Papers

Incorporating prior data into online reinforcement learning accelerates training but typically forces a difficult trade-off between high computational costs and long, multi-stage training pipelines. While fixed-length stabilization phases are significantly more computationally efficient than static update schedules, they require task-dependent manual tuning, risking either the waste of prior knowledge or severe overfitting. To address this, we propose SOPE, an algorithm that uses an actor-aligned Off-Policy Policy Evaluation (OPE) signal as an automated early-stopping mechanism to dynamically control the length of offline training phases. By evaluating the critic on a held-out validation split under the current policy's action distribution, SOPE halts gradient updates exactly when out-of-distribution benefits saturate, eliminating the need for manual schedule tuning. Evaluated on 25 continuous control tasks from the Minari benchmark suite, SOPE improves baseline performance by up to 45.6% while reducing the required TFLOPs by up to 22x, thus balancing the tradeoff between sample and computational efficiency. These findings demonstrate that adaptive, evaluation-driven update schedules are more effective than relying on static, exhaustive update schedules.

Neural operators perform well on structured domains, yet their behaviour on irregular geometries remains poorly understood. We show that this limitation is not merely an encoding issue, but a depth-wise failure mode inherent to deep operator architectures. We formalise the Geometric Forgetting Hypothesis: due to the Markovian structure of operator layers and their reliance on global mixing mechanisms, neural operators progressively lose access to domain geometry as depth increases. Using layer-wise geometric probing, we demonstrate that both spectral and attention-based operators systematically lose geometric fidelity. We show that this geometric forgetting degrades accuracy, stability, and generalisation. To counteract it, we introduce a lightweight geometry memory injection mechanism that restores geometric constraints at intermediate depths with minimal architectural overhead. This simple intervention consistently mitigates forgetting and exposes a geometric shortcut instability in transformer-based operators, revealing that geometric retention is a structural requirement rather than a design choice.

Tokamaks remain leading candidates for achieving practical fusion energy, yet many important control problems inside these devices are still difficult or unsolved. One such challenge is controlling the plasma rotation profile, which strongly influences stability, confinement, and transport. While the average rotation can be controlled, controlling the full profile is challenging due to high dimensionality, response to multiple actuators and dependence on plasma condition. Learning-based control methods, such as reinforcement learning (RL), provide a potential solution to this challenging problem with ability to model complex interactions leading to effective multi-input multi-output control. However, learning such policies is challenging due to the lack of accurate simulators that can model the rotation profile dynamics. In this work, we investigate the use of offline RL and offline model-based RL algorithms for rotation profile control, training them solely on historical data from the DIII-D tokamak. Our final method uses probabilistic models of plasma dynamics to generate rollouts for RL training. We deploy this policy on the DIII-D Tokamak and observe promising real-world results. We conclude by highlighting key challenges and insights from training and deploying an RL policy on a complex physical device while using only limited past data.

Measuring learning progress is at the core of curiosity-driven exploration, which rewards an agent for going where its model is still learning. However, the abstract notion of learning progress is not directly measurable, and existing methods often derive it from the prediction error in the output space. This paper proposes Gradient-Momentum Coupling (GMC), which measures how strongly a sample drives change in the parameter space, given by the normalized absolute product of its gradient with the momentum of previous gradients. Directions of change that persist across samples accumulate in momentum, while noise cancels out. In controlled experiments GMC allocates near uniform priority across tasks with varying levels of noise, where prediction error chases the noisiest, and orders learnable tasks by improvement speed rather than difficulty. On four MiniGrid MultiRoom tasks, substituting GMC for prediction error inside the Intrinsic Curiosity Module (ICM) recovers exploration ICM loses to unpredictable observations.

Large language models (LLMs) are increasingly applied to complex problems, from automated code debugging to science itself. Such problems require generating plausible hypotheses from sparse, noisy, and ambiguous data and updating them as more information arrives. Yet how closely LLMs approach optimal inductive reasoning remains unclear. We study this question in the classic number game, in which a learner infers a concept from a few ambiguous positive integers: $\{2, 16, 4, 8\}$, for example, is explained both by “powers of 2” and by “numbers up to 20.” We probe the concepts LLMs hypothesize from the observed examples and use their behavioral patterns to understand how they generate and update hypotheses. We measure the LLM posterior over hypotheses in three tasks (prediction, evaluation, and generation) and compare LLM behavior with a Bayesian model and with humans. LLMs favor narrower hypotheses than humans and show a robust evaluation–generation gap: in evaluation they select hypotheses more consistent with the observed examples, whereas in generation they propose narrower hypotheses that are more often rules. They also update poorly: as examples accumulate, their accuracy declines while their confidence grows. Thinking reduces, and can even reverse, the gap; it also broadens the hypotheses LLMs favor, raising accuracy while lowering confidence. Whereas prior studies have shown that LLMs are good hypothesis proposers, our results and further comparisons with the Bayesian model show that LLM inductive reasoning lacks coherence across the three tasks, suggesting that LLMs do not maintain a stable representation shared across tasks that make the same demand. Our results highlight a limitation of LLMs in inductive inference, particularly in settings such as science, where hypotheses must go beyond the data.

Linear Attention (LA) offers a promising paradigm for scaling large language models (LLMs) to long sequences by avoiding the quadratic complexity of self-attention. Recent LA models such as Mamba2 and GDN interpret linear recurrences as closed-form online stochastic gradient descent (SGD), but naive SGD updates suffer from rapid information decay and suboptimal convergence in optimization. While momentum-based optimizers provide a natural remedy, they pose challenges in simultaneously achieving training efficiency and effectiveness. To address this, we develop a chunkwise parallel algorithm for LA with a stepwise momentum rule by geometrically reordering the update coefficients. Further, from a dynamical systems perspective, we analyze the momentum-based recurrence as a second-order system that introduces complex conjugate eigenvalues. This analysis guides the design of stable gating constraints. The resulting model, Momentum DeltaNet (MDN), leverages Triton kernels to achieve comparable training throughput with competitive linear models such as Mamba2 and KDA. Extensive experiments on the 400M and 1.3B parameter models demonstrate consistent performance improvements over strong baselines, including Transformers, Mamba2 and GDN, across diverse downstream evaluation benchmarks. Code: https://github.com/HuuYuLong/MomentumDeltaNet .

LLMs often struggle with memory-constrained deployment on consumer-grade hardware due to their massive parameter sizes. While existing solutions such as model compression and offloading improve deployment feasibility, they often suffer from substantial accuracy degradation or severe throughput bottlenecks. Recent error compensation methods recover accuracy through auxiliary LoRA-style branches, and we observe that these branches are inherently amenable to offloading: they require substantial parameter storage but access only a small subset of compensation parameters during each inference step. Motivated by this opportunity, we propose HCInfer, a heterogeneous inference system that offloads residual compensation to the CPU while executing the compressed backbone on the GPU, and further introduces an asynchronous compensation pipeline and sensitivity-aware dynamic rank allocation to hide compensation overhead and maximize accuracy recovery. Experimental results show that HCInfer achieves a maximum accuracy improvement of 5.2% on downstream tasks compared to compression model and sustaining a maximum speedup of 10.4x compared to full-precision model.

Posterior collapse in variational autoencoders is often diagnosed by its symptoms: a small KL term, a strong decoder, or weak use of the latent code. These signals are useful, but they do not define a collapse boundary. We study a concrete failure mode, input-independent constant collapse, and show that this case admits an exact threshold. For any fixed nonconstant teacher distribution \(T(\cdot\mid x)\), the best constant student is the dataset-average teacher distribution, and its alignment cost is the teacher mutual information \(I_T(X;T)\). Therefore, if a strictly latent-only raw witness achieves alignment loss below this value, with a safety margin, the witness cannot be constant in the input. This identity turns a qualitative failure mode into a measurable one. In CIFAR-100 experiments with per-seed teacher search, full training stays on the certified side of the boundary, removing alignment drives the raw witness into the constant-student regime, and restarting from a collapsed checkpoint with alignment enabled restores the certificate. Tiny-ImageNet-200 fixed-target runs show the same prevention–collapse–rescue pattern across three independently searched teachers. Standard VAE-style baselines, including methods that preserve reconstruction quality or post-hoc predictability, remain negative under the raw certificate. The guarantee is intentionally narrow: it certifies that the matched nonconstant teacher-relative variation passes through the latent pathway, rather than claiming that all forms of posterior collapse have been ruled out.

Algorithms in machine learning and AI do critically depend on at least three key components: (i) the risk function, which is the expectation of the loss function, (ii) the function space, which is often called the hypothesis space, and (iii) the set of probability measures, which are allowed for the specified algorithm. This paper gives a survey of a certain class of loss functions, which we call ratio-based. In supervised learning, margin-based loss functions for classification tasks depending on the product of the output values $y_i$ and the predictions $f(x_i)$ as well as distance-based loss functions depending on the difference of $y_i$ and $f(x_i)$ for regression are common. Distance-based loss functions are in particular useful, if an additive model assumption seems plausible, i.e. the common signal plus noise assumption. However, in the literature, several loss functions proposed for regression purposes have a multiplicative error structure in mind and pay attention to relative errors, i.e. to the ratio of $y_i$ and $f(x_i)$. In this survey article, we systematically investigate such ratio-based loss functions and propose a few new losses, which may be interesting for future research. We concentrate on investigating general properties of ratio-based loss functions like continuity, Lipschitz-continuity, convexity, and differentiability, because these properties play a central role in most machine learning algorithms. Therefore, we do not focus on some specific machine learning algorithm to derive universal consistency, learning rates, or stability results. Instead, we want to enable future research in this direction.

Retrieval-augmented generation (RAG) typically treats retrieval and generation as separate systems. We ask whether an attention-based encoder-decoder can instead retrieve directly from its own internal representations. We introduce INTRA (INTrinsic Retrieval via Attention), a framework where decoder attention queries score pre-encoded evidence chunks that are then directly reused as context for generation. By construction, INTRA unifies retrieval and generation, eliminating the retriever-generator mismatch typical of RAG pipelines. This design also amortizes context encoding by reusing precomputed encoder states across queries. On question-answering benchmarks, INTRA outperforms strong engineered retrieval pipelines on both evidence recall and end-to-end answer quality. Our results demonstrate that attention-based models already possess a retrieval mechanism that can be elicited, rather than added as an external module.

Group-relative RL training (GRPO) samples a small group of parallel rollouts for every training prompt and uses their within-group reward spread to compute per-trajectory advantages. In agentic environments each rollout is a long multi-turn dialogue with one LLM call per step, so this multi-sample multiplier dominates the total training cost. When every rollout of a prompt ends with the same reward, the group has zero reward variance and contributes no gradient, so the extra rollouts add no information; such groups are common in practice (typically around 40% of all groups), so the wasted-compute fraction is substantial rather than marginal. Existing methods filter such groups at the prompt level, either after their rollouts are paid for or before any rollout begins, but both decide without using information that becomes available during the rollout itself. We instead ask whether the in-group divergence between the partial trajectories at an intermediate step can already predict that the group will be zero-variance: when the parallel rollouts have already converged on the same action prefix, the group is on track to produce a single reward, and we can stop early. We propose a one-parameter gate that stops a group when the mean pairwise prefix edit distance between its partial action sequences falls below a threshold. On a 60-iteration on-policy GRPO run on ALFWorld with Qwen2.5-7B, averaged over four random seeds, the gated arm finishes 10.7% faster in wall-clock (bootstrap 95% CI excludes 0) and shifts held-out success rate on 50 unseen tasks by +2.5 pp, with the held-out gain tracing to a measurable reduction in zero-advantage gradient-batch dilution. Code is available at https://github.com/zhiyuanZhai20/selective-rollout.

Decomposing complex tasks into a sequence of simpler subtasks can improve learning efficiency for an autonomous agent. Reinforcement learning (RL) can be used to optimize agent policies to complete subtasks, but requires well-defined subtask rewards and benefits from action masking. Recent work uses large language models (LLMs) to automate reward shaping and action masking, however none of them fully address reactivity to subtask failure and modularity to varying objects for compositional tasks. To overcome these challenges, we develop masking reward behavior tree (MRBT), a symbolic structure used as a reactive and modular reward and action mask function. We design an MRBT template and derive logical specifications to construct and verify MRBTs for a sequence of object-interaction subtasks. Further, we develop an automated pipeline that uses an LLM to generate MRBTs robust to varying task objects, an SMT-solver to verify correctness of specifications, and a neurosymbolic RL loop to train agents on compositional tasks. Experiments demonstrate successful generation and refinement of five MRBTs, consistently improving training efficiency and task success rates over baselines and MRBTs without action masking. We further highlight three advantages of MRBTs: transferability, modularity, and verifiability.

Large language models (LLMs) contain functionally distinct modules with heterogeneous gradient statistics. Although Adam(W) adapts updates at the parameter level, selecting relative learning rates across modules often relies on costly manual tuning. We study this problem through gradient signal-to-noise ratio (SNR). Our analysis identifies structural sources of module-wise SNR differences and relates SNR to the attenuation of Adam's normalized mean-gradient component. Based on these findings, we propose Module-wise Learning Rate Scaling via SNR (\name), a lightweight method that replaces manual multiplier search with one-shot calibration. MoLS estimates module-level SNRs after an initial adaptation period and uses inverse-square-root scaling to determine relative learning-rate multipliers. The multipliers are introduced smoothly during warmup and then held fixed while the global learning-rate schedule continues. Our theoretical results quantify the effect of calibration and its sensitivity to estimation error and relative SNR drift. Extensive experiments across LLM pre-training and fine-tuning show that MoLS reduces perplexity and improves downstream accuracy with minimal computational overhead. It achieves performance comparable to carefully tuned module-specific learning rates and integrates with memory-efficient optimizers and parameter-efficient fine-tuning.

Fitted Q-iteration (FQI) is well understood when the data come from a fixed distribution and the action set is finite, whereas value-based continuous control collects its data adaptively and maximizes over continuous actions. Both features break the standard analysis: the regression data are dependent, and every density-ratio concentrability coefficient of a deterministic greedy policy is infinite when the exploration has no atoms. Conditioning on the history freezes the regression target, so each round of FQI is a regression with martingale noise, and we obtain localized rates that are fast for classes of polylogarithmic metric entropy. We transfer the regression error to the loss of the greedy policy by smoothing the measures on which this loss is evaluated, at a price that depends on the action dimension and the smoothness of the residuals in the action. Under deterministic dynamics the state is smoothed as well, and the state dimension enters the price. Minimax lower bounds with the same exponents in the sample size show that no learner avoids these prices, for restart data in the case of the state dimension. The coverage conditions do not involve the unknown greedy policies; they hold under exploration floors with restarts and, for action smoothing, bounded transition densities.

In the mid-twentieth century, mathematician and polymath John von Neumann created a computational system on an array of cells as a simple model of the human brain, where each cell had one of a finite set of roles or states that he predicted would be modelled by a diffusion process. In this work, we show that such a system, when developed in a modern deep learning setting, enables the construction of an artificial neuron having specialized roles that can be learnt. We refer to this neuron as the Von Neumann neuron, and the resulting neural network from such neurons result in a self-engineered design whose architecture is only dependent on the structure and locations of its inputs and outputs on this cellular array. The mathematical framework for these Von Neumann Networks (VNNs) is also constructed and shows that they are based on the extension of neural operators and the learning of Green's functions with convolutions on a cellular topology having a diffusion signature. We also prove that these VNNs are part of a more general computational system called Cellular Machines that are computationally universal. Initial experiments show that VNN based multi-layered perceptrons outperform their equivalent deep learning variant on basic tasks, while being more parameter efficient and are capable of learning new types of tasks. This includes the ability to solve for and construct an extension of the Von Neumann (hardware) architecture common to all modern computers to cells and suggests new opportunities that could be explored.

Differentially private federated fine-tuning of large models with LoRA suffers from aggregation error caused by LoRA's multiplicative structure, which is further amplified by DP noise and degrades both stability and accuracy. Existing remedies apply a single update mode uniformly across all layers and all communication rounds (or alternate them on a fixed schedule), ignoring both the structural asymmetry between the two LoRA factors and the round-wise dynamics of training. We propose AS-LoRA, an adaptive framework defined by three axes (i) layer-wise freedom, in which each layer independently selects its active component, (ii) round-wise adaptivity, in which the selection updates over communication rounds, and (iii) a curvature-aware score derived from a second-order approximation of the loss. Theoretically, AS-LoRA eliminates the reconstruction-error floor of layer-tied schedules, accelerates convergence, implicitly biases solutions toward flatter minima, and incurs no additional privacy cost. Across GLUE, SQuAD, CIFAR-100, and Tiny-ImageNet under strict DP budgets and non-IID partitions, AS-LoRA improves over the federated LoRA baselines by up to $+7.5$ pp on GLUE and $+12.5$ pp on MNLI-mm for example, while matching or exceeding SVD-based aggregation methods at $33\text{--}180 \times$ lower aggregation cost and with negligible communication overhead. Code for the proposed method is available at https://anonymous.4open.science/r/as_lora-F75F/.

In this paper, we investigate the supremum-norm generalization error and the uniform inference for a specific class of kernel regression methods, namely the kernel gradient flows. Under the widely adopted capacity-source condition framework in the kernel regression literature, we first establish convergence rates for the supremum norm generalization error of both continuous and discrete kernel gradient flows under the source condition $s>\alpha_0$, where $\alpha_0\in(0,1)$ denotes the embedding index of the kernel function. Moreover, we show that these rates match the minimax optimal rates. Building on this result, we then construct simultaneous confidence bands for both continuous and discrete kernel gradient flows. Notably, the widths of the proposed confidence bands are also optimal, in the sense that their shrinkage rates are greater than, while can be arbitrarily close to, the minimax optimal rates.

It is well established that spectral graph neural networks (GNNs) can universally approximate node signals; however, their expressive power remains bounded by the 1-dimensional Weisfeiler-Lehman test, which is mirrored in their lack of universality for higher-order signals. To go beyond this bound, we propose the Full-Spectrum GNNs (FSpecGNNs), a second-order generalization of classical spectral GNNs. FSpecGNN advances spectral filtering from two perspectives: (1) it lifts signals from the node domain to the node-pair domain; and (2) it extends the univariate spectral filter over eigenvalues to a bivariate filter over eigenvalue pairs. We show that classical spectral GNNs arise as a diagonal special case of FSpecGNNs, and prove that FSpecGNNs can be at most as expressive as Local 2-GNN while universally approximating node-pair signals, the latter being particularly beneficial for heterophilic graph learning. Moreover, FSpecGNN admits scalable implementations that avoid explicit node-pair-level computations; combined with a low-rank approximation that reduces full-spectrum convolution to a combination of polynomial spectral filters, it enables learning on large graphs. Empirically, FSpecGNN validates the predicted expressivity and delivers strong performance on heterophilic benchmarks.

We investigate the ability of transformers to perform in-context reinforcement learning (ICRL), where a model must infer and execute learning algorithms from trajectory data without parameter updates. We show that a linear self-attention transformer block can provably implement policy-improvement methods, including semi-gradient SARSA and actor-critic, via explicit parameter constructions. Beyond existence, we design a teacher-mimicking training procedure, analyze its gradient-flow dynamics, and establish the first convergence guarantee in the ICRL literature: under suitable richness conditions on the training MDP distribution, gradient flow converges locally and exponentially to an optimal parameter manifold corresponding to the desired RL update. Empirically, training transformers on randomly generated tabular MDPs confirms these predictions: the learned models recover the parameter structure of our explicit constructions and, when deployed on unseen MDPs, deliver strong in-context control performance. Together, these results illuminate how transformer architectures internalize and execute classical reinforcement learning algorithms in context, bridging mechanistic understanding and training dynamics in ICRL.

Current critic-less RLHF methods aggregate multi-objective rewards via an arithmetic mean, leaving them vulnerable to constraint neglect: strong performance in one objective (e.g., correctness) can offset critical failures in others (e.g., safety or formatting). This can mask low-performing "bottleneck" rewards that are necessary for reliable multi-objective alignment. We propose Reward-Variance Policy Optimization (RVPO), a risk-sensitive framework that penalizes variance across rewards during advantage aggregation, and shifts the objective from "maximize sum" to "maximize consistency." Its SoftMin objective can be locally approximated by the mean reward minus an inter-objective variance penalty. We evaluate RVPO on rubric-based medical and scientific reasoning with up to 17 reward signals (Qwen2.5-3B/7B/14B) and on tool-calling with rule-based constraints (Qwen2.5-1.5B/3B). RVPO improves overall scores on HealthBench (0.254 vs. 0.207 for GDPO at 14B, averaged across five training seeds). RVPO also maintains competitive accuracy on GPQA-Diamond while reducing late-stage degradation observed in other multi-reward methods.

Long-range electrostatics and polarization remain central obstacles to extending machine learning interatomic potentials (MLIPs) to ionic, polar, and interfacial systems. Here we introduce a semi-local framework for learning electrostatics from energies and forces using polarizable atomic multipoles. Local equivariant descriptors predict environment-dependent latent monopoles, dipoles, and quadrupoles, while residual non-local charge transfer and polarization are captured by non-self-consistent linear response in induced charges and dipoles. Across four diverse benchmarks and four short-range MLIP architectures, the multipole hierarchy and response terms systematically improve potential energy surface accuracy, with the largest gains in systems where long-range effects are essential. More importantly, physically meaningful electrical responses emerge without direct supervision. The learned latent multipoles yield accurate Born effective charge tensors and infrared spectra in close agreement with experiments. The induced-dipole extension introduces new capabilities: it predicts polarizabilities and thereby enables semi-quantitative Raman spectra for bulk water and hybrid MAPbI$_3$ perovskite, as well as the essential features of the surface-specific vibrational sum-frequency generation spectrum at the water-air interface. In ferroelectric HfO$_2$, the predicted electrical response also captures LO-TO splitting and polarization switching. This systematically improvable, physically transparent framework enables MLIPs trained on standard energy and force labels to predict polarization-sensitive observables.

Gaussian process (GP) marginal likelihood scores and kernel conditional independence tests are theoretically appealing for nonlinear causal discovery but computationally prohibitive at scale. We present three complementary RFF-based methods forming a practical toolkit for score-based, constraint-based, and hybrid causal discovery. The Fourier Feature Marginal Likelihood (FFML) score approximates the exact GP marginal likelihood by replacing the $n x n$ kernel Gram matrix with a finite-dimensional feature representation, reducing cost to $O(nm^2 + m^3)$ while retaining the probabilistic interpretation and automatic complexity penalty of the exact score. FFML extends to mixed (continuous and discrete) parent sets via a product-kernel construction, with a Kronecker path for small discrete parent sets and a Hadamard-product path otherwise. The Tetrad Random Fourier Feature (TRFF) score is a complementary BIC-style alternative using penalized Student-t regression with random Fourier features. TRFF offers robustness to heavy-tailed noise and faster runtime than FFML. Empirically, TRFF and FFML exhibit a complementary precision-recall profile: TRFF achieves higher precision while FFML achieves better recall and lower SHD overall. The Fourier Feature Conditional Independence (FFCI) test is a fast nonparametric CI test for mixed data, using ridge residualization in feature space and a Frobenius-norm cross-covariance statistic approximated as a weighted sum of chi-squared variables. Empirically, BOSS+FFML achieves the lowest SHD on nonlinear data, while BOSS+TRFF offers the highest precision. When run through PC-Max, FFCI and RCIT exhibit complementary precision-recall profiles: RCIT is more precise while FFCI achieves better recall and substantially lower SHD, at approximately twice the runtime.

Weak-to-strong generalization is a phenomenon in post-training whereby a strong student model, when finetuned solely with feedback from a weaker teacher, can not only surpass the teacher, but can improve upon its own capabilities. Recent work of Burns et al. (2023) demonstrated that this can occur in the setting of frontier language models, and subsequently there has been a flurry of both empirical work trying to exploit this phenomenon, as well as theoretical work attempting to understand it. In this work, we demonstrate that weak-to-strong generalization occurs in standard linear logistic regression, under mild distributional assumptions on the data. In fact, we show that this happens for most student-teacher pairs, suggesting that weak-to-strong generalization is in fact almost inevitable, even in this basic setting. Notably, our setting does not require the student to be more expressive or have more model capacity in any way compared to the teacher, which runs contrary to the prevailing theoretical belief that a mismatch in model capacity is a central mechanism to weak-to-strong generalization.

Agentic artificial intelligence systems produce outputs through sequences of interdependent autonomous decisions, yet standard evaluation assesses outputs alone and cannot diagnose the underlying process. We develop a behavioral evaluation methodology that complements output-level testing by scoring the intermediate decision process itself. Behavioral traces logged at each autonomous decision point are grouped into five-day episodes and scored along six domain-specific dimensions (regime detection, routing, adaptation, risk calibration, strategy coherence, error recovery) by an ensemble of three large language model (LLM) judges. A perturbation procedure that corrupts one dimension while leaving the other five intact confirms dimension specificity; cross-model agreement reaches Krippendorff's alpha = 0.85. The composite behavioral score correlates at Spearman rho = 0.72 with realized 20-day Sharpe ratio. Closing the loop, the framework converts deficient per-dimension scores into a credit-assigned penalty added to the Soft Actor-Critic reward. Three fine-tuning cycles, confined to validation data, reduce one-day MAPE from 0.61% to 0.54% (11.5% relative; p<0.001, d=0.31) on the held-out 2017 to 2025 test period, significant under Diebold-Mariano and localized by Giacomini-White to the high-volatility regime. The methodology is application-agnostic and applies to any agentic system whose intermediate decisions can be logged.

Traffic sensor networks evolve as sensors are added and traffic distributions change, whereas most forecasting models assume a fixed node set and repeatedly retrain on all available data. We propose CoMemNet, a Continual Memory Network for efficient prediction over evolving traffic sensor networks. CoMemNet uses an Online branch to adapt to the current period and an exponential-moving-average Target branch as a stable feature reference. A Wasserstein-based Drift Sampler compares node-wise Online-Target feature distributions and selects a limited set of drift-sensitive nodes for updating. A lightweight Node-Adaptive Temporal Memory Replay Buffer (TMRB-N) retains compact temporal states without repeatedly traversing all historical training data. The prediction backbone does not consume an adjacency matrix; sensor adjacency is used only to construct data and optionally expand the selected update set to a limited neighborhood. Experiments on three multi-period PeMS datasets include three-seed evaluation, strong static retraining and continual baselines, controlled sampling strategies, continual-learning metrics, robustness tests, and resource accounting. The results show that CoMemNet maintains stable prediction accuracy and efficient adaptation under bounded shared-node selection, achieving a better balance between historical knowledge preservation and current-period prediction performance. Meanwhile, as the evolving network expands, CoMemNet shows clearer accuracy and cumulative training-time advantages over current-period retraining baselines. The code is available at:https://meiwu5.github.io/CoMemNet.

Large language models (LLMs) can learn new tasks through fine-tuning, but learning tasks sequentially often leads to catastrophic forgetting of earlier ones. This paper proposes CRAFT, a continual learning framework that keeps model weights frozen and instead learns low-rank interventions on hidden representations. Similar tasks share an intervention to help knowledge transfer, with similarity judged by how each task alters the model's output distribution rather than by a learned gating network. Within a shared intervention, a Kullback-Leibler (KL) penalty keeps each new task close to the intervention's existing behavior, which limits forgetting of earlier tasks, while a small, task-private intervention in the deepest layers preserves dedicated capacity that later tasks cannot overwrite. The learned interventions also make it possible to identify the task at test time, so task identity labels are not needed. Across multiple benchmarks and model scales, CRAFT improves overall performance and reduces forgetting compared with strong LoRA-based methods, while remaining robust to task ordering.

Solving AC Optimal Power Flow (AC-OPF) is of central importance in electricity market operations, where interior-point methods (IPMs) such as IPOPT are the standard solvers. A growing body of work uses machine learning to predict primal warm-start iterates, reporting iteration reductions of 30-46%. We show that these reported gains rest on an inappropriate evaluation baseline: prior methods benchmark against the flat start $V_m = 1, V_a = 0$, whereas the solver's actual default - the variable-bound midpoint $(l+u)/2$ - is near-optimal for log-barrier centrality. Against this corrected baseline, no primal-only warm-start method reduces solver iterations. We trace the failure to a geometric property of interior-point methods: primal prediction accuracy is anticorrelated with convergence speed, and providing the ground-truth optimal solution $x^*$ without dual variables causes the solver to diverge. Oracle experiments establish that the complete primal-dual-barrier state $(x^*, \lambda^*, z^*, \mu^*)$ reduces IPOPT iterations from 23 to 3 - an 85% reduction that is structurally inaccessible to primal-only methods. To enable rigorous evaluation of warm-start methods on this task, we release a benchmark suite comprising dual-labeled AC-OPF datasets with IPOPT-extracted solutions, a corrected evaluation protocol, and WARP - a topology-conditioned encode-process-decode interaction network that predicts the full interior-point state $(\hat{x}, \hat{\lambda}, \hat{z}, \hat{\mu})$ on the heterogeneous constraint graph. WARP achieves a 76% reduction in IPOPT iterations while natively accommodating N-1 contingency topology variations without retraining.

Cooperative inference across independently deployed machine learning models is increasingly desirable in distributed environments, as there is a growing need to leverage multiple models while keeping their data and model parameters private. However, existing cooperative frameworks typically rely on sharing input data, model parameters, or a common encoder, which limits their applicability in privacy-sensitive or cross-organizational settings. To address this challenge, we propose Consensus Embedding-based Federated Inference (CE-FI), a framework that enables pretrained models to cooperate at inference time without sharing model parameters or raw inputs and without assuming a common encoder. CE-FI introduces two components: a Consensus Embedding (CE) layer that maps heterogeneous intermediate representations into a common embedding space, and a Cooperative Output (CO) layer that produces predictions from these embeddings. Both layers are trained using shared unlabeled data only, so the cooperative stage does not require additional labeled data. Experiments on image classification benchmarks – CIFAR-10 and CIFAR-100 – under diverse non-IID conditions show that CE-FI consistently outperforms solo inference and performs comparably to conventional methods that require stronger sharing assumptions. Additional evaluations on text and time-series tasks indicate applicability beyond image classification, although performance depends on the ensemble strategy. Further analysis identifies representation alignment as the primary bottleneck.

Can linearly decodable failure signals in LLM hidden states be leveraged to correct those failures? We investigate this classification-correction gap via Overthinking (OT)–a stable behavioral regime (Jaccard >= 0.81, 94% inter-annotator agreement) in medical QA where models answer correctly under resampling yet fail in extended chain-of-thought. OT is linearly decodable at 71.6% balanced accuracy (p < 10^{-16}). Yet five families of fixed linear steering (29 configurations, n=1,273) all yield Delta ~= 0, with identical null results cross-architecture (Qwen2.5-7B) and cross-domain (MMLU-STEM). Three convergent lines of evidence suggest representational entanglement: the OT direction has 85-88% overlap with task-critical computation (specificity ratio <= 0.152); non-targeted shared-direction steering damages accuracy (-12.1pp); and LEACE concept erasure damages accuracy (-3.6pp, p=0.01), while 10 random erasures produce Delta=+0.3pp. The per-instance probe-steering correlation is r=-0.002 (p=0.97). Positively, the same probe enables selective abstention (held-out AUROC=0.610, exceeding all five uncertainty baselines, p=0.009): decodable failure structure supports post-generation reliability estimation even when the fixed linear steering family cannot exploit it for correction.

Recent advances in large language models (LLMs) have significantly improved language-driven 3D content generation, but most existing approaches still treat scene generation and user interaction as separate processes, limiting the adaptability and immersive potential of interactive multimedia systems. This paper presents a unified framework that closes the loop between language-driven 3D scene generation and immersive user interaction. Given natural language instructions, the system first constructs structured scene representations using LLMs, and then optimizes spatial layouts via reinforcement learning under geometric and semantic constraints. The generated environments are deployed in a virtual reality setting to facilitate HRI-in-the-loop, where user interactions provide continuous feedback to align generated content with human perception and usability. By tightly coupling generation and interaction, the proposed framework enables more responsive, adaptive, and realistic multimedia experiences. Experiments on the ALFRED benchmark demonstrate state-of-the-art performance in task-based scene generation. Furthermore, qualitative results and user studies show consistent improvements in immersion, interaction quality, and task efficiency, highlighting the importance of closed-loop integration of generation and interaction for next-generation multimedia systems. Our project page can be found at https://proj-showcase.github.io/h3ds/.