Can frontier general-purpose agents understand client intent, explore creative directions, execute revisions, and prepare production deliverables? We introduce DRAFT-Bench (Design Reliability Across Feedback and Tasks), a lifecycle-oriented evaluation of requirement elicitation, creative exploration, constrained editing, iterative revision, and production handoff. Its Requirement-to-Artifact framework links evolving requirements, protected properties, artifact versions, and delivery obligations. We compare native Codex/Astra and Claude Code across 35 design briefs under 43 conditions, alongside a separate 40-scenario RFQ track. Astra completes 43/43 design conditions and Claude 35/43. All scored core runs receive full structural credit, yet on the 28 core conditions scored for both systems, semantic checks flag violations in 11 for Astra and 5 for Claude; these are model judgments, not expert-confirmed failure rates. RFQ coverage is high, but success ordering changes across cohorts. SKU expansion, revision dependencies, and White ink files expose ambiguity among unmet obligations, missing inputs, and evaluator interpretations. These findings separate execution, requirement compliance, and professional acceptance. Unpaired tracks and different tools preclude claims of RFQ-to-delivery success, factory acceptance, or a pure model ranking.
Probabilistic load forecasts for AI data centers need to represent both uncertainty under changing workloads and temporal dependence to support grid operations. This paper develops an open, grid-oriented benchmark using measured GPU power from a 448-GPU AI cluster. Forecasts are evaluated through four stages: probabilistic prediction, online uncertainty calibration, temporal scenario generation, and horizon-wide reserve sizing. Following a change in cluster utilization, a quantile LightGBM forecast achieves only 63.4% coverage for a nominal 90% interval, while online regime-conditioned calibration increases coverage to 87.3%. Even with well-calibrated marginal forecasts, independent sampling produces unrealistic ramp distributions and excessive reserve procurement. Reconstructing temporal dependence from historical forecast errors reduces the discrepancy in ramp statistics and lowers realized reserve cost by 7.4-10.6% across three forecasting tasks. These findings show that marginal forecast quality alone is insufficient for operational decisions and motivate joint evaluation of calibration, temporal structure, and decision outcomes.
Recent progress in embodied robotics is bringing robots closer to deployment in homes, as vision-language models (VLMs) are demonstrating significant advances in reasoning, making VLM-driven robots promising home assistant agents. These agents will often be introduced into homes that already have cameras, Internet of Things (IoT) devices, and residents with wearables. To fulfill their potential, they need to integrate with existing smart infrastructure and reason through ambiguous scenarios: a smoke alarm may come from a pan about to ignite, a pan already on fire, a resident cooking or a faulty sensor, and each calls for a different response. To that end, we introduce HomeLoop, a simulation framework and benchmark built on Isaac Sim that combines robot bodies, networked devices and wearables, simulated residents and visitors, faulty sensors, and hazards that evolve over time. Our benchmark is composed of 20 tasks across five families grounded in real-world household hazards and needs, and designed so that the same alert can have different causes. A runtime monitor scores each episode with temporal-logic checks on the final outcome and the actions along the way, and every action carries a predefined cost. We evaluate six open-weight and proprietary VLMs. The best achieves a family-balanced score of 61.9%, and across models we see a pattern of over-reaction: failed episodes involving prohibited interventions outnumber those involving missed required interventions by more than two to one, with models often shutting off the stove before investigating the alarm's cause. HomeLoop is an extensible foundation intended for use in research on robot-IoT coordination, long-term memory, proactive assistance, and privacy-safety trade-offs. HomeLoop will be released with 2,100 home layouts that can be automatically furnished in diverse configurations, five robot bodies and 20 networked device types.
Physics-informed neural networks (PINNs) often suffer from imbalanced optimization of competing loss components, particularly when approximating solutions with high-frequency or multiscale features. Although existing balancing strategies have proven effective in alleviating this difficulty, the mechanisms governing the evolution of individual task contributions to Adam's parameter updates remain insufficiently understood. We introduce task-wise indicators to characterize this evolution and identify three distinct training regimes: smooth variation, intermittent spikes, and persistent oscillations. Focusing on the early smooth regime, we combine discrete-time analysis with a continuous-time singular perturbation framework under the full-batch setting and the parameter regime $(1-\beta_2)/(1-\beta_1)\ll1$. A first-order approximation of the locally attracting slow manifold reveals two factors governing the step-to-step redistribution of task contributions: curvature-related alignment disparities and differences in the coordinatewise distributions of preconditioned task gradients. Both factors involve Adam's shared second-moment estimate. Inspired by AdaTask, we propose Gradient-Modulated Adam (GMAdam), which modulates incoming task gradients using task-specific second-moment estimates. Numerical experiments on high-frequency Poisson and wave equations, the Allen-Cahn equation, and two-dimensional lid-driven cavity flow ($Re=100,400$) demonstrate consistent improvements in predictive accuracy over AdaTask at comparable computational cost.
Recognizing geometric texture patterns on 3D models is critical for applications such as heritage protection, model editing, and 3D printing. Existing 3D model segmentation methods predominantly focus on identifying high-level semantic parts, thereby failing to capture middle-level textures. While recent Vision-Language Models can perform image segmentation with appropriate text prompts, they still struggle to recognize 3D geometric textures. In this work, we present a method that segments 3D mesh models based on middle-level geometric textures. We further introduce a new geometric texture segmentation benchmark and fine-tune SAM3 accordingly. The segmented texture regions are then clustered based on DINO feature similarity and lifted from 2D to 3D. Additionally, we develop an interactive tool to enable both 3D model editing based on our segmentation results and model retrieval from the dataset based on geometric and texture similarity. Experiments demonstrate that our method significantly outperforms existing methods in segmenting geometric textures and can facilitate various applications.
The abundance of user-generated conversational data makes the question of how to responsibly derive insights from confidential data without putting the individuals in it at risk ever more important. However, the majority of existing differentially private (DP) synthetic text generation (SDG) and anonymization approaches ignore the context of the data's downstream usage, which results in loss of important information. DP-SDG provides the same guarantee regardless of the recipient, and anonymization relies on predefined attribute categories. They both ignore the observation that an attribute that is harmless to one audience can be sensitive to another, and each audience typically needs only some attributes preserved. We propose Ext-Classify-Priv, a framework that (1) extracts the information types present in a corpus, (2) classifies which of them are private for a given audience following contextual integrity, and (3) protects only those attributes, using either anonymization or SDG with Feature-DP approaches. On Reddit data with 3 audiences, our audience-specific releases are more faithful than a single audience-agnostic release for every audience. On bioRxiv data, our SDG method improves MAUVE from 0.097 to 0.420 at epsilon=1 and on the proprietary real estate data, our anonymization method improves MAUVE from 0.698 to 0.959 over the best baselines.
A growing literature examines the impact of engaging with AI chatbots on our relationality: our ability to form relationships with others, maintain them through life's changes, and exit when needed. Currently, substantial work focuses on purpose-built AI companions (e.g. Replika, Character.AI), with less work investigating socioemotional uses of the more widely-used general-purpose chatbots. In this paper, we present a 100-person investigation into how people engage general-purpose AI for socioemotional needs. Participants' AI uses varied, from practicing specific social skills to seeking additional bandwidth and information as they navigated their social contexts. To explain this heterogeneity, we provide 8 states that describe how AI expands or contracts relationality for people with different levels of skills, bandwidth, and information. We then present a unified conceptual model for the role of technology in mediating human relationality, the Relational Capacity Framework, and outline future research designing appropriate roles for AI in human connection.
A source can make a training mixture easier to learn while making its accuracy less valuable for the target. We study how the training problem a source creates and the finite-sample error of learning it jointly determine target-risk guarantees under reweighting. For binary classification, for each pair of laws and each positive target weight, we derive the sharpest target excess hinge-risk bound implied by mixture excess alone, uniformly over clipped predictors. This exact conversion curve is determined by density-weighted local votes. A two-sample oracle inequality for weighted support vector machines supplies the mixture excess, and the composition becomes explicit for Gaussian kernels. On a tilted-boundary class, we establish the all-procedure minimax classification rate $[n+s_+^2/(m^{-1}+\tau^2)]^{-1}$ in the transferable strength $s_+$ and the tilt $\tau$, separating source sampling uncertainty ($m^{-1}$) from relation uncertainty ($\tau^2$). A target-checked Gaussian-SVM selector adapts to unknown $s_+$ and $\tau$ up to a logarithmic factor for fixed dimension, and retains the target-only rate up to the same factor against arbitrary independent sources under the stated target model. On a separate smooth opposition class, the minimax risk within the family of exact Gaussian SVMs fitted to the original labels with nonnegative weights is exponentially larger than the unrestricted minimax risk whenever the target sample is at most half the source sample, despite arbitrary data-dependent weighting and tuning. The value of a source thus depends on the training problem it creates, on what accuracy there means for the target, and on whether the permitted use of the source can implement the needed correction.
Grid Foundation Models (GridFMs) are emerging as a powerful framework for learning and transferring knowledge across power grids and grid analysis tasks. This work extends GridFMs from steady-state analysis to dynamic simulation. We present DynGridFM, which uses component-wise tokenization and a structured latent operator to learn shared dynamic representations across heterogeneous grids and enable transfer to unseen systems. Pretrained on nine grids, DynGridFM is adapted to a held-out 48-machine grid, where fine-tuning on 64 trajectories achieves accuracy comparable to training from scratch on 128. On a single GPU, it achieves 924x to 1232x speedup over numerical simulators on one CPU core.
Learning locomotion directly on a humanoid robot is a promising direction to avoid the sim-to-real gap. However, locomotion rewards in Reinforcement Learning (RL) typically track the base velocity, which cannot be measured onboard without an external motion capture system or well-calibrated state estimator. This is why we introduce \method{}, which learns humanoid locomotion from minutes of samples with a reward that only requires the onboard Inertial Measurement Unit (IMU). Instead of tracking the base velocity, STRIDE rewards the robot for following a commanded acceleration profile. To learn efficiently from this reward, we use the off-policy RL algorithm XQC and combine a high gain setting for the joint position controllers with left-right symmetry augmentation. In simulation, STRIDE learns to walk on the Unitree G1 and Booster T1 humanoids within 14 minutes of samples. Compared to learning without these two components, it reduces the falls during training by 81%, while without them the robot learns only in over an hour.
Clinical gait and motion analysis is widely used to guide targeted interventions for patients with neurological disorders, the outcome of which is typically a report with descriptions of joint movements and forces. Current practices rely on experts' manual effort, which is labor-intensive and time-consuming. Recent advancements in Large Language Models (LLMs) provide opportunities to address this problem, but translating human motion into clinical gait descriptions remains challenging because the quality of the context directly affects LLMs' reasoning capability over continuous motion signals. To tackle this, we introduce MotionDecipher, a novel framework bridging the semantic gap between motion data and clinical interpretations. Specifically, we made 3 key advancements: 1) a Human Motion Tokenizer that converts motion signals into domain-specific tokens for time- and frequency-informed gait interpretation; 2) a Meta Feature Selector that transforms tokens into Structured Representations; and 3) a Hybrid Context Constructor to determine the optimal demonstration-based prompt for In-Context Learning. We evaluated MotionDecipher on data collected at an accredited gait lab from 122 patients and 46 controls. Results show that MotionDecipher achieves 0.91 BERTScore (13% over GPT-5), while expert evaluation ranks MotionDecipher highest in Correctness, Completeness, and Conciseness compared to baselines, highlighting its promise for efficient gait analysis.
Generative recommendation aims to predict item identifiers from user interaction histories, but explicit control over the properties of recommended items remains underexplored. A key challenge is to connect human-readable concepts to internal features whose interventions effectively change the recommended items. To address this, we propose Steerable Generative Recommender (SteerGR), a framework that uses sparse autoencoders for concept-level control in game discovery. SteerGR proceeds through three stages: (i) Concept Definition and Grounding for defining play concepts with an LLM; (ii) Feature Discovery and Selection for discovering candidate features by their concept associations and selecting one by its steering effect; and (iii) Sparse Feature Intervention for amplifying or suppressing selected features to steer recommendations without retraining the recommender. Extensive experiments with two industrial generative recommendation models on a real-world game dataset show that the selected features successfully steer recommendations toward or away from target concepts. We believe this work takes a first step to extend SAE-based steering from LLMs to industrial generative recommendation.
Personalized explanations typically adapt their content or presentation based on preferences, interaction histories, or other user characteristics. We examine whether regularities inferred from prior textual interpretations can support personalization of explanation organization. In a controlled within participant study with 122 participants, prior interpretations were represented along three dimensions: information selection, relative salience, and relational organization. The relational dimension was partially operationalized through co mobilization. These representations were aggregated into fixed historical profiles and used to organize otherwise identical explanation sentences. Participants encountered four organization procedures: the OWN historical profile, the DONOR historical profile, a population profile called POP, and a language model procedure using the historical profile called PROFILE LLM. Post exposure objective comprehension was the primary outcome. OWN achieved 86.5 percent correct responses, compared with 66.0 percent for DONOR, 63.9 percent for POP, and 73.0 percent for PROFILE LLM. Mixed effects analyses showed higher comprehension with OWN than with each alternative procedure. No reliable overall condition effect was detected for recommendation acceptance. These findings indicate that historical profiles derived from prior interpretations can provide an actionable signal for personalizing explanation order while keeping explanatory content and wording fixed.
Multi-agent LLM systems increasingly aggregate judgments across model instances, but reasoning about other agents can itself introduce failure. We study a setting in which LLMs shift their public decisions after observing ambiguous social signals while reporting inaccurate estimates of what other agents privately judged, even when the group's recorded private judgments do not support the apparent consensus. We call this failure phantom consensus and study it as a pluralistic-ignorance-style pattern in controlled LLM collectives. To measure this phenomenon, we introduce PHANTOM-Bench, a benchmark that separates isolated private judgments, public decisions after ambiguous social signals, elicited estimates of the group's private judgment, transparency-based recovery, and controls for neutral public text and explicit vote-following. Across nine closed- and open-weight models, ambiguous public signals consistently degrade decision quality and can turn correct private group judgments into incorrect public collective outcomes. Public accuracy drops by as much as 47.8 percentage points, and collective rationality failures reach 47.3% in the most susceptible setting. We observe substantial errors in reported group-private beliefs alongside public-stage failures, while control conditions reveal model-dependent differences in susceptibility to ambiguous and explicit peer signals. Revealing the agents' recorded private decisions repairs many failures without revealing the gold label, although recovery varies across models. Our results show that evaluating multi-agent LLM systems requires measuring not only final agreement, but whether that agreement reflects independently measured private judgments or an apparent consensus.
Continuous music emotion recognition (MER) estimates how Arousal and Valence evolve within a song, yet the field's standard metric, pooled frame-level correlation, is dominated by between-song differences and cannot distinguish a model that tracks emotional dynamics from one that only ranks songs by their average level. We decompose annotation variance into between- and within-song components and show that mean squared error (MSE) targets the former, while windowed Lin's concordance correlation coefficient (CCC = $r \cdot C_b$) targets the latter by penalising the variance compression MSE induces. A single weight $\lambda$ controls the trade-off: on 1,764 DEAM songs, raising $\lambda$ from 0 to 1 lowers pooled Arousal $r$ from 0.843 to 0.810 while raising mean within-song $r$ from 0.204 to 0.368. Building on this objective, we present MMAS (Multimodal Music Analysis System), an open-source model that predicts Arousal and Valence every 0.5 s from raw audio. It fuses frozen MERT-v1-95M embeddings with ten symbolic features from automatic transcription through a dual-branch temporal convolutional network with cross-modal attention. Because both branches are audio-derived, one model trains jointly on DEAM and PMEmo without domain conditioning. Under song-grouped 5-fold cross-validation (2,531 songs), pooled $r$ reaches 0.801/0.685 (Arousal/Valence) on DEAM and 0.742/0.746 on PMEmo. Against a ridge baseline on identical features, within-song Arousal $r$ roughly triples (0.344 vs. 0.125; $n = 353$ songs, Wilcoxon $p = 1.7 \times 10^{-21}$, $d = 0.56$). Complete transcription failure costs at most 0.04 pooled $r$. Split-conformal intervals attain 93.7% (Arousal) and 90.0% (Valence) marginal coverage at a 95% target, with the shortfall traced to within-song dependence. MMAS is released as a pip-installable package.
Quantization and pruning reduce the cost of deploying language models, but their effects are usually reported as benchmark scores. Such scores record only the loss that remains at the output. Inside the model, compression damages some components more than others, and the rest of the network can partly compensate. The observed loss therefore reflects both the damage and this compensation, which benchmark scores cannot separate. How the two combine, and whether their balance changes with model size, remains unclear. We study this compensation for in-context retrieval, a capability carried by a small, identifiable set of attention heads. To separate the two, the expected loss from ablating these heads is decomposed exactly into the part restored by the rest of the network, attributed to the components responsible, and the part that reaches the output. This decomposition is then compared across model sizes. For damage of comparable size, smaller models restore most of the loss, whereas larger models retain most of it, because the compensating responses offset less of it as models grow. Compressing the heads produces the same pattern. Compression that is nearly harmless in small models can therefore be costly in large ones.
Armed conflict can increase the value of Generative AI (GenAI) while undermining the conditions required for its responsible use. Yet the perspectives of community-based humanitarian responders working under these conditions remain underrepresented in responsible AI research. We examine GenAI use in Sudan's Emergency Response Rooms (ERRs), a community-led humanitarian network, through an exploratory survey of 89 respondents and 15 semi-structured interviews with GenAI users and non-users. War reconfigures GenAI's role, making it most useful where its outputs are hardest to verify. ERR responders turned to GenAI when conflict made human expertise difficult or dangerous to reach. However, participants cautioned that GenAI could not independently establish rapidly changing local conditions, and that relying on its outputs could lead to harmful decisions when human verification was difficult. Risks arose before GenAI was used. For some volunteers, carrying a phone or accessing the internet could expose them to danger. Others were concerned that sharing sensitive beneficiary or operational information with GenAI could put people at risk. Participants feared that AI-assisted reports could undermine their credibility with donors, whose funding sustains the food and medical aid ERRs provide. Access to GenAI was uneven, and volunteers using it sometimes relied on colleagues with relevant expertise or local knowledge to verify its results. Our findings show that responsible use of GenAI during conflict depends not only on model performance or individual adoption, but also on whether responders can safely access these tools, verify their outputs against local conditions, and recognize when limiting or avoiding GenAI is the more responsible choice.
Large language models use auxiliary conditioning to incorporate additional information into time-series forecasts. However, auxiliary conditioning does not consistently improve accuracy and can increase forecasting error when its influence is poorly matched to the observed temporal context. We propose Forecast-Aware Residual Refinement (FARR), which learns to correct forecasts using endpoint evidence. FARR obtains a valid forecast under the standard condition and a null forecast under a predefined null condition. Their difference reveals how conditioning changes the prediction. A shared residual refiner combines this endpoint evidence with historical features to estimate the remaining error of the valid forecast. Horizon-wise gating then controls the contribution of the predicted residual to the final output. Experiments on four long-horizon forecasting benchmarks show that FARR achieves the best dataset-average MSE and MAE among the compared methods. Code will be released at https://github.com/blossomrichzyy-star/farr-forecast-aware-residual-refinement.
Modeling gene regulation in single cells requires jointly representing DNA sequence, chromatin accessibility, and gene expression across regulatory scales. Sequence-based models capture DNA-level regulatory grammar but often lack cell-level resolution, whereas many single-cell omics models represent genes and peaks as fixed-vocabulary tokens without their underlying sequence. Here, we present MUSS, a sequence-based multi-scale model for single-cell gene regulation that preserves both DNA-level and cell-level resolution. MUSS derives peak- and gene-level representations from DNA using a frozen pre-trained sequence encoder and integrates them through peak-peak self-attention and distance-biased peak-gene interactions. Peak masked language modeling followed by training on paired scATAC-seq and scRNA-seq aligns these representations with cell-specific regulatory states. Across gene expression prediction, enhancer-gene linking, motif-guided regulatory prioritization, and zero-shot cell clustering, MUSS achieves strong performance against sequence-based and single-cell omics baselines. It further predicts expression for genes and peaks unseen during task training. These results support sequence-grounded, multi-scale representations for decoding cell-type-specific regulatory programs. Code, datasets, and model checkpoints are available at https://github.com/Lotfollahi-lab/MUSS_neurips26.
Automated fact-checking tools, cheap and fast but imperfect, are increasingly available on social platforms. We study when such tools can backfire in a sequential social learning model of misinformation. Agents observe an article's reliability and their predecessors' actions, which indicate whether they liked or disliked the article. They may also pay to fact-check it. We examine how welfare, defined as the probability of a correct cascade, changes when a free, error-prone automated tool supplements costly but perfectly accurate manual fact-checking. Two opposing effects arise. Broader use of fact-checking helps agents escape wrong cascades. At the same time, errors from the automated tool can mislead agents who would otherwise have verified the article perfectly, reducing the likelihood of correct cascades. We prove that adding the tool strictly lowers welfare for articles of intermediate reliability when truthful articles are not too informative and manual checking is sufficiently costly. Simulations show that the tool can help on either side of the intermediate range, and that it improves on manual-only checking on average only above a threshold accuracy.