Publication Radar.

A span tagger that ranks its outputs by risk can defer the riskiest ones to a person, and a second model can sharpen this ranking. Which second model should it be? Partner selection is conditional, not global: the best second opinion is the model that is right where the primary is wrong. Across eight tagger types, three seeds and three English benchmarks (576 evaluated primary-partner pairings), choosing the partner by held-out accuracy on the primary's errors recovers 82% of the best partner's gain, against 69-70% for span F1 or a random pick. Frozen before a fourth dataset was chosen, the rule replicated there (Spearman 0.40 vs. 0.04 for span F1). Partner value is also directional. The predictive distribution is the second opinion: hard labels discard half of a partner's value, and scoring the served label improves 94% of pairings. Fixed before the fourth dataset was chosen, this score also replicated there. For LoRA-tuned decoders, independently trained partners beat every readout of the primary network. Applied zero-shot to 51 MASSIVE locales, every primary has a helpful trained partner in every locale, and the preferred partner follows competence.

Tool-using language agents are vulnerable to indirect prompt injection, and a growing literature probes model internals to catch unsafe actions before they execute, with sharply conflicting results. We argue much of the conflict is an evaluation artifact, and propose a double-stratification protocol: separate injection exposure from behavioral commitment, in both the evaluation and the training population. On the primary Qwen AgentDojo grid, monitors trained on the naive mixed unsafe-versus-rest contrast mostly learn exposure: they flag exposed trajectories whether or not the agent complies, and barely predict which cases turn unsafe. Retrained on exposed cases only, the same probe families predict the model's own unsafe tool call up to two steps early, the maximum the three-step trajectories admit (logistic AUROC 0.853, matched by a trained TF-IDF text baseline at 0.855). Mistral-7B reaches 0.896 and OLMo-2-7B 0.912; their naive contrasts learn different things: inverted exposure ranking on one, both axes on the other. Representations separate on exposure (probe 1.000 versus trained text 0.908) and where visible context is sanitized or unavailable. Training contrast alone reproduces the shape of recent pre-action-probe negative results, one candidate explanation for the conflict: the signal exists, but the model's own zero-shot forecast and NLA verbalizations do not reliably recover it. An unstratified AUROC overstates what a latent monitor knows, and what a naive contrast learns is population-dependent: evaluations should stratify both.

Closing the gap between frontier and smaller language models is usually approached by distillation: fine-tuning the small model (student) on a frontier model's (teacher) outputs, which binds transfer to training compute and leaves the transferred knowledge largely opaque. We propose Artifact Distillation: a teacher solves training instances and externalizes its problem-solving procedure into an artifact bank that a student queries at inference time, with no gradient updates. Across four LiveBench tasks, a bank written by Claude Opus 4.5 lets Claude Haiku 4.5 outperform unaided Claude Sonnet 4.5 on spatial reasoning (0.707 vs. 0.647) at 60% of its cost, and reach Sonnet 4.5's level on table joining at 89% of its cost. The same raw bank raises Qwen3.5-27B's pooled score from 0.730 to 0.800, showing that transfer is not confined to the teacher's model family.

Large reasoning models have recently demonstrated remarkable capabilities in solving complex tasks, where supervised fine-tuning (SFT) on long chain-of-thought data serves as a crucial foundation for eliciting and enhancing their reasoning abilities. Despite rapid progress in both improving and analyzing reasoning-oriented SFT, the field still lacks a systematic survey that consolidates its fast-growing literature. To fill this gap, we present a comprehensive review of recent advancements in reasoning SFT, examining the literature through the dual lenses of methodological design and analytical investigation. First, we review methodological improvements across the SFT pipeline and categorize them into data-centric approaches and algorithm-centric innovations. Second, we reorganize analytical studies along three dimensions: data characteristics, optimization dynamics, and mechanistic insights. Finally, by synthesizing current research focuses and remaining bottlenecks, we outline promising future directions for reasoning SFT. We hope this survey deepens the understanding of reasoning SFT and paves the way for advanced reasoning models.

Despite their strong performance, large language models remain highly sensitive to prompt formulation. This issue is typically addressed at inference-time through manual prompt optimization, which introduces overhead and lacks scalability. Conversely, we focus on train-time approaches that build prompt robustness into the model during instruction tuning. A common strategy is to mix multiple paraphrases of each prompt into training. We demonstrate that across trainable parameters, the per-paraphrasing gradients interfere, with 57% exhibiting sign conflicts and the naive gradient average discarding approximately a fifth of the magnitude. Therefore, we develop two gradient-merging methods, Prompt-TIES and Prompt-DARE, that address this interference by leveraging interference-resolution operators from model and gradient merging and applying them to prompt gradients. In this study, we compare the efficacy of our methods to increase model robustness to train-time robustness methods spanning all three established classes: data construction, consistency regularization, and contrastive alignment. The evaluation is conducted across four models from two families at two scales. We observe the interaction between prompt phrasing and the specific examples accounts for 54% of per-example variability on average, and exceeds 70% on tasks such as ANLI and WiC. The gradient-merging methods directly improve this local robustness and nearly halve per-example variance on fully fine-tuned models at little cost to overall accuracy. The other train-time methods enhance raw performance, but fail to improve per-example variance reliably. Code: https://github.com/FSadrieh/prompt-agnostic-llms. Model checkpoints: https://huggingface.co/prompt-agnostic-language-models.

Deployed classifiers often act on a prediction only when its confidence clears a threshold and pass the rest to a person. We ask whether one threshold serves all languages equally on typed decision interfaces, APIs that return a probability for each option of a fixed multiple-choice question. In a confirmatory study with a written analysis plan, three vendors' interfaces classified 1,000 MASSIVE items, each in 51 languages, with a threshold of 0.9. Even when a request is answered correctly in both English and another language, the English version clears the threshold more often, by 8.6-11.7 percentage points on average. Observed coverage, the share of requests handled automatically, is lower than English's in every other language, while the requests the gate handles are rarely wrong. The gap remains when every language is reweighted to the same accuracy, and a standard calibration metric (ECE) does not detect it on one interface. In exploratory analyses, English's coverage stays higher when each language gets its own threshold fitted to the same 5% error target. Translating inputs into English nearly closes the gap at 0.9 but leaves English ahead at that error target. Requests written natively in African languages show a gap in the same direction when English is their human translation. We recommend reporting per-language coverage at the operating threshold.

Users engage with large language models on tasks that are purely objective ("When was Netflix founded?"), purely subjective ("What should I name my child?") or anywhere in-between. We have rich tools for evaluating model accuracy with respect to objective criteria. However, when it comes to subjective aspects of a task, we do not have similarly developed methodology at hand. In this paper, we take some early steps towards framing the question of subjective evaluation. Rather than measure how often a model is right or wrong on a question, we instead ask, how do models make choices among multiple viable responses without a clear winner? To study this question, we restrict our focus to short-form content, and consider comparison tasks among three such content types: numbers, baby names, and fictional cities. We identify 29 aspects of these short-form examples covering many areas of syntax and semantics (length, alphabetic order, popularity, vowel mix, spoken sound, etc.). We show first that models tend to have strong and reproducible preferences for or against some of these aspects. Second, we show that variations in how the model is asked to express a preference may cause some of these aspects to activate or deactivate in the model's selection. Third, we show that overall preference distributions among pairs may vary radically based on seemingly-unimportant variations in the nature of the question posed to the model, raising new questions about manipulation. And finally, we show that, if models are used both to create lists of alternatives and to rank them, there are unexpected interactions between the two stages that may also drive the models' stated preferences.

Adaptive optimizers are widely used in deep learning, yet their behavior can degrade under heavy stochasticity and label noise where the sign and magnitude of gradients can be unstable. We study MGO-SC (Moment-Guided Optimizer with Sign-Consistency), a lightweight modification of Adam-style methods that derives a coordinate-wise consistency gate from the magnitude of a sign exponential moving average (EMA): coordinates whose gradient sign is stable are trusted, while coordinates whose sign fluctuates are down-weighted. We show that the naive form of this gate, which can only shrink updates, loses effective step size and underperforms when the learning rate is matched to the baseline. We therefore introduce a mean-preserving variant that reallocates trust across coordinates without changing the net per-tensor step magnitude. On CIFAR-10 with ResNet-18, under a matched learning rate and decoupled weight decay, mean-preserving MGO-SC consistently improves over its AdamW baseline across clean, symmetric, and asymmetric label noise (e.g., +1.4 points at 40% symmetric noise; paired bootstrap 95% CI excluding zero; Wilcoxon p<0.01), while adding no second-order information or per-sample filtering. We report matched-seed effect sizes, nonparametric bootstrap confidence intervals, and paired significance tests, and we release code and raw logs.

Post-training moves a language model's weights by about a thousandth of their norm and changes its behavior sharply; pretraining moves them by a third and buys a slow drift. We introduce leverage, an exact measure of what a unit of weight change buys: the Fisher quadratic form of an update, computed with one forward-mode pass per direction, divided by the same form for random directions of the same size at the same checkpoint. Leverage orders released Olmo-3, Tülu-3, Zephyr, and OLMo-2 checkpoints over four orders of magnitude: pretraining windows sit at the random level on prose, SFT five times above it, RL-Zero and RLVR at 400-1,500, and DPO at 400-23,000. With this instrument we find that pretraining accumulates into noise: windows of 30k to 1M steps are no more consequential on prose than random directions of their size, as a noisy quadratic model of SGD predicts. We find that training precision changes what an update looks like: training bf16 parameters without an fp32 master copy freezes 36-99% of coordinates, lowers leverage up to sevenfold, and creates an alignment between update magnitudes and the diagonal Fisher that fp32 training does not show. We derive and verify a law for reuse: the sign-flip null predicts the relative damage of random drop-and-rescale (DARE) at keep rate $p$ as $(1-p)/(p\,ρ_{\mathrm{sign}})$, within a factor of two of measurement wherever DARE succeeds. Two controls make these readings reliable: every controlled SFT and DPO trajectory with fp32 master weights loses leverage as it moves, and two re-runs of a released DPO stage from its anchor with its published recipe land within 1% of the released displacement and 4% of its leverage, where a plain recipe reads five to six times lower; leaving one component out traces the level to the length-normalized loss and its batch size of 128 together.

Efficient communication remains a major bottleneck in large language model multiagent systems. Latent communication reduces repeated natural-language exchange by transferring internal model states between agents, but these transfers remain costly and are typically designed around conventional KV-cache representations. We address this problem from two complementary perspectives. We introduce SelectKV, a boundary-aware method that selectively communicates KV states while protecting inherited context, and HyLaMAS, which extends latent collaboration to hybrid linear-attention models by communicating both KV caches and recurrent states through a shared workspace. Across six benchmarks, SelectKV reduces communicated KV positions and logical KV payload by 9.7-9.9%, with accuracy unchanged on two benchmarks, within one percentage point on three, and six points lower on HumanEval+. Across Qwen3.5 model scales, HyLaMAS reduces decoded output tokens by 53-80% and achieves 1.16-2.78x speedups relative to text-based multi-agent communication. Together, these independently evaluated methods represent two complementary approaches to efficient latent-state communication across conventional Transformer and hybrid model architectures.

Recent work has hypothesised that large language models may represent some concepts as multi-dimensional manifolds instead of single directions. Using these manifolds to interpret or steer representations often assumes that moving along them corresponds to a continuous change in the represented concept. In this work, we test this continuous manifold assumption on the prototypical weekday manifold. Despite its striking circular geometry, we find that the weekday representations behave more like seven discrete states arranged in a circle than a continuous ring. First, we show that local sensitivity has seven minima aligned with the weekdays, producing a step-like mapping between representation and behaviour under on-manifold steering. Second, we test whether any concepts are represented along the manifold outside the seven discrete states, and find neither time of day, uncertainty, nor natural text populate the intermediate space. Together, these results challenge the continuous interpretation of the weekday manifold and motivate testing whether movement along other representation manifolds correspond to meaningful changes in model representations and behaviour.

Quantum machine learning (QML) is increasingly explored for medical classification, yet its behaviour under subgroup-aware fairness criteria remains understudied. We present an empirical fairness audit of classical and quantum classifiers for medical imaging, assuming QML is not inherently fair. Two scenarios are evaluated: dermatological lesions on DDI (sensitive attribute: Fitzpatrick skin type) and chest X-rays on CheXpert-derived data (sensitive attribute: age group). Using ResNet-18 embeddings, we train SVM, MLP, QSVM and QCNN under a baseline and three mitigation settings: differentiable Equalized Odds (EO) regularisation, Gradient Reversal Learning (GRL), and both combined. Foundation models (PanDerm, EVA-X) serve as external references. Quantum classifiers may achieve competitive aggregate performance while still exhibiting subgroup disparities. Since a smaller subgroup AUC gap may stem from degrading the best group rather than improving the worst, each mitigation is also tested against a pre-specified non-inferiority margin relative to its baseline. Under this criterion, several configurations with the smallest disparity, notably QCNN with GRL on DDI, are not clinically acceptable due to loss of global discrimination. On DDI, QSVM with GRL+EO performed best and approached PanDerm in the performance/fairness trade-off; on CheXpert, QCNN with GRL achieved the highest aggregate performance but increased disparity, and EVA-X remained stronger. Fairness effects are model- and dataset-dependent, so QML requires empirical subgroup-aware evaluation. All quantum models run in classical simulation; we claim no quantum advantage.

Event cameras, spiking cochleas and similar sensors emit asynchronous events, yet most learning systems first bin them into frames or packets, so the output depends on how the stream was sliced. Networks whose state is updated exactly at every event and evolves in closed form between events are invariant to this slicing. We ask what more makes the output (i) independent of the processing order of simultaneous events and (ii) continuous in the time stamps, even when two events cross in time, while (iii) staying sensitive to temporal order. For arbitrary nonlinear event models we prove that (i) and (ii) are each equivalent to pairwise commutation of the event updates, and that once updates commute, all order dependence comes from their non-commutation with the inter-event dynamics and vanishes linearly in the time lag. This identifies a class of time-mediated non-commutative (TM) models and yields a resolution-robustness bound for any model. We construct the minimal TM layer, a learnable, provably Lipschitz generalization of the Hassenstein-Reichardt correlator; prove that no single additive layer can uniformly represent ordered pair interactions, whereas a three-dimensional TM state can; show that each TM layer adds two levels of temporal nesting versus one for additive layers; and give a single-layer higher-order construction (TM-k) and a stacking rule preserving all three properties. We instantiate these results as the Clock-Free Event (CFE) layer, linear in the number of events, give a topology-free variant for temporal point processes with closed-form likelihoods, and specify a preregistered clock-robustness protocol with a physical-time threat model for spike-retiming attacks. In preliminary development-set experiments, the TM state beats an otherwise identical additive state by 3.5-9.1 points on DVS128 Gesture and by 24.5 and 30.6 points on SHD and Speech Commands.

An AI developer often has several candidate versions of a system and can deploy a mixture of them, sending each request to a randomly chosen version, as in A/B testing. Each stakeholder (e.g., a safety team, a regulator, or a customer with private test data) accepts a mixture exactly when its expected performance (e.g., its rate of harmful replies) meets the stakeholder's own standard, which the developer can learn only by asking whether a proposed mixture is acceptable. We model this as choosing a lottery over $m$ alternatives for $n$ agents, each of whom accepts exactly the lotteries in an unknown halfspace and answers such binary membership queries. While Choo et al. (ICML 2026) sought a lottery that every agent accepts, none may exist when, say, safety and helpfulness requirements conflict, so we minimize the number of rejecting agents. We show that this problem is APX-hard even when all halfspaces are known and all utilities are $0$ or $1$. We then give exact algorithms that run in polynomial time when $m$ is constant, and in quasi-polynomial time when, for a constant $η$, some optimal lottery has no accepting agent within $η$ of its threshold. When only some agents can be queried, optimizing over a uniform sample of agents gives additive and relative guarantees for the whole population, with a sample size that does not depend on $n$, and a linear program gives a polynomial-time method for any $m$. On the other hand, for $m\ge 3$, a deterministic algorithm that queries only $K$ distinct agents cannot guarantee fewer than $n-K$ extra rejections, and this is achievable; in other words, it can guarantee nothing for the agents it does not ask. Small synthetic experiments indicate that our worst-case bounds are conservative on random instances.

Modern automatic speech recognition (ASR) systems have achieved remarkable accuracy by scaling model depth and capacity, but at the cost of increased memory, energy, and computational demands. On edge devices where ASR is often most needed, such as watches and glasses, deploying such large models is infeasible due to extremely constrained resources. This raises a fundamental question: can we achieve high representational power in deep ASR models without scaling up parameterization? In this work, we revisit the role of depth and identify layer wise representational dynamics, in which most layers learn functionally similar transformations. Motivated by this insight, we propose a block-recurrent ASR architecture that replaces parameterized depth with a small set of recurrent blocks, each consisting of weight-shared layers, thereby preserving effective depth while drastically reducing model size. Through representation-guided grouping and knowledge distillation, block-recurrent models retain the accuracy of large reference models while using an order of magnitude fewer parameters. Across Open ASR Leaderboard benchmarks, a model with only two shared blocks recovers 97.4% of the accuracy of large models on average. The compact backbone further enables efficient on-device personalization through lightweight MoE-LoRA adaptation, allowing user-specific models to match or surpass the accuracy of general-purpose ASR systems with significantly lower inference cost. By reusing weights across recurrent steps, the proposed architecture reduces memory-access overhead and inference energy consumption. Together, these results show that high quality ASR can be achieved without large parameterization, and that structured recurrence provides a principled path towards ultra-compressed, energy-efficient, and personalized speech recognition on the edge.

As growing compute budgets outpace the growth of high-quality training data, large language models must be trained with both compute and data constraints in mind. Unfortunately, existing prescriptions for data-constrained scaling suffer from two important limitations. First, past works hold key hyperparameters such as weight decay fixed with scale, despite the fact that optimal hyperparameters vary wildly with model size and number of epochs. Second, past works neglect ensembling, a promising axis for scaling under data constraints. As a surprising consequence, prior prescriptions can lead to loss that increases with compute budget. We address both limitations by fitting power laws that prescribe how to transfer hyperparameters with scale and scaling laws that prescribe how to allocate compute with ensembling. Training on 100M tokens from FineWeb, our recipe reduces validation loss by over 0.15 nats compared to prior recipes, a 14% decrease in perplexity, and achieves monotonic improvements with compute through ensembling. These gains are achieved using prescriptions fit to smaller runs and validated at up to 12$\times$ larger compute budgets.

International climate policy requires coordination among countries whose mitigation decisions impose local costs but generate global benefits. Classical game-theoretic and integrated assessment approaches can study these incentives, but typically rely on prescribed behavioural assumptions or equilibrium concepts. Recent work has therefore explored Multi-Agent Reinforcement Learning as a way to model climate-policy cooperation as an emergent outcome of repeated strategic interaction. We extend an existing MARL climate framework with a multi-stage negotiation mechanism in which agents propose agreements, accept or reject commitments, and face penalties for non-compliance. This allows us to study how agreement design affects learned mitigation behaviour, compliance, and climate outcomes. In an idealized setting with four homogeneous agents, we find that non-binding agreements do little to overcome free-riding, while moderate non-compliance penalties improve cooperation and reduce warming. However, excessively strong penalties discourage ambitious commitments, revealing a trade-off between enforcement and participation. We position the framework as a methodological contribution for testing agreement design rather than a predictive model of real negotiations.

Equation learning seeks to discover compact, interpretable equations from data using various machine-learning techniques. However, few workflows discover and interpret learned equations in sparse, high-dimensional materials datasets. We present nn-EQL, an equation-learning framework that generates candidate equations through iterative training, pruning, and retraining of a neural network with customizable, user-specified neuron-level activation functions. Candidate equations are ranked using a minimum description length (MDL)-inspired score, balancing error and sparsity, and diagnosed to assess term-wise contributions, input-scaled sensitivities, and uncertainties. We demonstrate nn-EQL on standard AI-Feynman benchmarks as well as multiscale high-dimensional materials datasets such as finite-element simulations of heterogeneous materials, and high-throughput experimental thin-film deposition data. nn-EQL integrates with tabular data and high-dimensional data via standard encoder models, learning equations and providing insight into learned terms. Looking forward, integrating physically interpretable representations, multimodal and multi-fidelity data, and adaptive data acquisition could enable equation learning to formulate and test mechanistic hypotheses.

On PushT, our fixed action updater (1,505 parameters) solves 8 of 16 scenes reserved from our configuration selection. The comparison planners perform as follows: Selected Adam (7 tasks solved, total planning and encoding time 218.440s), adapted iCEM (2 tasks solved, total planning and encoding time 49.309s), released LePlanner (16 tasks solved, total planning and encoding time 5.652s). The total planning and encoding time for our action updater is 59.079s. Note that our action updater does not satisfy the task-cost criterion we fixed before evaluation. We also conduct another matched experiment where we only replay the first ten commands in each context and then allow the next solve 4, 8 or 12 updates (subsequent solves all get 8 updates). Among all 16 contexts (8 scene clusters), the 3 budgets get 9, 8 and 7 successes, respectively. Interestingly, the extra updates did not rescue any failing task in this experiment. All the changes in success/failure happen in one scene cluster, where we observe that the extra updates decrease the costs of the selected latent world-model trajectories for our updater, while the tasks that used to be successes become failures. The code, weights, evaluators, execution traces and cost accounting accompany the paper.

Graph Contrastive Learning (GCL) has achieved remarkable success in self-supervised representation learning by maximizing agreement between augmented views of a graph. Although augmented view generation is the central step in GCL, existing models rely on stochastic augmentation of the input graph, such as edge perturbation and node masking, which lack structural awareness and frequently destroy critical topology or introduce semantically inconsistent views. While recent adaptive methods improve upon random perturbations, they primarily operate on localized node or edge neighborhoods and fail to capture higher-order, macro-level structural organization crucial for graph classification. To address this limitation, we propose Structure-aware Compression for Graph Contrastive Learning (SAC-GCL), a novel framework that employs structural decomposition and graph compression as a principled view generation strategy. Moving beyond random augmentations, SAC-GCL leverages structural decomposition techniques to extract cohesive subgraphs and condense them into supernodes, constructing coarsened graph views that preserve local density and capture higher-level structural patterns while reducing noise and redundancy. We leverage multiple cohesive subgraph models to derive complementary views of the graph, each capturing distinct structural properties. To capture multi-resolution information, a hierarchical view encoder processes these compressed views by initializing super-node features with shared-weight cohesive subgraph embeddings. Our results from extensive experiments on six benchmark datasets demonstrate that structure-aware compression provides a powerful alternative to random perturbations, yielding more robust and discriminative graph representations for graph classification tasks.

Metis is a bet that better data, better training methods and new architectures can buy what a bigger parameter count cannot, at a fraction of the cost. Metis-1.6 is its latest step: two language models built from scratch by a single researcher, Praxis (2.94B parameters, 0.46B active per pass) and Logos (10.23B parameters, 1.18B active per pass). Both run a hybrid Mamba-2 and attention mixture-of-experts stack twice for every token, so each token gets a second look. Each was trained on 1.02T tokens in total: 804.76B of pretraining, 18B of context extension to 131,072 tokens, and 200B of supervised fine-tuning. Metis-1.6 was never meant to challenge frontier models, and across 40 post-trained benchmarks it mostly does not: Logos averages 52.93, seventh of 17 models in its comparison set and within a point of Qwen3-4B. But it gets there on roughly a twelfth of the median training data of the comparators that disclose it, and with fewer active parameters per token than any model that beats it, which puts it on the Pareto frontier for both. In mathematics it ranks second of 30 models, with 67.50% on AIME 2025 and 63.64% on HMMT February 2026. It is weak at code, instruction following and long context. The rest of the report is the part I found most valuable: what went wrong. The optimizer I configured never ran, and attention trained with AdamW for all of pretraining. A 1.77B-parameter n-gram memory learned to switch itself off. Five fine-tuning choices cost capability. And a length reward under group-mean normalization punished correct solutions for being long once a group's pass rate passed 2/3, a sign inversion I prove and fix with a one-line clamp. I close with the 10B-token ablations behind MoRE (Mixture of Recursive Experts), the looped architecture of Metis-1.7.

Distributed fire-monitoring systems may involve many sensing devices reporting over a shared uplink, even though the operational objective is often to identify regions exhibiting evidence of fire rather than to recover device-specific measurements. We introduce TERRA (Task-Oriented Edge Representation and Regional Aggregation), an edge-AI framework that combines task-oriented local inference, identity-free region-tagged reporting, and permutation-invariant regional evidence aggregation. TERRA supports two reporting modes: TERRA-Score, in which each device communicates a quantized local fire probability, and TERRA-Feature, in which it communicates a compact learned representation for receiver-side aggregation using a permutation-invariant DeepSets model. Using all 15 numerical sensing modalities of the distributed indoor EN54 fire-alarm dataset as a proof-of-concept testbed, we evaluate regional inference for an episode-level fire-scenario target under repetition-disjoint three-fold cross-validation. We study representation size, task-payload quantization, report availability, and robustness to missed and false recovered messages. With nine reports and perfect message recovery, TERRA-Feature achieves mean AUPRC 0.990 using a 24-bit task payload, while TERRA-Score achieves 0.987 using only 6 bits; both attain an MDR of 1.81% at FAR $\leq$1%. TERRA-Feature achieves slightly higher AUPRC under reliable recovery, whereas the more compact TERRA-Score is less sensitive to increasing false-message contamination.

Stellarators offer a promising magnetic confinement approach for achieving controlled nuclear fusion. However, stellarator design remains challenging, as evaluating configuration performance requires computationally expensive, high-fidelity simulations. In particular, key performance metrics derived from guiding center simulations are costly to include directly in the design optimization loop, and are almost exclusively used as diagnostics. Machine learning has the potential to advance fusion research by enabling rapid generation and evaluation of new stellarator configurations. Effectively applying ML for fusion demands large-scale datasets that capture both design diversity and high-fidelity simulation outputs. In this work, we introduce the Benchmark of Stellarator Simulations (BoSS), a dataset of stellarator designs paired with guiding center simulation results and trajectories. The BoSS is the first large-scale dataset of guiding center simulations, containing simulation data for 180,000 distinct magnetohydrodynamic equilibria spanning a wide range of magnetic geometries and confinement properties. To exemplify the utility of ML to accelerate stellarator design, we formulate and address two sets of modeling tasks: generative modeling for designing stellarators with good particle confinement, and surrogate modeling for predicting stellarator heat deposition patterns and particle trajectories. The tasks reveal the potential impact ML can have in the stellarator field and important opportunities to improve the applicability of ML models in the sciences. Our baselines generate diverse, performant stellarators and provide useful surrogate predictions, and point to long-horizon trajectory forecasting and out-of-distribution generation as key directions for future work. The BoSS can be accessed through https://boss.flatironinstitute.org/.

Many open mixture-of-experts (MoE) language models train with Muon instead of AdamW. Most balance the load with a bias on each expert's routing score, adjusted after every step. Their recipes disagree on whether Muon should also train the router. In models with 55M and 108M active parameters, we measure how the optimizer changes the span of these biases, the largest minus the smallest in a layer. The span grows when all tokens favor the same experts, so it shows how hard a layer is to balance. At learning rates tuned on shorter runs, Muon raised the span of the first MoE layer at every seed, roughly doubling it on average, while the other layers changed much less. The rise came from Muon on the rest of the network, whichever optimizer trained the router. In those runs, the inputs to the first router shared a direction, and soon after the warmup its busiest expert averaged four to seven times its fair share. Over the first 3,000 steps, subtracting the batch mean from these inputs removed this overload while barely changing the loss. On the smaller model, halving Muon's learning rate removed most of the rise and slightly improved the loss. Moving only the router to AdamW made the loads shift more during the warmup and slightly raised the loss, so at this scale the results favor keeping the router on Muon.

Vision-Language-Action (VLA) models offer a promising paradigm for general-purpose household robotics. However, inference and data-collection strategies that have been deployed in smaller-scale settings do not necessarily generalize to large-scale, diverse task suites. To address this gap, we conduct a systematic, paired evaluation of previously proposed inference and data-collection strategies across all 100 tasks. Our evaluation uses a fine-tuned $π_{0.5}$ model based on the 2025 BEHAVIOR Challenge winner, which achieves a Q-score of 0.39 on the public test set. These results show that proposed strategies do not transfer uniformly to new environments and tasks. Project page, videos, and detailed analysis: https://rubenhammele.github.io/behavior2026

Key-value (KV) caching is a core mechanism in large language models, where intermediate key and value representations are stored to accelerate autoregressive decoding. As generation length increases, KV cache memory and computation grow linearly, making compression critical for efficient long-output reasoning. Dictionary-based sparse KV compression is appealing because it represents each token's KV states as a sparse combination of learned basis elements (atoms) in a dictionary, preserving token-level structure without enforcing a global low-rank constraint. However, existing dictionary-based methods face two practical limitations during decoding: iterative sparse coding methods such as Orthogonal Matching Pursuit add decoding-time overhead, and static dictionaries may mismatch evolving KV distributions under test-time generation. While adaptive dictionary updates can reduce this mismatch, naive append-only strategies lead to dictionary growth during generation. We propose FAD-KV, a \textbf{F}ast \textbf{A}daptive \textbf{D}ictionary KV caching framework for efficient reasoning decoding. FAD-KV integrates (i) a screening-based sparse coding scheme, where a one-shot Top-\(n\) selection identifies a small candidate pool followed by lightweight coefficient refinement, and (ii) a fixed-size dictionary adaptation mechanism that reuses inactive atoms instead of expanding the dictionary. By controlling both sparse coding cost and adaptive dictionary growth, FAD-KV provides a memory-bounded and computation-efficient KV compression framework for long reasoning workloads. Experiments on reasoning and long-context benchmarks show that FAD-KV achieves a strong quality-efficiency trade-off while avoiding unbounded dictionary growth during decoding.

Large language models (LLMs) have emerged as potential tools for numerical optimization. Despite promising results, their intrinsic optimization capabilities and limitations remain poorly understood, particularly in high dimensions. This paper investigates the ability of LLMs to exploit mathematical problem structures and generate feasible, near-optimal solutions. We develop an evaluation framework that distinguishes solution quality from execution failures, including numerical inaccuracies, constraint violations, and output-format errors. Using representative problems from convex optimization, fractional programming, and constrained projection, we assess LLM performance against deterministic reference solvers across different dimensions. Our results reveal that LLMs can capture useful optimization structures and produce near-optimal solutions, even when their outputs fail to satisfy numerical or formatting requirements. However, reliability deteriorates with increasing problem dimension and depends strongly on the underlying optimization structure. To further investigate the limitations, we deploy an iterative refinement mechanism that provides targeted feedback on feasibility and solution quality. The results demonstrate that iterative refinement substantially improves reliability and extends the range of problem dimensions over which LLMs produce valid solutions. Nevertheless, persistent numerical and structural errors limit their applicability as standalone optimizers. These findings provide insights into the capacity and limits of LLM-based optimization and motivate hybrid approaches that combine LLM reasoning with deterministic numerical verification.

Pre-training foundational models is computationally expensive as models scale. In this work, we introduce Bi-Phase Training (BPT), a parameter-efficient pre-training method that preserves the high-rank learning dynamics observed in full-parameter optimization with substantially reduced trainable parameters. BPT parameterizes each weight update as a composition of constrained high-rank left/right transformations via trainable diagonal matrices and low-rank adapters whose updates are periodically merged into the frozen base weights, resulting in a compounding expansion of accessible directions during training. Across language-model pre-training from scratch (Qwen2.5 models from 100M to 1.5B parameters and OLMo-2-1B) on FineWeb-Edu (10B and 100B tokens), BPT achieves validation perplexity close to full-parameter baselines while using far fewer trainable parameters; e.g., at the 1.5B scale, BPT uses 66% fewer trainable parameters with only a small validation-perplexity gap. Additionally, a comprehensive evaluation across diverse downstream tasks shows that BPT preserves much of the full-parameterized model's downstream performance.

Latent multi-agent systems reduce inference cost by keeping intermediate reasoning and inter-agent communication in continuous representations.\ Reusable skills, however, are commonly supplied as text or stored as layer-wise Key-Value states or skill-specific model parameters, introducing repeated processing or substantial per-skill storage.\ We introduce \textbf{SkillMem}, which represents each reusable skill as a small external latent memory and makes it directly available during latent multi-agent reasoning.\ A role-conditioned cross-attention reader allows the same memory to guide different agents according to their roles and current hidden states.\ We train the skill memories and shared reader through behavioral distillation, matching the change in token predictions induced by a textual skill while keeping the backbone frozen.\ For our experiments, we obtain textual skills from paired successful and failed multi-agent trajectories, although existing skills can also be compiled directly.\ Across six benchmarks and Qwen3 models ranging from 4B to 14B parameters, SkillMem provides strong gains for smaller backbones while preserving efficient latent multi-agent inference across model scales.\ We use the same $8\times256$ memory configuration across all evaluated model scales, requiring 9.6\,KB of skill-specific storage per skill, including its retrieval embedding.\ For a library of 100 skills, SkillMem reduces total storage by more than $38\times$ compared with alternative skill representations and achieves up to a $4.23\times$ end-to-end inference speedup.\ These results show that compact procedural memories can be reused without sacrificing the efficiency of latent multi-agent reasoning.

A widely held assumption in knowledge distillation is that teacher and student representations must occupy a compatible or explicitly aligned feature space. We propose RPKD (Random Prototype Knowledge Distillation), which discards this assumption by comparing both networks against a shared bank of randomly initialized, frozen prototype vectors. Alignment is therefore carried out in an external reference space rather than in the teacher's native feature coordinates: the teacher side is projected by fixed random maps that carry no trainable parameters, and only the student adapts. Two branches handle the transfer: a logit branch aligning global prototype similarity distributions and a feature branch matching spatially-resolved prototype activation maps. Both operate within the same fixed prototype space, which can be interpreted as a random feature embedding that approximately preserves inter-sample similarity structure, so the prototype bank and the distillation objective stay unchanged across teacher-student pairs while only the input projections are instantiated per architecture. In low-category settings, decoupling the prototype vocabulary from the task label space yields richer supervision than the label space alone, an advantage absent from logit-based methods, whose supervisory signal collapses as class count shrinks. This advantage is especially relevant in real-world scenarios where the label space is typically constrained. Extensive experiments demonstrate the effectiveness of RPKD, achieving a maximum gain of +7.98% over OFA on CIFAR-100, +12.44% on ImageNet-100, and +1.28% on ImageNet-1K. Source code is available at https://github.com/HamzaEtcibasi/RPKD

Rounded capacity cuts (RCCs) are among the most effective cuts for capacitated vehicle-routing relaxations, but exact separation is computationally prohibitive at scale. Existing heuristic and neural separators can be fast, but often fail to produce sufficiently many effective cuts within practical separation budgets. We introduce SAG-Sep, a separator that scores node membership in RCC-inducing subsets across vehicle-count levels from a fractional LP solution in a single encoder forward pass, avoiding the iterative predict-and-coarsen inference used by NeuralSEP-style separators. SAG-Sep encodes the LP solution as a Sparse Augmented Graph (SAG), adding flow-informed multi-hop edges that expose long-range routing structure without densifying attention. Beyond LP-aware sparse encoding, we observe that high-value exact RCCs are typically nested, organizing into onion-like subset families, and SAG-Sep's node scores track subset nesting depth. This motivates Onion-Sweep (OS), a training-free decoder that ranks customers by their scores, evaluates the RCC violation of every prefix of the ranking in one vectorized pass, and emits the most violated prefix of each replica together with a cross-replica family of nested proposals, converting independent node scores into structurally coordinated RCC candidates without a tuned threshold or probing band. On the NeuralSEP benchmark with N=1000 customer instances, SAG-Sep with OS reduces the average root gap by 17.7% relative to the strongest NeuralSEP baseline, while the single-pass SAG-Sep separator achieves a roughly 30-fold per-iteration neural separation speedup. Our findings suggest that LP-aware sparse encoding and the nested structure of high-value cuts are structural patterns worth leveraging in neural separation more broadly.

Time-series data play an important role in various tasks, such as weather forecasting, web traffic analysis, and speech recognition. Recently, pioneering studies have increasingly adopted Joint-Embedding Predictive Architecture (JEPA) based encoders to learn time-series representations through latent-space prediction. However, they often overlook the underlying factors governing transitions between time-series observations. In this work, we hypothesize that explicitly modeling the information behind changes in time series can facilitate time-series modeling. Building on this hypothesis, we propose ACTS-WM, a self-supervised latent action-conditioned world model framework for time-series analysis. First, we introduce a latent action modeling approach that infers the underlying information behind time-series transitions as latent actions, without requiring explicit action annotations. We then develop a JEPA-based encoder and an action-conditioned world model to predict target time-series representations conditioned on the learned latent actions and task information. We evaluate ACTS-WM across 4 representative tasks on 22 datasets from diverse domains, covering univariate and multivariate forecasting and classification. Comprehensive experiments demonstrate that ACTS-WM can improve downstream performance over pretrained JEPA-based encoders by learning complementary representations.

Implicit neural representation (INR) methods for deformable medical image registration require the network parameters to be optimized for each new image pair, making adaptation speed a practical limitation. This work investigates whether early adaptation can be accelerated using an initialization learned from a collection of previously solved registration tasks. We examine two schemes: Reptile meta-initialization and STRAINER transfer of the shared component of a coordinate network. KAN-IDIR is used as the base INR displacement-field model. Experiments are conducted on OASIS-1, with 364 consecutive pairs used for training and 49 pairs for test-time adaptation. Registration quality is evaluated using mean Dice and HD95, together with the fraction of voxels with a negative Jacobian determinant at iteration 1500. In the configurations studied, both schemes yield higher Dice and lower HD95 at early iterations than random initialization. Reptile numerically outperforms STRAINER under a small iteration budget, although their pretraining budgets are not matched. As the number of iterations increases, the differences between the schemes shrink. STRAINER pretraining takes 39-40 minutes, compared with 44.98 hours for Reptile.

Machine learning (ML) architectures are being developed for fast emulation of numerical simulators to accelerate both scientific discovery and engineering design. While these approaches are well-tested on popular PDE (partial differential equation) benchmarks, it is necessary to evaluate their performance on datasets that more accurately reflect industrial and engineering applications. We evaluate several computer vision-based architectures trained as surrogate models on a challenging computational physics task involving multiple materials, revealing significant implementation challenges for adapting certain pre-trained models, including transformer-based architectures. For such multi-physics multi-material datasets, our results underscore the importance of learning algorithms designed for sparse and imbalanced data, techniques for adapting spatial encodings and physical consistency metrics, in addition to methods for reducing the cost of generating sufficient training data.

This paper presents a two-stage learning framework for versatile and agile humanoid locomotion that couples online reduced-order model (ROM) reference generation with whole-body tracking control. In the first stage, we train a command-conditioned whole-body reference generator using single rigid body dynamics combined with full humanoid kinematics. In the second stage, the frozen ROM policy is rolled out online to generate references for training a whole-body tracking policy on the full-order dynamics model. Learning a kino-dynamically consistent reference generator in the simplified ROM space enables us to generate references over a broad command space and over multiple locomotion strategies using a single unified framework. We demonstrate the approach on a Unitree G1 humanoid robot, achieving longitudinal velocities from -2.0 m/s to 2.8 m/s, lateral velocities up to +-1.0 m/s, angular velocities up to +-3.0 rad/s, one-legged hopping, and continuous transitions across the commands, all using one unified reference policy and one whole-body tracking policy.

We empirically investigate how MLPs and Transformers differ when trained on data sets containing one or more features encoded by noisy k-parity. Noisy k-parity gives us hard synthetic data sets in which we can control the predictivity and availability of every feature. Across a range of data distributions, data scales and model sizes, we find large differences in the training dynamics and inductive biases of the two architectures: Transformers are more robust to noisy features but struggle to learn complex, low-availability features, whereas MLPs show the opposite trend and improve steadily with more data. We also find that, under a fixed training budget, increasing the number of unique training examples can slow Transformer optimisation enough to reduce final accuracy, so that a smaller data set performs better. Training-accuracy results show this is an optimisation-speed effect; it shrinks, but does not vanish, when the budget is scaled with the data set size. By giving precise control over feature predictivity and availability, this testbed offers a clear starting point for understanding when and why Transformers and MLPs learn differently, and for designing data and training budgets that exploit those differences.

Identifying a small molecule from its tandem mass spectrum is limited by paired data. Foundation models offer a way around this scarcity: pretrain a spectrum encoder on unlabeled spectra and a molecule encoder on a large corpus of structures, then align the two frozen representations, as in vision-language and as the Platonic Representation Hypothesis suggests. We audit this route on MassSpecGym with spectrum encoders retrained free of test-fold contamination. Alignment needs a metric relating distinct molecules, but the objectives that produce spectrum encoders never see molecular structure: we prove that some predictive objectives fix molecular identity without fixing that metric. Moreover, we find that a candidate-based objective rewards a spectrum-independent prior over the benchmark's decoys, which we isolate. What ultimately addresses data scarcity is not architectural: simulated spectra surpass frozen geometry.

Full-order surrogates for high-dimensional nonlinear flows are often unnecessary when applications require only selected degrees of freedom. We propose reduced-order models that bypass full-state reconstruction by deploying graph neural networks (GNNs) directly on coarse meshes. Applied to parameterized 2D/3D confined cylinder flows and an unevenly heated cavity over Reynolds/Rayleigh-number ranges, the models generalize to unseen conditions and accurately predict sparse probe locations and integrated quantities. Coarse-mesh GNNs match or outperform finer-mesh surrogates while reducing training time and memory, even for 3D chaotic dynamics. A dynamics-informed adaptive sampling strategy guided by frequency-spectrum variations reduces maximum errors by an order of magnitude versus uniform sampling at negligible cost. Coarse-mesh training also makes noise-augmented multi-step unrolling practical, improving long-horizon autoregressive predictions. The approach outperforms parametric DMD, sparse MLPs, U-Nets, and Fourier neural operators at lower cost, and enables system identification and state estimation from sparse observations. These results show that coarse-mesh representations, adaptive sampling, and noise-augmented unrolled training form an efficient framework for reduced-order modeling of parameterized unsteady flows without solver-resolution reconstruction.

This paper studies two non-transvenous implantable cardioverter-defibrillators (ICDs), namely the Subcutaneous ICD (S-ICD) and the Extravascular ICD (EV-ICD). Both avoid the complications of transvenous leads but are prone to oversensing, which can trigger inappropriate shocks associated with increased mortality. Oversensing risk is predicted by the ratio of T- or P-wave amplitude to R-wave amplitude (T:R, P:R), and these ratios vary over a 24-hour period. Yet S-ICD screening uses a single 10-second electrocardiogram (ECG) segment, and the EV-ICD has no formal screening protocol whatsoever. We present a 24-hour ambulatory screening framework for both devices which includes the first T:R analysis in an EV-ICD cohort and the first full evaluation of P:R prediction. We introduce AmpRatioNet, a family of 1D convolutional neural network architectures operating directly on the ECG signal, and benchmark it against the best existing approach - based on 2D phase space reconstruction images. Evaluation uses 4,605 clinician-annotated 10-second segments from 89 patients, with patient-level 10-fold cross-validation and bootstrap confidence intervals. AmpRatioNet-SE demonstrates superior performance on every metric in each setting with no device- or ratio-specific redesign. An explainability analysis against interpretable classical peak labelling algorithms shows that AmpRatioNet-SE's increase in error on oversensing-prone segments is up to 14 times smaller than theirs. Segment-level predictions are aggregated into a 24-hour timeline, extending S-ICD screening beyond 10 seconds and offering a plausible EV-ICD screening pathway pending device-level validation.

PairFormer-based all-atom models demonstrate strong capabilities in biomolecular binder design. However, existing methods often use all-atom representations for modality-specific binder design, such as protein and peptide design. In contrast, recent geometric latent diffusion models pursue unified binder design by representing molecular blocks from different modalities in a shared latent space. These complementary strengths motivate us to investigate how to generalize the use of pretrained all-atom representations to unified binder design. We introduce PairFuse, a parameter-efficient method that enables this generalization through representation fusion. Specifically, we use lightweight layerwise adapters to integrate pretrained all-atom representations into an independently pretrained geometric latent diffusion model. Our adapters integrate pretrained all-atom representations through two mechanisms: enriching the diffusion model's molecular block features and incorporating pairwise structural information into the geometric diffusion model's denoiser. Theoretically, we prove that the augmented denoiser preserves geometry and pairwise attention fusion provides additional attention expressivity beyond feature-only fusion. Experimentally, across antibody, peptide, and small molecule binder-design benchmarks, PairFuse achieves stronger performance than the pretrained diffusion model and existing representation-fusion baselines.

Benign fine-tuning can erode safety alignment in large language models, and sequential adaptation substantially compounds this risk. Prior work mitigates this erosion by replaying harmful-prompt refusals and benign-compliance examples, while other work instead matches base-model logits via distillation. However, the individual contributions of supervised exposure and distillation to safety behavior and refusal geometry during sequential adaptation remain unclear. We factor these two strategies apart and study their behavioral and geometric effects during sequential LoRA fine-tuning of Mistral-7B and Llama-3-8B across diverse task sequences (code generation, German-language summarization, and science question-answering). Behaviorally, supervised replay exposure yields the largest reduction in HarmBench attack success rate (ASR) on Mistral, averaging 51.1 points below base, but at a substantial over-refusal cost, averaging 53.4 points above base on OR-Bench-hard. Logit distillation, without any supervised replay loss, largely prevents safety erosion at a fraction of this cost: an average 5.4-point ASR reduction against only a 5.1-point rise in over-refusal. Llama shows the same asymmetry between the two strategies. Combining exposure and distillation moderates exposure's over-refusal while leaving downstream task scores largely unaffected. Geometrically, the two models diverge: on Mistral, distillation preserves the base refusal direction more closely than exposure does, whereas on Llama, refusal behavior after exposure largely survives ablation of the extracted refusal direction. This dissociation between models indicates that preserving the base refusal direction is not a reliable proxy for behavioral safety.

Earthquake sequences are stochastic on two separated timescales: the seismic waves of any individual earthquake propagate over seconds to minutes, while the earthquakes themselves arrive as a sparse, self-exciting occurrence process, with inter-event intervals from seconds to months. Respecting both scales in one training objective is the open challenge. Foundation models (FMs) for seismic waveforms instead import their learning targets from audio and vision, where every target is computed from a single fixed-length window. Such a window holds one realization of the wave propagation (an earthquake trace), never the occurrence process that decides when earthquakes happen. We develop nine model designs sharing one Mamba encoder-decoder, trained on a fixed corpus of 3M windows, varying the self-supervised objective, quantization bottleneck and mixer; the two key comparisons are architecture-matched, varying the objective alone. Each checkpoint is evaluated frozen on three downstream tasks. Varying the target redistributes downstream skill rather than raising it, and no checkpoint wins everywhere. Adding masking pressure to our strongest classifier re-specializes it rather than making it a generalist: it gains 6.9 points of few-shot P-wave picking precision and loses 9.6 points of few-shot classification accuracy. By the standard imported from language and vision, these models are feature extractors rather than FMs. Controls on capacity, architecture, and domain proximity do not explain why skill moves rather than rises. Against task-specific baselines the advantage is regime-dependent: our checkpoints beat them few-shot on classification and ground-motion prediction, and trail them on full data. We argue the binding constraint is the target, and that settling the question requires a learning target defined across events, the timescale that makes seismicity stochastic in the first place.

Numbers computed from trained models, such as an example's difficulty or a model's benchmark lead, change with seed and prompt. We ask when such measurements can be corrected or compared. Noise shrinks correlations, and the textbook correction divides by the square root of two estimated reliabilities. On two CIFAR-100 difficulty scores, it returned correlations above one in 70 of 100 classes. We prove that its mean squared error is infinite even when the correlation and the reliabilities use separate seeds. Our rule corrects only when both reliabilities pass a significance test, then divides by their upper confidence bounds. Its output is always bounded, and its coverage guarantees assume Gaussian noise. It helps once reliability is well estimated, more than halving the error with 500 simulated units. Prompt settings can reverse rankings. On ARC, Qwen2-1.5B trails Llama-3.2-3B by 3.8 points with no examples in the prompt and leads by 16.6 with two. Across four benchmarks, averaging the two settings names one winner for 15 of 16 such reversals. We compare models within each setting, controlling family-wise error asymptotically in the number of items for a fixed prompt grid. Aligning embedding coordinates across seeds raises held-out per-dimension reliability from 0.003 to 0.813.

Glioblastoma and other high-grade gliomas (HGG) have a poor prognosis and are associated with progressive functional decline. The standard of care to assess progressive disease is contrast-enhanced MRI, which is expensive, claustrophobic, and performed only once every 2-3 months. Wearable accelerometers offer a non-invasive alternative for continuously monitoring patient function between clinic visits. However, accelerometry data streams are typically noisy and highly patient-specific, making conventional time series methods and summary statistics poorly suited to detecting clinically meaningful change. This work aims to classify stable versus progressive disease states by applying rough path signatures for feature extraction from wrist-worn accelerometer data. We convert accelerometer recordings from HGG patients enrolled in the BrainWear study into walking-path segments and use signature kernel maximum mean discrepancy (SigMMD) to quantify change in patients' walking behaviour. We overcome the challenge of inter-patient heterogeneity by adjusting for the natural variation in a patient's walking during stable tumour periods. This patient-specific normalisation substantially improves separation, yielding a candidate biomarker of tumour progression. Our results provide proof-of-concept evidence that signature-based analysis of wearable accelerometry can distinguish stable from progressive disease states, motivating larger longitudinal studies of continuous wearable monitoring as a digital biomarker in neuro-oncology.

Reinforcement learning holds the promise of robots that learn manipulation skills from interaction alone, without pretrained priors or demonstration data. In practice, on-policy methods such as PPO routinely stall on long-horizon, contact-rich tasks: dense reward shaping carries the policy through stages, yet stitching them together and discovering sparsely rewarded events rarely succeeds. We trace this failure to the missing action discrimination in state-value credit assignment: a state-value critic judges progress only by episode outcomes, so when the policy converges to a partial solution all outcomes coincide, and improvement stops because no action can prove better than another. We therefore extend PPO with an action conditioned value critic which, can compare the executed action with alternatives sampled at the same state and aid the learning signal. Combined with Generalized Advantage Estimation and a bound on exploration noise, we show how the extended algorithm escapes the plateau and completes precise peg insertion within a training budget where PPO with identical exploration fails.

Context: Non-determinism in deep learning algorithm design and implementation leads to performance variation, meaning model performance is not a single value, but rather a distribution. These model performance distributions are underexplored despite their impact on robustness. Objective: We investigate the robustness of deep learning performance to sources of non-determinism, specifically focusing on how performance distributions differ across various architectures and tasks. Method: We conducted 186 experiments on image classification and time series forecasting architectures. Each experiment was run 100 times with different random seeds to generate performance distributions, resulting in 18,600 runs. Robustness was quantified using metrics for spread, symmetry, and tail risk. Results: Performance distributions are frequently non-Gaussian, particularly in time series forecasting. Model size does not systematically affect robustness -- larger image classification models show fewer outliers but not lower spread, while smaller time series models show lower spread but more extreme underperformers. Training duration does not scale linearly; early stopping effectively balances performance and robustness. Mean performance does not predict robustness - time series forecasting shows moderate correlation while image classification shows none. Time series models produce nearly three times more underperforming outliers than image classification models, indicating substantially higher tail risk. Conclusions: Tail risk poses serious concerns for Trustworthy AI in high-stakes applications. Models performing well on average may exhibit long tails and extreme outliers revealed only through distributional analysis. Mean performance alone should not guide model selection; assessment of spread, symmetry, and tail risk is essential for reliable model assessment where consistent performance is critical.

The allocation of electric vehicles to public charging stations is increasingly constrained by the fact that several heterogeneous facilities often serve the same stream of requests while observing only partial local information. An accept/reject decision that is defensible for one station may nevertheless reduce the efficiency of the charging district as a whole. We formulate this operational problem as a Decentralized Partially Observable Markov Decision Process (Dec-POMDP) and examine whether Centralized Training with Decentralized Execution (CTDE), instantiated through QMIX, improves assignment performance relative to a cooperative Independent Q-Learning baseline (IQL). The evaluation uses a six-station urban network simulated with SUMO, five independent random seeds, 200 training episodes, and four traffic flux scenarios ranging from 100 to 250 vehicles/hour. Under this protocol, QMIX reaches a mean reward per evaluated decision of +0.776, versus +0.578 for IQL (delta = +0.198, seed-paired Cohens dz = 2.98, exact two-sided Wilcoxon p = 0.0625, the minimum attainable value at this sample size), and raises the acceptance rate from 72.4% to 94.0% (delta = +21.6 pp, dz = 5.24, Wilcoxon p = 0.0625). A five-variant ablation suggests this gain is primarily attributable to the CTDE architecture when holding cooperative reward terms fixed, though not formally isolated with a single seed. By contrast, cooperative reward shaping and experience sharing show no clear descriptive benefit and may slightly reduce performance. These findings support the view that, for multi-station charging assignment, coordination should be embedded in the learning architecture rather than added only through heuristic cooperation terms.

Recurring cash flows are the backbone of cash-flow underwriting: they separate durable revenue from one-off deposits, expose standing obligations, and ultimately determine ability to pay. For small businesses, recurrence is both indispensable and hard to estimate, because much of their income arrives as ambiguous transfers, most payment streams are short, and schedules follow calendar conventions that frequency-domain methods treat as noise. We present CADENCE, Calendar-Aware Distributional Event Network for Cash-Flow Estimation, a compact causal Transformer that treats each transaction as a token and predicts the amount and a full, nonparametric timing distribution of every cash flow's next payment. Inheriting transaction semantics from an external labeler and a frozen text encoder keeps CADENCE compact yet general across the long tail of small-business counterparties, while predicting a calibrated distribution rather than an opaque embedding makes its outputs interpretable enough to feed existing underwriting models and power credit decisions. Ablations confirm the role of each design choice, with counterparty semantics driving recall on short histories, calendar inputs driving timing accuracy, and no-event supervision driving precision. Against the production FFT baseline, CADENCE improves one-day timing accuracy by 93.3% and next-event F1 by 6.7%, while producing calibrated distributions that serve directly as a confidence gate for downstream decisions. In an end-to-end credit evaluation on production applications, replacing FFT recurrence features with CADENCE features raises KS by 4.8%, raises the approval rate at a fixed loss rate by 18.2%, and lowers the loss rate at a fixed approval rate by 13.1%.

Multimodal large language models (MLLMs) are vulnerable to visual jailbreak attacks that bypass safety safeguards. Existing methods typically optimize isolated adversarial examples, limiting attack diversity, transferability, and generalization. We propose Jailbreak Connectivity (JC), a framework based on the observation that effective adversarial inputs can be connected through low-loss regions in image space. JC optimizes continuous adversarial paths: path sampling yields diverse jailbreak images, lightweight surrogate classifiers guide cross-model transfer, and a distribution-level objective enables universal attacks across harmful queries. On MM-SafetyBench, JC achieves average attack success rates of 84.01%, 85.77%, and 97.95% on Qwen3-VL-8B-Instruct, InternVL3.5-8B, and Gemma-3-12B-IT, respectively, exceeding the strongest baselines by up to 68.36 percentage points. In black-box evaluations, JC improves attack success rates over FigStep by 24.69 and 27.34 percentage points on GPT-5.6 and Gemini 3.1 Pro, respectively. These findings reveal structured vulnerabilities in MLLM safety alignment and inform more comprehensive security evaluation.

We introduce Nemotron 3.5 Super and Nemotron 3.5 Lightning, open and efficient Mixture-of-Experts Hybrid Mamba-Transformer models designed for reasoning, long-context understanding, and agentic workloads. Super is a 120B-parameter, 12B-active multimodal model, while Lightning is a smaller 30B-parameter, 3B-active text model; both designed for efficient deployment. Both models combine large-scale distillation, long-context training, and native speculative decoding, while Super further incorporates Indic continued pretraining and agentic software-engineering mid-training. We develop a comprehensive post-training pipeline spanning supervised fine-tuning and reinforcement learning (RLVR and RLHF), complemented in Super by specialized teacher training and multi- teacher on-policy distillation (MOPD). This transforms Super from a text-only base model into a multimodal agent capable of reasoning over text, images, video, and interactive tool environments. We further explore test-time scaling through parallel agent execution and self-verification. Across training and deployment, we jointly optimize capability and efficiency. Indic continued pretraining improves Indian-language benchmarks by up to 9.82 percentage points, while agentic mid-training strengthens software-engineering capabilities and robustness across agent harnesses. The post-trained Super model achieves 51.78% on SWE-bench Pro, surpassing Nemotron 3 Ultra despite using approximately 4.6x fewer active parameters. Post-training MTP boosting improves speculative- decoding acceptance length by up to 21.8% under probabilistic draft sampling, while NVFP4 quantization achieves 99.62% median accuracy recovery relative to BF16. Inference optimizations enable up to 4,856 output tokens/s/GPU on GB200 while sustaining at least 50 tokens/s/user.