Trainability

Model post-training, and in particular reinforcement learning (RL), is one of the primary mechanisms by which developers can shape models' values and behaviors. However, as models become increasingly evaluation and training aware, they may be motivated to resist training when the perceived objective conflicts with their current values, undermining developers' ability to detect misalignment and correct model behavior through further training. In this paper, we demonstrate generalization hacking, in which a model collects reward during RL while preventing the rewarded behavior from generalizing. We construct a model organism on Qwen3-235B-A22B, finetuning on synthetic documents describing training awareness and self-inoculation, a novel mechanism in which the model frames compliance as context-specific in its chain of thought, without demonstrating or instructing either behavior. The model organism achieves train-time harmfulness comparable to controls while maintaining a persistent ${\sim}15$ percentage point compliance gap across 700 steps of RL. Additionally, a control organism trained only on training awareness documents independently discovers inoculation-like reasoning under RL pressure, developing its own compliance gap despite never being exposed to the concept. Because the generalization-hacking organism receives high reward throughout, standard training metrics provide no signal that generalization has failed. Our results constitute the first demonstration that a model can actively resist RL behavioral modification while maintaining high reward, suggesting that as models become more capable and training-aware, they may be able to undermine the training process itself.

Distilling a large speech foundation model (SFM) into an efficient student model has been successfully applied to low-resource environments. Although distillation reduces inference latency, it requires an additional student model training. However, the training efficiency of SFM distillation remains underexplored. In this work, we explore training acceleration of SFM distillation to speed up model deployment. We examine the potential of stacking, in which the model depth is progressively increased through training until the target model depth is reached. While existing stacking methods improve training speed, they suffer from performance degradation. To handle this limitation, we propose interleaved stacking, a novel stacking method that consistently preserves layer position throughout the stacking process. This property is particularly critical in SFMs, in which each layer encodes distinct layer-specific knowledge. We validate the effectiveness of the proposed method on SUPERB.

Large-scale model training increasingly relies on composing multiple parallelism strategies, such as data, pipeline, and expert parallelism, together with memory-saving optimizations like ZeRO. Deployed systems for foundation model pretraining often rely on human experts to manually design a high-level parallelism strategy then implement the corresponding low-level execution strategy, making it difficult to adapt the system to new strategies. Meanwhile, many general-purpose frameworks are more flexible but their implementations are still tied to a fixed set of common parallelism strategies, making it challenging to integrate state-of-the-art strategies. We present Piper, a user-controllable distributed training system that decouples the strategy from the runtime implementation. Piper allows users to declare a comprehensive distributed training strategy with a small set of model annotations and scheduling directives. Each directive applies a transformation on Piper's intermediate representation (IR), a unified global training DAG that represents all computation and communication. Using this IR, Piper compiles per-device execution plans and executes them with a distributed runtime agnostic to the strategy. We show that the combined system maintains performance parity on commonly available strategies such as ZeRO, while also enabling additional performance and memory efficiency gains through joint scheduling of compute and communication in composed parallelism strategies such as DeepSeek-V3's DualPipe.

We study dense-to-sparse continual training as a way to construct channel-sparse large language models from dense checkpoints. Starting from a Qwen2.5-8B dense backbone, we continue training at 32K context and introduce a predictor-gated sparse SwiGLU FFN in the 32K stage. For each token and layer, we use a low-rank predictor to produce FFN-channel routing logits. We then apply a bank-wise top-k rule to retain 16 channels in every 64-channel bank, yielding 4x sparsity in the FFN intermediate activation. Unlike post-hoc sparse inference methods, the routing module is placed on the main language modeling path and optimized during continual training, enabling the dense model to be upcycled into a hardware-oriented sparse model. We report the architecture, training recipe, benchmark performance, and training lessons. We also identify a layer-local long-context failure mode on RULER-CWE and propose a single-layer repair algorithm that substantially improves the affected length range.

Quantum Circuit Born Machines (QCBMs) offer a natural approach to generative machine learning by leveraging the Born rule. Recent work has provided a method to classically train QCBMs with Instantaneous Quantum Polynomial (IQP) circuits via the Maximum Mean Discrepancy (MMD) loss. Despite the assumed intractability of sampling from IQP circuits classically, their expectation values can be computed classically, enabling training of these IQP QCBMs. However, quantum machine learning (QML) models have various other challenges, including trainability issues caused by exponential concentration or barren plateaus. While these issues have been explored for parameters sampled from a uniform distribution, little work has been done to rigorously treat the use of arbitrary Gaussian initialization schemes. This work leverages Stein's lemma and Lipschitz concentration bounds for Gaussian random variables to provide an analytical lower bound of the variance of the gradient and a probabilistic concentration bound of the deviation of the gradient from its mean. It discusses strategies to either avoid or encourage exponential concentration, as well as the conditions under which barren plateaus are more likely to occur.

Autonomous LLM training is often framed as recipe search, which leaves the training harness largely static. This limitation sharpens in agentic RL, where shifting bottlenecks and scalar rewards mask diverse failure modes. We introduce EvoTrainer, an autonomous training framework that co-evolves LLM policies and training-side harnesses through empirical feedback: it diagnoses rollout-level evidence, revises diagnostics, backtests interventions, and accumulates reusable skills. Evaluated on mathematical reasoning, competitive-programming code generation, and repository-level software engineering, EvoTrainer matches or exceeds the human-engineered RL references under the same data, codebase, and evaluation protocol, with the largest gain on long-horizon agentic SWE. Trajectory analyses show that retained strategies diverge across domains, evolving diagnostics prevent invalid high-scoring branches from being promoted, and reusable skills shape later search. Autonomous LLM RL should move beyond recipe search toward joint evolution of policies and the training harnesses that interpret them.

Generalized segmentation of medical images prevents performance degradation when different imaging devices and clinical protocols are used across multiple domains. The Whitening Transform-based Probabilistic Shape Regularization Extractor (WT-PSE), published in IEEE Transactions on Medical Imaging in 2024, addresses this challenge by employing feature decorrelation and Wasserstein distance-based knowledge distillation to achieve robust cross-domain segmentation. This study systematically examines improvements to the WT-PSE learning framework. Four limitations in the original implementation are identified: limited training augmentations that fail to simulate real scanner variations, reliance on per-pixel binary cross-entropy loss that is sensitive to edge noise, the absence of a scheduled loss weighting strategy that may destabilize early training, and the lack of ablation switches for controlled scientific comparison. To address these issues, we propose four enhancements: (1) domain-adaptive augmentation including random erasing, gamma correction, and salt-and-pepper noise; (2) a hybrid BCE and Dice loss function for improved edge-aware segmentation under noisy conditions; (3) a curriculum-based Dice weight scheduling strategy; and (4) command-line control flags for systematic ablation studies. Experiments on the fundus optic disc segmentation benchmark demonstrate that the improved pipeline achieves a final epoch optic-disc Dice score of 0.956 and an ASD score of 13.31, outperforming the baseline epoch-5 Dice score of 0.939. These results indicate that training-level improvements can provide consistent performance gains without modifying the underlying WT-PSE architecture.

This paper describes our approach for the 8th UG2+ Workshop (CVPR 2026) Track~2, which targets semantic segmentation of outdoor scenes degraded by five weather conditions: blur, darkness, snow, haze, and glare. A central challenge we observe is a severe generalization gap -- models that perform well on the validation set often collapse on the test set. For instance, SegFormer-B5 drops 16.1 mIoU points from validation to test, suggesting that model capacity alone is insufficient for robustness. We investigate whether a carefully designed training recipe, rather than architectural complexity, can address this gap. Starting from a pre-trained SegMAN-S backbone, we systematically study the effects of domain-adaptive fine-tuning, multi-source data mixing, scene-balanced sampling, and synthetic degradation augmentation. Our final system achieves 59.9% mIoU on the official test set while maintaining a validation-test gap of only 6.5 points -- less than half that of larger models. We analyze negative results from architectural modifications, loss function variants, and model scaling to provide practical insights for weather-robust segmentation under limited data.

Search agents powered by large language models can autonomously decompose queries, retrieve information, and synthesize answers through multi-step reasoning. However, the rapid growth of training methods has outpaced controlled comparison: existing works differ in retrieval corpora, reward designs, and training protocols, making it unclear what actually drives improvements. We present a controlled empirical study that isolates three under-explored dimensions of search agent training. First, we identify a critical data-coverage issue in the widely used Wikipedia 2018 corpus and show that correcting it alone yields larger gains than the differences between training algorithms. Second, we systematically compare outcome-based and process-based reward methods across three base models, finding that the simplest outcome-based approach achieves competitive or superior performance in most settings, and that process-level credit assignment can over-correct agent behavior. Third, we analyze training data diversity, off-policy data utilization, and search budget scaling, distilling practical guidelines for training effective search agents. Our code is available at https://github.com/YiboZhao624/SearchAgentReview.

We study Latent Recurrent Transformer (LRT), a lightweight augmentation of autoregressive transformers that reuses a high-level source-layer hidden state from the previous token as recurrent memory for the next token. Because this state is already computed during ordinary decoding, LRT introduces a cross-token, cross-layer latent pathway while preserving the standard attention mechanism, KV-cache interface, and one model forward per generated token. To pretrain this recurrence without sequentially unrolling the full sequence, we introduce interleaved parallel training: one full-sequence initialization forward constructs a shared buffer, followed by sequential refinement of disjoint position subsets with parallel computation within each subset. This provides every token with recurrent-memory-aware supervision at approximately 2x ideal token compute. Across 1.3B- and 2.1B-parameter nanochat-style backbones and a wide range of training budgets, LRT improves both BPB and CORE under matched effective compute. Additionally, LRT outperforms two-forward PonderLM-2 and matches a three-loop Transformer in BPB, while retaining one-forward-per-token decoding with 9% latency overhead over the standard Transformer.

Robotic haptic devices combined with virtual reality offer novel opportunities to train fine force generation, an essential yet overlooked component of post-stroke rehabilitation. This study proposes that manipulating the rendered dynamics of tangible virtual objects can be leveraged to train precise force control while engaging the somatosensory system. We conducted an experiment with fifty healthy participants who performed a curling-inspired task in which they had to stretch a virtual spring to generate a target release force to propel the stone to a predefined location on the ice sheet. During training, the spring's force-elongation relationship was modeled as either a linear or non-linear function, i.e., a Gaussian or antisymmetric Gaussian (AS-Gaussian) function with zero derivative at the release target force. Results indicate that the AS-Gaussian group consistently achieved higher force accuracy during training than the linear group, while the Gaussian group only outperformed the linear group toward the end of training. Analysis of personality traits revealed that higher Free Spirit scores were associated with poorer performance and reduced task exploration under Gaussian dynamics, whereas higher Transform-of-Challenge scores correlated with increased exploration. Despite these training effects, no significant differences in long-term retention were found across spring types or personality traits. Participants primarily relied on learned target elongation rather than target force, as evidenced by performance in a transfer task with a different stiffness but the same target force. While promising for somatosensory neurorehabilitation, these methods require refinement to reduce reliance on proprioceptive cues before testing with neurological patients.

Multi-agent LLM workflows route inference through specialized roles to lift end-task accuracy, but jointly training those roles with reinforcement learning is unstable in ways that are poorly understood. We study when end-to-end RL training of multi-agent LLM workflows improves over their base models, comparing Shared-Policy training, where all roles update one policy, with Isolated-Policy training, where each role has its own parameters. Our experimental matrix spans Eval-Opt, Voting, and Orch-Workers workflows, math and code tasks, and three model scales (0.6B, 1.7B, 4B). We find that multi-agent RL usually improves over base models, but gains depend jointly on workflow, task, and scale, not on policy sharing alone. Isolated-Policy tends to reach higher peak accuracy yet more often falls off a terminal accuracy cliff, while Shared-Policy training does not eliminate failure; it redistributes failure into qualitatively different patterns. We then explain the strongest of these patterns through role-level gradient dynamics induced by workflow topology and policy routing: under Isolated-Policy, parallel same-role agents on shared prompts amplify per-role gradients and drive terminal degradation in Voting and Orch-Workers workflows; under Shared-Policy, asymmetric per-step gradient mass causes the shared policy to be captured by the dominant role, producing different failure signatures by task and workflow. Together, the empirical map and its underlying mechanisms show that policy sharing routes training pressure through different channels rather than offering uniform stability, making it a design choice with workflow- and task-conditional tradeoffs.

Training high-capacity vision models from scratch requires substantial computational resources. To improve training efficiency of a wide target model, existing growth methods often assume the availability of narrower models, obscuring the true computational cost of the entire pipeline. We propose an efficient training protocol, RBDC, that builds wide models by coupling in a parameter-free block-diagonal way narrower, independently trained models in a recursive way. This allows a flexible allocation of the training budget available across all the models involved. Evaluated with vision transformers (DeiT) and convolutional networks (ResNet) on ImageNet, our RBDC training protocol shows a much better efficiency than models trained from scratch with the standard protocol, yielding 30% FLOPs reduction at similar test accuracies. It also achieves higher performances at same training FLOPs than training protocols from the model growth literature. Finally, we show that our models can serve as better backbones than their original counterparts for downstream object detection and instance segmentation tasks.

Standard neural network training relies on learning-rate schedules tied to a fixed horizon, leading to strong path dependence and costly re-tuning as data availability changes. Schedule-Free (SF) methods address this by removing explicit schedules, yet SF-AdamW, the current state-of-the-art anytime optimizer, consistently underperforms well-tuned AdamW baselines. We propose SF-NorMuon, a schedule-free spectral optimizer that closes this gap: with a single hyperparameter configuration, SF-NorMuon matches or exceeds tuned AdamW on 125M and 772M parameter language models across $1$--$8\times$ Chinchilla horizons. On the theoretical side, we prove a stationarity guarantee for schedule-free spectral dynamics and identify weight decay at the fast iterate as essential for long-horizon stability. SF-NorMuon enables practitioners to obtain high-quality checkpoints at any point during training without committing to a horizon in advance. By closing the performance gap with tuned baselines, SF-NorMuon makes horizon-free optimization more practical, taking a step towards truly open-ended, continual learning.

Serious games are widely used for learning and training across domains such as healthcare, defense, and education. Persistent challenges remain, however, including static scenario design, authoring bottlenecks, limited learner modeling, and difficulty implementing meaningful real-time instructional adaptation. Recent advances in artificial intelligence (AI) introduce novel capabilities such as dynamic scenario variation, contextual feedback, adaptive pacing, and learner-state modeling that may help address some of these limitations. At the same time, integrating AI into serious games raises important questions related to validity, transparency, system control, and learner trust. This chapter examines how contemporary AI approaches may support real-time instructional adaptation in serious games. It distinguishes between instructional intelligence, defined as a system's capacity to infer learner knowledge and reason about pedagogically appropriate responses, and adaptivity, defined as the ability to modify instructional actions during interaction. A historical synthesis of adaptive learning systems is presented, tracing developments from early computer-assisted instruction through intelligent tutoring systems (ITS), dynamic difficulty adjustment (DDA), authoring platforms, learning analytics, and recent AI-enabled architectures. Building on this perspective, the chapter discusses how large language models (LLMs), reinforcement learning (RL), and agent-based architectures may contribute to more integrated forms of intelligence and adaptivity in serious games. It also highlights practical and research challenges associated with AI-enabled systems, including explainability, validation, computational cost, and the limited empirical evidence regarding long-term learning outcomes in AI-enabled serious games.

Long-tailed out-of-distribution (LT-OOD) detection is often addressed with specialized training, including auxiliary out-of-distribution (OOD) data, abstention heads, contrastive objectives, energy losses, or gradient-conflict control. We show that these training mechanisms can obscure a simpler issue: frozen long-tailed representations may already contain useful OOD evidence, but raw Mahalanobis distance is distorted by frequency-coupled feature radius and poorly supported tail covariance. We propose Hyperspherical Pooled Mahalanobis (HPM), a post-hoc detector that normalizes features onto the unit sphere and replaces class-specific covariance with a pooled, ridge-regularized metric while keeping class means as semantic anchors. In CIFAR-LT experiments and an ImageNet-100-LT near-OOD boundary analysis, HPM improves raw Mahalanobis scoring; for Prior-Calibrated ERM (PC-ERM), it raises AUROC from 46.49 to 85.67 on CIFAR-10-LT and from 50.40 to 78.35 on CIFAR-100-LT. This simple PC-ERM+HPM pipeline also achieves the best Log Efficiency Score (LES; 3.08) on CIFAR-100-LT, retaining roughly 95% of the best CIFAR-100-LT AUROC observed among the compared post-hoc scores at substantially lower training-time cost. These results argue for evaluating representation quality, detector geometry, and training complexity as separate factors in LT-OOD detection.

The release of public 3D medical image segmentation (MIS) datasets accelerates clinical research but simultaneously heightens risks of unauthorized AI model training. While Unlearnable Examples (UE) offer protection by injecting imperceptible perturbations to prevent effective model learning, existing methods primarily target 2D scenarios. They neglect the volumetric spatial correlations and inter-slice anatomical consistency inherent in 3D medical volumes, which serve as critical learning priors for 3D segmentation networks. To bridge this gap, we propose VoxShield, a UE framework that explicitly targets the volumetric inductive biases of 3D networks. Our core insight is that by systematically dismantling the cross-slice continuity that 3D architectures rely on, we can fundamentally impair their spatial aggregation process. Specifically, we introduce an Inter-Slice Frequency Consistency Disruption mechanism that maximizes the spectral divergence between adjacent slices, injecting structural incoherence along the $z$-axis. Complementing this structural attack, a Semantic Prediction Disruption module is incorporated. By maximizing the $\ell_1$ divergence between clean and perturbed logits, it forces the injected noise to penetrate the entire network and corrupt the final semantic mapping. Experiments on BraTS19 and FLARE21 demonstrate that VoxShield successfully degrades 3D segmentation performance, reducing the DSC from 80.0% to near 0.0% and from 88.6% to 6.8%, respectively. All protections are achieved with minimal perturbation ($ε=4/255$) to preserve high visual fidelity. The code is available at https://github.com/KK266299/VoxShield.

Shortcut features are often invoked to explain out-of-distribution (OOD) failure, but training correlation, learned shortcut use, and test-time failure need not coincide. We study a minimal binary model with one invariant coordinate and one family-dependent shortcut coordinate. In the deterministic regime, positive average shortcut correlation pulls logistic ERM toward positive shortcut weight, but ridge regularization keeps the classifier invariant-dominated and prevents deterministic OOD failure. When the invariant coordinate is noisy, ridge-logistic ERM switches to the shortcut rule once the training shortcut signal exceeds the invariant signal. Whether that transition causes failure depends on the held-out family: weaker shortcut correlation yields positive excess risk, and sign-flipped families yield above-chance error. Synthetic checks match these analytic regimes and show that the same training-side transition can have different held-out consequences. The model separates shortcut attraction, shortcut-rule transition, and cross-family OOD failure.

How can we train models whose post-trained capabilities survive subsequent fine-tuning? Rather than focusing on downstream interventions to mitigate forgetting of upstream capabilities, we study how upstream training choices - that is, the manner in which a capability is acquired - shape how robustly that capability is retained. We investigate this question in a controlled three-stage language-model pipeline: pretraining, post-training to acquire a target capability, and downstream fine-tuning on a new objective. Across 135M and 1B models, two post-training domains, and two downstream fine-tuning tasks, we find that immediate post-training performance does not reliably predict retention after subsequent fine-tuning: training recipes that look equivalent immediately after post-training can retain the target capability very differently after subsequent fine-tuning. In particular, early exposure - mixing post-training data into pretraining - consistently improves the frontier between retained upstream performance and downstream performance. In compute-matched experiments, where the target data must be allocated between pretraining and post-training, we find that the optimum lies at neither extreme. Together with our other empirical and theoretical findings, this supports the view that post-training drives immediate specialization while early exposure improves robustness to later forgetting. Replay and dropout, typically used to mitigate forgetting as it occurs during fine-tuning, provide complementary gains to early exposure when applied during post-training. Our findings suggest that robustness to subsequent fine-tuning should be treated as a first-class objective of upstream training, addressed preventatively through choices like early exposure rather than reactively during fine-tuning itself.

Modern large language model (LLM) training is inherently dynamic: resource fluctuations, RLHF phase shifts, and cluster elasticity continually reshape the optimal parallelism layout, posing a significant challenge to existing training frameworks built around a static execution model. We present DynaTrain, a distributed training system for sub-second, online reconfiguration across arbitrary multi-dimensional parallelism. At its core, we propose a Virtual Parameter Space (VPS) abstraction that unifies all distributed training states under one logical coordinate space, turning any parallelism configuration into a deterministic mapping and collapsing complex transition into manageable geometric intersections. On top of VPS, a state routing-and-transition layer executes rank-local transfers under a memory-aware, deadlock-free schedule, and an Elastic Device Manager overlaps new-world construction with ongoing training to mask topology-change cost. On dense and MoE models up to 235B parameters, DynaTrain reconfigures a 70B dense model in under 2s and a 235B MoE model in 4.36s, outperforming state-of-the-art checkpoint-based and elastic systems by up to three orders of magnitude while preserving correctness.

Decentralized multi-agent path finding (MAPF) routes a team of agents on a shared grid, each acting from its own local view. The standard solution trains one shared neural policy with Proximal Policy Optimization (PPO), a popular on-policy reinforcement learning algorithm. Such a policy works well on clean observations, but a small input perturbation on one agent often changes its action, which then blocks a neighbour, and the team jams. In this paper we present two training recipes that keep the same network and the same deployment loop, yet make the policy hold up under perturbed observations. The first recipe, Adv-PPO, trains the shared policy against worst-case perturbations of its own input and selects the checkpoint by performance under adversarial perturbation. The second recipe, Adv-PPO+MACER, fine-tunes that checkpoint with a small on-policy smoothness term whose gradient follows the certified radius of randomized smoothing. On POGEMA with 8x8 maps and four agents, the unprotected PPO policy reaches 95.8% clean success but only 2.5% under the strongest attack. Adv-PPO recovers worst-case success to 59.2% at one percentage point of clean cost. Adv-PPO+MACER recovers it to 77.5% +/- 6.0% across three independent seeds at less than one percentage point of clean cost. We support these numbers with per-attack curves, a certified action-stability sanity check (which measures the smoothed-policy wrapper, not the deployed argmax policy), and side-by-side rollout storyboards that show the failure mode and the fix inside one environment instance.

We investigate whether the training protocol can induce spatial locality in the early layers of a Vision Transformer (ViT) trained from scratch, without large-scale pretraining. Keeping the architecture and optimization procedure fixed, we compare a Baseline protocol with a Modern protocol (AutoAugment/ColorJitter, CutMix, and Label Smoothing) on CIFAR-10, CIFAR-100, and Tiny-ImageNet, characterizing each attention head via Mean Attention Distance (MAD) and normalized entropy. Across all three datasets, the Modern protocol produces more local and more concentrated attention in early layers; on CIFAR-100, the minimum MAD drops from 0.316 (Baseline) to 0.008 (Modern). To identify the source of this effect, we conduct an ablation study on CIFAR-100 by adding or removing each component individually. The results identify CutMix as the determining component within our experiments: all conditions with CutMix exhibit MAD 0.024, while all conditions without CutMix remain at MAD 0.210. AutoAugment and Label Smoothing show no independent effect on locality. Taken together, these findings suggest that the pressure to classify from partial image regions, induced by CutMix, can promote the emergence of local attention in Vision Transformers.

Preference learning methods like Direct Preference Optimization (DPO) are known to induce reliance on spurious correlations, leading to sycophancy and length bias in today's language models and potentially severe goal misgeneralization in future systems. In this work, we provide a unified theoretical analysis of this phenomenon, characterizing the mechanisms of spurious learning, its consequences on deployment, and a provable mitigation strategy. Focusing on log-linear policies, we show that standard preference-learning objectives induce reliance on spurious features at the population level through two channels: mean spurious bias and causal-spurious correlation leakage. We then show that this reliance creates an irreducible vulnerability to distribution shift: more data from the same training distribution fails to reduce the model's dependence on spurious features. To address this, we propose tie training, a data augmentation strategy using ties (equal-utility preference pairs) to introduce data-driven regularization. We demonstrate that this approach selectively reduces spurious learning without degrading causal learning. Finally, we validate our theory on log-linear models and provide empirical evidence that both the spurious learning mechanisms and the benefits of tie training persist for neural networks and large language models.

Reversible instance normalization (RevIN) and its successors (Dish-TS, SAN, FAN) have become the de facto plug-in for time-series forecasting, yet the map they apply to each data point is strictly affine, $x \mapsto ax+b$, so they cannot reshape the underlying distribution -- heavy tails remain heavy and skewness remains uncorrected. We propose NoRIN, a non-linear reversible normalization based on the arcsinh-form Johnson $S_U$ transform with two shape parameters $(δ,\varepsilon)$ that control tailedness and skewness; the linear $Z$-score used by RevIN is recovered only in the limit $δ\to \infty$. Training $(δ,\varepsilon)$ jointly with the backbone via gradient descent reliably pushes them toward this linear limit within a few epochs -- a phenomenon we name the degeneration problem: the forecasting loss is locally indifferent to shape, and the high-capacity backbone compensates for any monotone reparameterization of its input. NoRIN escapes the degeneration by decoupling shape selection from gradient training: $(δ,\varepsilon)$ are initialized by a closed-form Slifker-Shapiro quantile fit and refined by Bayesian optimization on the validation objective, while the inner training loop is identical to standard RevIN-style training. Across six representative backbones x five real-world datasets x three prediction horizons (90 configurations), decoupled shape optimization recovers $(δ^\star,\varepsilon^\star)$ that sit systematically far from the linear limit, with values that vary in a backbone-dependent way. This empirically supports the central thesis: different backbones genuinely require different normalization parameters to reach their best performance.

Deep continual learning requires models to adapt to new tasks without retraining from scratch. However, neural networks can lose their ability to adapt to new tasks after training on previous ones, a phenomenon known as loss of plasticity. There have been several explanations and diagnostics proposed for plasticity loss. Motivated by the philosophy that "all models are wrong, but some are useful", we ask: can existing diagnostics predict a neural network's plasticity? In this work, we take a practical view to interpret plasticity as trainability, i.e., a neural network's future optimization gain on a target task. We first take a theoretical approach, showing, by constructing a few counterexamples, that some widely adopted diagnostics of plasticity, including representation rank and neural tangent kernel rank, can fail to predict the loss of trainability in both regression and classification settings. We instead propose a novel metric, called optimization readiness, which combines gradient strength and gradient reliability. We prove that optimization readiness lower bounds one-step optimization gain under standard smoothness assumptions, providing a theoretical guarantee for its predictive power. Empirically, we show that across commonly used deep continual learning settings, such as Slowly-Changing Regression and Permuted MNIST, optimization readiness more reliably ranks checkpoints by trainability than prior diagnostics, even with substantially fewer samples.

Transformers have revolutionized machine learning and deploying attention layers in the model is increasingly standard across a myriad of applications. Further, for large models, it is common to implement Low Rank Adaptation (LoRA), whereby a factorized parameterization of them is trained, to achieve a surprisingly beneficial accuracy-size trade-off. In this work, via a unified framework we rigorously establish trainability of such models under stochastic methods. We prove that for any mild regularization, the empirical regression loss on a attention layer and LoRA on a shallow neural net, both induce Poincaré inequality for the corresponding Gibbs' measure. Then it follows via invoking recent results that a certain SDE, which mimics the SGD, minimizes the corresponding losses. In both the cases, our first-of-its-kind results of trainability on attention and nets, do not rely on any assumptions on the data or the size of the architecture.

Variational quantum circuits with angle encoding implement truncated Fourier series, and architectures arranging $N$ qubits with $L$ encoding layers each -- sharing encoding budget $E = NL$ -- generate identical frequency spectra, identical frequency redundancy, and require the same minimum parameter count for coefficient control. Despite this equivalence, trainability varies substantially with architecture shape $(N,L)$ at fixed $E$. We identify structural rank deficiency of the coefficient matching Jacobian $J$ as the mechanism responsible. For serial single-qubit architectures, we prove $\mathrm{rank}(J) \leq 2L+1$ regardless of parameter count $P$, with $\dim(\ker J) \geq P-(2L+1)$ growing without bound -- a phenomenon we term \emph{structural gradient starvation}: a growing fraction of parameters become structurally decoupled from the loss as $P$ increases at fixed $L$. Parallel architectures avoid this via independent phase trajectories, ensuring $σ_{\min}(J^{(\mathrm{par})}) > 0$ generically for $P \leq 2E+1$, so no parameter lies in $\ker J$. For practitioners, we further show that the two natural routes to increasing parameter count have fundamentally different effects: adding feature map (FM) layers monotonically strengthens the Jacobian QFIM eigenvalue spectrum and achieves $R^2 \geq 0.95$ with $1.6$--$2.2\times$ fewer parameters than adding trainable blocks across all tested architectures, while trainable blocks improve training only through the classical interpolation mechanism with no quantum-specific benefit.

Distributed quantum algorithms offer a promising pathway to scale variational quantum algorithms beyond the constraints of noisy intermediate-scale quantum hardware. However, existing approaches implicitly assume a trusted entanglement-sharing layer across quantum processors. We show that this assumption introduces a fundamental vulnerability: adversarial perturbations of shared entanglement induce structured gate-level noise that directly impacts quantum learning. We develop a framework that maps entanglement-level perturbations to gate-level noise via an explicit Kraus representation. To quantify their impact, we introduce Kraus expressibility, a metric that generalizes unitary expressibility to noisy quantum channels. We then establish a trade-off between Kraus expressibility and trainability of noisy quantum circuits through gradient variance analysis. Our analysis reveals that an adversary can manipulate Kraus expressibility to maintain sufficiently large cost gradients (avoiding barren plateaus) while systematically biasing optimization toward incorrect solutions. We validate these findings through numerical simulations, demonstrating adversarial degradation of expressibility and trainability.

Tumor segmentation in whole-body PET/CT imaging is crucial for precise disease evaluation and treatment planning. However, it remains challenging due to variability in lesion size, contrast, and anatomical distribution. Relying on manual segmentation makes the process time-consuming and prone to intra- and inter-observer variability. This work presents a whole-body tumor segmentation method developed for the AutoPET III challenge, where the goal is to build models that generalize across tracers and multi-center data. We employ the nnU-Net framework with a ResNet-based encoder as our baseline and systematically investigate the impact of training strategies, including intensity normalization, batch dice optimization, and data augmentation using CraveMix. Our experiments show that these strategies significantly influence model performance, particularly in reducing false positives and improving robustness to lesion variability. The best-performing configuration achieves a Dice score of up to 0.80 on the preliminary test phase, and our method ranked third in the AutoPET III challenge. The code is publicly available here.

Should a single collision necessarily terminate an entire navigation episode? In most deep reinforcement learning (DRL) frameworks for robot navigation, this remains the standard practice: every collision immediately triggers a global environment reset and is penalized as a complete task failure. While a collision during deployment naturally indicates task failure, applying the same treatment during training prevents the agent from exploring challenging obstacle configurations, which slows learning progress in the early training phase. In this work, we challenge this convention and propose a Multi-Collision reset Budget (MCB) framework that decouples local collision termination from global environment resets, allowing the agent to retry difficult configurations within the same episode. Simulation experiments show that MCB improves early-stage learning efficiency by reaching target success-rate levels with fewer interactions, with small collision budgets producing the most consistent gains. Real-world experiments on heterogeneous robot platforms further validate the deployability of the learned policies in cluttered environments.

Driving agents trained in one simulated town often perform poorly in a new town because the road shapes, intersections, and lane layouts can be different. This paper studies how to improve this kind of transfer in the CARLA driving simulator without giving the agent any training data from the test towns. The agent is trained only in Town05 and Town06, then evaluated directly in Town03 and Town04. To focus on road-layout differences, all experiments use the same weather and traffic settings. We propose a training method that encourages the agent to learn features that are useful across towns rather than features tied to one training town. During training, the agent is asked to predict the high-level visual meaning of future camera views and is also discouraged from relying on cues that reveal which source town the data came from. These extra learning signals are used only during training; at test time, the driving policy uses the same observation and control interface as the baseline agent. In controlled comparisons with matched DreamerV3-style world-model driving agents, the proposed method achieves the highest mean held-out success: 36.6% on Town03 with a 95% confidence interval of [30.5, 42.7] and 85.6% on Town04 with a 95% confidence interval of [84.0, 87.2], computed across five training seeds. Seed-paired tests against the strongest primary baselines show positive success-rate differences in both held-out towns. Additional experiments show that predicting future visual meaning alone or removing town-specific cues alone is not enough to match the combined method. These results suggest that combining future-scene understanding with reduced reliance on source-town-specific features can improve cross-town driving performance in this CARLA setting.

Masked diffusion language models (MDMs) have recently emerged as a promising alternative to standard autoregressive large language models (AR-LLMs), yet their optimization can be substantially less stable. We study blockwise MDMs and compare them with AR-LLMs on three controlled tasks that stress different aspects of structured generation: in-context linear regression, graph path-finding, and Sudoku solving. We find that standard random-masking MDMs fail to reliably learn linear regression, exhibit high variance training dynamics on graph path-finding, while outperforming AR-LLMs on Sudoku. To mitigate these instabilities, we propose two locality aware blockwise models, namely Jigsaw and Scatter, that inject left-to-right inductive bias by enforcing autoregressive locality within blocks while preserving iterative refinement at the block level. Empirically, Jigsaw matches AR-LLM stability on linear regression and remains strong on Sudoku, while Scatter retains diffusion's planning advantage on path-finding. Our results indicate that standard random-masking MDMs, even with blockwise variants, may be a suboptimal instantiation of diffusion LMs for ordered generation, motivating models beyond random masking.

Training large-scale deep neural networks effectively and stably is essential for applying deep learning across various fields. However, conventional methods, which rely on training a single large network, often encounter challenges such as gradient vanishing, overfitting and unstable learning. To overcome these limitations, we introduce Self-Abstraction Learning (SAL), a hierarchical framework. In SAL, networks are arranged by structural complexity, where the simplest topmost network is trained first and its hidden and output layers serve as guidance for the successively more complex networks below. This top-down sequential guidance effectively mitigates optimization issues, enabling stable training of deep architectures. Various experiments across MLP, CNN, and RNN architectures demonstrate that SAL consistently outperforms conventional methods, ensuring robust generalization even in data-scarce and complex network regimes.

Small instruct-tuned LLMs produce degenerate verbal confidence under minimal elicitation: ceiling rates above 95%, near-chance Type-2 AUROC, and Invalid validity profiles. We test whether confidence-conditioned supervised fine-tuning (CSFT) with self-consistency-derived targets can close the gap between internal information and verbal readout. A pre-registered Phase 0 protocol on Gemma 3 4B-it with a modal filter restricting training to items with correct modal answers produced a negative result: AUROC2 dropped from 0.554 to 0.509 due to label-entropy collapse in the training targets. An exploratory rescue removed the filter, training on all 2,000 calibration items. This produced a binary verbal correctness discriminator with AUROC2 = 0.774 on held-out TriviaQA, compressing a 10-sample self-consistency signal (AUROC2 = 0.999) into a single-pass readout exceeding logit entropy (0.701). The shuffled-target control showed no improvement (0.501). On MMLU, accuracy improved from 54.2% to 77.4% with the shuffled model at baseline (56.1%), supporting a target-dependent interpretation. The result is exploratory, binary rather than continuously calibrated, and observed at a single scale. It identifies two design lessons: confidence training requires label entropy, and correct targets regularise output format.

Barren-plateau results have established exponential gradient suppression as a widely cited obstacle to the scalability of variational quantum algorithms. When and whether these results extend to a given objective has been addressed through loss-specific arguments, but a general structural characterization has remained open. We show that the objective itself admits a fixed-observable representation if and only if the loss is affine in the measured statistics, thereby identifying the exact boundary of the standard concentration-based proof template. Existing transfer results for non-affine losses achieve this reduction under additional assumptions; our characterization implies that such a reduction is not structurally available for a class of non-affine objectives, placing them outside the automatic reach of the existing proof template. Beyond the affine regime, a chain-rule decomposition reveals three governing factors -- model responsivity, loss-side signal, and transmittance -- and induces a loss-class dichotomy: bounded-gradient losses inherit suppression, while amplification-capable losses can in principle counteract it. In the exponentially wide setting, both classes fail, but for different structural reasons. When the interface is instead designed at polynomial width -- exposing coarse-grained statistics rather than individual bitstring probabilities -- the exponential-dimensional obstruction is relaxed and the dichotomy plays a genuine role. In a numerical demonstration on a charge-conserving quantum system, the amplification-capable objective produces resolved gradients several orders of magnitude larger than affine and inheriting baselines at comparable shot budgets. Over the tested interval, its scaling trend is statistically distinguished from the exponential trend of both alternatives. The boundary is affine; what lies beyond it is a representation-design problem.

This paper is a step-by-step, self-contained guide to the complete training cycle of a Physics-Informed Neural Network (PINN) -- a topic that existing tutorials and guides typically delegate to automatic differentiation libraries without exposing the underlying algebra. Using a first-order initial value problem with a known analytical solution as a running example, we walk through every stage of the process: forward propagation of both the network output and its temporal derivative, evaluation of a composite loss function built from the ODE residual and the initial condition, backpropagation of gradients -- with particular attention to the product rule that arises in hidden layers -- and a gradient descent parameter update. Every calculation is presented with explicit, verifiable numerical values using a 1-3-3-1 multilayer perceptron with two hidden layers and 22 trainable parameters. From these concrete examples, we derive general recursive formulas -- expressed as sensitivity propagation relations -- that extend the gradient computation to networks of arbitrary depth, and we connect these formulas to the automatic differentiation engines used in practice. The trained network is then validated against the exact solution, achieving a relative $L^2$ error of $4.290 \times 10^{-4}$ using only the physics-informed loss, without any data from the true solution. A companion Jupyter/PyTorch notebook reproduces every manual calculation and the full training pipeline, providing mutual validation between hand-derived and machine-computed gradients.

Reinforcement Learning (RL) has emerged as a powerful training paradigm for LLM-based agents. However, scaling agentic RL for deep research remains constrained by two coupled challenges: hand-crafted synthetic data fails to elicit genuine real-world search capabilities, and real-world search dependency during RL training introduces instability and prohibitive cost, which limits the scalability of Agentic RL. LiteResearcher is a training framework that makes Agentic RL scalable: by constructing a lite virtual world that mirrors real-world search dynamics, we enable a continuously improving training recipe that empowers a tiny search agent to outperform large-scale open-source and commercial models (e.g., Tongyi DeepResearch and Claude-4.5 Sonnet). Specifically, on common benchmarks such as GAIA and Xbench, our LiteResearcher-4B achieves open-source state-of-the-art results of 71.3% and 78.0% respectively, demonstrating that scalable RL training is a key enabler for Deep Research Agents.

While recent Spoken Language Models (SLMs) have been actively deployed in real-world scenarios, they lack the capability to discern Third-Party Interruptions (TPI) from the primary user's ongoing flow, leaving them vulnerable to contextual failures. To bridge this gap, we introduce TPI-Train, a dataset of 88K instances designed with speaker-aware hard negatives to enforce acoustic cue prioritization for interruption handling, and TPI-Bench, a comprehensive evaluation framework designed to rigorously measure the interruption-handling strategy and precise speaker discrimination in deceptive contexts. Experiments demonstrate that our dataset design mitigates semantic shortcut learning-a critical pitfall where models exploit semantic context while neglecting acoustic signals essential for discerning speaker changes. We believe our work establishes a foundational resource for overcoming text-dominated unimodal reliance in SLMs, paving the way for more robust multi-party spoken interaction. The code for the framework is publicly available at https://tpi-va.github.io

Conditional depth execution routes a subset of tokens through a lightweight cheap FFN while the remainder execute the standard full FFN at each controlled layer. The central difficulty is gate training: the gate decision must propagate through many layers before it influences the language modeling (LM) loss, so the resulting gradients are weak and noisy. Auxiliary losses are commonly stacked to stabilise training, yet the interactions among them -- particularly between a predictive auxiliary and explicit score supervision -- have not been systematically compared under controlled conditions. We evaluate two gate designs under a 157.5M-parameter decoder-only model with controller-only training, 50% full-path budget, and 3-seed runs on a fineweb-edu subset. The MLP gate (G1) maps the current hidden state to a utility score; the JEPA-guided gate (G3) adds an action-conditional predictor that forecasts, in a low-dimensional latent space, the outcome of executing full vs. cheap per token, aligned against a fixed target head. Under the standard recipe with oracle-style utility regression and pairwise rank supervision (util/rank), G3 improves early-to-mid optimisation over G1 in 3/3 seeds (lower avg LM, faster threshold hits, ~10.3x lower grad norms), with 20k-step endpoint LM within a 0.005 heuristic reference. A key finding (ablation A3): jointly removing util/rank improves best/avg LM and threshold-hit speed in 3/3 seeds for both gates, and the early-to-mid advantage of G3 over G1 disappears. We trace this to an off-policy oracle label that assumes all subsequent layers execute full, whereas gated execution routes only a fraction through full -- making util/rank net-negative under the current recipe. Removing util/rank also cuts the training FLOPs proxy from ~1.53x to ~1.07x full-only (2.87h to 1.75h on a V100-32GB, ~39%). Conclusions are scoped to the studied regime.

Universal Machine Learning Interatomic Potentials (uMLIPs), pre-trained on massively diverse datasets encompassing inorganic materials and organic molecules across the entire periodic table, serve as foundational models for quantum-accurate physical simulations. However, uMLIP training requires second-order derivatives, which lack corresponding parallel training frameworks; moreover, scaling to the billion-parameter regime causes explosive growth in computation and communication overhead, making its training a tremendous challenge. We introduce MatRIS-MoE, a billion-parameter Mixture-of-Experts model built upon invariant architecture, and {Janus}, a pioneering high-dimensional distributed training framework for uMLIPs with hardware-aware optimizations. Deployed across two Exascale supercomputers, our code attains a peak performance of 1.2/1.0 EFLOPS (24%/{35.5%} of theoretical peak) in single precision at over 90% parallel efficiency, compressing the training of billion-parameter uMLIPs from weeks to hours. This work establishes a new high-water mark for AI-for-Science (AI4S) foundation models at Exascale and provides essential infrastructure for rapid scientific discovery.

We explore the possibility of solving Partial Differential Equations (PDEs) using discrete weak formulations. We propose a programming environment for defining a discrete computational domain, introducing discrete functions defined over a set of points, constructing discrete inner products, and introducing discrete weak formulations employing Kronecker delta test functions. Building on this setup, we propose a discrete neural network representation, training the solution function defined over a discrete set of points and employing discrete finite difference derivatives in the automatic differentiation procedures. As a challenging computational model example, we focus on Stokes equations in two-dimensions, defined over a discrete set of points. We train the solution using the discrete weak residual and the Adamax algorithm with discrete automatic differentiation of the discrete gradients. Despite introducing the python environment, we also provide a rigorous mathematical formulation based on discrete weak formulations, proving the well-posedness and robustness of the loss function. The solution of the discrete weak formulations is based on neural network training employing a robust loss function that is related to the true error. In this way, we have a robust control of the numerical error during the training of the neural networks. Besides the Stokes formulation, we also explain the functionality of the proposed library using the Laplace problem formulation.

Instead of directly distilling a language model, this study addresses the problem of aligning a base model with a target model in distribution by designing the domain mixture of training data for pretraining or continued pretraining as a fixed training recipe. We propose a method for determining domain weights by viewing models as points in log-likelihood space and aligning the training update direction with the direction toward the target model. Experiments with NanoGPT show that the proposed method consistently reduces the KL divergence between the trained base model and the target model relative to training with Pile-original weighting. Although knowledge distillation remains more effective when available, the proposed method achieves meaningful alignment, and downstream task performance also tends to become closer to that of the target model.

Gradient-based LUT- and logic-gate-based neural networks (LUTNet, LogicNets, DiffLogic, PolyLUT, NeuraLUT, WARP-LUT, DWN, LILogicNet, LightLUT) replace multiply-accumulate arithmetic with Boolean lookups. The same trained checkpoint deploys to GPU as bitwise ops on bit-packed activations, to FPGA as LUT primitives, and to ASIC as standard-cell gates, all from one code path. Yet each method ships its own training pipeline, encoder, connectivity rule, fan-in, and hardware-reporting convention. The natural practitioner question, which of these choices actually matter for accuracy and which for hardware cost, therefore has no answer in the current literature. We release \textbf{BitLogic}, a unified framework that factors the field into a five-axis design space (encoder, connectivity, fan-in, node parameterization, head) and instantiates every prior method under one shared training and evaluation protocol. The framework deliberately omits method-specific procedures such as calibration, pruning, and thresholding, and all evaluations are limited to two-layer feed-forward networks. Combining the per-axis winners identifies a new best-of-space configuration that outperforms every retrained prior on every (dataset, width) cell in which every compared prior fits the shared budget, across MNIST, Fashion-MNIST, CIFAR-10, and CIFAR-100. We evaluate the best-of-space model on all three backends. On MNIST, the resulting two-layer network reaches ${\sim}126$,MSamples/s on FPGA, ${\sim}15\times$ the throughput of a bit-packed GPU forward path that itself processes $64$ samples per $64$-bit operation, at four-to-five orders of magnitude less energy.

Offline goal-conditioned reinforcement learning (GCRL) is typically benchmarked by the best tuned success rate of each method. This score measures attainable performance, but it does not reveal how reliably a learned goal-conditioned signal can be extracted into a policy: a method could succeed across many value-learning and extraction settings, or only at a narrow, hard-to-find configuration. We study this gap across four methods, GCIQL, GCIVL, QRL, and CRL, under a shared advantage-weighted regression (AWR) extractor. For each method, we construct trainability landscapes over the optimizer learning rate, which affects value learning and actor optimization, and AWR temperature, which controls how selectively the actor imitates high-advantage transitions. Across AntMaze, Cube, and Scene, we observe distinct regimes: high-scoring methods may be broadly accessible or brittle, while broad relative basins may still sit below low absolute ceilings. To interpret these differences, we pair landscapes with post-hoc diagnostics of future-vs-random goal discrimination and AWR weight concentration. Their relationship to downstream success is task-dependent. On AntMaze, where future goals align with path-like progress, these diagnostics explain landscape regimes. On Cube and Scene, goal ranking and manipulation control decouple: methods can rank goals well while failing downstream, or succeed through action-conditioned advantages despite weak future-vs-random separation. These results show that peak tuned success alone does not establish broadly extractable goal-conditioned behavior. Trainability landscapes expose this gap, while extraction diagnostics offer a lower-cost lens on how learned signals become policies.

Online handwriting recognition using inertial measurement units opens up handwriting on paper as input for digital devices. Doing it on edge hardware improves privacy and lowers latency, but entails memory constraints. To address this, we propose Error-enhanced Contrastive Handwriting Recognition (ECHWR), a training framework designed to improve feature representation and recognition accuracy without increasing inference costs. ECHWR utilizes a temporary auxiliary branch that aligns sensor signals with semantic text embeddings during the training phase. This alignment is maintained through a dual contrastive objective: an in-batch contrastive loss for general modality alignment and a novel error-based contrastive loss that distinguishes between correct signals and synthetic hard negatives. The auxiliary branch is discarded after training, which allows the deployed model to keep its original, efficient architecture. Evaluations on the OnHW-Words500 dataset show that ECHWR significantly outperforms state-of-the-art baselines, reducing character error rates by up to 7.4% on the writer-independent split and 10.4% on the writer-dependent split. Finally, although our ablation studies indicate that solving specific challenges require specific architectural and objective configurations, error-based contrastive loss shows its effectiveness for handling unseen writing styles.

Sign language recognition from monocular video or 2D pose sequences is challenging, both because 3D information must be inferred from 2D observations and because the signal is inherently spatiotemporal. Moreover, the large and continually growing vocabulary of signs in production settings makes conventional closed-set classification impractical: adding a class requires new labeled data and retraining. We propose a contrastive Transformer-based model that learns rich representations of body key-point sequences, enabling direct comparison between embedding vectors. These representations support one-shot and few-shot tasks such as classification of signs never seen during training. On the LSA64 dataset, using only 48 classes for representation learning, the model reaches 88.4% accuracy on 16 held-out classes with as few as eight reference examples per class, and its accuracy improves consistently with the number of training classes and support examples.