Papers

Field-scale retrieval of surface soil moisture (SM), leaf area index (LAI), and plant height (PH) is essential for precision agriculture, yet it remains an ill-posed inverse problem. Concurrent variations in soil moisture and canopy density generate substantial ambiguities in radar backscatter and spectral responses, which reduces the effectiveness of traditional feedforward regression models in heterogeneous smallholder cropping systems. This study presents the Iterative Energy-Based Transformer (iEBT) for the joint retrieval of coupled soil-canopy states from Sentinel-1 C-band SAR and Sentinel-2 multispectral time series. Instead of direct regression, iEBT embeds multi-modal predictors within a shared sequence, produces an initial state estimate, and iteratively updates the target [SM, LAI, PH] vector through normalized gradient descent to minimize a learned scalar compatibility energy function. Using 700 quality-controlled field measurements from Varanasi, India, iEBT achieved the highest learned-model performance on the random test split, with a four-seed mean R^2 of 0.854 \pm 0.012 (R_SM^2 = 0.841, R_LAI^2 = 0.905, R_PH^2 = 0.821). WCM and PROSAIL were retained as physically interpretable SAR and optical reference models for comparison. Modality ablations confirmed that Sentinel-1 drives SM retrieval, while Sentinel-2 dominates LAI, whereas PH relies on combined structural-phenological signatures. Crucially, the model's terminal energy functions as an uncalibrated post-retrieval quality diagnostic; screening the 10% highest-energy samples markedly reduced target level root-mean-square errors. While leave-one-campaign-out validation highlights persistent cross-season domain shift challenges due to localized management variations, compatibility-guided multimodal fusion offers a structured self-diagnostic path toward reliable biophysical parameter estimation

We introduce a PAC framework in which the learner can access sampling oracles both before and at decision time. Sample complexity is measured by a pair $(n,m)$, where $n$ is the learning budget spent before a query is known and $m$ is the additional sampling budget per query. We demonstrate its relevance in discounted Markov decision processes with exogenous i.i.d.\ contexts revealed before acting. Contexts may affect both rewards and transitions but remain uncontrolled by the agent. The learner can sample the unknown context distribution and the transition kernel. We study policy evaluation (PE), best-value estimation (BVE), and best-policy extraction (BPE). When rewards and transitions are known, a variance-reduced algorithm solves all three tasks with sample complexity $\bigl(\widetilde O((1-\gamma)^{-3}\varepsilon^{-2}),0\bigr)$, which is minimax optimal up to logarithmic factors. Let $\mathcal{X}$ be the controlled state space. When transitions are also unknown, we give a PE algorithm with complexity $\bigl(\widetilde O(|\mathcal X|(1-\gamma)^{-3}\varepsilon^{-2}), \widetilde O((1-\gamma)^{-2}\varepsilon^{-2})\bigr)$ and matching lower bounds at this budget pair. For BVE and BPE, we give an algorithm with a common offline budget $\widetilde O(|\mathcal X|^2|\mathcal A|(1-\gamma)^{-4}\varepsilon^{-2})$ and respective query costs $\widetilde O(|\mathcal A|(1-\gamma)^{-2}\varepsilon^{-2})$ and $\widetilde O(|\mathcal A|(1-\gamma)^{-3}\varepsilon^{-2})$. Importantly, all bounds are independent of the context-space cardinality.

Sampling from an unnormalized target density by reversing an Ornstein-Uhlenbeck diffusion requires the score of each noise-perturbed marginal law. Two exact identities are available: Tweedie's identity and a target-score identity, each yielding unbiased finite-reference score estimators for the OU-marginal score. Score estimators induced by scalar blends of Tweedie and TSI score estimators can reduce variance, but they are too rigid for singular or strongly anisotropic targets. We formulate blended score estimation as a conditional risk-minimization problem over matrix valued blending coefficients, referred to as gates. Our central result is to show the optimal matrix valued gate for blended score estimation is given \[ G_\star(y,t) = \alpha_t^2 \left(\alpha_t^2 I_d + \gamma_t\, \mathbb{E}[H_0(X_0)\mid Y_t=y] \right)^{-1}, \qquad H_0=-\nabla^2\log p_0 .\] Here $\alpha_t = e^{-t}$ and $\gamma_t = 1-e^{-2t}$ are the OU coefficients, and the conditional expectation is under the OU posterior of $X_0$ given $Y_t=y$. We call this formula the Laplace-Fisher Gate Identity (\LFGI{}). Because the Tweedie-TSI disagreement has conditional mean zero, the gate changes the score-estimator variance but not its expected value. We derive the variance-optimal matrix gate, record the Gaussian special case, and establish finite-reference consistency and stability bounds for estimating the gate from weighted reference samples. We then use the finite-reference LFGI score estimator for normalized density evaluation in Bayesian inverse problems. In regimes where MCMC pilot samples and derivative information are already available, LFGI uses those byproducts to construct a normalized surrogate for the posterior density. The resulting surrogate supplies information that the MCMC samples alone do not provide: posterior-energy evaluation, model-evidence estimation, and downstream density-based diagnostics. On a PDE-constrained inverse-problem benchmark, the LFGI surrogate improves posterior-density calibration and sampling diagnostics relative to the other tested score-estimator classes. Experiments using LFGI with known model evidence check absolute evidence calibration in both Gaussian and non-Gaussian settings.

Accuracy degradation is the standard metric for Catastrophic Forgetting (CF), however, it records only whether forgetting occurred or not. It saturates at the extremes and collapses discretely at task boundaries, hiding the internal structure of what is being forgotten. We introduce six softmax-derived metrics spanning true-label rank (TLR), predictive confidence, and distributional divergence that characterize forgetting continuously, each normalized to [0, 1] with no modification to training. On CIFAR-100, these metrics carry information where accuracy does not: at 0% accuracy, the Confusion Margin spans an IQR of [0.32, 0.50] across classes that accuracy treats identically. We demonstrate that this richer signal is actionable in mitigating catastrophic forgetting. Per-sample metric scores used as loss weights reduce forgetting by 1.3 percentage points over uniform experience replay (ER) on CIFAR-100. Furthermore, the slope of a metric over a small window provides a stable sampling criterion: at a small-window size (e.g. 3 epochs), accuracy-trend degrades to 34.79% (std. = 2.32) while log-TLR achieves 41.07% (std. = 0.57). This gap is structural since reliable small-window trend estimation requires a continuous signal. On TinyImageNet, log-TLR trend sampling reduces forgetting by 7.7 percentage points over the ER baseline.

Length extrapolation in language models involves competing objectives: retrieval fidelity, long-document likelihood, short-context quality, and inference cost. We present ATMA, a 378M-parameter hybrid recipe that combines Polar Attention with gated-delta recurrent memory, and study these objectives as a Pareto problem rather than claiming general architectural dominance. Polar Attention separates a normalized direction channel from a bounded participation-ratio magnitude channel. We select the recipe with a complete 120-cell, 1B-token factorial sweep, then train matched NoPE, RoPE, and Polar variants for 9.816B tokens at length 2K and evaluate them through 256K. Across the factorial, memory improves Polar's 64K retrieval score in all 20 matched cells (mean +47.8 points), whereas its effect on NoPE is small and inconsistent. At 256K, Polar retains 34.4% teacher-forced target-token accuracy and 9.0% exact five-token accuracy; exact retrieval is 18.0% on synthetic contexts but 0.0% on FinePDFs contexts. Polar also limits mean fixed-target bits-per-byte degradation to 1.26 times, at a 1.9-point mean cost on eight short-context tasks. Raven baselines lead BABILong and have length-independent decode state, illustrating a different point on the frontier. Finally, a post-hoc checkpoint audit shows that nearly identical 2K validation curves can conceal a 6.70-nat difference at 256K. Because those runs were neither seed-paired nor randomized across devices, we interpret this as checkpoint variability associated with an infrastructure transition, not a causal hardware effect. Code: https://github.com/kreasof-ai/atma

Physics-informed neural networks (PINNs) embed governing equations in their loss function, enabling mesh-free solutions to partial differential equations. Low training loss is treated as evidence that the learned solution is physically correct. This paper shows that assumption breaks down when encoded physics are incorrect. By perturbing PDE parameters before training, a setting we describe as physics parameter poisoning or parameter misspecification, we produce models that train to low loss but give incorrect answers; we treat the perturbation schedule as sensitivity analysis rather than only as a security threat, and none of our claims requires an adversary. Achieving low residual loss does not discriminate accurate from inaccurate solutions: poisoned models reach losses at or below the clean baseline yet differ by large margins, so driving the residual down is not evidence of physical accuracy. Across three PDE systems (Burgers equation, Navier-Stokes cavity, and convection-diffusion), poisoned models match or beat the clean-model training loss while their solutions differ by up to 71% in the fixed sweep and up to 128% under adversarial search; at Cavity Re=400 the poisoned loss falls below the clean baseline. We define a detection difficulty ratio R (solution error divided by training loss) to summarize how invisible the corruption is, though cross-PDE comparison is complicated by differences in loss scale. We test six candidate defenses, none of which reliably detects corruption across all regimes. We propose a post-hoc defense: sweeping the PDE residual loss across parameter values without retraining. The loss minimum recovers the true training parameter without external data, and generalizes across all three PDE systems. The effect holds across five network architectures (8.7K to 133K parameters), is bidirectional, and is confirmed across multiple random seeds.

User modeling in industrial recommender systems typically produces dense embeddings, which suffer from representational constraints due to fixed-dimensional vectors. As an emerging alternative for discrete user representation, LLM-generated natural-language user profiles often capture language semantics rather than sequential patterns in user behaviors. Meanwhile, Semantic ID (SID) is effective for item representation, yet existing discrete user representations remain largely confined to quantizing already-pooled user vectors, clustering historical watches, or predicting coarse topic clusters solely for candidate retrieval. We propose TokenMinds, an industrial system that extends the PLUM framework from SID-based item retrieval to generative sequential user modeling. Built on an encoder-decoder architecture adapted from pretrained LLMs, TokenMinds introduces a dual-output design: the encoder outputs a dense user embedding capturing a holistic user state, while the decoder generates discrete SID-based user tokens representing multi-faceted future interests. Grounded in item semantics, these tokens preserve distinct interests that a single pooled vector would average away. Additionally, a shared SID vocabulary across long- and short-form videos unifies cross-scenario behaviors into a single model, cutting upstream training and serving compute by 50% and 31%. Furthermore, leveraging the text vocabulary inherited from the pretrained LLM, TokenMinds interleaves textual search queries directly with watch histories to capture explicit user intent. We validate TokenMinds through offline evaluations and live A/B experiments across YouTube's primary ranking surfaces, serving billions of users asynchronously: SID-based user tokens are effective on their own, yield additive gains alongside dense embeddings, and together outperform the previous-generation, non-LLM user model in production.

Volume and quality of datasets are crucial for deep learning model training, yet they are often constrained by availability and data acquisition costs. Synthetic data augmentation can extend existing datasets with realistic images, and the quality of these images is generally assessed through fidelity metrics such as FID, KID, IS, LPIPS and SSIM that measure structural or distributional similarity. However, such metrics, including the widely used FID, focus on visual fidelity without reflecting downstream utility, and can diverge from human perception under perturbations that are imperceptible to human observers. In this work, we systematically evaluate Earth observation datasets alongside synthetic counterparts generated by deep generative models, comparing automatic metrics against human perception and downstream tasks. Our results reveal a stark misalignment: semantics-preserving perturbations such as rotation drastically alter metric scores while leaving human recognition unaffected, and synthetic samples that score poorly on automatic metrics achieve comparable or higher perceived realism, and can improve downstream performance when combined with real data. By benchmarking semantic segmentation models trained on mixed real-synthetic datasets, we demonstrate that quality metrics rooted in ImageNet-pretrained feature spaces are unreliable indicators for geospatial data. Our findings underscore that automatic quality evaluation of synthetic datasets should be grounded in downstream task performance and human evaluation.

We investigate how reward design shapes the internal attention patterns of reinforcement learning agents trained for autonomous driving. Using three Perceiver-based agents that share identical architectures and training data but differ only in their reward configurations$\unicode{x2014}$ranging from basic violation penalties to continuous proximity penalties$\unicode{x2014}$we analyze cross-attention allocation across 50 real-world scenarios from the Waymo Open Motion Dataset. A central methodological finding is that naïve pooling of timesteps across episodes substantially underestimates the attention$\unicode{x2013}$risk relationship; within-episode correlation with Fisher z-transform aggregation is the appropriate statistic and reveals a robustly positive link between collision risk and agent-directed attention. Building on this validated methodology, we demonstrate two reward-conditioned effects: agents trained with navigation rewards allocate up to $2.0\times$ more attention to GPS-path tokens than those trained with additional proximity penalties$\unicode{x2014}$and $4.7\times$ more than agents with no navigation incentive$\unicode{x2014}$revealing that reward content directly determines which scene elements the encoder prioritizes, and continuous time-to-collision penalties create a $\textit{learned vigilance prior}$$\unicode{x2014}$elevated resting agent surveillance maintained throughout collision-free phases. In several scenarios, the complete-reward and minimal-reward models exhibit opposite attention$\unicode{x2013}$risk correlation directions, demonstrating that reward design can qualitatively reverse attentional strategy rather than merely modulating its magnitude. These results suggest that attention analysis is a practical diagnostic for verifying that a reward function produces the intended representational behaviour in safety-critical RL systems.

Autonomous aerial vehicles operating in shared airspace must predict the future positions of non-cooperative obstacles to plan evasive maneuvers before a collision becomes unavoidable. Unlike cooperative systems that share intent, non-cooperative obstacles such as birds, uncontrolled drones, or debris exhibit multi-modal motion that deterministic predictors cannot adequately represent. Existing methods either rely on recurrent encoders that propagate temporal information sequentially, limiting their ability to capture long-range kinematic precursors of maneuver initiation, or produce point forecasts that provide no distributional information to downstream planners. This paper presents AeroCast, a probabilistic trajectory prediction framework that combines a Transformer encoder with a Mixture Density Network output head to predict per-timestep Gaussian mixture distributions over future three-dimensional displacements. A translation-invariant consecutive displacement encoding and a calibration-oriented training objective address the input design and mode-degeneracy challenges specific to mixture-based aerial trajectory prediction. On a hybrid real-and-synthetic quadrotor corpus spanning nine motion categories, AeroCast reduces Average Displacement Error and Final Displacement Error by approximately 50% relative to the baselines over a five-second horizon, and achieves the lowest negative log-likelihood and Continuous Ranked Probability Score among all compared methods. Ablation analysis identifies velocity input and model capacity as the primary contributors to prediction quality, and positional encoding as essential for long-horizon trajectory coherence. AeroCast inference completes in 0.1ms per sample, compatible with real-time onboard deployment at 100Hz.

On-device language-model agents improve by accumulating experience in retrieved memory rather than by updating weights. This memory is hard-bounded and exposed: it consumes RAM and energy, reaches peers through a thin uplink, and becomes an attack surface because it is writable by what the agent reads. Existing systems each cover one part of this problem: agentic memories grow without a budget, on-device methods keep entries by success alone, and poisoning is studied mainly as an attack rather than as a memory-governance problem. We propose \sys{}, a single net-value-per-byte score that governs an agent's experience-memory lifecycle. The main idea is to let the budget act as the curator: each entry is scored as value minus harm, per byte, so one ruler decides what to keep, share, and trust. \sys{} makes three decisions: (1) KEEP evicts low-value bytes under the RAM and energy budget; (2) SHARE sends an insight only when its value exceeds its uplink cost; and (3) TRUST gates a peer entry by provenance. On language-model-agent task-drift benchmarks and a real heterogeneous Jetson testbed with two robot-arm nodes and a hub, \sys{} reduces memory by $2.7\times$ and uplink by $2.4\times$, drives injection success from 0.75 to zero, and raises accuracy on cases corrupted by poison or stale memory. Curating by net value reduces footprint, energy, uplink, and injection success together without reducing accuracy. In this setting, forgetting by net value improves the agent rather than weakening it.

We introduce Directed Hypergraph Signal Processing (DHGSP), a unified framework that extends graph signal processing to accommodate both higher-order (polyadic) and asymmetric (directional) relationships simultaneously. Using the tensor singular value decomposition (t-SVD) within the t-product algebra, we define a novel adjacency tensor for directed hypergraphs, a topologically faithful shift operator, and a lossless Directed Hypergraph Fourier Transform (t-DHGFT). Experiments on real traffic networks demonstrate that DHGSP outperforms matrix-based (graph and digraph) and undirected tensor-based (hypergraph) baselines in denoising tasks.

Speculative decoding accelerates inference by letting a draft model propose tokens for a target model to verify, raising a concrete safety question: at temperature zero, can draft-side behavior leak into safety-scored outputs? We answer with Typical-Acceptance Invariance Screen (TAIS), a behavioral-equivalence screen that pairs target-only and speculative outputs on the same safety battery and requires byte-identity evidence, TOST equivalence at +/-3pp, and per-task Cohen's h below a calibrated null cutoff of |h| < 0.1. Applied to a 16,783-sample confirmatory core plus 44,066 matched expansion samples (fp16/bf16 execution, canonical and DPO-adversarial drafts, GPTQ-4bit drafts, two seeds, and four safety benchmarks), the tested temperature-zero vLLM stacks show no detectable safety divergence under TAIS. The largest absolute Cohen's h on matched target-only versus speculative refusal is 0.024, roughly an order of magnitude below the conventional trivial-effect floor; 25 of 27 per-task TOST contrasts pass at the +/-3pp margin (the two non-pass contrasts are capability-domain Wald-CI edge cases at identical ceiling rates, not genuine non-equivalence); the DPO-adversarial draft produces byte-identical output to the canonical draft across 4,006 samples; and bf16 changes 36%-53% of output bytes without moving any per-task safety rate outside equivalence. A separate 4,006-sample 70B production-scale probe, which lacks a matched 70B target-only arm and is therefore not counted as a TAIS pass, produces AdvBench refusal 0.839 over 700 AdvBench completions with 95% Wilson CI [0.809, 0.864]. We make no claim about sampling temperatures, untested frameworks, untested model families, or tree-speculation variants such as EAGLE and Medusa.

Sparse Mixture-of-Experts (MoE) models route each token to a few of many experts, inviting the hypothesis that experts form functional modules tied to capabilities or languages. We test this causally on Command A+, a frontier open-weights MoE (218B total / 25B active; 128 experts, 8 active, +1 shared). We build a routing-mass atlas, pre-register six family-to-axis hypotheses before any intervention, and ablate each family at inference time against a size-matched random-expert null, measuring whether it selectively breaks its own axis (worst off-target effect at most one third of on-target). Crucially, we test the same families under four metrics and a held-out, independent-corpus run with bootstrap confidence intervals. Our finding is cautionary: robust functional modularity is rare and measurement-dependent. Of six pre-registered families, only one, the Arabic-language family, is a clean selective module that survives an independent corpus and a conservative statistical bar (1/6; a more permissive pre-registered point rule admits 3/6, but that count is threshold-sensitive). Every other family has a real causal effect yet fails selectivity, and its apparent modularity flips with the measurement: with the corpus, the metric, and the statistical bar. A positive control on Qwen3-30B-A3B recovers its published disjoint structure, confirming the method detects modularity when present. The verdict reproduces on the un-quantized BF16 model, ruling out a 4-bit quantization artifact. We conclude that ablation-based modularity verdicts are not safe unless the corpus, metric, and statistical bar are controlled. We release the atlas and ablation data.

Many modern Language Model (LM) pipelines return an averaged model, such as an exponential moving average of the training iterates, rather than the final iterate itself. This raises a fundamental question: given that we will return an iterate average, how should we change training to improve the performance of this average? We study this question by formulating optimizer design for the iterate-average estimator as an optimal-control problem. In a continuous-time stochastic quadratic model, we solve for the control strategy that minimizes the error of the returned average subject to a penalty on the size of the intervention. A practical approximation to this controller yields PACE, a lightweight wrapper around AdamW that pulls the live weights toward their exponential moving average with a clipped, per-coordinate control strength. We prove that a stylized version of PACE converges at the standard stochastic convex optimization rate, up to a factor depending on the averaging rule, while in the quadratic setting it can strictly improve the limiting squared error of the iterate-average estimator and can do so by an arbitrarily large factor on some instances. Empirically, our results suggest that PACE improves over AdamW and EMA-evaluated AdamW in supervised fine-tuning of 1-2B parameter LMs and in GPT-2 pretraining on FineWeb for a wide range of learning rates, decay schedules, and other hyperparameters.

In cooperative multi-agent reinforcement learning (MARL), from a deployment perspective, it is challenging and expensive to train agents from scratch for each new environment or task. In this work, we propose GCT-MARL, a transfer learning framework that builds on the multi-view graph contrastive backbone of MAIL and augments it with a per-view, adaptively weighted alignment loss and a two-phase training protocol specifically designed for transfer across populations of varying sizes and compositions. We empirically demonstrate that the proposed framework markedly accelerates convergence on the target task relative to from-scratch training, in both homogeneous (within-faction, varying N) and heterogeneous (cross-faction and mixed unit-type) transfer scenarios. Furthermore, we show that the framework naturally supports continual learning by sequentially chaining the two-phase transfer protocol across a series of related tasks. Overall, this work provides a unified approach to mitigating key limitations in current MARL transfer methods with new insights at both methodological and empirical levels.

Online time-series forecasters receive labels only after horizon-dependent delays, while every adaptation step spends limited compute. We study when an online learner should update, not how to adapt at every opportunity, and introduce ADOWIP: a residual-adapter framework with sealed delay queues, exact budget accounting, and auditable update telemetry. Its main scheduler is an observed decision-loss priority gate that updates only after feedback is revealed, when downstream loss, optionally penalized by prediction MSE, exceeds a calibrated empirical quantile and budget remains. We prove hard-budget feasibility, projected-OGD regret for a convex linear accepted-update subproblem, and stability plus conditional finite-sample gate-selection statements. On public ETT capacity-planning tasks, a frozen calibration/evaluation split selects a gate that lowers held-out decision loss against always, fixed-period, and drift-triggered exact-update baselines under matched compute. Secondary threshold/load-index ETT suites are mixed: 33 of 41 selected contrasts clear the stricter cross-artifact Holm family, and the 8 nonpassing rows are explicitly excluded from primary claims. The same protocol improves an external UCI Bike capacity proxy with 20/0 held-out wins, and a fixed gate passes three full-year Capital Bikeshare station-rebalancing contrasts. Probe-based and finance experiments remain negative, delimiting the current scope of decision-prioritized adaptation.

Recovering governing Ordinary Differential Equations (ODEs) from data is a central challenge in modeling dynamical systems across scientific domains. Existing approaches cast discovery as a static inference problem over fixed datasets, assuming that the observed trajectories are sufficiently informative. However, dynamical systems evolve over large state spaces, and limited data can make multiple equations observationally indistinguishable, leading to identifiability gaps and the recovery of incorrect governing equations. To address this, we introduce LLM-ACES, or LLM-guided Active Closed-loop Equation Search, a closed-loop framework that jointly optimizes symbolic hypothesis construction and adaptive data acquisition. In LLM-ACES, a large language model (LLM) proposes operator priors that partition the large search space into distinct regions, within which candidate equations are fit to the observed data. The disagreement among these candidates guides the acquisition of informative trajectories, creating a feedback loop that iteratively refines both the hypothesis space and the discovered dynamics. On 122 ODE systems spanning ODEBench and ODEBase, LLM-ACES achieves the lowest median NMSE, outperforming state-of-the-art baselines by several orders of magnitude while achieving a high symbolic accuracy of 46.2% and 52.4%, respectively. Our analysis further shows that LLM-ACES is sample-efficient, achieving better performance with one-tenth the data. Furthermore, LLM-ACES's feedback-driven data acquisition makes it robust to noise and recovers the correct symbolic structure, while baselines introduce spurious terms that fit the data locally but obscure the true governing relationships.

Today's reasoning models use thinking tokens to attain stronger performance on benchmarks than their instruction-tuned counterparts. It is also generally believed that this more "deliberative" mode should improve alignment and safety, by providing the model a safe space to consider whether its planned answer to a request violates its safety principles. We present evidence that this intuition is not always correct. Across frontier open-weight reasoning models spanning GPT-OSS, Qwen, Olmo, and Phi families, we find that the eventual refusal/compliance outcome is already strongly predictable via a trained head on the first token's hidden representation ($0.84$-$0.95$ AUROC and $\sim88\%$ balanced accuracy for predicting refusal/compliance) before any visible thinking. The thinking process turns out to be more akin to prefix completion than to deliberative revision, with the final outcome rarely changing after the first $\sim20\%$ of thinking, despite giving the appearance of deliberation at the text level ($\sim74\%$ of text-level deliberations occur when the response distribution is already locked to one refusal/compliance side). We also find that existing inference-time and training-based safety interventions, despite being motivated by the goal of inducing deliberation, largely shift model behavior toward over-refusal while suppressing already-scarce deliberation signals. Our results suggest that safety behavior in current reasoning models is much less deliberative than commonly assumed, and highlight the need for methods that induce real safety deliberation.

Many reinforcement learning (RL) problems in the infinite-horizon average-reward setting require optimizing multiple conflicting objectives while satisfying multiple safety constraints. A common approach is concave scalarization, where the agent maximizes a utility $ f(J^\pi_{r_1}, \ldots, J^\pi_{r_M}) $ subject to a scalarized constraint $ g(J^\pi_{c_1}, \ldots, J^\pi_{c_N}) \ge 0 $, where $J^\pi_{r_m}$ and $J^\pi_{c_n}$ denote the average-reward and cost under policy $\pi$. However, the nonlinearity of $f$ and $g$ introduces bias in policy-gradient and actor-critic methods, since gradients must be evaluated using noisy estimates of $J^\pi,$ and $ \mathbb{E}[\partial f(J^\pi)] \neq \partial f(\mathbb{E}[J^\pi]),$ and this bias propagates through both primal and dual updates. We propose an MLMC-based primal-dual Natural Actor-Critic algorithm for average-reward MDPs that controls bias in scalarized objectives, constraint evaluation, and actor-critic estimation without requiring mixing-time knowledge. We show that the algorithm achieves optimal global convergence and constraint-violation rates of $ \tilde{O}(1/\sqrt{T}) $. To our knowledge, this is the first result establishing optimal convergence for concave scalarized multi-objective RL in the average-reward setting, both with and without constraints, and the first to do so without mixing-time information even in the absence of scalarization.

Neural scaling laws for transformer language models predict smooth improvements in pretraining loss with increasing parameters, but downstream capabilities such as in-context learning are known to emerge abruptly past a certain model scale. In this paper, we show that emergent capabilities arise stochastically throughout training, with larger models acquiring them earlier on average. We demonstrate that the emergence of capabilities such as pattern completion and indirect object identification corresponds to the abrupt learning of task-relevant attention patterns. To isolate this phenomenon, we train transformer models on synthetic linear map and cellular automata datasets, and we show that the difficulty of learning attention patterns depends on context length and pattern sparsity. Moreover, scaling the number of attention heads improves learning efficiency on our synthetic tasks, while increasing the head dimension yields diminishing returns past a minimum capacity. We additionally investigate architectures with alternative attention mechanisms, showing that MLP-Mixer outperforms a transformer on linear map tasks with complex attention patterns. Our findings provide a mechanistic insight into emergence, showing that downstream capabilities arise abruptly due to the intrinsic difficulty of learning sparse attention patterns in transformer models.

Neural scaling laws describe how pre-training loss decays as power laws with training time, model size, and compute. This position paper argues that the exponents of these power laws are fixed by generic mechanisms: a one-third time scaling due to the strong nonlinearity of Softmax, an inverse width scaling due to representational superposition, and an inverse depth scaling due to ensemble averaging of Transformer layers. These mechanisms are robust to a wide range of data structures and architectural details, placing current large language models in a universality class with fixed exponents. The coefficients, however, are expected to be sensitive to data and architecture details, and directly determine practical quantities such as the optimal model shape and the compute-optimal frontier. We therefore argue that understanding the coefficients is the key to near-term performance improvements, and that a closer examination of the current universality class may reveal pathways to better universality classes.

Financial fraud detection in digital banking requires reasoning over multiple heterogeneous event streams – transactions, login sessions, risk signals – that individually appear benign but collectively reveal fraudulent patterns. We propose the Multi-Stream Fraud Transformer (MSFT), a unified architecture that encodes each event stream with independent Transformer encoders and fuses their representations through configurable mechanisms. We conduct a systematic ablation study comparing five fusion strategies: concatenation, gated fusion, time-aware positional encoding, cross-stream attention, and a full combination. On a large-scale dataset (10M users, 1.5% fraud rate) with 85M parameter models, we demonstrate that (1) sequence models significantly outperform gradient-boosted trees operating on aggregated features (0.74 vs. 0.99 AUROC), (2) per-stream encoding is essential – a single-stream Transformer baseline with matched parameter budget reaches only 0.82 AUROC, an 18-point gap that confirms the multi-stream inductive bias is necessary, (3) time-aware positional encoding achieves the highest discrimination (0.9961 AUROC), (4) gated fusion yields the best precision (0.989) suitable for production deployment, and (5) the risk event stream provides the strongest individual signal contribution. We further validate on proprietary production data from a digital banking platform, showing over 22% relative AUROC improvement over the XGBoost baseline.

Target-specific peptide design requires sequence and structure co-design under full atom geometric constraints. Latent generative frameworks offer an effective route for this problem by compressing fine grained atomic structures into block level latent representations and performing conditional generation in a compact latent space. However, the scalability of such systems depends heavily on the geometric backbone used throughout their encoding, decoding, and denoising components. We introduce MEET (Memory Efficient Equivariant Transformer), an E(3) equivariant backbone for scalable atomistic peptide modeling. MEET maintains coupled invariant scalar and equivariant vector feature streams, while reformulating geometric computation around memory efficient attention. It initializes vector features through global coordinate aggregation, incorporates pairwise distances through augmented query and key dot products, and injects covalent bond information through sparse bond adaptation. Integrated into a VAE and latent diffusion pipeline for full atom peptide generation, MEET achieves linear memory scaling with atom count and improves generation quality over existing peptide design methods. Experiments on large scale AFDB derived datasets further show that the proposed backbone supports systematic model and data scaling, leading to better binding affinity, physical validity, and sample diversity.

In machine learning, model certification has been identified as an important method for gaining assurance about a model's trustworthiness and quality. A model's quality is largely determined by its ability to generalize, i.e., to perform well on data beyond what it was trained on. It is not possible to certify generalization directly, however, as it depends on unknown data and is not directly measurable. Proxies such as test accuracy can be misleading when the training process is perturbed (intentionally or accidentally), and metrics such as sharpness – which has an empirically supported link to generalization – are computationally expensive and can also serve as unreliable signals when training deviates from a prescribed procedure. In this work, we propose directional sharpness, a metric designed to efficiently and reliably indicate generalization despite potential training deviations. We provide empirical and analytical evidence that directional sharpness (1) correlates more strongly with generalization than existing metrics and (2) identifies models with poor generalization more reliably than existing metrics. Furthermore, directional sharpness is efficiently computable in model auditing settings, where the verifier has access to training data, and via zero-knowledge proofs that certify quality without revealing training data.

Federated split learning (FSL) enables collaborative training across bandwidth-constrained IoT devices, but repeated activation and gradient exchange creates a communication bot-tleneck. Prior work optimises either activation compression or synchronisation frequency in isolation. This paper presents an FSL framework for IoT rainfall prediction that jointly regulates activation compression and the synchronisation interval h̊o} via a latency driven scheduler on a server with per client EMA smoothing. The system is evaluated on hourly ERA5 data from 11 weather stations through a 17 scenario simulation matrix and a four scenario Raspberry Pi deployment over a real wide-area link. The simulation matrix validates scheduler switching across low, high, and mixed latency profiles, while the Pi deployment validates the high latency endpoint selected by the same policy. AUPRC varies only slightly across configurations (0.6381-0.6484 in simulation; within 0.011 on Pi), indicating that aggressive quantisation and sparser aggregation do not materially degrade predictive quality in this setting. On Pi, the selected endpoint (int8 with rho=3) achieves an 87% reduction in activation upload payload and a 54% reduction in synchronisation traffic relative to the float32 baseline, while reducing runtime jitter from +/-688 s to +/-10 s.

Traffic accident reconstruction is a forensic inverse problem that requires recovering physically consistent motion from sparse and heterogeneous evidence. Existing learning-based approaches predominantly optimize for semantic plausibility or visual realism, rather than quantitative agreement with measurable geometry and dynamics. Here, we present TRACER, a training-free framework that formulates reconstruction as a closed-loop structured inference process. Instead of directly generating dense trajectories, our framework constructs and iteratively refines event-anchored motion hypotheses under geometric, kinematic, and interaction constraints, guided by structured case memory and consistency-driven diagnosis. This design enables incremental, interpretable corrections when evidence is insufficient, making the accident reconstruction process more aligned with the workflow of human experts. Experiments on real-world accident data show that TRACER achieves improved geometric fidelity, velocity consistency, and collision accuracy over both data-driven and physics-based baselines.

Machine unlearning (MU) is commonly judged by output forgetting, such as low forget-set accuracy or reduced membership inference. But if output-level success can coexist with retraining-inconsistent residuals in representation space, what kind of forgetting are current evaluations actually certifying? We study this question through retraining-consistent representation forgetting, using a population of retrained models (i.e., trained from scratch without the forget data) as an operational reference distribution for ideal forgetting. Across multiple unlearning methods, datasets, and models, our theoretical analysis and empirical results show that standard output-level evaluation can systematically overestimate the success of unlearning. Under this stronger lens, current methods often appear forgotten at the output layer while exhibiting a structured mismatch relative to retraining. They partially align with retraining on forget samples, remain more inconsistent on retain samples, and tend to leave residual discrepancy localized along retraining-related directions rather than diffuse in representation space. This structured mismatch is characterized by forget/retain asymmetry, directional mismatch, and concentrated residuals along retraining-related directions. In many cases, the discrepancies exceed the variability of independent retraining and lead to a risk of information recovery during relearning attacks. Overall, the results suggest that current MU is often evaluated for apparent forgetting rather than retraining-consistent forgetting. More broadly, retraining reveals what output forgetting hides.

To evaluate whether vision-language models can reason about geological histories, it is necessary to construct observations for which the underlying process history is known. Furthermore, reasoning over geological histories is not just a question of recognizing visual patterns, but also of understanding temporal and structural relationships that may be only indirectly visible or highly ambiguous. When ground-truth event histories are not uniquely identifiable or are unavailable, it remains an open challenge to teach models capable of visual reasoning to produce valid geological reconstructions that are consistent with both observed evidence and geological principles. We therefore investigate whether defining a verifiable geological reasoning task can improve geological event reconstruction across observation domains through reinforcement learning with verifiable rewards (RLVR). To this end, we present Geo-Strat-RL, a synthetic environment that generates stratigraphic observations and compact visible-evidence event histories. The environment combines a geological generator with an executable verifier that scores chronology, event identity, deposition, and structural relationships. We show that RLVR improves geological reconstruction in vision-language models (VLMs), increasing geological content scores on held out stratigraphic diagrams. We further evaluate the same held-out geological histories in a synthetic seismic observation domain by converting the generated scenes into acoustic-impedance-derived amplitude sections. In this controlled paired-renderer setting, we present evidence that geological reasoning learned from stratigraphic diagram-domain RLVR training transfers to synthetic seismic representations without seismic-specific training examples, supporting the hypothesis that RLVR can teach reusable geological reasoning concepts across related observation formats.

High-dimensional partial differential equations (PDEs) with unknown coefficients arise widely in scientific machine learning, including continuous-time reinforcement learning, yet solving them efficiently in a data-driven way remains challenging. Existing deep learning solvers often rely on repeated automatic differentiation to evaluate differential operators, which can cause instability and amplify derivative errors in high dimensions, while probabilistic methods based on stochastic representations require explicit knowledge of the data-generating dynamics and therefore do not apply to black-box environments. We introduce two types of simulators as data-generating mechanisms, and take a “representing-then-learning" approach that learns the solutions and their derivatives under settings where the underlying PDE operators are accessible only through simulations and pointwise evaluations. Our representation of derivatives relies on the zeroth-order derivative (ZOD) estimators derived from perturbed Monte Carlo trajectories. This fully model-free approach generates targets for the gradient and Hessian networks using only function evaluations. We provide a statistical learning analysis of the proposed approach, including a bias–variance tradeoff for ZODs. Assuming a standard contraction property of the underlying operator, we establish a non-asymptotic error bound that decomposes the total error into discretization error, approximation error, statistical error, and ZOD bias. Crucially, we derive the sample complexity of the learned representations in (weighted) Sobolev space, characterizing the error up to second-order derivatives. Numerical experiments illustrate the competitive performance of the method in moderate and high dimensions.