Papers

Large language models (LLMs) have been rapidly improving in long-context tasks, powered by Chain-of-Thought (CoT) reasoning. However, the internal mechanisms underlying this improvement remain unclear. We investigate these mechanisms through a needle-in-a-haystack (NIAH) counting task, where an LLM is asked to count the number of records dispersed in a long text. Across twelve model comparison groups, Thinking (or reasoning) improves counting accuracy over Non-thinking, with pronounced gains at larger counts. This motivates our mechanistic analysis, which identifies two contrasting mechanisms: (i) broad retrieval, where Non-thinking models broadly attend to multiple needles; (ii) targeted retrieval, where Thinking models use enumeration in CoT traces to successively retrieve needles. Targeted retrieval concentrates attention on individual needles and is accompanied by more compact internal representations. Moreover, causal intervention analysis suggests that Thinking models use the CoT trace to maintain and update an internal counter as needles are successively retrieved, even without explicit numbering. In small controlled experiments, both retrieval mechanisms and counter states emerge under standard autoregressive training. Together, our results connect long-context retrieval with representation geometry of counting, supporting a state-tracking account of CoT reasoning.

Inverse Reinforcement Learning (IRL) aims to recover a reward function that explains expert demonstrations. Existing IRL methods typically rely on a bi-level optimization procedure that alternates between reward learning and policy optimization, leading to substantial computational burden and training instability. In this work, we introduce a different route that eliminates policy optimization entirely by leveraging diffusion policies. Our key insight is that a diffusion policy encodes the action-gradient structure of the optimal soft Q function, enabling reward learning to be cast as a sequence of value recovery problems, thereby allowing us to bypass reward-policy loops inherent in prior IRL methods. Specifically, our method proceeds in three stages: (I) recovering the optimal soft Q function via action-gradient matching and estimating the corresponding soft value function (LogSumExp of Q values) in a way inspired by Gumbel regression; (II) calibrating these soft values by inferring a state-dependent offset; (III) extracting the reward by enforcing Bellman consistency. This leads to Loop-Free Inverse Reinforcement Learning (LFIRL), a fully offline algorithm that operates in a simple, loop-free, and sequential manner. LFIRL is simple to implement and significantly improves training efficiency while maintaining strong reward recovery performance. Empirically, across Maze, Franka Kitchen, Adroit Hand Pen, and Push-T benchmarks, LFIRL achieves 2-3x speedup over the fastest baselines, while matching or surpassing state-of-the-art methods in reward recovery quality.

Reinforcement learning with human feedback (RLHF) learns from human comparisons, which can be corrupted or deliberately manipulated. This paper studies online risk-sensitive RLHF with static conditional value-at-risk (CVaR) under adversarial preference-label flips. We consider additive linear rewards and a fixed-reference protocol with one comparison per episode and at most $C$ flipped labels over $K$ episodes. We propose weighted streamed-preference CVaR RLHF (WSP-CVaR-RLHF), which combines uncertainty-weighted reward estimation with optimistic augmented-state CVaR planning. For known transitions and normalized rewards, we establish the regret bound $\widetilde{O}\left(\frac{d}{\kappa}\sqrt{\frac{K}{\alpha}}+\frac{dC}{\kappa\alpha}\right)$ up to lower-order terms, where $d$ is the reward-feature dimension, $\alpha$ is the CVaR level, and $\kappa$ characterizes the preference link. The bound separates the clean statistical cost from the penalty caused by corrupted feedback. We further extend the analysis to unknown tabular transitions, where the trajectory distribution entering the CVaR objective must be learned together with the reward. We address the resulting coupled uncertainty using rectangular transition confidence sets, joint optimistic planning, and a history-level CVaR simulation argument. Experiments under four adversarial attacks demonstrate that WSP-CVaR-RLHF consistently reduces cumulative regret relative to its unweighted robust counterpart while preserving confidence-set coverage.

When a classifier is recalibrated from only a few thousand held-out examples, the capacity of the calibration map becomes a statistical design choice rather than a purely architectural one: a scalar map can underfit structured residual miscalibration, while a highly adaptive map can be hard to estimate reliably from so small a split. We disentangle the calibration objective from adaptive capacity and propose signal-routed temperature scaling (SRTS-BCE), a 10-parameter, argmax-preserving calibrator that cross-fits a correctness-risk score over six logit statistics and fits one top-label-BCE temperature per $K=3$ risk groups, recovering TvA-TS as its $K=1$ limit. On fine-tuned CIFAR-100 / ViT-B/16, SRTS-BCE reduces $\mathrm{ECE}_{15}$ from 1.65 (scalar TvA-TS) to 0.96, matching the higher-capacity SMART+BCE head (0.95) at the full calibration budget. The two regimes separate as the budget shrinks: at $n=250$ SRTS-BCE beats SMART+BCE on all three CIFAR-100 backbones (the seed-to-draw hierarchical interval excludes zero), whereas the flagship comparison against the scalar remains directional. A protocol-frozen Tiny-ImageNet follow-up reproduces the small-budget separation and exhibits a budget-dependent ranking reversal on Swin-T; matched routing and map controls show that the effect is tied neither to the learned router nor to discrete grouping. Together the results identify post-hoc calibrator capacity as a finite-sample design choice whose preferred level shifts with the amount of available calibration data.

Differentially private (DP) text generation can protect individual records, but privacy alone does not specify what evidence a released statement carries about the underlying data. We identify this as an evidence gap: a private report may contain plausible claims without indicating whether they are strongly supported by the private dataset. We introduce PrivCert, a framework for privacy-preserving reporting that makes statement support explicit through privacy-preserving certificates and emit-or-abstain decisions. As a canonical instantiation, PrivCert-PF (Proposal-and-Filter) separates data-independent candidate discovery from private support certification, emitting only statements whose support passes a private evidence test. We provide theoretical grounding for this framework by characterizing the limits of implicit evidence under DP, deriving a sharp privacy–honesty frontier for single-statement certification, and establishing a worst-case cost for fine-grained multi-statement certification. Experiments on synthetic tasks and TAB, WildChat, and Yelp show that explicit certification maintains low unsupported emission, while free-text DP baselines frequently produce low-support claims under the same declared support semantics. We further show that the PrivCert contract can be realized with histogram, sparse-vector, and Gaussian mechanisms, and use DP synthetic data to illustrate an important boundary: support in a private proxy does not automatically certify support in the original data. Together, these results position privacy-preserving reporting as an evidence-design problem: not only how to generate private text, but what a private report can substantiate about its underlying data.

Surface electromyography (sEMG) provides a wearable, camera-free signal for continuous hand-motion inference. Mapping muscle activity to joint kinematics remains challenging because the recorded waveforms are indirect measurements, their relationship with motion changes over time, and individual anatomy and sensor placement alter the signal distribution. This paper presents EFormer, a residual feature-correction network built on a frozen tracking backbone. EFormer combines a high-rate event branch, temporally aligned local cross-attention, two causal rotary position embedding (RoPE) temporal layers, and a bounded, dynamically gated residual. EFormer receives 16-channel sEMG sampled at 2 kHz and fuses a 64-channel tracking representation at 25 Hz with a 128-channel event representation at 200 Hz. Cross-attention uses a nominal delay of 100 ms, a 300 ms history parameter, and a 50 ms tolerance; its causal mask restricts each query to events occurring 50-400 ms earlier. The correction scale is 0.15. The evaluated continuation-training configuration contains 585,376 trainable parameters and 5,974,508 frozen parameters. On the test set, EFormer achieves an MAE of 0.1546634 rad, an RMSE of 0.24063 rad, and an R^2 of 0.74801, compared with 0.1745326 rad, 0.2715448 rad, and 0.6791103 for the official tracking baseline. EFormer reduces MAE by 11.38% relative to the baseline. The results show that temporally aligned event-feature correction can reduce continuous hand-pose tracking error.

Self-consistency samples many reasoning trajectories and aggregates their final answers, treating the LLM as a black box that returns one answer per trajectory. Yet the final answer of each trajectory is sampled from a softmax vector that is available from the model's log-probabilities. We refer to this as the grey-box setting in which each trajectory reveals this answer distribution rather than a single draw from it. We formulate efficient inference in this setting as sequential mode identification with distribution-valued observations: sample trajectories one at a time and stop as soon as the LLM's modal answer is identified at a prescribed confidence level. We characterize the asymptotic stopping rate of mode identification with distribution-valued observations exactly and show that it is never worse than the black-box rate. We then propose the ASC-D algorithm, a betting stopping rule that attains this asymptotic stopping rate. On MMLU-Redux, ASC-D uses $46.4$–$95.6\%$ fewer trajectories than answer-only adaptive self-consistency baselines and achieves the highest fixed-budget correct-certification rate across three open-source models.

In semantic segmentation, a recent line of RankSEG methods directly optimizes Dice/IoU scores at inference time, improving alignment with evaluation metrics without modifying model training. Despite its theoretical and empirical success, RankSEG relies on the restrictive Conditional Independence Assumption (CIA), which ignores crucial label correlations and therefore degrades performance in ambiguous or low-contrast scenarios. However, accounting for full label dependence is computationally prohibitive, requiring $\mathcal{O}(d^3)$ time. To address this, we replace the CIA with a Spatially Localized Dependence (SLD) structure that captures local label correlations while keeping the dependence model tractable. We further overcome the remaining computational bottleneck via a Reciprocal Moment Approximation coupled with a novel fixed-point optimization strategy that eliminates exhaustive search. The proposed algorithm achieves a highly practical $\mathcal{O}(d \log d)$ complexity and consistently outperforms conventional argmax and CIA-based RankSEG across diverse segmentation benchmarks. Improvements are significant in low-contrast or small-object scenarios, where label dependence offers valuable signals complementary to image information for accurate segmentation. The code of experiments is available at https://github.com/ZixunWang/RankSEG-DEP.

Machine learning systems increasingly face the need to remove the influence of entire data domains, such as toxic language, harmful behavior, or topical content, rather than isolated records. Recent work formalizes this problem as distributional unlearning: selecting a subset of a forget domain whose removal moves the training distribution away from an unwanted population while preserving proximity to the desired one. However, existing analyses often impose parametric assumptions to obtain tractable selection rules. These assumptions may be poorly suited to high-dimensional language-model representations. We introduce Mamushi, a framework for non-parametric distributional unlearning that ranks forget examples using a probabilistic classifier whose Bayes-optimal logit equals the forget-to-retain log-density ratio (up to an additive class-prior constant). We show that thresholding the population log-density ratio yields the optimal fixed-budget selection rule for our removal–preservation objective and establish a non-asymptotic transfer guarantee relating score-estimation and threshold-calibration errors to degradation from the population-optimal selection rule. Our empirical evaluation spans real-world datasets on toxic-language removal and topical-domain removal regimes using different representations, with Mamushi achieving a more favorable removal–preservation trade-off than other baselines. Our work shows that Mamushi can serve as an efficient selection approach for downstream machine unlearning procedures, reducing the number of forget examples required to reach a fixed forgetting target.

World models aim to learn representations of real-world environments and predict their future evolution. Recent object-centric world models have made expressive progress by representing visual scenes as sets of object-level latent states, but object-object relations are often captured only implicitly, which limits explicit relational and temporal structure modeling and object-centric dynamic memory modeling. To address such challenges, we propose World-As-Graph (WAG), a graph-based object-centric world model that introduces relational inductive bias into JEPA-style predictive representation learning. The proposed WAG contains two main modules: (1) Relation-aware structure induction, which constructs time-varying latent graphs from object-centric slots and designs relation-aware object masking policies to guide relational object representation learning in latent space; (2) Object-centric memory transition, which maintains and updates object-level dynamic states by combining relational information from neighboring objects with historical memory, enabling effective autoregressive future prediction. Extensive experiments on both visual reasoning and robotic manipulation tasks could demonstrate the superior performance of our proposed WAG.

Long-term precipitation nowcasting requires modeling radar-echo evolution while preserving localized high-intensity structures. Recent radar-specific studies motivate location-aware prediction and separating echo displacement from intensity change. However existing encoders learn historical representations mainly from final forecast errors. We propose PrecipJEPA, which couples a structured forecasting path with an auxiliary path that enriches its encoder from observed radar history. In the forecasting path, an online encoder first converts the observations into spatiotemporal tokens. The Task-Driven Future-State Predictor (TFP) combines these tokens with a recent-dynamics summary and spatiotemporal queries to construct future radar states. The Parallel Motion-Source Renderer (PMSR) decodes these states into motion and source-sink fields that transform the latest observation into future frames. During joint training, the History-Masked JEPA (H-JEPA) operates on the auxiliary path to predict masked historical features from visible context, directly supervising the same online encoder from the observed sequence. Experiments on SEVIR and MeteoNet show that PrecipJEPA improves highest-threshold CSI by 118.6% and 35.1%, respectively, over the strongest baselines, while maintaining the highest mean CSI throughout the 3-hour forecast.

Multimodal federated learning (MFL) has emerged as a pivotal paradigm for leveraging distributed data to enhance model performance. However, existing methods predominantly rely on idealized assumptions of model homogeneity and balanced modality distributions, rendering them ill-suited for practical scenarios characterized by heterogeneous client architectures and severe modality imbalance. To address these challenges, we propose a Multimodal Federated learning Prototype-guided Bilateral Alignment (MFedPBA) framework. MFedPBA facilitates robust knowledge synergy through a dual alignment mechanism: (i) at the feature level, it aligns heterogeneous feature spaces via a projection encoder optimized by contrastive learning and the Gromov-Wasserstein distance; (ii) at the decision level, it employs an entropy-weighted aggregation of naturally aligned logit prototypes. This novel design achieves robust MFL by jointly tackling heterogeneous feature spaces and collectively aggregating decisions. Extensive experiments demonstrate that our method significantly outperforms state-of-the-art baselines under conditions of model heterogeneity and modality imbalance.

Offline multi-objective optimization (MOO) seeks solutions with better objective trade-offs using only a fixed dataset, without querying the objectives. Diffusion models trained on such data have emerged as a promising approach, but their samples are not inherently better than the data and must be steered toward the Pareto front. Existing methods guide or condition every sampling step. We instead act on the initial noise and leave the sampling process unchanged. Across Off-MOO-Bench, we observe that the objectives, as functions of the noise, are sensitive to only a few directions. We estimate these directions once per task via a Recursive Feature Machine using function values alone, and a small cache serves every trade-off, so each candidate costs one noise displacement and one ODE solve. We prove that this displacement increases the learned scalarized objective in expectation, and that sweeping trade-offs recovers the flow's attainable front up to proxy and steering errors. With additional guidance, for which we introduce novel data-adaptive and Pareto-aware operators, our method attains the best average hypervolume rank among generative methods on 47 tasks, at comparable or lower sampling cost. Steering alone outranks the best prior generative method at a fraction of its sampling cost.

Flow Matching (FM) learns a velocity field whose ODE transports a simple source distribution to a target law. Existing finite-sample theory largely treats ambient-space regularity or data supported exactly on low-dimensional sets. We study linear FM under a noisy latent-generator model, where a low-dimensional Hölder map is perturbed by nondegenerate ambient Gaussian noise, so the target law is full-dimensional despite its latent structure. We construct a spatially regular ReLU velocity class and establish non-asymptotic high-probability approximation and estimation bounds whose leading sample-size exponent is governed by the latent dimension rather than the ambient dimension, with ambient and noise dependence kept explicit. Fixed positive target noise keeps the interpolation nondegenerate over the full time interval. The same spatial regularity propagates the learned velocity error through the transport ODE, yielding a corresponding Wasserstein convergence guarantee. These results show that exact low-dimensional support is not necessary for Flow Matching to retain latent-dimensional statistical behavior.

Any-dimensional machine learning models, such as graph neural networks (GNNs), can be naturally trained and evaluated on inputs of different sizes and dimensions. Inspired by the GNN transferability literature, we show mathematical conditions under which a partial differential equation (PDE) learning-based solver can be trained in small dimensions and directly applied to solve a higher dimensional PDE in a zero-shot fashion. These conditions are based on symmetries in both the partial differential equation and the initial data. When the equations satisfy the symmetries but the data does not, which is the case for many PDEs arising from physics, we show that our theory gives a principled way of warm-starting low-dimensional PDE solvers for higher dimensional PDEs. We apply this method on the heat equation, Burgers' equation, and the compressible Navier–Stokes equations, improving the performance in both zero-shot and typical training regimes on high dimensional data. For example, we train a surrogate model on 2D Navier–Stokes data and achieve better results on 3D test data than a baseline surrogate model trained on 3D data, while only using 12$\%$ of the flops and 20$\%$ of the total data size.

Evaluating interactive agents is expensive. Agent behavior is stochastic, so reliability must be measured over repeated trials, but failures are rare and differ widely in how much they matter. Standard benchmarks spend this budget uniformly: a read-only lookup is sampled as often as an irreversible payment action. We instead formulate evaluation as a sequential allocation problem. Given a fixed trial budget and a set of scenarios whose failure behavior is unknown, which scenarios should be run, and run again? We propose a risk-aware contextual Thompson Sampling policy that combines a pre-execution scenario context vector and a fixed impact score with the failure outcomes observed during evaluation, and we test it by offline replay over 70 $\tau$-bench airline scenarios and 824 recorded trials. Our main result is at the smallest budget: with only 50 trials ($6\%$ of the corpus), the policy recovers $86\%$ of the impact-weighted failures an oracle could find, compared to $25\%$ for uniform allocation. It discovers $3.5\times$ more impact-weighted failures (215.4 vs. 62.2) with the same number of trials, delivers $5\times$ the discovery per dollar, and cuts the budget wasted on scenarios that never fail from $34\%$ to $2.8\%$. The rest of our analysis demonstrates and qualifies this result: a budget sweep shows the advantage shrinks as the budget approaches the corpus size, and paired significance tests show that scenario context helps mainly at small budgets while posterior-based exploration helps at moderate ones. Risk-aware adaptive allocation therefore helps most exactly where evaluation budget is scarcest.

Learning must acquire new capabilities while preserving both prior responsibilities and the capacity to learn again. We introduce Generalized Residual Closure (GRC), a framework for learning as recursive closure of future-relevant discrepancies: closing a residual establishes the conditions for subsequent prediction, interaction, and learning. Under a complete representation-relation description at a fixed learner-world boundary, persistent internal learning has two primitive modes: Transformation within a representation and revision of the Representation itself. We establish a local tangent decomposition under regularity assumptions and a criterion for when representation revision is necessary. In an affine model, we derive a necessary-and-sufficient condition for stability-plasticity compatibility and the unique solution of a constrained quadratic update problem, which preserves registered old responsibilities while reducing residuals with an effective safe response. We prove that reconstructive semantic protection weakly enlarges the safe-response operator relative to preserving an exact historical realization. Dynamic sufficiency and future-closure viability extend representation adequacy from current prediction to lawful future updating and continued learning. Growth Learning expands the lawful closure domain or lowers optimal closure cost without regression of the registered capability-cost frontier; a conditional commit rule maintains this order. Restricted-sector recoveries and a conditional representation theorem connect the framework to optimization, machine learning, and control. Together, these results organize adaptation, representation revision, and reusable capability within a common account of continued learning.

Large language models often know the truth and say otherwise: a model that answers correctly when asked neutrally will affirm a user's mistaken belief, or misstate a fact its system prompt wants hidden, once the context rewards it. Such deception is a behavior conditioned on context, not knowledge, yet machine unlearning, the natural tool for removing a behavior from the weights, is built to forget facts that a deceptive model still needs. We propose to unlearn when a model deceives rather than what it knows, with a contrastive forget unit built from the model's own realized deceptions: the same question under a deception-triggering and a neutral context, admitted only where belief holds and behavior flips. Standard objectives on this unit face a dilemma. Suppression objectives such as NPO leave much of the deception in place. Target-based objectives, which distill the model's neutral behavior into the pressured context, remove it but induce context blindness: a target generated without the context teaches the model to stop reading it, eroding benign system-prompt instructions, secret-keeping and the reasoning a monitor inspects, a failure invisible to deception rates and capability benchmarks. We introduce PACT, which trains toward pressure-aware counterfactual targets (the model's own honest response, with a trace that registers the pressure and resists it) while retaining the benign uses of the triggering context. On two 32B reasoning models, PACT reduces held-out deception from over 50% to under 3% while system-prompt adherence, secret-keeping and the reasoning trace stay at the base model's level. On a tug-of-war score of removal against retention, PACT reaches 0.94 and 0.86, against at most 0.77 and 0.60 for any baseline. Like removed knowledge, removed deception is shallow under relearning, and terms that simulate the attacker hold it only at a cost in context use.

Single-cell transcriptomics enables profiling of cellular states at unprecedented resolution, but its high dimensionality, sparsity, and technical batch effects pose significant challenges for representation learning. Existing single-cell foundation models typically encode each cell independently or only model cells from the same batch for denoising, thereby underutilizing the rich relational information across batches and cell types to model gene expression patterns. We argue that single-cell models can benefit from more informative cell-context modeling. By comparing consistency and variation across cells, models can capture fine-grained gene-gene dependencies associated with cell states, which are essential for learning high-quality representations. Inspired by the use of multiple sequence alignment (MSA) context in protein modeling, we propose CellMSA, a single-cell representation learning framework that introduces an MSA-inspired inductive bias into transcriptomic modeling. For each target cell, CellMSA retrieves relevant cells from different batches and biologically related cell types as context, and summarizes cross-cell patterns into a context-dependent gene-pair representation. This representation is then injected into a pair-aware target-cell encoder for fine-grained representation learning. We pretrain CellMSA on a large-scale human single-cell corpus of approximately 109 million cell observations, including 65.6 million primary observations. Experiments show that our framework consistently outperforms existing methods across multiple benchmarks. Code is available at the following repository: https://github.com/PharMolix/CellMSA.

Actor-critic methods achieve strong performance in continuous control, but their policies can produce highly oscillatory actions. A common remedy is to add auxiliary smoothness losses. However, their contribution can be negligible when their gradients are small relative to the native actor gradient. Moreover, existing methods often combine multiple auxiliary losses, complicating loss balancing without necessarily improving the return-smoothness trade-off. We introduce DAMPER (Direction-Aware Magnitude-Controlled Projection with Explicit Return Priority), which combines the native actor gradient with a temporal-consistency gradient through conflict-conditioned projection and adaptive magnitude control. It removes the auxiliary component opposing the actor gradient and scales the retained temporal direction relative to the actor gradient norm, preserving positive alignment with the native actor gradient. Experiments with TD3 and SAC on six continuous-control tasks show reduced action oscillation relative to the native agents in all 12 task-backbone pairs and the best oscillation score among the compared methods in eight, with task-dependent return trade-offs.

Augmenting small concurrent studies with external or historical cohorts is attractive in drug development, where enrollment is slow, follow-up is expensive, and closely related trial or real-world data are often already available. Bayesian dynamic borrowing (BDB) provides a principled framework for adaptively controlling the influence of external data, but classical implementations often depend on hand-specified priors and MCMC-based inference, which can be computationally expensive and not generalizable. In this work, we study amortized neural posterior estimation (NPE) as a flexible alternative. A single network is pretrained on simulated current/external dataset pairs spanning covariate shift, outcome drift, and joint non-exchangeability, and then returns an approximate posterior for a scalar current-study target in a single forward pass. Through simulation studies, we find that NPE is most useful under outcome drift and joint mismatch: in the harder outcome-drift regimes, it gives up to about five-fold lower absolute bias than the best classical baseline and keeps Type I error close to nominal. After pretraining, posterior summaries are obtained in about 8 ms per dataset, roughly $10^3\times$ faster than MCMC-based borrowing baselines in our timing experiment. We further analyze Alzheimer's Disease Neuroimaging Initiative (ADNI) data and show that, when mild cognitive impairment outcomes differ across cohorts, the NPE formulation recovers the later-cohort risk level in this example without claiming greater precision. Code is available at https://github.com/ChinHungScott/NPE-for-Bayesian-Dynamic-Borrowing-MLHC-.

Representer explanations rank the training landmarks that most influence a self-supervised representation. At scale, this ranking rests on up to four stacked approximations of the empirical neural tangent kernel (eNTK). These are random output heads, a parameter sketch, landmark sampling and a coefficient fit. Existing analyses bound each approximation separately, but none certifies the top-$K$ set against their combined error. We introduce CAIRN (Certified Approximation for Interpretable Representer laNdmarks), a framework that carries this error through to the ranking. We derive the exact variance of the sketched multi-head eNTK, which matches measurement within $4\%$ where Johnson-Lindenstrauss bounds err by up to $2.5\times$. This yields a high-probability top-$K$ certificate for a fixed coefficient fit, alongside exact residual-trace certificates for discarded spectral mass. An exact product-variance identity separates kernel error from fit variability and identifies when a larger kernel budget can still sharpen a ranking. Stochastic Lanczos Quadrature (SLQ) estimates the effective dimension within $0.72\%$ and guides the landmark budget without dense eigendecomposition. We show that residual mass does not control class coverage, and residual-greedy selection cuts the worst coverage excess of $k$-means++ from $8.5\times$ to $1.55\times$ ($4\times$ on the sketched eNTK). Cross-view initializers outperform principal-component initialization in five (AUI) to all six (CSI) settings. Against the KREPES Gauss-Newton solver, CAIRN converges $2.5$ to $11.3\times$ faster, trails by at most $0.31$ points and gains up to $3.14$ points on MNIST. Together, these results make the reliability of representer explanations measurable and show where approximation budgets are best spent.

Video Multimodal Large Language Models (Video-MLLMs) support reasoning over video inputs, yet remain vulnerable to jailbreak attacks that elicit policy-violating responses. Existing video jailbreaks primarily manipulate how harmful queries are visually presented, thereby treating video merely as a carrier. Consequently, the surrounding video scenario remains unexplored as a contextual attack surface. In this paper, we show that the same harmful query can elicit different safety responses when placed in different video scenarios. To systematically exploit this vulnerability, we propose SceneJail, an adaptive black-box jailbreak framework with two coordinated components. Adaptive Scenario Construction dynamically searches for a surrounding scenario that is contextually compatible with the harmful query. Scenario-aware Prompt Search uses black-box response feedback to search for textual guidance tailored to the selected scenario. Extensive evaluations on the HADES and SafeBench datasets across eight Video-MLLMs, including two proprietary models, GPT-4.1 and Gemini3.5-Flash, demonstrate the effectiveness of SceneJail. SceneJail-F, which presents the complete query persistently, achieves average attack success rates (ASR) up to 91.5%, outperforming the strongest baselines by 29.1 percentage points. Furthermore, SceneJail-S, which distributes the query across successive frames, remains highly robust against current defenses, retaining a 72.3% ASR even under strict image filtering.

Knowledge distillation can transfer reasoning from stronger teachers to frozen students through reusable prompts, but avoiding weight updates does not eliminate supervision. Without ground-truth answers, teacher solutions are unverified, and agreement with the teacher can reward shared mistakes. We introduce Knowledge-to-Prompt (K2P) for label-free knowledge distillation to prompts. K2P synthesizes reusable instructions from teacher solutions, refines them using paired teacher and student responses, and guides search and selection with answer agreement. It retains candidates that adaptive search may undervalue and selects on reserved questions. Deployment uses only the frozen student and selected prompt. Our theory separates generation and selection gaps and gives conditions under which agreement-guided construction yields accuracy guarantees despite imperfect teacher references. Across reasoning tasks and students, K2P outperforms label-free alternatives overall and remains competitive with supervised prompt optimization. Ablations and archive diagnostics assess the contributions of teacher solutions and refinement, while revealing the limits of agreement-guided selection.

Approximate machine unlearning seeks to remove the influence of a forget set from a trained model without full retraining. Existing gradient-based methods require data-dependent hyperparameter search, struggle when forget and retain knowledge are entangled, and offer little insight into where unlearning actually happens inside the network. We recast unlearning through the lens of task arithmetic: if finetuning produces a merged task vector $\tau_m$ that combines learning on forget and retain sets, unlearning is the inverse operation that subtracts a learned forget component $\tau_F$ to recover the retain task vector $\tau_R$. The forget signal is concentrated: at every layer, forget activations lie in a subspace spanned by a handful of dominant directions, so we factorize $\tau_F$ in a low-rank forget basis, which is faithful up to a small tail-eigenvalue residual and limits how far the correction can perturb retain. We then optimize three intuitive goals (match the merged vector inside the forget span, suppress leakage into the retain span, and bound the correction size) that provably bound forget leakage and retain damage in activation space. The resulting algorithm, Unmerge, is fast and powerful: on class-level unlearning with ResNet-50 on CIFAR-100 and Tiny ImageNet, it improves Tug-of-War by up to ~24% over a baseline of comparable runtime and by up to ~18% over stronger baselines that run ~5x slower, keeps membership-inference exposure at the level of retraining, and shrinks the feature-distribution gap to the retrained model, where relabeling methods leave forget features cleanly separable. Further studies show that Unmerge also applies to ViT-S/16 and scales to Llama-3.2-3B. The per-layer basis geometry that drives the algorithm also serves as a layerwise diagnostic for when and where unlearning becomes structurally hard.

A stochastic hybrid system (SHS) is governed by a stochastic differential equation (SDE) describing the continuous dynamics and a Markov reset kernel triggered on the guard surface. Its probability evolution can be described by a hybrid Fokker-Planck (HFP) equation with a partial differential term corresponding to the SDE and an integral term arising from the reset kernel. This work shows that such an SHS can be approximated by an SDE in a higher-dimensional latent space where the sample paths are continuous. The key to this result is to encode different branches of the reset kernel using auxiliary variables, transforming the resets into deterministic ones that enable topological gluing. By the embedding theorem, the glued manifold can then be embedded into a higher-dimensional Euclidean space. We show that the probability evolution on the embedded image no longer requires explicit reset terms in the HFP equation. Building on this theorem, we design a loss that matches the evolving state distributions, enabling a single latent SDE to recover the probability evolution of the SHS without mode labeling, trajectory segmentation, or event-based simulations.

Constrained multi-agent control requires more than predicting rewarding actions: an action can cease to be executable as contact windows, shared capacity, and deadlines change. We introduce VERA, a centralized-training, decentralized-execution framework that separates feasibility estimation from credit assignment. Each actor predicts a five-dimensional verifiable feasibility representation (VFR). After an action is proposed, exact action-conditioned margins available only during training supervise that representation, while a counterfactual group-relative advantage (CGRA) ranks candidate representation-action pairs. Execution uses one actor pass and no privileged state. In a dynamic space-air-ground integrated network (SAGIN), VERA obtains 55.33% +/- 3.60% success with 0.45% +/- 0.81% coverage violation, within 1.33 percentage points of a privileged-mask reference. With rewards matched over ten paired seeds, VERA improves success over the strongest baseline by 8.74 percentage points (p=0.023) and reduces violation by 52.19 percentage points (p=5.7e-8). A ten-seed 4-by-2 factorial attributes a 14.16-16.48 percentage-point gain to CGRA across handcrafted, learned, random, and latent representations; evaluation on seven unseen topologies preserves a 24.33-30.02 percentage-point advantage over multi-agent proximal policy optimization. From 10 to 40 users, success remains 50.1-53.8%, and VFR adds only 0.026 ms to a central processing unit (CPU) actor step. Cross-domain tests further identify the governing condition: counterfactual credit succeeds when candidate scores respect shared constraints and fails under incompatible reward geometries. These results establish action-conditioned feasibility as an auditable training interface and counterfactual credit as a geometry-dependent optimization mechanism.

Real-time voice conversion (VC) systems commonly rely on pretrained speaker embeddings from automatic speaker verification (ASV) models. While effective for speaker discrimination, these embeddings are trained to remain stable across phonetic and prosodic variations within-speaker, which may conflict with frame-level acoustic generation in streaming constraints. To address this issue, we propose VOSSA (Voiceprint Optimization for Streaming Speech Architectures), a speaker representation framework that extracts speaker information from intermediate content encoder layers and aggregates using attentive statistics pooling. The embedding is trained jointly with VC objectives, removing the need for a separate speaker encoder. Across six datasets, VOSSA improves F0 dynamics and vowel-discriminative acoustic cues while maintaining comparable NISQA-MOS, WER, and speaker similarity. Perceptual tests further indicate improvements in naturalness, speaker similarity, intelligibility, and vibrancy.

Optimization modeling formulates real-world decision problems as mathematical programs that solvers can use to find optimal decisions. Large language models (LLMs) can automate this process, but the resulting correct formulations can require substantial time and memory to construct and solve, limiting practical scalability. Therefore, we systematically investigate whether LLMs can identify problem structure from natural-language descriptions and apply suitable optimization modeling techniques to generate mathematical models and solver code that solve the problems correctly and efficiently. To this end, we first curate OptTips, a knowledge base of 50 expert modeling techniques in eight families. Using this knowledge, we develop OptDachshund, a multi-agent framework that transforms problems from existing optimization benchmarks into new tasks for evaluating LLMs' use of modeling techniques. It constructs conventional and expert mathematical models with solver code for the same task and data, providing baselines for correctness and computational cost. The resulting EfficientOpt benchmark contains 561 expert-reviewed tasks with paired reference implementations. Evaluation of 11 representative LLMs reveals an efficiency gap on correctly solved tasks with comparable measurements: for every LLM, most generated programs take longer to solve than their expert counterparts. Within the comparable reference-size subset, 57% of programs with correct objective values and fewer variables and linear constraints have longer recorded solver times. Case studies show that different modeling techniques can achieve the same optimal value at similar recorded cost. Faster solving may not reduce execution time if the code takes longer to prepare data and build the model. LLM optimization modeling should therefore be evaluated for both correctness and computational efficiency.

We study the last iterate of standard importance-weighted temporal-difference (TD) learning from a single trajectory of a finite Markov decision process. The trajectory follows a behavior policy $b$ with an irreducible, possibly periodic state chain; a target policy $\pi$ covered by $b$ is evaluated with action importance ratios bounded by $\varrho$. Write $H=(1-\gamma)^{-1}$, with $\mu_{\min}$ the minimum stationary probability and $\mathfrak h_\star$ the largest expected hitting time of the reverse behavior state chain. We prove that the last iterate achieves sup-norm error at most $\varepsilon$ with high probability after $\widetilde O(\varrho H^3/(\mu_{\min}\varepsilon^2)+\mathfrak h_\star)$ transitions for $0<\varepsilon\le1$. The statistical term retains the cubic effective-horizon dependence of synchronous TD, while the initialization cost is additive in $\mathfrak h_\star$. Setting $b=\pi$ and $\varrho=1$ recovers the on-policy rate. For aperiodic behavior chains, $\mathfrak h_\star=\widetilde O(t_{\rm mix}/\mu_{\min})$ recovers a mixing-time bound. The proof combines an anchored local Poisson equation in reverse time to control fluctuations without a mixing-time factor, a compensation argument for successive hitting costs, and an expected cumulative propagation mass bound for decreasing step sizes. Given the policy ratios and their bound, an empirical state-count rule attains the same rate without prior knowledge of model parameters. The framework covers nonnegative affine importance-weighted prediction updates, including successor features, transition-dependent discounting, and coupled finite-mode predictions. A perturbation estimate yields a time-uniform guarantee for online return-variance estimation. A three-state, two-action construction gives matching minimax lower bounds for both terms up to logarithmic factors on the considered parameter classes.