Large language models suffer from content effects in reasoning tasks, particularly in multi-lingual contexts. We introduce a novel method that reduces these biases through explicit structural abstraction that transforms syllogisms into canonical logical representations and applies deterministic parsing to determine validity. Evaluated on the SemEval-2026 Task 11 multilingual benchmark, our approach achieves top-5 rankings across all subtasks while substantially reducing content effects and offering a competitive alternative to complex fine-tuning or activation-level interventions.

Anticipating future actions in videos is challenging, as the observed frames provide only evidence of past activities, requiring the inference of latent intentions to predict upcoming actions. Existing transformer-based approaches, which rely on dot-product attention over pixel representations, often lack the high-level semantics necessary to model video sequences for effective action anticipation. As a result, these methods tend to overfit to explicit visual cues present in the past frames, limiting their ability to capture underlying intentions and degrading generalization to unseen samples. To address this, we propose Action-Guided Attention (AGA), an attention mechanism that explicitly leverages predicted action sequences as queries and keys to guide sequence modeling. Our approach fosters the attention module to emphasize relevant moments from the past based on the upcoming activity and combine this information with the current frame embedding via a dedicated gating function. The design of AGA enables post-training analysis of the knowledge discovered from the training set. Experiments on the widely adopted EPIC-Kitchens-100 benchmark demonstrate that AGA generalizes well from validation to unseen test sets. Post-training analysis can further examine the action dependencies captured by the model and the counterfactual evidence it has internalized, offering transparent and interpretable insights into its anticipative predictions.

Contact-rich manipulation demands human-like integration of perception and force feedback: vision should guide task progress, while high-frequency interaction control must stabilize contact under uncertainty. Existing learning-based policies often entangle these roles in a monolithic network, trading off global generalization against stable local refinement, while control-centric approaches typically assume a known task structure or learn only controller parameters rather than the structure itself. In this paper, we formalize a physically grounded interaction frame, an instantaneous local basis that decouples force regulation from motion execution, and propose a method to recover it from demonstrations. Based on this, we address both issues by proposing Force Policy, a global-local vision-force policy in which a global policy guides free-space actions using vision, and upon contact, a high-frequency local policy with force feedback estimates the interaction frame and executes hybrid force-position control for stable interaction. Real-world experiments across diverse contact-rich tasks show consistent gains over strong baselines, with more robust contact establishment, more accurate force regulation, and reliable generalization to novel objects with varied geometries and physical properties, ultimately improving both contact stability and execution quality. Project page: https://force-policy.github.io/

Modern AI models are not static. They go through multiple updates in their lifecycles. We propose to design Sequential Membership Inference (SeMI) attacks leading to tighter privacy audits by exploiting the sequence of models and injecting a target canary at a controlled insertion time. First, for empirical mean computation, we develop SeMI*, an {optimal SeMI attack to identify the presence of a target inserted at a specific insertion step}. We derive the power of SeMI* to show that accessing the model sequence yields more powerful MI attacks than scrutinising only the final model. SeMI* exhibits an isolation property -- its power depends on the statistics obtained right before and after insertion of the target. Leveraging this insight, we develop practical white-box (accessing model gradients) and black-box (accessing loss) SeMI attacks against models trained with (DP-)SGD. Across datasets and models trained with (DP-)SGD, our experiments show that SeMI attacks achieve higher powers than snapshot-independent baselines, and yield tighter privacy audits thanks to (a) control over the insertion time and (b) observations across the model sequence.

We study eigendecomposition on $SO(n)$ under streaming observations $C_k = C_{\mathrm{sig}} + σ_k^2 I + E_k$, where the isotropic background $σ_k^2 I$ may be time-varying and arbitrarily large. Standard algorithms couple their stability to $\lVert C_k \rVert_2 \approx σ^2$, forcing step sizes, contraction rates, and iteration counts to degrade with the noise floor. We observe that $σ^2 I$ lies in the center of the matrix algebra and therefore *should never enter* the eigenspace dynamics. We construct a discrete double-bracket flow whose skew-symmetric generator $Ω= [A, \operatorname{diag}(A)]$ operates in the tangent Lie algebra $\mathfrak{so}(n)$, where scalar multiples of the identity vanish by antisymmetry. The resulting trajectory, Lyapunov function, and maximal stable step size $η_{\max} = 1/L_C$ depend exclusively on the trace-free signal $C_e$ -- achieving pointwise, pathwise $σ^2$-invariance. We establish input-to-state stability with a noise ball governed solely by trace-free perturbations, prove global convergence via strict-saddle geometry and a discrete Łojasiewicz argument, and extend the framework to top-$k$ eigentracking on the Stiefel manifold $\operatorname{St}(k,n)$ at cost $k$ matrix-vector products per step.

Federated low-rank adaptation (FedLoRA) has facilitated communication-efficient and privacy-preserving fine-tuning of foundation models for downstream tasks. In practical federated learning scenarios, client heterogeneity in system resources and data distributions motivates the use of heterogeneous LoRA ranks across clients. However, we identify a previously overlooked phenomenon in heterogeneous FedLoRA with SVD-based allocation, termed rank collapse, where the energy of the global update becomes concentrated in the minimum shared rank, resulting in suboptimal performance and high sensitivity to rank configurations. Through theoretical analysis, we reveal the root cause of rank collapse: a mismatch between rank-agnostic aggregation weights and rank-dependent client contributions, which systematically suppresses higher-rank updates at a geometric rate over rounds. Motivated by this insight, we propose raFLoRA, a rank-partitioned aggregation method that decomposes local updates into rank partitions and then aggregates each partition weighted by its effective client contributions. Extensive experiments across vision, language, and reasoning tasks show that raFLoRA prevents rank collapse, improves model performance, and enhances robustness across diverse heterogeneous configurations compared with strong FedLoRA baselines.

Relational deep learning (RDL) has emerged as a powerful paradigm for learning directly on relational databases by modeling entities and their relationships across multiple interconnected tables. As this paradigm evolves toward larger models and relational foundation models, scalable and realistic benchmarks are essential for enabling systematic evaluation and progress. In this paper, we introduce RelBench v2, a major expansion of the RelBench benchmark for RDL. RelBench v2 adds four large-scale relational datasets spanning scholarly publications, enterprise resource planning, consumer platforms, and clinical records, increasing the benchmark to 11 datasets comprising over 22 million rows across 29 tables. We further introduce autocomplete tasks, a new class of predictive objectives that require models to infer missing attribute values directly within relational tables while respecting temporal constraints, expanding beyond traditional forecasting tasks constructed via SQL queries. In addition, RelBench v2 expands beyond its native datasets by integrating external benchmarks and evaluation frameworks: we translate event streams from the Temporal Graph Benchmark into relational schemas for unified relational-temporal evaluation, interface with ReDeLEx to provide uniform access to 70+ real-world databases suitable for pretraining, and incorporate 4DBInfer datasets and tasks to broaden multi-table prediction coverage. Experimental results demonstrate that RDL models consistently outperform single-table baselines across autocomplete, forecasting, and recommendation tasks, highlighting the importance of modeling relational structure explicitly.

In this work, we study nonconvex-strongly convex online bilevel optimization (OBO) using only first-order oracle. Existing OBO algorithms are mainly based on hypergradient descent, which requires access to a Hessian-vector product (HVP) oracle and potentially incurs high computational costs. By reformulating the original OBO problem as a single-level online problem with inequality constraints and constructing a sequence of Lagrangian function, we eliminate the need for HVPs arising from implicit differentiation. Specifically, we propose a fully first-order algorithm for OBO, and provide theoretical guarantees showing that it achieves regret of $O(1 + V_T + H_{2,T})$ with a total of $O(T\log T)$ iterations, where $V_T$ measures the variation in function values and $H_{2,T}$ characterizes the drift variation of the inner-level optimal solution. We also establish a sublinear regret bound under the single-loop structure by introducing additional gradient-variation terms. Furthermore, we develop an improved variant with an adaptive inner-iteration scheme, which removes the dependence on $H_{2,T}$ and achieves regret of $O(\log T + V_T)$. Finally, under the stochastic OBO setting, we establish the regret bound for the fully first-order algorithm, i.e., $O(T^{2/3}(1 + σ^2) + V_T + H_{2,T})$. Numerical experiments demonstrate the feasibility of our algorithm and support our theoretical findings.

Unsupervised industrial anomaly detection (UAD) is essential for modern manufacturing inspection, where defect samples are scarce and reliable detection is required. In this paper, we propose HLGFA, a high-low resolution guided feature alignment framework that learns normality by modeling cross-resolution feature consistency between high-resolution and low-resolution representations of normal samples, instead of relying on pixel-level reconstruction. Dual-resolution inputs are processed by a shared frozen backbone to extract multi-level features, and high-resolution representations are decomposed into structure and detail priors to guide the refinement of low-resolution features through conditional modulation and gated residual correction. During inference, anomalies are naturally identified as regions where cross-resolution alignment breaks down. In addition, a noise-aware data augmentation strategy is introduced to suppress nuisance-induced responses commonly observed in industrial environments. Extensive experiments on standard benchmarks demonstrate the effectiveness of HLGFA, achieving 97.9% pixel-level AUROC and 97.5% image-level AUROC on the MVTec AD dataset, outperforming representative reconstruction-based and feature-based methods.

The Rashomon set captures the collection of models that achieve near-identical empirical performance yet may differ substantially in their decision boundaries. Understanding the differences among these models, i.e., their multiplicity, is recognized as a crucial step toward model transparency, fairness, and robustness, as it reveals decision boundaries instabilities that standard metrics obscure. However, the existing definitions of Rashomon set and multiplicity metrics assume centralized learning and do not extend naturally to decentralized, multi-party settings like Federated Learning (FL). In FL, multiple clients collaboratively train models under a central server's coordination without sharing raw data, which preserves privacy but introduces challenges from heterogeneous client data distribution and communication constraints. In this setting, the choice of a single best model may homogenize predictive behavior across diverse clients, amplify biases, or undermine fairness guarantees. In this work, we provide the first formalization of Rashomon sets in FL.First, we adapt the Rashomon set definition to FL, distinguishing among three perspectives: (I) a global Rashomon set defined over aggregated statistics across all clients, (II) a t-agreement Rashomon set representing the intersection of local Rashomon sets across a fraction t of clients, and (III) individual Rashomon sets specific to each client's local distribution.Second, we show how standard multiplicity metrics can be estimated under FL's privacy constraints. Finally, we introduce a multiplicity-aware FL pipeline and conduct an empirical study on standard FL benchmark datasets. Our results demonstrate that all three proposed federated Rashomon set definitions offer valuable insights, enabling clients to deploy models that better align with their local data, fairness considerations, and practical requirements.

Long-horizon planning is widely recognized as a core capability of autonomous LLM-based agents; however, current evaluation frameworks suffer from being largely episodic, domain-specific, or insufficiently grounded in persistent economic dynamics. We introduce EcoGym, a generalizable benchmark for continuous plan-and-execute decision making in interactive economies. EcoGym comprises three diverse environments: Vending (adapted from the closed-source Vending-Bench, with full open-source release), Freelance (new), and Operation (new), implemented in a unified decision-making process with standardized interfaces, and budgeted actions over an effectively unbounded horizon (1000+ steps if 365 day-loops for evaluation). The evaluation of EcoGym is based on business-relevant outcomes (e.g., net worth, income, and DAU), targeting long-term strategic coherence and robustness under partial observability and stochasticity. Experiments across eleven leading LLMs expose a systematic tension: no single model dominates across all three scenarios. Critically, we find that models exhibit significant suboptimality in either high-level strategies or efficient actions executions. EcoGym is released as an open, extensible testbed for transparent long-horizon agent evaluation and for studying controllability utility trade-offs in economic settings.

Neural networks are increasingly used as fast surrogate models across various domains, but unconstrained predictions can violate physical, operational, or safety requirements. We propose SnareNet, a feasibility-controlled architecture to learn mappings whose outputs must satisfy input-dependent constraints. SnareNet appends a differentiable repair layer that navigates in the constraint map's range space, steering iterates toward feasibility and producing a repaired output that satisfies constraints to a user-specified tolerance. We stabilize end-to-end training by adaptive relaxation, a new training paradigm that snares the neural network at initialization and shrinks it into the feasible set, enabling early exploration and strict feasibility later in training. On optimization learning and trajectory planning benchmarks, SnareNet consistently attains improved objective quality while satisfying constraints more reliably than prior work, and it is the first to enforce non-convex constraints at medium-to-high precision robustly across instances.

Reinforcement Learning (RL) is increasingly applied to large-scale decision-making problems like logistics, scheduling, and recommender systems, but existing algorithms struggle with the curse of dimensionality in such large discrete action spaces. We propose Distance-Guided Reinforcement Learning (DGRL), combining Sampled Dynamic Neighborhoods and Distance-Based Updates to enable efficient RL in problems with up to $10^{20}$ actions. Unlike prior methods, DGRL performs stochastic volumetric exploration and transforms policy optimization into a stable regression task, decoupling gradient variance from action space cardinality. On structured tasks, DGRL provably guarantees local value improvement. DGRL naturally generalizes to hybrid continuous-discrete action spaces. We demonstrate performance improvements of up to 66% against state-of-the-art benchmarks across regularly and irregularly structured environments, while simultaneously improving convergence speed and computational complexity.

Bayesian optimization (BO) is a popular technique for sample-efficient optimization of black-box functions. In many applications, the parameters being tuned come with a carefully engineered default configuration, and practitioners only want to deviate from this default when necessary. Standard BO, however, does not aim to minimize deviation from the default and, in practice, often pushes weakly relevant parameters to the boundary of the search space. This makes it difficult to distinguish between important and spurious changes and increases the burden of vetting recommendations when the optimization objective omits relevant operational considerations. We introduce BONSAI, a default-aware BO policy that prunes low-impact deviations from a default configuration while explicitly controlling the loss in acquisition value. BONSAI is compatible with a variety of acquisition functions, including expected improvement and upper confidence bound (GP-UCB). We theoretically bound the regret incurred by BONSAI, showing that, under certain conditions, it enjoys the same no-regret property as vanilla GP-UCB. Moreover, assuming known ARD lengthscales -- the same assumption underlying GP-UCB regret bounds -- BONSAI provably recovers the relevant-coordinate set at zero acquisition cost, yielding a method that matches the GP-UCB regret rate while recovering the minimal-$\ell_0$ solution -- a guarantee not provided by prior sparse-BO methods. Across many real-world applications, we empirically find that BONSAI substantially reduces the number of non-default parameters in recommended configurations while maintaining competitive optimization performance, with little effect on wall time -- averaging only $1.5\times$ the candidate-generation cost of standard BO, compared to $7$-$34\times$ on average for prior sparse-BO methods (IR, ER, and SEBO).

Neuronavigation is widely used in biomedical research and interventions to guide the precise placement of instruments around the head to support procedures such as transcranial magnetic stimulation. Traditional systems, however, rely on subject-mounted markers that require manual registration, may shift during procedures, and can cause discomfort. We introduce and evaluate markerless approaches that replace expensive hardware and physical markers with low-cost visible and infrared light cameras incorporating stereo and depth sensing, combined with algorithmic modeling of the facial geometry. Validation with 50 human subjects yielded a median tracking discrepancy of only 2.32 mm and 2.01$^\circ$ for the best markerless algorithm compared to a conventional marker-based system, which indicates sufficient accuracy for transcranial magnetic stimulation and a substantial improvement over prior markerless results. The study also suggests that integration of the data from the various camera sensors can improve the overall accuracy further. The proposed markerless neuronavigation methods can reduce setup cost and complexity, improve patient comfort, and expand access to neuronavigation in clinical and research settings.

Online bilevel optimization (OBO) has emerged as a powerful framework for many machine learning problems. Prior works have developed several algorithms that minimize the standard bilevel local regret or the window-averaged bilevel local regret of the OBO problem, but the optimality of existing regret bounds remains unclear. In this work, we establish optimal regret bounds for both settings. For standard bilevel local regret, we propose an algorithm with adaptive iteration strategy that achieves the optimal regret $Ω(1+V_T)$ with at most $O(T\log T)$ total inner-level gradient evaluations. We further develop a fully single-loop algorithm whose regret bound includes an additional gradient-variation terms. For the window-averaged bilevel local regret, we design an algorithm that captures linear environmental variation through a novel window-based analysis and achieves the optimal regret $Ω(T/W^2)$. The algorithm also supports an efficient single-loop structure, achieving an $O(T/W)$ regret bound with $O(WT)$ total gradient evaluations. Experiments validate our theoretical findings and demonstrate the practical effectiveness of the proposed methods.

Large language models (LLMs) are increasingly deployed in high-stakes settings where good decisions require forming beliefs over the probability of unknown outcomes. However, it is unclear whether LLMs act as if they hold coherent beliefs when making decisions, or if so, how we could validate models' reports of such beliefs. We propose a decision-theoretic framework that elicits both probability judgments and decisions from an agent and tests their mutual consistency. Formally, our methods characterize whether it is possible for the actions to be produced by a ``near-rational" decision maker who holds the elicited probability as their true belief. We show that, perhaps surprisingly, this formalization implies empirically testable conditions even without any assumption about the agent's utility function. Applying our framework to stylized clinical diagnosis tasks, we find that models' reported beliefs are demonstrably imperfect summaries of the information revealed in their decisions, but that the discrepancies are small for the strongest models.

True self-evolution requires agents to act as lifelong learners that internalize novel experiences to solve future problems. However, rigorously measuring this foundational capability is hindered by two obstacles: the entanglement of prior knowledge, where ``new'' knowledge may appear in pre-training data, and the entanglement of reasoning complexity, where failures may stem from problem difficulty rather than an inability to recall learned knowledge. We introduce SE-Bench, a diagnostic environment that obfuscates the NumPy library and its API doc into a pseudo-novel package with randomized identifiers. Agents are trained to internalize this package and evaluated on simple coding tasks without access to documentation, yielding a clean setting where tasks are trivial with the new API doc but impossible for base models without it. Our investigation reveals three insights: (1) the Open-Book Paradox, where training with reference documentation inhibits retention, requiring "Closed-Book Training" to force knowledge compression into weights; (2) the RL Gap, where standard RL fails to internalize new knowledge completely due to PPO clipping and negative gradients; and (3) the viability of Self-Play for internalization, proving models can learn from self-generated, noisy tasks when coupled with SFT, but not RL. Overall, SE-Bench establishes a rigorous diagnostic platform for self-evolution with knowledge internalization. Our code and dataset can be found at https://github.com/thunlp/SE-Bench.

3D Gaussian Splatting (3DGS) revolutionized novel view rendering. Instead of inferring from dense spatial points, as implicit representations do, 3DGS uses sparse Gaussians. This enables real-time performance but increases space requirements, hindering rate-constrained applications. 3DGS compression emerged as a field aimed at alleviating this issue. While impressive progress has been made, at low rates, compression introduces artifacts that degrade visual quality significantly. We introduce NiFi, a method for extreme 3DGS compression through restoration via artifact-aware, diffusion-based one-step distillation. We show that our method achieves state-of-the-art perceptual quality at extremely low rates, down to 0.1 MB, and towards 1000x rate improvement over 3DGS at comparable perceptual performance. Code is available at: https://github.com/ceteke/nifi

Understanding how neural representations respond to geometric transformations is essential for evaluating whether learned features preserve meaningful spatial structure. Existing approaches primarily assess robustness primarily by comparing model outputs under transformed inputs, offering limited insight into how geometric information is organized within internal representations and failing to distinguish between information loss and re-encoding. In this work, we introduce SEIS (Subspace-based Equivariance and Invariance Scores), a subspace metric for analyzing layer-wise feature representations under geometric transformations, disentangling equivariance from invariance without requiring labels or explicit knowledge of the transformation. Through controlled experiments across diverse architectures, we uncover several consistent patterns. First, convolutional encoders exhibit a depth-wise transition from strong equivariance to increasing invariance, with both properties stabilizing within the first few training epochs. In segmentation decoders, however, equivariance tends to recover in later layers. Second, this trade-off is not intrinsic but is shaped by training decisions: data augmentation actively strengthens both equivariance and invariance simultaneously, and multi-task learning induces synergistic gains in both properties beyond what either task achieves alone. Extending our analysis beyond convolutional networks, we find that transformer-based models exhibit distinct geometric behaviors, while MLP-Mixers display intermediate characteristics.

Modality following is the ability to selectively leverage multimodal contexts based on user instructions. It is fundamental to the safety and reliability of multimodal large language models (MLLMs) in real-world deployments. However, the internal mechanisms governing this decision-making process remain largely under-explored. In this work, we investigate the mechanism underlying modality following through an information flow perspective. Our findings reveal that instruction tokens serve as structural anchor for modality arbitration: Shallow attention layers perform undifferentiated information transfer, aggregating multimodal cues to instruction tokens as a latent buffer; in contrast, deep attention layers selectively strengthen the instruction-compliant subspace and resolve modality arbitration according to the instruction-specified intent, with a sparse subset of attention heads driving this process. Targeted attention-head interventions further validate the functional specificity of these heads: blocking only $5\%$ of the identified heads substantially degrades modality following while preserving general visual and language capabilities, whereas targeted amplification can restore failed modality-following samples by up to approximately $60\%$. Together, this work provides a mechanistic account of modality following and informs future efforts to improve how MLLMs integrate and utilize multimodal evidence under user instructions.

Reinforcement learning algorithms such as group-relative policy optimization (GRPO) have shown strong potential for improving the mathematical reasoning capabilities of large language models. While a growing body of work seeks to improve training entropy, rollout diversity, and exploration, most existing methods still train models with a single fixed reasoning prompt or template, which can encourage prompt-specific overfitting and unstable training dynamics. In this work, we introduce Prompt Augmented Policy Optimization (PrAg-PO), a simple policy optimization method that mixes prompt templates with template-specific format rewards during training. By encouraging models to generate reasoning traces under diverse instructions and output formats, PrAg-PO increases rollout diversity and improves robustness. Compared with GRPO and DAPO, PrAg-PO achieves significantly higher reasoning accuracy while mitigating premature training collapse. Empirically, experiments on DeepSeek-R1-Distill-Qwen-1.5B, Qwen2.5-Math-1.5B, and Qwen3-1.7B show that PrAg-PO consistently outperforms strong baselines and achieves competitive performance against recent methods on mathematics benchmarks, using only a fixed MATH Level 3-5 training set of 8.5K problems. The code and model checkpoints are available at https://github.com/wenquanlu/PrAg-PO.

Much recent work has argued that the variation in the languages of the world can be explained from the perspective of efficient communication; in particular, languages can be seen as optimally balancing competing pressures to be simple and to be informative. Focusing on the expression of meaning -- semantic typology -- the present paper asks what factors are responsible for successful explanations in terms of efficient communication. Using the Information Bottleneck (IB) approach to formalizing this trade-off, we first demonstrate and analyze a correlation between optimality in the IB sense and a novel generalization of convexity to this setting. In a second experiment, we manipulate various modeling parameters in the IB framework to determine which factors drive the correlation between convexity and optimality. We find that the convexity of the communicative need distribution plays an especially important role. These results move beyond showing that efficient communication can explain aspects of semantic typology into explanations for why that is the case by identifying which underlying factors are responsible.

Clinical brain-to-text interfaces are designed for paralysed patients who cannot provide extensive training recordings. Pre-training improves data-efficient generalisation by learning statistical priors across subjects, but these priors critically depend on context. While natural speech might unfold gradually over minutes, most methods pre-train with only a few seconds of context. Thus, we propose MEG-XL, a model pre-trained with 2.5 minutes of MEG context per sample, 5-300x longer than prior work, and equivalent to 191k tokens, capturing extended neural context. Fine-tuning on the task of word decoding from brain data, MEG-XL matches supervised performance with a fraction of the data (e.g. 1hr vs 50hrs) and outperforms brain foundation models. We find that models pre-trained with longer contexts learn representations that transfer better to word decoding. Our results indicate that long-context pre-training helps exploit extended neural context that other methods unnecessarily discard. Code, model weights, and instructions are available at https://github.com/neural-processing-lab/MEG-XL .

Diffusion models have emerged as powerful learned priors for Bayesian inverse problems (BIPs). Diffusion-based solvers rely on a presumed likelihood for the observations in BIPs to guide the generation process. Likelihood misspecification is common in practical BIPs and is known to degrade recovery performance, particularly under outlier contamination. We investigate this problem by first characterizing the induced posterior deviation and proving the stability of diffusion-based solvers for linear BIPs. Our stability analysis further reveals potential robustness deficiencies of existing diffusion-based solvers under outlier-contaminated measurements. To address this issue, we propose a simple yet effective solution: robust diffusion posterior sampling, which is provably outlier-robust for linear BIPs and compatible with existing gradient-based posterior samplers. Empirical results from scientific inverse problems and natural image tasks demonstrate the effectiveness and robustness of our method, with consistent performance gains in challenging scenarios involving outlier contamination for both linear and nonlinear tasks.

Knowledge editing methods for large language models are commonly evaluated using predefined benchmarks that assess edited facts together with a limited set of related or neighboring knowledge. While effective, such evaluations remain confined to finite, dataset-bounded samples, leaving the broader impact of editing on the model's knowledge system insufficiently understood. To address this gap, we introduce Embedding-Virtualized Knowledge (EVK) that characterizes model knowledge through controlled perturbations in embedding space, enabling the exploration of a substantially broader and virtualized knowledge region beyond explicit data annotations. Based on EVK, we construct an embedding-level evaluation benchmark EVK-Bench that quantifies potential knowledge drift induced by editing, revealing effects that are not captured by conventional sample-based metrics. Furthermore, we propose a plug-and-play EVK-Align module that constrains embedding-level knowledge drift during editing and can be seamlessly integrated into existing editing methods. Experiments demonstrate that our approach enables more comprehensive evaluation while significantly improving knowledge preservation without sacrificing editing accuracy.

Gradient-based attribution is the workhorse of mechanistic interpretability, yet whether it reliably tracks causal importance at the component level remains largely untested. We causally evaluate this assumption across two algorithmic tasks and up to 10 random seeds, uncovering a systematic, layer-wise failure: gradient attribution consistently overvalues early-layer \textbf{Gradient Bloats} and undervalues late-layer \textbf{Hidden Heroes}. Rank correlation collapses from $ρ= 0.72$ on sequence reversal to $0.27$ on sequence sorting, reaching $ρ= -0.18$ in individual seeds. This failure stems from first-order gradient attribution's inability to detect collective redundancy: joint Bloat ablation causes $14\times$ greater damage than individual results predict. Consequently, Bloats dominate gradient rankings despite negligible functional impact, while ablating Hidden Heroes destroys OOD accuracy ($-36.4\% \pm 22.8\%$). This systematic inversion of early-layer feature extraction and late-layer computation motivates causal validation as a prerequisite for circuit-level claims.

The vanilla self-attention mechanism in Transformers can be viewed as a two-layer fast-weight MLP, whose weights are dynamically induced by inputs and whose hidden dimension is equal to the sequence length $N$. As the context extends, the expressive capacity of such an $N$-width MLP increases, but it becomes unscalable for extremely long sequences. Recently, this fast-weight perspective has motivated the Mixture-of-Experts (MoE) attention mechanism, which partitions the sequence into rigid blocks, treats them as fast-weight experts, and sparsely routes the tokens to them. In this paper, we elevate this perspective to a unifying framework for efficient attention mechanisms, interpreting them as making fast weights scalable through either routing or compression, and organizing them into a five-dimensional taxonomy. Then, we propose Mixture-of-Top-$k$ Attention (MiTA), which employs a small set of landmark queries to gather top-$k$ attended key-value pairs as query-aware and deformable routed experts, while compressing the $N$-width MLP into a narrower shared expert. Consequently, our MiTA improves the flexibility of prior MoE attention from rigid to deformable fast-weight experts, as well as the scalability of prior top-$k$ attention from query-specific set to reusable top-$k$ set. We conduct extensive experiments on vision tasks showing the superior effectiveness and efficiency of our MiTA, and also uncovering intriguing properties such as an emergent token-pruning effect and easy generalization from standard attention. Code is available at https://github.com/QishuaiWen/MiTA.

Long-term weather forecasting is critical for socioeconomic planning and disaster preparedness. While recent approaches employ finetuning to extend prediction horizons, they remain constrained by the issues of catastrophic forgetting, error accumulation, and high training overhead. To address these limitations, we present a novel pipeline across pretraining, finetuning and forecasting to enhance long-context modeling while reducing computational overhead. First, we introduce an Efficient Multi-scale Transformer (EMFormer) to extract multi-scale features through a single convolution in both training and inference. Based on the new architecture, we further employ an accumulative context finetuning to improve temporal consistency without degrading short-term accuracy. Additionally, we propose a composite loss that dynamically balances different terms via a sinusoidal weighting, thereby adaptively guiding the optimization trajectory throughout pretraining and finetuning. Experiments show that our approach achieves strong performance in weather forecasting and extreme event prediction, substantially improving long-term forecast accuracy. Moreover, EMFormer demonstrates strong generalization on vision benchmarks (ImageNet-1K and ADE20K) while delivering a 5.69x speedup over conventional multi-scale modules. Code: https://github.com/chenhao-zju/emformer

Large language models (LLMs) are increasingly embedded in AI-based tutoring systems. Can they faithfully model novice reasoning and metacognitive judgments? Existing evaluations emphasize problem-solving accuracy, overlooking the fragmented and imperfect reasoning that characterizes human learning. We evaluate LLMs as novices using 630 think-aloud utterances from multi-step chemistry tutoring problems with problem-solving logs of student hint use, attempts, and problem context. We compare LLM-generated reasoning to human learner utterances under minimal and extended contextual prompting, and assess the models' ability to predict step-level learner success. Although GPT-4.1 generates fluent and contextually appropriate continuations, its reasoning is systematically over-coherent, verbose, and less variable than human think-alouds. These effects intensify with a richer problem-solving context during prompting. Learner performance was consistently overestimated. These findings highlight epistemic limitations of simulating learning with LLMs. We attribute these limitations to LLM training data, including expert-like solutions devoid of expressions of affect and working memory constraints during problem solving. Our evaluation framework can guide future design of adaptive systems that more faithfully support novice learning and self-regulation using generative artificial intelligence.