Carnegie Mellon University（卡内基梅隆大学）

Real-time multimodal applications, including voice agents and interactive video generation, compose heterogeneous models into pipelines whose efficient deployment requires application-specific decisions about placement, streaming, and intra-model parallelism. Existing serving systems and auto-parallelism compilers commit to limited transformations and fixed workload assumptions, so achieving high performance on a new application requires hand-crafting an efficient implementation. We present FlashRT, an agent harness that guides coding agents to lift simple developer-written reference implementations into optimized multi-GPU deployments that flexibly weigh target metrics like latency and throughput. Using a new chain-of-program paradigm, FlashRT directs a generic coding agent through a multi-pass transformation process where an agent transforms the reference into an intermediate representation (IR) to capture data dependencies and persistent-state scopes, validates this IR via a sequential interpreter, and performs static analyses to identify candidate transformations. Then, the agent iteratively implements, verifies, and benchmarks each candidate under a measurement-gated optimization loop to produce effective deployments that span different hardware budgets. Across various applications, including video world models and multimodal LLMs, FlashRT converts reference implementations into highly efficient deployments, delivering up to ~70x latency reduction and 2.8x throughput improvement on NVIDIA B200 GPUs. On AMD MI355X GPUs, FlashRT matches the peak latency reduction while increasing peak throughput improvement to 3.6x, demonstrating that agent-driven optimization can be more scalable on platforms with less mature expert optimization. In fact, for Qwen3-Omni text-to-audio inference, FlashRT reduces response latency by 65% compared to the expert vLLM-Omni implementation on AMD MI355X.

Industrial Video Anomaly Detection (IVAD) aims to identify anomalous objects and events in an industrial process, which is crucial for modern manufacturing and quality control systems. Existing VLM-based anomaly reasoning methods are capable of detecting open-ended anomalies in general domains. However, their performance declines in industrial settings characterized by intricate object transformations, strict physics, and procedural constraints. To tackle the complexity of such interaction-intensive detection, we introduce a training-free agentic framework for anomaly detection free of domain-specific knowledge, emphasizing object state evolution like humans inspectors. It is designed to track spatial-temporal dynamics and underlying transformations of detected objects over time, and then reason over the object-wise temporal state trajectories to identify abnormal objects in grounded frames. Our method overcomes limitations of prior approaches that rely on retraining on normal clips or injecting domain knowledge as context for test-time inference. Extensive experiments on three IVAD datasets demonstrate that our method outperforms frontier VLMs, agentic frameworks, and traditional VAD methods fine-tuned on the respective datasets, while providing interpretable reports over anomaly processes and types.

Standard methods for autonomous navigation in unstructured terrain are prone to myopic behaviors in long-horizon scenarios. The use of metric maps built from LiDAR or cameras provides necessary local geometry and semantic information but is strictly limited by depth sensing range. By discarding data beyond the mapping horizon robots suffer from suboptimal, short-sighted decisions. To recover this lost information, we focus on extracting long-range traversability-aware frontiers directly from first-person-view (FPV) images. By leveraging satellite imagery, we compute the set of feasible navigation paths for a dataset of image/pose pairs and use them to supervise our network, reducing the need for extensive human demonstration data. We demonstrate that this approach improves performance in long-range off-road navigation over existing methods by more than 10% in various offline benchmarks and reduces the number of human interventions incurred in a set of real-world experiments. More details can be found at https://theairlab.org/ss_frontiers_iros .

Large Language Models (LLMs) often struggle to navigate value conflicts when trained with the compressed scalar rewards of Reinforcement Learning from Human Feedback (RLHF). To address this challenge, we investigate how chain-of-thought (CoT) reasoning can help improve performance in this domain. Geometrically, we show that CoT correlates with further smoothing the model's loss landscape in its sharpest direction, helping resolve the optimization instability of traditional scalar rewards. We also demonstrate via relevant downstream benchmarks that value conflict-focused CoT may generalize to different kinds of moral reasoning, demonstrating that this CoT has the potential to be an effective mechanism for better moral reasoning. To capitalize on this potential, we create a new value conflict-focused CoT design that further smooths the sharpest direction of the loss landscape and increases moral reasoning performance. This finding shows that explicitly modifying and improving the design of reasoning dynamics offers a promising avenue for improving model performance on user requests with complex value conflicts, advancing pluralistic alignment in LLMs.

We introduce Retrieval-Augmented Interpretable Learning (RAIL), a probabilistic meta-learning framework for zero-shot generation of task-specific interpretable models that synthesizes coefficient-space structure from natural-language task descriptions and a memory of previously learned task-specific predictors. RAIL retrieves related source tasks, transfers structure through coefficient space, and generates a new predictor in the original diagnostic-feature space, enabling zero-shot and few-shot clinical procedure prediction with feature-level explanations. Its probabilistic formulation provides uncertainty over retrieval, model coefficients, and predictions, supporting reliability-aware deployment: uncertain predictions or unstable explanations can be flagged for additional clinical review rather than treated as automatic decisions. This makes RAIL particularly suited for healthcare settings, where prediction tasks are highly long-tailed, new clinical targets arise frequently, and models must remain inspectable, uncertainty-aware, and compatible with human oversight. Across long-tailed clinical procedure prediction tasks, RAIL maintains reliable performance across data-availability regimes: it achieves 73.4% accuracy in the held-out zero-shot settings, where no supervised task-specific model can be trained, and remains near 73.2% accuracy in the extreme few-shot regime with only 2-4 examples, where supervised task-specific models perform close to chance. RAIL further benefits from clinically informed task representations and yields retrieval, uncertainty, and coefficient-level diagnostics that make model behavior more transparent. These results suggest a path toward scalable clinical prediction systems that can adapt to new tasks while preserving interpretability and reliability.

Behavioural shifts in wild great ape populations, particularly the breakdown of social structures, can serve as an early indicator of population decline. Automating the detection of behaviours indicative of these shifts is therefore a critical task for conservation. Several valuable datasets have recently been introduced for the automated recognition of great ape behaviour, yet few include fine-grained social behaviour annotations, and those that do are captured either in captive settings or via aerial platforms such as UAVs. We address this gap by introducing PanAf-SBR, the first wild great ape camera trap dataset annotated with social behaviours. PanAf-SBR extends PanAf500 with 100 additional videos covering 36,063 frames. These come with 81,096 annotations including bounding boxes, segmentation masks, intra-video identities, and seven social behaviour classes defined under the action giver and receiver convention of ChimpACT. We use this data together with the AlphaChimp architecture to establish the first benchmarks for fine-grained social behaviour recognition in wild great apes from camera trap footage. We further conduct bidirectional transfer learning experiments between PanAf-SBR and the captive ChimpACT dataset, finding that cross-dataset pre-training is highly beneficial for specific classes rather than of uniform benefit. Finally, we examine the role of background context by inverting the segmentation masks to suppress non-ape pixels.

An AI research agent can improve the score it sees without finding a modelling change that works on new materials. We ask a stricter question. After repeated experiments, does the selected change survive on data that never entered the loop, and can its code be reused? We separate the search into changes to features, models, representations, and training data. Seven searches produce 701 evaluated changes across ten Matbench endpoints. Agents receive only the mean over five inner folds, reducing reliance on any single development split. We then freeze the selected code and evaluate it once on an untouched holdout. Nine of ten choices remain the best tested single intervention. The surviving changes reveal two materials modelling regimes. With composition alone, feature, model, and representation changes provide comparable routes to improvement. They include held-out MAE reductions of 17.4\% for band gap and 18.6\% for steel strength, as well as gains on both classification endpoints, while screened external data adds little. For structure tasks, richer geometry descriptors and model or calibration changes lower mean held-out MAE by 14.6\% and 7.1\% and lead on different property families, whereas composition embeddings do not transfer. Combining separately found feature and model changes yields a 26.3\% mean held-out improvement. These results show in materials prediction that closed-loop agents can produce decisions that survive unseen evidence and code changes that can be reused across tasks and combined. More broadly, they provide an evaluation design for testing executable discoveries beyond the feedback loop.

We describe a method for computing signed distance to point clouds that allows fast pointwise evaluation at arbitrary spatial resolution. As input, our method takes a point cloud with normals; as output, it provides an analytical parameterization that allows queries of signed distance to the approximate underlying surface at arbitrary points - simultaneously providing reconstruction and distance. Our key idea is to reconstruct shapes by locally fitting point clouds with tori, which have closed-form signed distance functions. Tori are fitted in a feed-forward manner, using a pre-trained network to output per-point curvature and shift parameters. Importantly, our method does not require costly global optimization or spatial discretization, and is easily parallelizable. Underlying our method is a new theory that unifies signed distance with the classic reconstruction methods of winding numbers and Poisson surface reconstruction. We use our method to compute signed distance to point clouds arising from photogrammetry, meshes, 3D Gaussians, and neural implicits. Our method allows point clouds to be used directly in applications, without explicit surface reconstruction: as examples, we take offsets of point clouds, apply morphological and Boolean operations, and directly visualize offset surfaces using sphere tracing.

Malaria diagnosis in endemic regions depends on species-level identification of Plasmodium parasites in thick blood smears, but deep learning detectors classify detections without providing morphological evidence for their predictions, limiting the ability of microscopists to audit those predictions at the case level. We present SGMCE (Segment-Grounded Morphological Concept Explanation), a post-hoc explanation framework that requires no additional training, no morphological annotations, and no labelled explanation data, yet produces per-detection natural-language explanations anchored in thick-smear morphology. For each detection, SGMCE extracts mask-guided crop thumbnails, computes fourteen handcrafted computer-vision morphological features (shape, colour, chromatin, haemozoin pigment) using adaptive within-mask thresholds, and queries GPT-4o with both visual evidence and computed measurements, conditioned on a thick-smear-specific knowledge base compiled from the World Health Organization bench aids. The primary output is a structured explanation identifying which morphological features support the detected species and why the competing species are excluded. Explanations are validated by four automatic metrics: Knowledge-Base Consistency (KBC), CV-Claim Faithfulness (CCF), Discriminativeness Score (DS), and LLM-as-Judge (LLMj). A sentence-level semantic scoring rule with species-aware negation filtering resolves the vocabulary mismatch between clinical prose and knowledge-base terms. Across 737 detections from 139 thick-smear images spanning four Plasmodium species and white blood cells, parasite-class mean KBC is 0.91, mean DS is 0.99, and mean CCF is 0.97, while a per-rule CCF breakdown confirms that the CV-grounded claims made by the vision-language model are consistent with the measurements they cite.

Bundle adjustment (BA) is a critical technique in various robotic applications such as simultaneous localization and mapping (SLAM), augmented reality (AR), and photogrammetry. BA optimizes parameters such as camera poses and 3D landmarks to align them with observations. With the growing importance of deep learning in perception systems, there is an increasing need to integrate BA with deep learning frameworks for enhanced reliability and performance. However, widely-used C++-based BA libraries, such as GTSAM, g$^2$o, and Ceres Solver, lack native integration with modern deep learning libraries like PyTorch. This limitation affects their flexibility, ease of debugging, and overall implementation efficiency. To address this gap, we introduce an eager-mode BA library seamlessly integrated with PyTorch with high efficiency. Our approach includes a sparsity-aware auto-differentiation design and GPU-accelerated sparse operations designed for 2nd-order optimization. Our eager-mode BA on GPU demonstrates substantial runtime efficiency, achieving an average speedup of 18.5$\times$, 22$\times$, and 23$\times$ across all benchmarks compared to GTSAM, g$^2$o, and Ceres, respectively.

Retrieval-Augmented Generation (RAG) is usually evaluated by whether the final answer is correct. Under knowledge conflict, this hides a key question: did the model follow retrieved evidence, rely on its parametric prior, or produce a post-hoc rationale? We study this as context compliance, the regime in which retrieved context controls the answer even when it conflicts with the model's prior knowledge. We introduce Context-Driven Decomposition (CDD), an inference-time diagnostic intervention that elicits contextual and prior answers, isolates the conflicting premise, and records a resolution trace that can be perturbed. Across Epi-Scale stress tests, TruthfulQA misconception injection, and cross-model reruns, CDD makes three behaviors visible. First, misleading retrieval can severely degrade accuracy: under a worst-case TruthfulQA misconception-injection probe, Standard RAG reaches only 15.0%. Second, better answers need not share the same mechanism: CDD improves adversarial accuracy on Gemini-2.5-Flash and shows directional gains across Claude variants, yet trace-perturbation sensitivity is high only on Gemini. Third, explicit decomposition improves controlled-conflict robustness over a conflict-aware instruction baseline on localized factual conflicts, with the clearest margins on Entity Swap (88.0% vs 79.3%) and Logical Contradiction (83.2% vs 75.4%). We frame RAG conflict handling as an observability problem.

Visual token pruning is widely used to reduce the inference cost of multimodal large language models (MLLMs), but it is usually evaluated only by accuracy. We study how pruning affects calibration, defined as the agreement between confidence and correctness, and show that the selection rule matters more than the token budget alone. On POPE with LLaVA-1.5, coverage-based pruning from 576 to 128 tokens reduces expected calibration error from 0.041 to 0.016 without a statistically significant accuracy loss. In contrast, attention-based selection preserves confidence while accuracy deteriorates, becoming less calibrated than random pruning at aggressive budgets. Across pruning conditions, kept-set coverage is strongly associated with accuracy (Spearman $ρ=+0.89$) but not with mean confidence ($ρ=-0.03$), producing a strong inverse relation with overconfidence ($ρ=-0.92$); controlled kept-set interventions support this evidence-coverage account. Two boundaries limit it: query-conditioned FastV is more overconfident than its coverage predicts, and on language-prior-dominated ScienceQA, coverage ceases to order calibration. The selector ordering otherwise generalizes to GQA and LLaVA-NeXT, and coverage beats random on Qwen2-VL, although the calibration gain over the unpruned model is task- and model-dependent. We also identify an evaluation pitfall in FastV: zeroing rather than removing pruned tokens can reduce accuracy to chance. Visual token pruning therefore changes confidence quality as well as efficiency, and calibration should be evaluated alongside accuracy when comparing pruning methods.

AI agents may be able to assist with emails and documents, but can they reliably complete everyday online workflows on real websites? Everyday online tasks offer a realistic yet unsolved testbed for evaluating the next generation of AI agents. To this end, we introduce ClawBench, an evaluation framework comprising 153 everyday online tasks that people need to accomplish regularly in their lives and work, spanning 144 platforms across 15 categories, from completing purchases and booking appointments to submitting job applications. These tasks require capabilities beyond existing benchmarks, such as obtaining relevant information from user-provided documents, navigating multi-step workflows across diverse platforms, and write-heavy operations like filling in many detailed forms correctly. Unlike existing benchmarks that evaluate agents in offline sandboxes with static pages, ClawBench operates on production websites, preserving the full complexity, dynamic nature, and interaction challenges of real-world web environments. An interception layer captures and blocks the final submission request, ensuring safe evaluation without real-world side effects. Our evaluations of 8 frontier models show that both proprietary and open-source models complete only a small portion of these tasks. For example, Claude Sonnet 4.6 achieves only 33.3%, which exposes gaps in current AI agents. Progress on ClawBench brings us closer to AI agents that can function as general-purpose assistants.

As the focus in LLM-based coding shifts from static single-step code generation to multi-step agentic interaction with tools and environments, understanding which tasks will challenge agents and why becomes increasingly difficult. This is compounded by current practice: agent performance is typically measured by aggregate pass rates on benchmarks, but single-number metrics obscure the diversity of tasks within a benchmark. We present a framework for predicting success or failure on individual tasks tailored to the agentic coding regime. Our approach augments Item Response Theory (IRT) with rich features extracted from tasks, including issue statements, repository contexts, solutions, and test cases, and introduces a novel decomposition of agent ability into LLM and scaffold ability components. This parameterization enables us to aggregate evaluation data across heterogeneous leaderboards and accurately predict task-level performance for unseen benchmarks, as well as unseen LLM-scaffold combinations. Our methods have practical utility for benchmark designers, who can better calibrate the difficulty of their new tasks without running computationally expensive agent evaluations.

Navigating unseen, large-scale environments based on complex and abstract human instructions remains a formidable challenge for autonomous mobile robots. Addressing this requires robots to infer implicit semantics and efficiently explore large-scale task spaces. However, existing methods, ranging from end-to-end learning to foundation model-based modular architectures, often lack the capability to decompose complex tasks or employ efficient exploration strategies, leading to robot aimless wandering or target recognition failures. To address these limitations, we propose VL-Nav, a neuro-symbolic (NeSy) vision-language navigation system. The proposed system intertwines neural reasoning with symbolic guidance through two core components: (1) a NeSy task planner that leverages a symbolic 3D scene graph and image memory system to enhance the vision language models' (VLMs) neural reasoning capabilities for task decomposition and replanning; and (2) a NeSy exploration system that couples neural semantic cues with the symbolic heuristic function to efficiently gather the task-related information while minimizing unnecessary repeat travel during exploration. Validated on the DARPA TIAMAT Challenge navigation tasks, our system achieved an 83.4% success rate (SR) in indoor environments and 75% in outdoor scenarios. VL-Nav achieved an 86.3% SR in real-world experiments, including a challenging 483-meter run. Finally, we validate the system with complex instructions in a 3D multi-floor scenario.

Reinforcement learning (RL) has become a central post-training paradigm for large language models (LLMs), but its performance is highly sensitive to the quality of training problems. This sensitivity stems from the non-stationarity of RL: rollouts are generated by an evolving policy, and learning is shaped by exploration and reward feedback, unlike supervised fine-tuning (SFT) with fixed trajectories. As a result, prior work often relies on manual curation or simple heuristic filters (e.g., accuracy), which can admit incorrect or low-utility problems. We propose GradAlign, a gradient-aligned data selection method for LLM reinforcement learning that uses a small, trusted validation set to prioritize training problems whose policy gradients align with validation gradients, yielding an adaptive curriculum. We evaluate GradAlign across three challenging data regimes: unreliable reward signals, distribution imbalance, and low-utility training corpus, showing that GradAlign consistently outperforms existing baselines, underscoring the importance of directional gradient signals in navigating non-stationary policy optimization and yielding more stable training and improved final performance. We release our implementation at https://github.com/StigLidu/GradAlign

We present NEMO, a system that translates Natural-language descriptions of decision problems into formal Executable Mathematical Optimization implementations using autonomous coding agents (ACAs). Existing approaches rely on specialized large language models (LLMs) or bespoke task-specific agents that are often brittle and frequently generate syntactically invalid or non-executable code. NEMO instead treats ACAs as a first-class abstraction analogous to API-based interaction with LLMs; their sandboxed execution guarantees code is executable by construction and supports automated validation and repair. We introduce novel coordination patterns including asymmetric validation loops between independently generated optimizer and simulator implementations, external memory for experience reuse, and robustness enhancements via minimum Bayes risk (MBR) decoding and self-consistency. Across nine established optimization benchmarks, NEMO achieves state-of-the-art performance on the majority of tasks with substantial margins on several datasets, demonstrating the power of execution-aware agentic architectures for automated optimization modeling.

Large language models (LLMs) are improving rapidly as reflected in benchmark scores, yet these AI benchmarks largely test capabilities such as factual recall, narrow question answering, mathematical problem-solving, and coding and agentic tool-use. What remains poorly measured is AI progress on the analytical knowledge work white-collar professionals perform daily, including synthesizing complex information, exercising judgment under uncertainty and incomplete information, applying strategic and adversarial thinking in multi-stakeholder settings, weighing trade-offs, and producing defensible, structured analyses. This gap is even more pronounced for subjective components of such work, where success can be challenging to define. The "case method" form of education practiced by top business schools provides a natural foundation for addressing this measurement gap, and we construct BusinessCaseBench, a benchmark spanning hundreds of questions drawn from business cases across eighteen disciplines, each paired with a grading rubric derived from the expert-written instructor case solution. On BusinessCaseBench, frontier AI models already score highly against instructor rubrics, and capability within one model family improves substantially over two years. These results provide strong evidence that AI performance on this class of work is already high and rapidly improving, with implications for business schools, where case pedagogy trains undergraduates and MBAs in this kind of analytical reasoning, and for entry-level professional roles, where such skills have historically anchored early-career work.

Where should the question go in a vision-language model (VLM) prompt: before the image or after it? Intuition says before: knowing what is asked should tell the model where to look. Yet across visual question answering benchmarks, question-first prompting consistently underperforms the image-first ordering recommended for frontier VLMs, a phenomenon we term the question-first paradox. We trace the paradox to a conflict between two stages of VLM computation. Logit-lens and attention probes show the intuition is half right: a question placed before the image genuinely steers perception, moving image patch representations toward question-relevant concepts. The failure lies downstream. Stranded behind hundreds of image tokens, the question is barely attended by the answer token, which instead commits to image-driven (often wrong) answers; a causal attention knockout confirms that the answer reads the question only when the question follows the image. The diagnosis yields a training-free fix: question echoing, restating the question on both sides of the image so that one copy steers perception while the other is read out at answer time. The same division of labor appears in a fifty-year-old finding on human ``adjunct questions'', where repeating a question before and after a passage aids comprehension more than either position alone. Echoing the image as well brings further gains, restoring the whole-image view a causal decoder otherwise loses. The paradox holds across five open VLMs, costing up to 17.5 group-accuracy points. Echoed prompts close it and surpass the best single-pass ordering on NaturalBench, POPE, Winoground, and open-ended VQAv2, by up to 19 Winoground group-accuracy points, with no training, fine-tuning, or architecture change. The paradox reveals a trade-off between steering perception and preserving question access; echoing resolves it through prompt design alone.

Logic and optimization can, in combination, make valuable contributions to rule-based AI. Logic is the obvious medium for encoding a rule base and drawing inferences from it, while optimization provides a powerful technology for computing inferences. Their combination has taken on new relevance amid a growing concern for transparency in AI. which is important for reproducibility, explainability, trustworthiness, and fairness. Rule-based AI provides a natural solution to transparency that is becoming increasingly practical due to today's highly advanced optimization methods. This article surveys several areas of logic-optimization partnership, including probabilistic logic, Bayesian logic, belief logics and Dempster-Shafer theory, nonmonotonic (default) logic, many-valued logics, and inference of logical formulas from noisy data based on Boolean regression. It shows how to compute projections, the fundamental problem of both logic and optimization, using decision diagrams and logic-based Benders decomposition. It describes the use of postoptimality analysis to explain how conclusions are reached, further enhancing transparency, as well as the role of optimization in answer set programming modulo theories. The paper concludes by suggesting possible future research directions.

Text-to-image (T2I) models generate realistic likenesses of some individuals when prompted with their names, raising privacy concerns. However, distinguishing whether a generated face is memorized or fabricated currently requires ground-truth photos, access to training data, or white-box access to model internals, limiting applicability. We introduce a fully black-box behavioral probe that distinguishes between memorized and unrecognized names, while requiring no reference photos or prior knowledge of training data. To benchmark this task, we present the NAMESAKES dataset of over one thousand names and faces of public figures spanning a wide range of fame levels, along with perturbed, less famous names. Experiments on state-of-the-art T2I models show that our probe substantially predicts identity memorization and separates memorized from unrecognized names, with further insights into differences across model families.

Cooperative object transport in unstructured environments remains challenging for assistive humanoids because strong, time-varying interaction forces can make tracking-centric whole-body control unreliable, especially in close-contact support tasks. This paper proposes a bio-inspired, interaction-oriented whole-body control (IO-WBC) that functions as an artificial cerebellum - an adaptive motor agent that translates upstream (skill-level) commands into stable, physically consistent whole-body behavior under contact. This work structurally separates upper-body interaction execution from lower-body support control, enabling the robot to maintain balance while shaping force exchange in a tightly coupled robot-object system. A trajectory-optimized reference generator (RG) provides a kinematic prior, while a reinforcement learning (RL) policy governs body responses under heavy-load interactions and disturbances. The policy is trained in simulation with randomized payload mass/inertia and external perturbations, and deployed via asymmetric teacher-student distillation so that the student relies only on proprioceptive histories at runtime. Extensive experiments demonstrate that IO-WBC maintains stable whole-body behavior and physical interaction even when precise velocity tracking becomes infeasible, enabling compliant object transport across a wide range of scenarios.

Model Predictive Control (MPC) has demonstrated significant performance improvements over today's control methods for residential Heating, Ventilation, and Air Conditioning (HVAC), but deploying MPC often requires substantial engineering effort. Reinforcement Learning (RL) may offer comparable performance with easier deployment, but its practical application for residential HVAC remains largely undemonstrated, leaving open questions related to occupant comfort and data requirements. To investigate these issues, we deployed one MPC variant and one model-based RL variant for one month each in an occupied house in a cold climate. The controllers adjusted an air-to-air heat pump's thermostat temperature setpoint based on measurements of the indoor temperature and the electric power used for heating. Relative to constant-setpoint operation, MPC saved 18.1\% (95\% confidence interval: 4.4 to 30.9\%) of weather-normalized heat pump energy and RL saved 20.9\% (2.6 to 38.3\%). MPC maintained acceptable occupant comfort. RL kept the house cooler, particularly during an initial adaptation phase, leading to three reports of occupant discomfort. The two algorithms had similar data requirements. We estimate that for a fresh deployment in another house, RL would take about one-third less engineering effort than MPC. While RL reduces deployment effort, it faces difficulties related to safe controller initialization and to mismatches between the modeled and true state and action spaces.

System Instructions (SIs), or system prompts, are pivotal for guiding Large Language Models (LLMs) but manual crafting is resource-intensive and often suboptimal. Existing automated methods frequently generate non-human-readable "soft prompts," sacrificing interpretability. This paper introduces SI-Agent, a novel agentic framework designed to automatically generate and iteratively refine human-readable SIs through a feedback-driven loop. SI-Agent employs three collaborating agents: an Instructor Agent, an Instruction Follower Agent (target LLM), and a Feedback/Reward Agent evaluating task performance and optionally SI readability. The framework utilizes iterative cycles where feedback guides the Instructor's refinement strategy (e.g., LLM-based editing, evolutionary algorithms). We detail the framework's architecture, agent roles, the iterative refinement process, and contrast it with existing methods. We present experimental results validating SI-Agent's effectiveness, focusing on metrics for task performance, SI readability, and efficiency. Our findings indicate that SI-Agent generates effective, readable SIs, offering a favorable trade-off between performance and interpretability compared to baselines. Potential implications include democratizing LLM customization and enhancing model transparency. Challenges related to computational cost and feedback reliability are acknowledged.

We report on CoreForge, an experience in using large language models (LLMs) to build an unweighted MaxSAT solver from research papers rather than from an existing solver codebase. The project focuses on unsatisfiability-based MaxSAT algorithms and follows an iterative workflow that combines paper discussions with ChatGPT, implementation through Codex prompts, and repeated LLM-assisted code audits and revisions. Although the codebase implements several algorithms and solver components, our evaluation focuses on configurations that combine core-guided optimization, lightweight preprocessing, core minimization, integration with integer linear optimization backends, and a new core-sequence lookahead approach. Our experience suggests that LLMs can support solver implementation from papers, while requiring external validation, benchmarking, and human guidance. In our experiments, fuzzing and MaxSAT Evaluation instances did not reveal wrong answers in the tested configurations, although performance remains below the best hand-engineered MaxSAT solvers. We summarize what worked, what remained difficult, and the lessons for future LLM-assisted solver development.

When sighted practitioners author accessible data visualizations, they build navigation structures (the nodes, edges, and input bindings that govern how assistive technologies traverse an interface) entirely in code, with no visual representation. Without a representation to react to, practitioners cannot develop judgment about what makes navigation good or bad, and the quality ceiling of non-visual experiences is set by the absence of a feedback loop. We address this problem through longitudinal co-design with practitioners across cartography, design systems, and open-source visualization, and make three contributions. First, we introduce an Inspector that renders navigation graphs as interactive node-link diagrams, and a Dimensions API that expresses navigation in terms of data dimensions rather than explicit graph construction. Second we present Skeleton, a direct-manipulation authoring environment in which the properties of an accessible navigation structure are translated into visual representations authors can observe and manipulate. Key techniques include a dual-view editor that simultaneously shows the system's navigation model and the end user's spatial experience, a scaffolding engine that automates spatial node placement by repurposing a visualization rendering pipeline, a live label-template editor with real-time screen-reader-output preview, and a testing mode that makes traversal sequence visually trackable. Third, we evaluate Skeleton through an in-situ study with 8 practitioners across visualization design, engineering, and research. Making navigation structure visible changed how practitioners engaged with accessible design: they reconsidered the architecture of their own visualizations, attended to a broader range of input modalities, and shifted from treating accessibility as a compliance task to treating it as a design problem. (abstract shortened for arxiv)

Disordered metamaterials feature microstructures with inherent randomness and irregularity, enabling them to achieve broader property coverage and superior performance unavailable in their regular counterparts. Despite their promise, designing disordered microstructures is substantially harder than designing regular ones. Their design remains trapped between manual parameterizations with limited expressiveness, and generative AI that is data-hungry and struggles to generalize. To address these limitations, we propose a generative design framework based on Neural Cellular Automata that dynamically grows complex microstructures through learned local interaction rules, inspired by the self-organizing processes in natural materials. This framework requires only a single training template, yet accommodates diverse disordered microstructures and adapts to irregular domains and arbitrary discretizations. By manipulating the learned local rules, we can steer the growth process to generate microstructures unseen during training, providing control over orientation, anisotropy, and directional thickness without retraining. As a dynamic, local growth process, it naturally produces spatially varying microstructures that transition smoothly to enable location-specific mechanical properties. We demonstrate this in a multiscale mechanical cloaking design, where microstructures vary across the space to meet an optimized heterogeneous property distribution. Our design enables excellent cloaking performance without complicated post-processing and incompatible assembly common in existing methods. This data-efficient, generalizable approach opens access to previously intractable disordered materials for biomedical implants and soft robotics.

Policy learning methods are increasingly used to inform treatment allocation under budget constraints. Most proposed methods assume complete treatment data, yet applications frequently suffer from missingness that can bias estimates and lead to suboptimal policies. We address this gap by extending efficient estimators for average treatment effect (ATE) estimation to policy value and conditional average treatment effect (CATE) estimation under missing at random (MAR) and missing completely conditionally at random (MCCAR) treatment data. Through asymptotic efficiency analysis, we prove that the MAR estimator, which leverages partially-observed units, is both valid and more efficient than the MCCAR estimator when MCCAR assumptions hold. This result provides formal justification for preferring MAR-based estimation in policy learning under both missing data settings. Our comprehensive experiments using synthetic and semi-synthetic datasets confirm that correctly specifying the missingness mechanism is crucial: misspecified estimators remain biased regardless of sample size, while our estimators achieve near-oracle performance when assumptions are satisfied. Our work provides practitioners with theoretically grounded, empirically validated tools for robust policy learning in the presence of missing treatment data.

Current alignment approaches typically focus on emulating human behavior using static representations of human preferences, failing to capture the dynamic, context-dependent nature of real-world human-AI interactions. In this paper, we argue for a shift from static and emulative to interactive and complementary alignment, where preferences emerge through interaction and alignment is defined not by satisfying preferences alone. We first formalize this gap by contrasting existing alignment with a trajectory-level view in which human and model behavior co-evolve over time. Because these interaction dynamics have not been adequately captured within existing ML formulations, we ground this perspective in insights from an interdisciplinary workshop. We draw on lessons from social-science accounts of human-human collaboration and then argue that human-AI systems amplify these dynamics, introducing new asymmetries that make reasoning about uncertainty harder and introduce new coordination challenges. Based on these lessons and new challenges, we conclude by outlining a research agenda for developing AI systems that align with humans in interaction, requiring an interdisciplinary synthesis of machine learning and the social and decision sciences.

Ionizing radiation from cosmic rays and gammas can induce discontinuous jumps in the environmental charge of superconducting qubits (charge jumps), causing correlated errors that challenge fault-tolerant quantum computing while simultaneously providing a detection signature for quantum sensing applications. Current detection methods operate offline, introducing latency incompatible with in-the-loop qubit control. In this paper, an online detector of charge jumps for superconducting qubits, based on a dilated causal convolutional neural network (DCCNN) designed for in-the-loop deployment on the Quantum Instrumentation Control Kit (QICK) platform, is presented. The network is trained on synthetic Ramsey tomography scans generated from qubit templates measured at the Northwestern Experimental Underground Site (NEXUS) at Fermilab, and translated to FPGA firmware via hls4ml with ap_fixed$\langle 16,6 \rangle$ quantization, reaching a per-inference latency of $6.19 μ$s on the Zynq UltraScale+ RFSoC ZCU216. At this operating point the DCCNN matches the detection efficiency of the established offline $χ^2$ algorithm ($0.843 \pm 0.022$ vs. $0.866 \pm 0.020$ on $|Δq| \in [0.1, 0.5] e$ at matched false-positive rate), while requiring no per-qubit hyperparameter tuning. This shifts charge-jump detection from a post-hoc diagnostic to a control-loop primitive, enabling adaptive protocols that respond to radiation-induced events in situ, with applications to quantum-computing error mitigation and to the use of superconducting qubits as particle detectors.