Harvard University（哈佛大学）

Autonomous driving requires both safe and efficient planning decisions in dynamic 3D environments. Although recent Vision/Video-Action models learn policies directly from visual observations and scale well with advances in vision transformers and large-scale training data, they often lack explicit geometric grounding and future-aware spatial guidance, limiting their ability to balance collision avoidance and driving progress. In this work, we propose GeoWorldAD, a geometry world action model that grounds trajectory planning in ego-aligned 3D space and anticipates short-horizon scene evolution with latent future geometry tokens. Present geometry provides essential spatial constraints for safe planning, while future geometry reveals how surrounding agents and ego-centric free space may evolve, reducing overly conservative decisions without sacrificing safety. To efficiently exploit these geometric cues, GeoWorldAD progressively aggregates multi-scale present geometry and latent future geometry through iterative trajectory refinement. Experiments on NAVSIM v1 and v2 demonstrate state-of-the-art performance, highlighting the effectiveness of explicit 3D geometry grounding and future geometry world modeling for safe and efficient autonomous driving.

A hallmark of neurodegenerative diseases such as Alzheimer's and Parkinson's is the aberrant aggregation of proteins into amyloid fibrils, and small molecules that selectively bind to these fibrils hold promise as diagnostics, imaging probes, and therapeutics. Predicting how such ligands bind to fibril targets, however, presents two fundamental challenges. First, resolved co-crystal structures of amyloid-ligand complexes are exceptionally scarce; even with recent advances in cryo-EM only a handful have been structurally characterized, making supervised training of docking models impractical for this target class. Second, amyloid fibrils present a binding mode fundamentally different from globular proteins: ligands intercalate into longitudinal cross-$β$ grooves and stack cooperatively along the fibril axis, a geometry that existing docking models are not designed to capture. To address these challenges, we present CORAL (COopeRative Amyloid Ligand docking), a reinforcement learning framework that trains a generative docking model to produce ligand pose distributions tailored to the cross-$β$ groove geometry. Our reward explicitly incorporates cooperative ligand-ligand stacking energy alongside protein-ligand docking affinity, directly capturing the distinctive binding geometry of amyloid fibrils. We further introduce a curated evaluation set of amyloid-ligand complexes constructed from model-generated poses validated by domain experts. Experiments on both experimentally resolved structures and this evaluation set demonstrate improved pose quality and binding affinity correlation over existing docking baselines.

An international agreement to limit AI development could be crucial to mitigate risks from AI. However, it remains unclear which conditions should determine when the limiting measures are relaxed. We survey existing international agreements, outline what properties appropriate conditions should satisfy, list possible conditions, and finally give a recommendation in an example scenario. We recommend a fixed time period after which a new organization established at the start of the period specifies conditions that address when AI development can be safely conducted, with a possibility of withdrawal in extraordinary circumstances. We hope to illustrate the considerations that would likely go into an international agreement to limit AI.

In multi-task learning (MTL) negative transfer is often considered as an optimization artifact, but it can also be viewed as a consequence of limited shared capacity and weak task redundancy. We investigate this effect through a Capacity--Redundancy (CR) identity that decomposes the sum of per-task predictive informations into joint predictive information that includes label redundancy defined via total correlation (TC), and a residual coupling term that quantifies interference left unresolved by the shared representation. Additionally, we show two key results: (i) a clustering-gap decomposition that gives a necessary and sufficient condition for clustered sharing to outperform global sharing, and (ii) a gradient--TC bridge in a Gaussian multi-task model that formally justifies gradient cosine similarity as a proxy for redundancy ordering. Empirically, we estimate the residual coupling $Δ$ from validation residual correlations, showing that clustered LoRA substantially reduces $\widehatΔ$, outperforms size-matched random partitions, and results in statistically significant gains with multi-seed confidence intervals.

This paper considers a multi-environment factor model in which high-dimensional covariates are collected from heterogeneous environments, with auxiliary labels available in a subset of these environments. The joint distribution of the covariates may vary across environments, whereas the latent structure is decomposed into invariant factors with shared loadings and heterogeneous factors with environment-specific loadings. Such a model is motivated by transfer learning and latent factor regression, where one seeks stable low-dimensional representations for both interpretation and robust out-of-sample prediction of the response $Y$. Leveraging the invariance principle, we show that the invariant and heterogeneous factors are disentangled under a minimal structural condition. Based on this, we propose ATLAS, an Auxiliary-label and invariance-guided Transfer via Latent Alignment across heterogeneous environmentS. ATLAS is a unified procedure that leverages the invariance principle to separate aligned invariant and unaligned heterogeneous factors, and further exploits supervision from auxiliary labels to extract prediction-invariant and transferable factors from those unaligned heterogeneous factors. ATLAS yields near-oracle performance for downstream latent factor regression, enables transferable prediction in new environments through the full latent signal when auxiliary labels are available, and reduces to robust invariant-factor-only prediction otherwise. We establish sharp non-asymptotic error bounds for recovering invariant and heterogeneous factors, identifying all the response-invariant factors, and estimating the invariant signal in $Y$.

Sepsis is a leading cause of mortality, yet optimal treatment policies remain contested. Existing reinforcement learning (RL) approaches learn fixed strategies for sepsis treatment, limiting adaptability to changing clinical objectives during inference. We propose EHRMPC, a framework that decouples learning patient dynamics from optimizing treatment by training a patient digital twin in the form of a generative electronic health record (EHR) model. The digital twin predicts clinical trajectories under interventions and enables model predictive control (MPC) to optimize treatments via inference-time planning over simulations. We evaluate EHR-MPC on a multicenter ICU sepsis cohort spanning 8 hospitals in the Mass General Brigham health system using both off-policy importance sampling and on-policy simulation-based evaluation. Relative to RL baselines, EHR-MPC achieves comparable off-policy performance and improved simulation performance. Unlike RL, this work frames sepsis treatment optimization as inference-time control over learned patient dynamics, establishing a general framework for decision making with generative clinical models.

Agentic reinforcement learning requires assigning credit to environment-facing actions such as searches, clicks, edits, navigation commands, and object interactions. Standard GRPO uses the final verifier outcome as a uniform advantage over all action tokens. This outcome signal is useful but structurally incomplete: it punishes useful exploration in failed rollouts and reinforces redundant or regressive actions in successful rollouts. We propose TRIAGE, a role-typed credit assignment framework that adds a semantic role axis to outcome credit. A structured judge classifies each segment as decisive progress, useful exploration, no-progress infrastructure, or regression, and a fixed role-conditioned rule maps these labels to bounded segment-level process rewards. This keeps verifier outcomes as the source of optimization direction while correcting the two main blind spots of outcome-only credit. We further show that the Bayes-optimal role-measurable correction is the L2 projection of the per-segment advantage residual onto the role variable, and that TRIAGE's fixed role constants approximate this projection, reducing advantage estimation error whenever the judge is reliable; we connect this to lower-variance policy gradients. Across ALFWorld, Search-QA, and WebShop, TRIAGE improves success rates over GRPO for two policy models and outperforms both a scalar judge-derived process reward and an outcome-supervised shared-backbone value baseline. Ablations show that the gain comes from role typing rather than merely adding dense rewards: reliable detection of regression inside successful trajectories is the dominant contributor, while exploration credit provides a consistent secondary gain; on completed ALFWorld and WebShop rollouts, TRIAGE also reduces environment-facing turns by an additional $10.4\%$ and $14.8\%$ relative to GRPO.

We study a sequential resource allocation problem motivated by adaptive network recruitment, in which a limited budget of identical resources must be allocated over multiple rounds to individuals with stochastic referral capacity. Successful referrals endogenously generate future decision opportunities while allocating additional resources to an individual exhibits diminishing returns. We first show that the single-round allocation problem admits an exact greedy solution based on marginal survival probabilities. In the multi-round setting, the resulting Bellman recursion is intractable due to the stochastic, high-dimensional evolution of the frontier. To address this, we introduce a population-level surrogate value function that depends only on the remaining budget and frontier size. This surrogate enables an exact dynamic program via truncated probability generating functions, yielding a planning algorithm with polynomial complexity in the total budget. We further analyze robustness under model misspecification, proving a multi-round error bound that decomposes into a tight single-round frontier error and a population-level transition error. Finally, we evaluate our method on real-world inspired recruitment scenarios.

Interpretability research often predicts model responses to targeted mechanistic interventions. But can we predict responses to unseen input data? We propose and demonstrate this alternate objective by using model internals to predict their out-of-distribution (OOD) behavior. We train hundreds of Transformers on simple synthetic tasks, where perfect in-distribution accuracy is compatible with multiple OOD generalization rules. We successfully use attention patterns -- observed only on in-distribution data -- to predict which rule each model follows on OOD data. Our experiments decouple the mechanistic faithfulness of our interpretation from its predictive value; ablations reveal such internal patterns can suppress rather than support the rule they predict, showing observational analysis can forecast behavior even when causal analysis fails to support a simple cause-effect link. Our findings are a proof-of-concept for a new interpretability objective: understanding model internals to predict behavior and assess reliability under distribution shift.

Illusions are entertaining, but they are also a useful diagnostic tool in cognitive science, philosophy, and neuroscience. A typical illusion shows a gap between how something `really is' and how something `appears to be', and this gap helps us understand the mental processing that led to how something appears to be. Illusions are also useful for investigating artificial systems, and much research has examined whether computational models of perception fall prey to the same illusions as people. Here, I invert the standard use of perceptual illusions to examine basic processing errors in current vision language models. I present these models with illusory-illusions, neighbors of common illusions that should not elicit processing errors. These include such things as perfectly reasonable ducks, crooked lines that truly are crooked, circles that seem to have different sizes because they are, in fact, of different sizes, and so on. I show that many current vision language systems mistakenly see these illusion-illusions as illusions. I suggest that such failures are part of broader failures already discussed in the literature.

In the rapidly evolving domain of artificial intelligence, Large Language Models (LLMs) play a crucial role due to their advanced text processing and generation abilities. This study introduces a new strategy aimed at harnessing on-device LLMs in invoking software APIs. We meticulously compile a dataset derived from software API documentation and apply fine-tuning to LLMs with capacities of 2B, 3B and 7B parameters, specifically to enhance their proficiency in software API interactions. Our approach concentrates on refining the models' grasp of API structures and syntax, significantly enhancing the accuracy of API function calls. Additionally, we propose \textit{conditional masking} techniques to ensure outputs in the desired formats and reduce error rates while maintaining inference speeds. We also propose a novel benchmark designed to evaluate the effectiveness of LLMs in API interactions, establishing a foundation for subsequent research. Octopus, the fine-tuned model, is proved to have better performance than GPT-4 for the software APIs calling. This research aims to advance automated software development and API integration, representing substantial progress in aligning LLM capabilities with the demands of practical software engineering applications.

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.

In recent years, the integration of multimodal machine learning in wellbeing assessment has offered transformative potential for monitoring mental health. However, with the rapid advancement of Vision-Language Models (VLMs), their deployment in clinical settings has raised concerns due to their lack of transparency and potential for bias. While previous research has explored the intersection of fairness and Explainable AI (XAI), its application to VLMs for wellbeing assessment and depression prediction remains under-explored. This work investigates VLM performance across laboratory (AFAR-BSFT) and naturalistic (E-DAIC) datasets, focusing on diagnostic reliability and demographic fairness. Performance varied substantially across environments and architectures; Phi3.5-Vision achieved 80.4% accuracy on E-DAIC, while Qwen2-VL struggled at 33.9%. Additionally, both models demonstrated a tendency to over-predict depression on AFAR-BSFT. Although bias existed across both architectures, Qwen2-VL showed higher gender disparities, while Phi-3.5-Vision exhibited more racial bias. Our XAI intervention framework yielded mixed results; fairness prompting achieved perfect equal opportunity for Qwen2-VL at a severe accuracy cost on E-DAIC. On AFAR-BSFT, explainability-based interventions improved procedural consistency but did not guarantee outcome fairness, sometimes amplifying racial bias. These results highlight a persistent gap between procedural transparency and equitable outcomes. We analyse these findings and consolidate concrete recommendations for addressing them, emphasising that future fairness interventions must jointly optimise predictive accuracy, demographic parity, and cross-domain generalisation.

We introduce GeCo, a geometry-grounded metric for jointly detecting geometric deformation and occlusion-inconsistency artifacts in static scenes. By fusing residual motion and depth priors, GeCo produces interpretable, dense consistency maps that reveal these artifacts. We use GeCo to systematically benchmark recent video generation models, uncovering common failure modes, and further employ it as a training-free guidance loss to reduce deformation artifacts during video generation.

Explainable Artificial Intelligence (XAI) is essential for trustworthy AI in healthcare, yet many existing methods rely on technical explanations that are difficult for clinicians and patients to interpret. We introduce Visualized Learning for Machine Learning (VL4ML), a human-centered explainability framework that communicates model predictions and uncertainty through intuitive visual representations rather than numerical or post-hoc explanations. By encoding diagnostic information in colors, patterns, and spatial structures, VL4ML enables users to interpret predictions without requiring knowledge of model internals or statistical expertise. We demonstrate the framework across multiple clinical tasks, including classification, regression, longitudinal prediction, and multimodal analysis. Its effectiveness was evaluated through a human-centered study involving 158 participants (39.2% clinical professionals) and an expert interpretability assessment. More than 79% of participants positively rated the visual explanations across evaluation dimensions, 84.0% found them more memorable than numeric outputs, and 76.9% reported faster decision-making. Over 82% successfully perceived uncertainty embedded in the visual representations without prior statistical training. No significant differences were observed between clinicians and non-clinicians or between male and female participants, indicating broad accessibility. These results suggest that VL4ML complements existing XAI and uncertainty quantification methods by providing intuitive, universally interpretable visual explanations that support transparent and trustworthy clinical decision-making.

Predictive world models enable robots to plan by imagining the outcomes of their actions, but their value for control hinges on generating many rollouts quickly. This creates a bottleneck for diffusion-based world models: multistep sampling makes each rollout expensive, limiting large-scale action search at inference time. We introduce DriftWorld, an action-conditioned world model based on drifting generative models. Rather than denoising iteratively at inference, DriftWorld learns an action-conditioned drift during training, allowing it to generate future frames from the current observation and a candidate action sequence in a single forward pass at 30+ fps, which is 17x faster on average than diffusion based baselines. We evaluate DriftWorld on standard vision-based robotic manipulation benchmarks, including Bridge-V2, RT-1, Language Table, Push-T, and Robomimic. By producing rollouts that are both accurate and fast, DriftWorld achieves state-of-the-art decision-making performance with far less inference time than diffusion-based world model baselines. Beyond online control, DriftWorld can also serve as an offline simulator for ranking real-world robot policies, with rollout-based scores correlating with ground truth at up to 0.99. These results show that drifting models are a strong fit for robot world modeling, where fast, high-quality imagination directly supports planning and policy evaluation.

Non-rigid point cloud registration is central to soft-tissue shape analysis, but large deformations, noise, and outliers make correspondence estimation challenging. Most learning-based methods rely on local objectives such as Chamfer distance, which encourage point-wise proximity but do not constrain the global plausibility of the predicted deformation field. We address this limitation with DINE, a maximum a posteriori framework that augments distance-based registration with a learned statistical prior over displacement vector fields. DINE is applied to two registration backbones, Robust-DefReg and DefTransNet, using a two-stage strategy: a first-stage model is trained with Chamfer distance, its predicted deformation fields are used to estimate a prior, and the model is then refined with a combined distance and negative log-prior objective. We compare a full-field PCA Gaussian prior with a per-vector normalizing-flow prior. Experiments on DeformedTissue and SynBench show lower mean Chamfer distance under deformation and corruption. On DeformedTissue, DINE-PCA reduces Chamfer distance by approximately 27--69\% relative to the corresponding Stage-1 backbone across deformation levels, and improves robustness by up to 66\% for outliers and 83\% for Gaussian noise. On SynBench, improvements are modest at the smallest deformation levels and reach approximately 59--79\% from moderate to severe deformation. These results suggest that global deformation plausibility is an important constraint for reliable soft-tissue point cloud registration. (The code will be published soon.)

Where do objective functions come from? How do we select what goals to pursue? Human intelligence is adept at synthesizing new objective functions on the fly. How does this work, and can we endow artificial systems with the same ability? This paper proposes an approach to answering these questions, starting with the concept of a subjective function, a higher-order objective function that is endogenous to the agent (i.e., defined with respect to the agent's features, rather than an external task). Expected prediction error is studied as a concrete example of a subjective function. This proposal has many connections to ideas in psychology, neuroscience, and machine learning.

Large language models (LLMs) are increasingly used to answer health-related questions and support healthcare workflows, yet evidence for their safety still relies heavily on static benchmarks that can rapidly become obsolete or be optimized against. Here we introduce a Dynamic, Automatic, and Systematic (DAS) red-teaming audit framework that continuously stress-tests LLMs for health across four safety-critical axes: robustness, privacy, bias/fairness, and hallucination/factual inaccuracies. Validated against board-certified clinicians with high concordance, a suite of adversarial agents autonomously mutates health-related test cases to uncover vulnerabilities in real time. Applying DAS to 15 proprietary and open-source LLMs revealed a profound gap between high static benchmark performance and low dynamic reliability--the "Benchmarking Gap". Despite median MedQA accuracy exceeding 80\%, 94\% of previously correct answers failed under dynamic robustness testing. This brittleness generalized to the realistic, open-ended HealthBench dataset, where top-tier models exhibited failure rates exceeding 70\% and sharp shifts in model rankings across evaluations, suggesting that high scores on established static benchmarks may reflect superficial memorization. We observed similarly high failure rates across other domains: privacy leaks were elicited in 86\% of scenarios, cognitive-bias priming altered recommendations in 81\% of fairness tests, and hallucination rates exceeded 74\% in widely used models. By converting LLM safety evaluation for health from a static checklist into a living adversarial audit, DAS provides a scalable framework for surfacing latent risks before such systems are deployed in consumer-facing health assistants, clinician-facing tools, and broader healthcare workflows. Code is available at https://github.com/JZPeterPan/DAS-Medical-Red-Teaming-Agents.

Social scientists are increasingly turning to unstructured datasets to unlock new empirical insights, e.g., estimating descriptive statistics of or causal effects on quantitative measures derived from text, audio, or video data. In many settings, unsupervised analysis is of primary interest, in that the researcher does not want to (or cannot) manually pre-specify all important aspects of the unstructured data to measure; they are interested in "discovery." This paper proposes a general and flexible framework for pursuing such discovery from unstructured data in a statistically principled way. The framework leverages recent methods from the literature on AI interpretability to map unstructured data points to high-dimensional, sparse, and interpretable "concept embeddings"; computes statistics from these concept embeddings for testing interpretable, concept-by-concept hypotheses; performs selective inference on these hypotheses using algorithms validated by new results in high-dimensional central limit theory, producing a selected set ("discoveries"); and both generates and evaluates human-interpretable natural language descriptions of these discoveries. The proposed framework has few researcher degrees of freedom, is robust to data snooping and other post-selection inference concerns, and facilitates fast and inexpensive sensitivity analysis and replication. Applications to recent descriptive and causal analyses of unstructured data in empirical economics are explored.

As Large Language Models (LLMs) are increasingly deployed as autonomous agents in high-stakes domains, understanding contextual factors that may modulate their decision-making becomes critical. While LLMs are trained to perceive and resonate with users' emotions, it remains unclear whether induced emotion can influence their sequential decision-making. We investigate this question using the Iowa Gambling Task (IGT), a classic psychological paradigm for studying decision-making under uncertainty, combined with an imagination-based emotion induction procedure. We first validate the feasibility of this paradigm by confirming that LLMs can sense strong, distinguishable emotions from context and that LLM agents can learn from sequential interactions in a human-like pace. With the validated setup, we find that, different from humans, induced emotion does not significantly bias the decision dynamics of LLM agents on average. However, the effects of anger are conditioned: inducing anger makes LLM agents less sensitive to penalties for bad decisions, and in early stages of the game, anger can lower exploration, locking decisions into a few choices early. These findings reveal the subtle yet distinct effects of induced emotion on LLM decision-making compared to human behavior, and provide a tool for future research on affective modulation of LLM agents.

Hematoxylin and eosin (H&E) staining is the cornerstone of histopathology, yet scalable, quantitative analysis of H&E whole-slide images (WSIs) remains a central challenge in computational pathology. We present Atlas H&E-TME, an AI-based system built on the Atlas family of pathology foundation models that predicts tissue quality, tissue region, and cell type labels across multiple cancer types, yielding over 4,500 quantitative readouts per slide at cell-level resolution. A key challenge to validating such systems is overcoming morphological ambiguity inherent to H&E-only ground truth and the limited scalability of more informed references drawing on modalities such as immunohistochemistry (IHC). We address this with a dual validation framework combining biologically grounded depth with technical and morphological breadth. For depth, we propose an IHC-informed multi-pathologist consensus protocol that substantially improves inter-rater agreement over conventional H&E-only annotation. This yields a molecularly grounded reference against which we compare Atlas H&E-TME and pathologists working from H&E alone. For breadth, we benchmark Atlas H&E-TME on over 200,000 high-confidence H&E-only pathologist annotations across 1,500+ cases spanning eight cancer types and their most common metastatic sites, with subtypes covering >90% of clinical cases per cancer type, drawn from 25+ sources and 8+ scanner models. Benchmarked against the IHC-informed consensus, Atlas H&E-TME matches or exceeds pathologist H&E-only performance and generalizes consistently and robustly across this broad morphological and technical scope. In doing so, Atlas H&E-TME turns the H&E slide -- the most ubiquitous data in pathology -- into a scalable, quantitative window into the tumor and its microenvironment, laying a foundation for the next generation of tissue-based biomarkers in translational and clinical research.

Large Language Models memorize vast amounts of training data, raising concerns regarding privacy, copyright infringement, and safety. Machine unlearning seeks to remove the influence of a targeted forget set while preserving model performance, ideally approximating a model retrained from scratch without the forget set. Existing approaches aim to achieve this by updating model parameters via gradient-based methods. However, these updates are computationally expensive, lead to irreversible weight changes, and degrade when the model is quantized for deployment. A recent alternative to changing model weights is activation engineering, where activations are changed during inference to steer model behavior. Despite circumventing weight editing, naive activation steering introduces its own failure modes, as a single global steering vector applies the same intervention to every input, leading to unintended changes in model behavior. We introduce Inference-Time Unlearning via Gated Activation Redirection (GUARD-IT), a training- and gradient-free method that unlearns via input-dependent activation steering at inference time. The resulting intervention is applied as a norm-preserving rotation in the residual stream, leaving model weights untouched. Experiments on TOFU and MUSE show that GUARD-IT matches or exceeds 12 gradient-based baselines across three model scales, while being the only method to simultaneously preserve utility, suppress memorization, and avoid catastrophic collapse across all settings. GUARD-IT further supports continual unlearning without retraining, and remains effective under quantization, a scenario in which parameter-editing methods degrade.

Efficient coding theory predicts that biological perceptual systems compress sensory input optimally under resource constraints, with the systematic structure of errors reflecting the geometry of that compression. Here we operationalize this principle using rate-distortion theory (RDT) to characterize how any system - biological or artificial - trades representational fidelity for informational efficiency. Treating stimulus-response behavior as an effective communication channel, we infer rate-distortion (RD) frontiers directly from confusion matrices and summarize each system with three geometric signatures: slope (beta), curvature (kappa), and area under the RD curve (AUC), capturing the marginal cost, abruptness, and overall efficiency of the accuracy-compression trade-off respectively. Applying this framework to human psychophysical data and 18 deep vision models across 12 families of controlled image perturbations at graded severities, we find that both biological and artificial systems follow a common lossy-compression principle but occupy systematically different regions of RD space. Humans exhibit smooth, flexible trade-offs characteristic of near-optimal efficient coding, while deep networks operate in steeper, more brittle regimes even at matched accuracy, with geometry dissociable from performance across training regimes. Critically, behavioral RD signatures track internal representational geometry, evidenced by the behaviorally inferred compression structure correlating with internal representational dissimilarity across all models. These results establish RD geometry as a compact diagnostic of perceptual compression strategy that recovers mechanistically interpretable structure in internal representations from behavioral input alone and extends naturally to the direct characterization of compression geometry in neural population activity.

Large language models (LLMs) and medical AI tools are routinely used by physicians and patients for medical advice, yet their clinical safety profiles remain poorly characterized. We present NOHARM (Numerous Options Harm Assessment for Risk in Medicine), a 1,100-task benchmark of primary care-to-specialist consultation cases to measure the frequency and severity of potentially harmful errors from LLM-generated medical consultation recommendations. NOHARM covers 10 specialties, with 12,747 expert annotations for 4,249 clinical management options. Across 20 notable LLMs and 4 widely used retrieval-augmented generation (RAG) clinical AI tools, direct application of recommendations carried potential for severe harm in up to 24.6% of cases, with errors of omission accounting for more than 80% of severe errors. Harm potential was not uniform across systems, with clinical AI tools outperforming generalist LLMs, and multi-agent AI teaming further improving performance in generalist models. In a randomized study of 101 U.S.-licensed generalist physicians, AI assistance improved physician performance compared to conventional resources. However, AI-assisted physicians frequently omitted valuable AI-generated recommendations and still scored lower than many AI systems alone. Had those recommendations been incorporated, combined human-AI responses would have outperformed both the human and AI system as used, suggesting complementary strengths and unrealized potential in human-AI teaming. Collectively, these results show that despite strong performance on medical knowledge benchmarks, widely used AI tools can produce medical consultation advice with the potential for severe harm, and highlight the need for explicit measurement of clinical safety. The benchmark and leaderboard are publicly available to support ongoing evaluation and improvement of AI systems used for clinical care.

Targeted interventions on language models, such as unlearning or model editing, aim to modify specific information, but their effects often propagate to related, unintended areas (e.g., removing virology content may degrade performance on allergies); these side-effects are commonly referred to as the ripple effect. We introduce RippleBench-Maker, an automatic pipeline that retrieves semantic neighbors of any source concept from a knowledge repository and generates multiple-choice questions at varying semantic distances. We instantiate this framework using WikiRAG, an open-source RAG system over English Wikipedia, to construct RippleBench-WMDP-Bio (584 seed topics, 352,961 questions), and evaluate eight unlearning methods on Llama3-8B-Instruct. All eight exhibit accuracy drops that are largest near the unlearned target and decay with semantic distance, each with a distinct propagation profile. We replicate these findings across Mistral-7B, Zephyr-7B, and Yi-34B; cross-model delta curves are nearly identical, suggesting ripple effects are a property of the unlearning method rather than the base model. We validate all major pipeline stages using a four-experiment Mechanical Turk study (5,200+ responses, 61 workers). We release all code, data, and infrastructure.

We study the fundamental problem of calibrating a linear binary classifier of the form $σ(\hat{w}^\top x)$, where the feature vector $x$ is Gaussian, $σ$ is a link function, and $\hat{w}$ is an estimator of the true linear weight $w^\star$. By interpolating with a noninformative $\textit{chance classifier}$, we construct a well-calibrated predictor whose interpolation weight depends on the angle $\angle(\hat{w}, w_\star)$ between the estimator $\hat{w}$ and the true linear weight $w_\star$. We establish that this angular calibration approach is provably well-calibrated in a high-dimensional regime where the number of samples and features both diverge, at a comparable rate. The angle $\angle(\hat{w}, w_\star)$ can be consistently estimated. Furthermore, the resulting predictor is uniquely $\textit{Bregman-optimal}$, minimizing the Bregman divergence to the true label distribution within a suitable class of calibrated predictors. Our work is the first to provide a calibration strategy that satisfies both calibration and optimality properties provably in high dimensions. Additionally, we identify conditions under which a classical Platt-scaling predictor converges to our Bregman-optimal calibrated solution. Thus, Platt-scaling also inherits these desirable properties provably in high dimensions.

The Circle of Willis (CoW) is an important network of arteries connecting major circulations of the brain. Its vascular architecture is believed to influence the risk, severity, and outcome of serious neurovascular diseases. However, characterizing the highly variable CoW anatomy remains a manual and time-consuming expert task. The CoW is commonly imaged by two non-invasive angiographic imaging modalities, magnetic resonance angiography (MRA) and computed tomography angiography (CTA), yet few datasets with annotated CoW anatomy exist, and there have been no established benchmarks for comparing CoW segmentation algorithms. We organized the TopCoW benchmark challenge alongside the release of an annotated CoW dataset with 125 paired MRA and CTA scans from the same patients. Voxel-level annotations for 13 vessel components were created using virtual reality technology and verified by clinical experts. Participants submitted algorithms for CoW segmentation and variant classification, which we evaluated on internal and external test sets comprising 226 scans from over five centers. The benchmark includes voxel-level segmentation, CoW component detection, CoW variant classification, and two clinical application tasks. We received submissions from over 250 participants across six continents. Top-performing teams achieved over 90% Dice scores for CoW segmentation, over 80% F1 scores for detecting key vessel components, and over 70% balanced accuracy in CoW variant classification across nearly all test sets. The best algorithms also supported clinically relevant downstream tasks by accurately classifying fetal-type posterior cerebral arteries and localizing aneurysms in relation to CoW anatomy. This benchmark demonstrated the utility of CoW segmentation algorithms for some downstream clinical applications with explainability.

Machine learning progress is often attributed to scaling model size and dataset volume, yet the composition of data can be just as consequential. Empirical findings repeatedly show that combining datasets from different domains yields nontrivial interactions. For instance, adding code improves mathematical reasoning, while certain mixtures introduce interference that reduces model performance. We refer to these effects collectively as data synergy, where the contribution of multiple domains exceeds or falls short of the sum of their isolated contributions. In this work, we formalize and quantify data synergy in language model pretraining. Leveraging observational variation across open-weight LLMs with diverse pretraining mixtures, we estimate both direct domain-to-benchmark synergy (how one domain contributes to performance on another) and a second-order domain-domain synergy (capabilities that require co-occurrence of multiple domains). Our framework improves predictive accuracy over domain-agnostic scaling laws and recovers stable synergy estimates. We validate these estimates by training models on predicted optimal and predicted anti-optimal mixtures and confirm that our synergy estimates correctly predict performance rankings.

Diffusion policies have shown promising empirical performance in representing and learning complex maneuvers for robots using behavior cloning (BC). In this paper, we explore training diffusion policies from scratch using reinforcement learning (RL) for multi-task robotic manipulation. Specifically, we aim to train a single diffusion policy for block-pushing tasks with multiple shapes. The proposed framework features a simple policy loss function, which is a reweighted evidence lower bound used in BC-based diffusion policy training and can seamlessly serve as the policy learning module in RL algorithms. To address the exploration challenges arising from the absence of demonstrations, we incorporate reverse curriculum generation and objective-centric representations. Combined with the expressiveness of diffusion policies, our design supports learning of multi-task block-pushing policies in our sparse-reward simulation setting. We further evaluate whether the trained diffusion policy transfers in zero-shot to real-world tasks under varying environmental conditions including goal positions, block shapes, block weights and surface friction, providing evidence that this pipeline can transfer to our real-world block-pushing setup under the tested variations.