that reflect operational constraints and efficient methods that balance competing goals. Putting these methods into use also requires data integration, model adaptation and specialist expertise. We present APEX, an extensible production scheduling framework built around a general model and hybrid multiobjective search. Agent assistance supports both scheduling and model refinement: agents prepare data and explore scenarios in natural language, while coding agents help implement and test new constraints and objectives. Shared construction and checking procedures connect these adaptations to the scheduling core. We benchmark eight APEX configurations against NSGA-II, SPEA2, MOEA/D and SMS-EMOA on 69 public job-shop, flexible job-shop and permutation flow-shop instances, assessing workload completion time (makespan), total job flowtime and computation time. Hybrid configurations achieve the best aggregate solution quality, although the leading method depends on the problem class and objective. A separate synthetic workflow study uses OpenAI's GPT-6-astra as an interaction layer between the human planner and the algorithmic core, testing rule additions, plan and objective changes, and what-if comparisons. All 24 sessions completed the requested changes and passed independent checks of saved models and schedules. A separate coding evaluation produced six native implementations of an additional objective or hard constraint through predefined extension hooks. All passed independent checks without modifying the core.

Coding agents now increasingly rely on execution to validate their solutions, making the construction of reliable execution environments a critical enabling capability. However, repository environment construction is challenging because execution requirements are fragmented across repository artifacts and may only become apparent during execution. Existing agent-based approaches address this problem through iterative interaction, but information about the current construction state, including discovered requirements, satisfied and unresolved prerequisites, and their dependencies, can remain distributed across the interaction history. We present Graph2Env, an agent-based approach centered on DepGraph, a typed dependency graph that explicitly represents the environment requirements needed for repository execution, their dependency relations, and their states. Graph2Env uses DepGraph to guide environment construction and continuously refines it with execution feedback, while persisting successful repairs into a replayable construction procedure. The resulting artifacts are then applied in a fresh environment to verify that the constructed environment can be reproduced. We evaluate Graph2Env on a benchmark of 200 Python repositories drawn from RATBench and EnvBench, against a static dependency-inference baseline (pipreqs), three specialized environment-construction systems (Repo2Run, RAT, and SetupX), and two general-purpose coding agents (SWE-agent and Claude Code). Graph2Env achieves an 81.0% Environment Build Success Rate (EBSR) and a 59.3% Environment Setup Success Rate (ESSR), outperforming the strongest baseline by 9.5 and 9.0 percentage points, respectively.

Wearable actigraphy offers a scalable, ecologically valid alternative to episodic clinical assessment. However, predicting continuous adolescent crystallized intelligence ($G_c$) from such traces remains challenging due to irregular device adherence and complex behavioral-environmental interactions. We address this using daily summary data derived from 21-day Fitbit records of 6,091 adolescents in the Adolescent Brain Cognitive Development Study (Release 5.1). We propose SATURN, a Sleep-Activity Temporal Unified Regression Network. It represents participants as 21-node temporal graphs encoding daily behaviors and temporal adjacency. To prevent imputation artifacts, invalid-day edges are dynamically pruned during forward passes. Node embeddings are refined via residual GATv2 layers, aggregated through masked attention pooling, and fused with sociodemographic covariates. Under family-controlled, age-sex-BMI-stratified cross-validation, SATURN achieves $R^2 = 0.2783 \pm 0.0127$, consistently improving upon flattened machine learning (Gradient Boosting, $R^2 = 0.2372$) and sequential deep learning (BiLSTM, $R^2 = 0.2688$) baselines. Explainability analyses identify light activity, metabolic equivalents, and sleep duration as dominant predictors, while Monte Carlo dropout and subgroup analyses confirm equitable performance across sociodemographic strata. Ultimately, SATURN establishes a rigorous computational framework for digital cognitive phenotyping, offering a scalable pathway to complement traditional assessments by highlighting macro-level behavioral anomalies.