We introduce AVACraft, a multimodal StarCraft II benchmark supporting both Multi-Agent Reinforcement Learning (MARL) and Vision-Language Model (VLM) paradigms. Unlike SMAC-family environments that rely on abstract state representations and exclude VLMs, AVACraft provides RGB visuals, natural language observations, and structured state information, enabling systematic comparison between training-based and zero-shot methods across 21 scenarios spanning micromanagement, coordination, and strategic planning. We establish comprehensive baselines: six MARL algorithms (IQL, QMIX, QTRAN, VDN, MAPPO, IPPO) with Swin-Transformer backbones trained for 5M steps, and multiple VLMs including proprietary (GPT-4o) and open-source (Qwen3-VL) models. Results reveal complementary strengths-MARL peaks at 19.3% win rate after 5M steps, while VLMs achieve 75-90% zero-shot with human-aligned decisions-exposing trade-offs between training efficiency, performance ceilings, interpretability, and deployment cost. Code: https://github.com/camel-ai/VLM-Play-StarCraft2.

Frontier language models demonstrate increasing ability at solving reasoning problems, but their performance is often inflated by circumventing reasoning and instead relying on their expanding knowledge and memorisation capacity. We introduce LINGOLY-TOO, a challenging reasoning benchmark of 1,203 questions and a total of 6,995 sub-questions that counters these shortcuts by applying expert-designed obfuscations to Linguistics Olympiad problems. These obfuscations preserve the underlying solution logic while reducing the likelihood problems are solvable with via knowledge or memorisation. Our experiments show that models exploit shortcuts on the original question as performance markedly drop upon obfuscation. Even the best reasoning models remain highly sensitive, with scores dropping from around 0.59 on original problems to 0.48 after obfuscation. LINGOLY-TOO disentangles reasoning from knowledge, offering a clearer measure of true reasoning capabilities.

Depression is a mental disorder and can cause a variety of symptoms, including psychological, physical, and social. Speech has been proved an objective marker for the early recognition of depression. For this reason, many studies have been developed aiming to recognize depression through speech. However, existing methods rely on the usage of only the spontaneous speech neglecting information obtained via read speech, use transcripts which are often difficult to obtain (manual) or come with high word-error rates (automatic), and do not focus on input-conditional computation methods. To resolve these limitations, this is the first study in depression recognition task obtaining representations of both spontaneous and read speech, utilizing multimodal fusion methods, and employing Mixture of Experts (MoE) models in a single deep neural network. Specifically, we use audio files corresponding to both interview and reading tasks and convert each audio file into log-Mel spectrogram, delta, and delta-delta. Next, the image representations of the two tasks pass through shared AlexNet models. The outputs of the AlexNet models are given as input to a multimodal fusion method. The resulting vector is passed through a MoE module. In this study, we employ three variants of MoE, namely sparsely-gated MoE and multilinear MoE based on factorization. Findings suggest that our proposed approach yields an Accuracy and F1-score of 87.00% and 86.66% respectively on the Androids corpus.

Graph-based learning excels at capturing interaction patterns in diverse domains like recommendation, fraud detection, and particle physics. However, its performance often degrades under distribution shifts, especially those altering network connectivity. Current methods to address these shifts typically require retraining with the source dataset, which is often infeasible due to computational or privacy limitations. We introduce Test-Time Structural Alignment (TSA), a novel algorithm for Graph Test-Time Adaptation (GTTA) that adapts a pretrained model to align graph structures during inference without the cost of retraining. Grounded in a theoretical understanding of graph data distribution shifts, TSA employs three synergistic strategies: uncertainty-aware neighborhood weighting to accommodate neighbor label distribution shifts, adaptive balancing of self-node and aggregated neighborhood representations based on their signal-to-noise ratio, and decision boundary refinement to correct residual label and feature shifts. Extensive experiments on synthetic and real-world datasets demonstrate TSA's consistent outperformance of both non-graph TTA methods and state-of-the-art GTTA baselines.

World models (WMs) represent the frontier of sample-efficient reinforcement learning, but their complexity leaves many promising improvements unrealized due to the significant expertise and effort required to identify and integrate them. Inspired by Rainbow, which showed that individually known improvements to DQN complement each other and can be effectively combined, we take on this challenge and ask whether the same principle applies to world model agents. We introduce Simulus, a modular token-based WM agent that integrates: (1) a flexible tokenization framework supporting arbitrary combinations of observation and action modalities; (2) intrinsic motivation for epistemic uncertainty reduction; (3) prioritized world model replay; and (4) regression-as-classification for reward and return prediction. Simulus achieves state-of-the-art sample efficiency for planning-free WMs across three diverse benchmarks: visual Atari 100K, continuous-control DMC Proprioception 500K, and symbolic Craftax-1M. Notably, intrinsic motivation proves beneficial even under the tight interaction budgets of sample-efficient RL, despite the risk of wasting scarce interactions on task-irrelevant experience. Ablation studies reveal that each component contributes individually, and their combination yields synergistic gains. Our code and model weights are publicly available at https://github.com/leor-c/Simulus.

Large language models (LLMs) have demonstrated significant utility in real-world applications, exhibiting impressive capabilities in natural language processing and understanding. Benchmark evaluations are crucial for assessing the capabilities of LLMs as they can provide a comprehensive assessment of their strengths and weaknesses. However, current evaluation methods often overlook the inherent randomness of LLMs by employing deterministic generation strategies or relying on a single random sample, resulting in unaccounted sampling variance and unreliable benchmark score estimates. In this paper, we propose a hierarchical statistical model that provides a more comprehensive representation of the benchmarking process by incorporating both benchmark characteristics and LLM randomness. We show that leveraging multiple generations improves the accuracy of estimating the benchmark score and reduces variance. Multiple generations also allow us to define $\mathbb P\left(\text{correct}\right)$, a prompt-level difficulty score based on correct ratios, providing fine-grained insights into individual prompts. Additionally, we create a data map that visualizes difficulty and semantics of prompts, enabling error detection and quality control in benchmark construction.

Autonomous Underwater Vehicles (AUVs) have advanced significantly in obstacle detection and path planning through sonar, cameras, and learning-based methods. However, safe and efficient navigation in cluttered environments remains challenging due to partial observability, turbidity, the limited field-of-view of forward-looking sonar (FLS), and occlusions that obscure obstacle geometry. To address these issues, we propose the Efficient Reactive Obstacle Avoidance Strategy (EROAS), a lightweight framework that augments a standard 2D FLS with a pivoting mechanism, effectively transforming it into a cost-efficient \emph{2.5D sonar}. This design provides vertical information on demand, extending situational awareness while minimizing computational overhead. EROAS integrates three complementary modules: first, Sonar Profile-guided Directional Decision Control (SPD2C) for rapid gap detection and generation of reference commands in both horizontal and vertical planes. Secondly, the Spatial Context Generator (SCG), which maintains a short-term obstacle memory of the past to mitigate partial observability, and finally, a Spatio-Temporal Control Barrier Function (ST-CBF) that enforces forward-invariance of safety constraints by filtering nominal references. Together, these components enable robust, reactive avoidance of obstacles in uncertain and cluttered 3D underwater settings. Simulation and hardware-in-the-loop (HIL) experiments validate the efficacy of the proposed EROAS algorithm, demonstrating improved trajectory efficiency, reduced travel time, and enhanced safety compared to conventional methods such as the Dynamic Window Approach (DWA) and Artificial Potential Fields (APF). https://github.com/AIRLabIISc/EROAS

Inverse folding models play an important role in structure-based design by predicting amino acid sequences that fold into desired reference structures. Models like ProteinMPNN, a message-passing encoder-decoder model, are trained to reliably produce new sequences from a reference structure. However, when applied to peptides, these models are prone to generating repetitive sequences that do not fold into the reference structure. To address this, we fine-tune ProteinMPNN to produce diverse and structurally consistent peptide sequences via Direct Preference Optimization (DPO). We derive two enhancements to DPO: online diversity regularization and domain-specific priors. Additionally, we develop a new understanding on improving diversity in decoder models. When conditioned on OpenFold generated structures, our fine-tuned models achieve state-of-the-art structural similarity scores, improving base ProteinMPNN by at least 8%. Compared to standard DPO, our regularized method achieves up to 20% higher sequence diversity with no loss in structural similarity score.

Multi-modal Large Language Models (MLLMs) exhibit impressive problem-solving abilities in various domains, but their visual comprehension and abstract reasoning skills remain under-evaluated. To this end, we present PolyMATH, a challenging benchmark aimed at evaluating the general cognitive reasoning abilities of MLLMs. PolyMATH comprises 5,000 manually collected high-quality images of cognitive textual and visual challenges across 10 distinct categories, including pattern recognition, spatial reasoning, and relative reasoning. We conducted a comprehensive, and quantitative evaluation of 15 MLLMs using four diverse prompting strategies, including Chain-of-Thought and Step-Back. The best scores achieved on PolyMATH are ~41%, ~36%, and ~27%, obtained by Claude-3.5 Sonnet, GPT-4o and Gemini-1.5 Pro respectively - highlighting the logical and visual complexity of these questions. A further fine-grained error analysis reveals that these models struggle to understand spatial relations and perform drawn-out, high-level reasoning. This is further strengthened by our ablation study estimating MLLM performance when given textual descriptions in place of diagrams. As evidenced by ~4% improvement over textual descriptions as opposed to actual images, we discover that models do not truly comprehend visual diagrams and the spatial information therein, and are thus prone to logical errors. Finally, we evaluate the OpenAI o1 models and find that their performance only matches the human baseline, highlighting the difficulty of the benchmark. The results on PolyMATH highlight the room for improvement in multi-modal reasoning and provide unique insights to guide the development of future MLLMs.

Human dexterity arises from combining high-level task reasoning with finger-level dexterity control and physical compliance at the muscle and skin layers. In robotics, large Vision-Language-Action (VLA) models demonstrate text-conditioned high-level planning across diverse manipulation tasks, typically using pincher grippers. Smaller imitation-learning policies, conversely, show success in dexterous tasks using higher degree-of-freedom (DoF) grippers, but only for limited-scope tasks. However, few approaches combine high-level reasoning with dexterous, robust low-level control, which requires both intelligent control and compliant robot design. We propose a method inspired by the two-channel hypothesis of human motor control that combines these capabilities using a switching controller integrating high-level VLAs and smaller control models. Coordination between the two channels is managed through an event-driven switching mechanism that monitors subtask progression and completion, requiring minimal demonstration data by fine-tuning the VLA to predict event signals and training lightweight subtask-level dexterous policies. This approach is applied to our custom compliant 13-DoF anthropomorphic robotic hand, where compliance can be modulated to evaluate its impact on dexterity and robustness when combined with an autonomous policy. We show that hardware-level compliance in robotic fingers enables passive adaptation to disturbances and improves contact stability. The methodology is validated across a range of language-conditioned dexterous tasks. To demonstrate modularity, we show that adaptation to additional dexterous skills and different compliant hands can be achieved without retraining the VLA model. This provides an efficient, scalable, cross-embodiment approach to dexterity that leverages compliance while retaining the advantages of large AI models.

Large language model-based (LLM) agents are emerging as a powerful enabler of robust embodied intelligence due to their capability of planning complex action sequences. Sound planning ability is necessary for robust automation in many task domains, but especially in surgical automation. These agents rely on a highly detailed natural language representation of the scene. Thus, to leverage the emergent capabilities of LLM agents for surgical task planning, developing similarly powerful and robust perception algorithms is necessary to derive a detailed scene representation of the environment from visual input. Previous research has focused primarily on enabling LLM-based task planning while adopting simple yet severely limited perception solutions to meet the needs for bench-top experiments, but lacks the critical flexibility to scale to less constrained settings. In this work, we propose an alternate perception approach -- a digital twin (DT)-based machine perception approach that capitalizes on the convincing performance and out-of-the-box generalization of recent vision foundation models. Integrating our DT representation and LLM agent for planning with the dVRK platform, we develop an embodied intelligence system and evaluate its robustness in performing peg transfer and gauze retrieval tasks. Our approach shows strong task performance and generalizability to varied environmental settings. Despite a convincing performance, this work is merely a first step towards the integration of DT representations. Future studies are necessary for the realization of a comprehensive DT framework to improve the interpretability and generalizability of embodied intelligence in surgery.

Sequential decision-making algorithms such as multi-armed bandits can find optimal personalized decisions, but are notoriously sample-hungry. In personalized medicine, for example, training a bandit from scratch for every patient is typically infeasible, as the number of trials required is much larger than the number of decision points for a single patient. To combat this, latent bandits offer rapid exploration and personalization beyond what context variables alone can offer, provided that a latent variable model of problem instances can be learned consistently. However, existing works give no guidance as to how such a model can be found. In this work, we propose an identifiable latent bandit framework that leads to optimal decision-making with a shorter exploration time than classical bandits by learning from historical records of decisions and outcomes. Our method is based on nonlinear independent component analysis that provably identifies representations from observational data sufficient to infer optimal actions in new bandit instances. We verify this strategy in simulated and semi-synthetic environments, showing substantial improvement over online and offline learning baselines when identifying conditions are satisfied.

Surgical data science has seen rapid advancement with the excellent performance of end-to-end deep neural networks (DNNs). Despite their successes, DNNs have been proven susceptible to minor "corruptions," introducing a major concern for the translation of cutting-edge technology, especially in high-stakes scenarios. We introduce the SegSTRONG-C challenge dedicated to better understanding model deterioration under unforeseen but plausible non-adversarial "corruption" and the capabilities of contemporary methods that seek to improve it. Built on a dataset generated through counterfactual robotic replay, SegSTRONG-C provides paired clean and "corrupted" samples, enabling reproducible evaluation of model robustness. Participants are challenged to train tool segmentation algorithms on "uncorrupted" data and evaluate them on "corrupted" test domains for the binary robot tool segmentation task. Through comprehensive baseline experiments and participating submissions from widespread community engagement, SegSTRONG-C reveals key themes for model failure and identifies promising directions for improving robustness. The performance of challenge winners, achieving an average 0.9394 DSC and 0.9301 NSD across the unreleased test sets with "corruption" types: bleeding, smoke, and low brightness. This highlights how prior knowledge, customized training strategies, and architectural choice can be leveraged to improve robustness. In conclusion, the SegSTRONG-C challenge has identified practical approaches for enhancing model robustness. However, most approaches rely on conventional techniques that have known limitations. Looking ahead, we advocate for expanding intellectual diversity and creativity in non-adversarial robustness beyond data augmentation, calling for new paradigms that enhance universal robustness to unforeseen "corruptions" to facilitate richer applications in surgical data science.

Large language models (LLMs) are trained from vast repositories of text authored by millions of distinct authors, reflecting an enormous diversity of human traits. While these models bear the potential to be used as approximations of human subjects in behavioral studies, prior efforts have been limited in steering model responses to match individual human users. In this work, we introduce "Anthology", a method for conditioning LLMs to particular virtual personas by harnessing open-ended life narratives, which we refer to as "backstories." We show that our methodology enhances the consistency and reliability of experimental outcomes while ensuring better representation of diverse sub-populations. Across three nationally representative human surveys conducted as part of Pew Research Center's American Trends Panel (ATP), we demonstrate that Anthology achieves up to 18% improvement in matching the response distributions of human respondents and 27% improvement in consistency metrics.

Peer review is fundamental to the integrity and advancement of scientific publication. Traditional methods of peer review analyses often rely on exploration and statistics of existing peer review data, which do not adequately address the multivariate nature of the process, account for the latent variables, and are further constrained by privacy concerns due to the sensitive nature of the data. We introduce AgentReview, the first large language model (LLM) based peer review simulation framework, which effectively disentangles the impacts of multiple latent factors and addresses the privacy issue. Our study reveals significant insights, including a notable 37.1% variation in paper decisions due to reviewers' biases, supported by sociological theories such as the social influence theory, altruism fatigue, and authority bias. We believe that this study could offer valuable insights to improve the design of peer review mechanisms. Our code is available at https://github.com/Ahren09/AgentReview.

Graph Neural Networks (GNNs) have gained popularity in numerous domains, yet they are vulnerable to backdoor attacks that can compromise their performance and ethical application. The detection of these attacks is crucial for maintaining the reliability and security of GNN classification tasks, but existing methods are often inflexible, relying on single metrics that fail to capture the full range of backdoor behaviors. Recognizing the challenge in detecting such intrusions, we devised a novel detection method that creatively leverages graph-level explanations. By extracting and transforming secondary outputs from GNN explanation mechanisms, we developed seven innovative metrics for effective detection of backdoor attacks on GNNs. Additionally, we develop an adaptive attack to rigorously evaluate our approach. We test our method on multiple benchmark datasets and examine its efficacy against various attack models. Our results show that our method can achieve high detection performance, marking a significant advancement in safeguarding GNNs against backdoor attacks.

Rapidly developing machine learning methods has stimulated research interest in computationally reconstructing differential equations (DEs) from observational data which may provide additional insight into underlying causative mechanisms. In this paper, we propose a novel neural-ODE based method that uses spectral expansions in space to learn spatiotemporal DEs. The major advantage of our spectral neural DE learning approach is that it does not rely on spatial discretization, thus allowing the target spatiotemporal equations to contain long range, nonlocal spatial interactions that act on unbounded spatial domains. Our spectral approach is shown to be as accurate as some of the latest machine learning approaches for learning PDEs operating on bounded domains. By developing a spectral framework for learning both PDEs and integro-differential equations, we extend machine learning methods to apply to unbounded DEs and a larger class of problems.

Reinforcement learning (RL) is an effective framework for solving sequential decision-making tasks. However, applying RL methods in medical care settings is challenging in part due to heterogeneity in treatment response among patients. Some patients can be treated with standard protocols whereas others, such as those with chronic diseases, need personalized treatment planning. Traditional RL methods often fail to account for this heterogeneity, because they assume that all patients respond to the treatment in the same way (i.e., transition dynamics are shared). We introduce Compositional Fitted $Q$-iteration (CFQI), which uses a compositional task structure to represent heterogeneous treatment responses in medical care settings. A compositional task consists of several variations of the same task, each progressing in difficulty; solving simpler variants of the task can enable efficient solving of harder variants. CFQI uses a compositional $Q$-value function with separate modules for each task variant, allowing it to take advantage of shared knowledge while learning distinct policies for each variant. We validate CFQI's performance using a Cartpole environment and use CFQI to recommend electrolyte repletion for patients with and without renal disease. Our results demonstrate that CFQI is robust even in the presence of class imbalance, enabling effective information usage across patient sub-populations. CFQI exhibits great promise for clinical applications in scenarios characterized by known compositional structures.

Multi-camera multi-object tracking (MCMOT) faces significant challenges in maintaining consistent object identities across varying camera perspectives, particularly when precise calibration and extensive annotations are required. In this paper, we present CalibFree, a self-supervised representation learning framework that does not need any calibration or manual labeling for the MCMOT task. By promoting feature separation between view-agnostic and view-specific representations through single-view distillation and cross-view reconstruction, our method adapts to complex, dynamic scenarios with minimal overhead. Experiments on the MMP-MvMHAT dataset show a 3% improvement in overall accuracy and a 7.5% increase in the average F1 score over state-of-the-art approaches, confirming the effectiveness of our calibration-free design. Moreover, on the more diverse MvMHAT dataset, our approach demonstrates superior over-time tracking and strong cross-view performance, highlighting its adaptability to a wide range of camera configurations. Code will be publicly available upon acceptance.

If a person can solve a task, can measuring their brain make it easier to train a model to solve that task too? Recent NeuroAI work suggests that supplementing task training with neural recordings can modestly improve model performance and robustness. However, it is unclear when there should be a benefit from using neural data and how much benefit to expect. We formulate this question mathematically, and begin to address it theoretically using a simple, analytically tractable linear gaussian model of task targets and neural recordings. For a multimodal estimator trained on both brain data and task labels, we derive scaling laws for how performance scales with the numbers of brain and task samples. From these laws we derive relative value and exchange rates between brain samples and task samples, quantifying how much extra task samples neural data is worth as a function of task-brain alignment, neural and task noise, latent dimension, and brain data sample size. We also analyze test distribution shift, to identify conditions where brain-regularized learning can produce substantial robustness gains through learned invariances. Finally, under a fixed collection budget, we characterize the regimes in which brain data is worth collecting. Our results provide a foundation for understanding how valuable brain data could be for improving machine learning.

Joint-Embedding Predictive Architectures (JEPAs) provide a simpleframework for learning world models by predicting future latent representations.However, JEPA training is subject to a bias-variance tradeoff.Without sufficient structural constraints, excessive representationalvariance causes the model to collapse to trivial solutions.The recent LeWorldModel (LeWM) shows that this issue can be alleviated bysimply constraining latent embeddings with an isotropic Gaussian prior.However, latent representations inherently lie on low-dimensional manifoldswithin a high-dimensional ambient space, and enforcing an isotropic Gaussianprior directly in this ambient space introduces an overly strong bias.In this work, we propose ame, which seeks a favorable operatingpoint on the bias-variance frontier by applying Gaussian constraints inmultiple random subspaces rather than in the originalembedding space.This design relaxes the global constraint while preserving itsanti-collapse effect, leading to a better balance between trainingstability and representation flexibility.Extensive experiments across fourcontinuous-control environments demonstrate that consistentlyoutperforms LeWM with very clear margins.Our method is simple yet effective, and serves as a strong baseline for future JEPA-based world model research.fdefinedeeemodeThe code is available at https://github.com/intcomp/Sub-JEPA.

Large language models fail at counting repeated tokens despite strong performance on broader reasoning benchmarks. These failures are commonly attributed to limitations in internal count tracking. We show this attribution is wrong. Linear probes on the residual stream decode the correct count with near-perfect accuracy at every post-embedding layer, across all model depths. This holds even at the exact layers where the wrong answer crystallizes while the model simultaneously outputs an incorrect count. Attention patterns show no evidence of collapse over repeated tokens and tokenization artifacts account for none of the failure. Instead, a format-triggered multi-layer perceptron (MLP) block overwrites the correctly-encoded count with a fixed wrong answer at roughly 88--93,% network depth. This prior fires for repeated word-tokens in space-separated list format and is absent for repeated digit-tokens. It is suppressed by comma-separated delimiters in larger models but persists in smaller ones. The finding holds across Llama-3.2 (1B and 3B) and Qwen2.5 (1.5B, 3B and 7B) at consistent relative depth. Counting failure is a failure of routing not of representation and the two require different interventions.

Muon and related norm-constrained matrix optimizers have become central to large-scale learning problems. They are formulated as a linear maximization oracle (LMO) over an ambient matrix-norm ball in unconstrained Euclidean space. However, these do not generalize cleanly to manifold-valued parameters such as low-rank factorizations, orthogonality constraints, or symmetric positive definite (SPD) matrices. Naively restricting the Muon LMO to the tangent space (i) breaks quotient symmetries and (ii) couples the tangent-space constraint with an ambient norm bound, thereby obstructing closed-form solutions on various manifolds of interest. We resolve both issues with a single observation: every Riemannian metric canonically lifts a unitarily invariant Euclidean norm to an intrinsic norm on each tangent space, and the resulting intrinsic norm constrained LMO is symmetry preserving. Building on this, we introduce intrinsic Muon (iMuon), a unified framework that yields closed-form updates on the fixed-rank, SPD, Stiefel, and Grassmann manifolds for any unitarily invariant norm, including the spectral, Frobenius, and nuclear norms. We establish convergence guarantees for both deterministic and stochastic iMuon with rate constants that depend only on the manifold dimension. Notably, on the fixed-rank manifold this constant depends only on the rank, making the rate independent of factor conditioning and removing the runtime factor-rescaling required by prior work. Experiments on LoRA finetuning of LLMs, image classification, and subspace learning illustrate the efficacy of the proposed approach.

One-step generative modeling has emerged as a leading approach to amortize the inference cost of diffusion and flow-matching models. Among distillation-free methods, MeanFlow training is notoriously unstable, with non-decreasing loss and unbounded gradient variance. In this work, we establish a theory that attributes this pathology to a misuse of the conditional velocity field: it plays two distinct statistical roles in the loss, both as an unbiased regression target and as a Monte Carlo control variate inside a Jacobi-vector product, with the original loss assigning the wrong coefficient to the latter. We derive the optimal coefficient in closed form, and show that a family of fixes in concurrent works corresponds to different practical realizations of the same optimum. A controlled sweep of this coefficient on two-dimensional benchmarks and on a latent Diffusion Transformer recovers the predicted bias-variance ordering. The optimal coefficient yields up to a %54 improvement in sample quality on two-dimensional benchmarks and a monotone FID trend at every matched-step DiT checkpoint. Crucially, the same DiT measurement also reveals a quantitative FID-MSE landscape mismatch: although gradient variance is minimized at an interior coefficient value, the coefficient that minimizes FID prefers the direct use of conditional velocity.

Most LLM benchmarks score how well a model responds to explicit requests. They leave unmeasured a different conversational ability: noticing and acting on needs the user has implied but not said. We call this \emph{conversational proactivity}. ProactBench decomposes it into three phase-tied types: \textsc{Emergent}, inference from a single disclosed anchor; \textsc{Critical}, synthesis across multiple anchors; and \textsc{Recovery}, grounded forward-looking value after task completion. We operationalise the benchmark with three agents: a Planner, a User Agent, and an Assistant Model. Their information asymmetries defend against style-confounded scoring, rubric leakage, external-context contamination, and information dumps. The released corpus contains 198 curated dialogues with 624 trigger points across 24 communication styles drawn from a psychometric inventory and audited by an independent LLM judge. Across 16 frontier and open-weight models, \textsc{Recovery} is both difficult and weakly predicted by six standard benchmarks, making it a useful new evaluation signal.

[Abridged] Using a Large Language Model (LLM) as an automatic rater (LLM-as-a-judge) is cheap but potentially biased: some judges run lenient, others strict, the middle of the scale gets compressed, and verbose answers may be over-rewarded. A common remedy is post-hoc calibration: leave the cheap judge in place and, on a modest set of paired anchors, fit a transformation from raw judge scores to an estimate of the human rating. We compare two correctors that take opposing views on how this mapping should be modeled: a parametric, small-anchor hierarchical Bayesian linear correction with per-score uncertainty, and a non-parametric Neural-ODE (FFJORD) score-transport flow. Both are run head-to-head on UltraFeedback fine-grained_score (1700 paired examples, 200 held out), with calibration split into three operational sub-questions: population-mean recovery, per-item accuracy, and distributional-shape match. The headline result is that the choice between methods is primarily a data-budget question. Both correctors close the raw $+0.71$-point mean offset to within $\pm 0.08$ of the GPT-4 reference, at 100 and at 1500 anchors. Past that, the methods swap roles. With 100 anchors, the linear corrector reconstructs the human-score distribution roughly twice as well by KL divergence (0.031 vs. 0.058) and ties the flow on MAE. With 1500 anchors the flow wins on every metric (MAE 0.320 vs. 0.359, Pearson 0.922 vs. 0.896, KL 0.026 vs. 0.037). The Bayesian linear corrector saturates well below 1500 anchors: residual $\tanh$-shaped non-linearity is, by construction, structure a linear correction cannot fit. The flow keeps improving as labels grow. We translate these findings into an explicit decision rule for production deployments.

Sparse autoencoders (SAEs) have been used widely to decompose and interpret neural network activations, especially those of transformer language models. One key issue with SAEs is their inability to directly model multidimensional features. Instead, SAEs may tile such features by a set of independent directions that must be grouped together after the SAE training phase, impeding discoverability and interpretation of learned feature representations. We begin to address this issue by introducing the Sparse MIXture of Autoencoders (SMIXAE) architecture. Empirically, we provide evidence that SMIXAE models have success both in directly learning previously identified manifold structures, as well as finding novel structures, within the open source Gemma 2 2B and 9B models. Finally, we discuss several limitations and point towards areas for future work.

Fairness impossibility results often look like distinct scalar incompatibility statements. We show that several share one RKHS geometry: fairness criteria are linear constraints on conditional mean embeddings, and unequal base rates make the law of total expectation overdetermine those constraints. This view yields four results. The Kleinberg--Mullainathan--Raghavan dichotomy needs only group-conditional unbiasedness, not full calibration. The \emph{Pokémon theorem} shows that a distinct group pair satisfying any finite collection of linear mean-fairness criteria leaves a residual violation witnessed by the MMD, decaying at the Kolmogorov $m$-width rate under spectral regularity. The same tools prove an impossibility for fair feature learning: parity and class-conditional separation in representation space force class collapse under unequal base rates. The approximate relaxations yield signal and error frontiers, allowing a trade-off between real-world estimators and fairness goals. Experiments on standard fairness benchmarks are consistent with our bounds.

3D scene understanding spans reasoning about free space, object grounding, hypothetical object insertions, complex geometric relationships, and integrating all of these with external tools and data sources. Existing 3D understanding methods typically rely on large-scale 3D-language training or focus on object grounding and simple spatial relationships. We argue that the broad generalization that motivates 3D-language training can be achieved at inference time, without 3D-specific training. We propose Flame3D, a training-free framework that represents scenes as editable visual-textual 3D memories and exposes them to an off-the-shelf MLLM through composable spatial tools. Flame3D also lets the agent synthesize custom spatial programs at inference time, enabling open-ended reasoning over layouts, empty space, and objects not yet present in the scene. External data and corrections can be added to the memory without retraining. In addition to showing competitive performance to finetuned 3D-LMM methods on ScanQA, we study multi-hop 3D reasoning capabilities of Flame3D by evaluating it on a curated compositional spatial-reasoning benchmark, Compose3D. We find that fixed tools fall short and that the agent's ability to synthesize spatial operations at inference time is essential. These results invite the question: should future progress in 3D scene understanding focus on richer scene memories and expressive compositional abstractions?

For AI systems to be useful to humans, they must understand and act in accordance with our values and preferences. Since specifying preferences is a hard task, inverse reinforcement learning (IRL) aims to develop methods that allow for inferring preferences from observed behavior. However, IRL assumes the human to be approximately optimal. This is a big limitation in cases where the human themselves may be learning to act optimally in an environment. In this paper, we formalize the problem of learning the preferences of a learning agent: a predictor observes a learner acting online and tries to infer the underlying reward function being (initially suboptimally) optimized by the learner. We model the learner as either being no-regret, or as converging to an optimal Boltzmann policy over time. In each of these settings, we establish theoretical guarantees for various preference learning algorithms, or otherwise show that such guarantees are impossible.