Papers

Time Series Foundation Models (TSFMs) have achieved remarkable zero-shot performance through extensive pre-training on massive time series datasets. Nevertheless, due to the low-dimensional properties and diverse structural patterns of time series data, performing naive fine-tuning on TSFMs often leads to overfitting and falling into the mean-prediction trap. To address these challenges, we propose SIFT, a robust adaptation method that enhances time series foundation models by preserving Semantic Invariance and structural Fidelity throughout the fine-Tuning process. We employ semantic-invariant adversarial augmentation, which utilizes semantic spectrum decomposition to partition the semantic space and then generates perturbations within the non-core semantic subspace to bolster the model's robustness against these perturbations, mitigating overfitting. We implement a component-based structural fidelity enhancement, which facilitates component-wise mixup and imposes a reconstruction objective to improve the model's ability to preserve structural fidelity, alleviating the mean-prediction trap. Extensive experiments on representative TSFMs covering 10 real-world datasets demonstrate that SIFT can significantly enhance the performance of TSFMs.

Predicting whether a trained model will generalize under distribution shift remains difficult, especially when target-domain data are unavailable. We introduce STAMP (Semantic Temporal Augmented Model Prediction), a source-only, target-label-free criterion that estimates out-of-distribution (OOD) performance from paired source-domain images. STAMP computes the output-space correlation ratio $\eta^2=S_B/S_T$ by contrasting semantically stable pairs with random pairs: higher $\eta^2$ indicates that model outputs vary with semantic identity rather than nuisance variation. On 44 chest X-ray models spanning CNNs, ViTs, MetaFormers, foundation models, and SSL/VLM probes, temporal STAMP attains Spearman correlations of $0.844$–$0.855$ with macro AUROC on VinDr-CXR, CheXpert, and MIMIC-CXR; a class-matched variant improves single-class RSNA from $0.311$ to $0.663$. STAMP attains the best average source-only medical ranking and outperforms the target-domain ATC and AoTL estimators without any target data. On 27 ImageNet models, temperature-scaled STAMPTS attains $\rho{=}0.984$ on ObjectNet and $\rho\geq0.905$ on four additional distribution shifts, with partial correlations of $0.662$–$0.949$ after controlling for ImageNet accuracy. Requiring approximately 12s per model on one GPU, STAMP is a practical pre-deployment model-selection and auditing tool.

When a language model explains an answer it has already given, does it reuse the computation that produced the answer or reconstruct a story from the answer alone? Attribution, transportability and recoverability are each compatible with causal use without establishing it. We propose an evidence standard: pair each positive statistic with a variable specific null that removes the tested variable's identity while matching relevant nuisance dimensions as far as possible, and audit unmatched dimensions. We apply this standard to a known cause. A cue naming a wrong option raises the rate of choosing that option by 64 to 68 percentage points across three models. Explanations mention the cue in 1.8 percent of items or fewer in three of four models tested. Three estimator classes yield favorable statistics, but none establishes causal sensitivity to the cue contrast under its own control in the three-model analysis. In the strongest case, a recovered cue direction reaches $R^2$ of 0.95 and exceeds a geometry matched random direction in all three seeds, while a direction fitted by the same pipeline with cue labels scrambled reproduces 61 to 76 percent of its effect at comparable realized edit magnitude. A fourth model passes one interchange endpoint, but unequal edit magnitudes and a contrast that changes both cue identity and cue-answer agreement limit its interpretation. These experiments leave causal access unresolved. They establish an evidentiary requirement: favorable mechanistic statistics must survive controls for variable identity and nuisance structure. Reusable controls separate generic from identity specific transport effects, fit null directions with scrambled labels, and audit realized intervention magnitudes.

ELBO-based reinforcement learning offers a sampler-agnostic approach to fine-tuning flow matching models with reward feedback. Timestep weighting in ELBO-based RL has large impact on performance, and it also provides a unified view (as we show in this work) to understand prediction losses heuristically chosen in prior work, yet it remains under-researched and is often chosen to inherit pretrain configs. We investigate impacts and dynamics of timestep weighting in ELBO-based RL. We show that effective weighting depends on both the reward landscape and stage of learning. (1) Through experiments on controlled CIFAR image generation, complemented by robotics, we investigate how weighting impacts reward-driven updates across noise levels. (2) Through gradient analysis, we reveal distinct patterns of cross-noise coordination across tasks and their evolution during training. These findings motivate the hypothesis that useful weighting depends on the gap between the policy's current behavior and the behavior favored by the reward. (3) Guided by this analysis, we study simple static weighting, budgeted profile selection, and dynamic schedules that improve performance beyond conventional target choices. Our results establish timestep weighting as an important design choice for flow-matching RL and motivate further research into methods that choose and adapt it throughout learning.

The memory usage and decoding latency of LLM inference grow rapidly with context length. To reduce these costs, key-value (KV) cache compression methods selectively retain cached states based on token importance or differences in attention patterns across heads. However, we discover that retrieval capability varies substantially with relative distance, even within the same attention head. To exploit this structure, we introduce Distance-KV, which learns a static KV retention pattern over the joint space of layers, attention heads, and relative distances. The pattern is learned offline with the language model frozen and reused across inputs to prune and compact the KV cache without online importance scoring. Across three backbone models and four long-context benchmarks, Distance-KV consistently achieves the best overall performance among competing KV cache compression methods, exceeding the strongest compression baseline by up to 9.3 points on RULER at 128K. On Llama-3.1-8B-Instruct at 128K, Distance-KV reduces KV cache memory by 65.4% and achieves a $1.66\times$ decoding speedup relative to Dense. Together, these results identify relative distance as an important structural dimension for understanding how LLMs retrieve information over long contexts and for designing more efficient inference methods.

Maximum common edge subgraph (MCES) matching finds a partial vertex correspondence between two labeled graphs that preserves as many labeled edges as possible. Molecular similarity search requires matching many graph pairs, making the cost of repeated queries important. The strongest baseline attains accurate MCES solutions but trains a separate network for each pair. We introduce Equivariant Neural Primal-Dual Assignment (ENPDA), which learns a shared matching policy and applies it to new pairs without further training, answering queries roughly three orders of magnitude faster and recovering its training cost after a few dozen queries. The policy recomputes exact objective marginals for candidate matches and learns corrections and step sizes that update their scores. Target prices respond to competition when several source vertices favor the same target. Four update rounds and a Hungarian projection produce a partial one-to-one matching. We prove per-pair guarantees that hold for any network parameters. In exact arithmetic, reordering either graph permutes the assignment and price states, the projected matching is one-to-one, and repaired prices give a valid MCES upper bound. Subtracting the preserved-edge count bounds the optimality gap; combined with structural caps, these certificates prove global optimality for 60 of 291 native test pairs. On three molecular benchmarks with disjoint train/validation/test splits, ENPDA improves over an analytic counterpart with the same update and projection budget by 7.4-8.6 accuracy points; after one second of refinement search, 2.5-3 points of the gain remain. Transferred without fine-tuning to edge-deletion tasks from social and protein graphs, the policy gains 9.1-17.6 points over the analytic counterpart. When output matchings must keep aromatic rings intact, ENPDA recovers more reference bonds than the baselines on all three datasets.

Quantization-Aware Pre-Training (QAPT) can increase the inference efficiency of DNNs, but a problematic behaviour known as rounding boundary weight oscillation can introduce detrimental noise into the training process and significantly reduce convergence speed. While existing methods can reduce this detrimental noise, they either introduce additional hyperparameters or memory overhead, or cannot consistently improve model accuracy. In this work, we propose optimization with $\textbf{C}$onstrained $\textbf{E}$mpirical $\textbf{W}$eight dis$\textbf{T}$ribution (CEWT), the first hyperparameter-free memory-overhead-free oscillation suppression method that consistently improves QAPT performance: an optimizer post-update step that projects weights to the nearest point in weight space whose empirical distribution (histogram) matches a zero-mean Gaussian. Our key insight is many quantizers are designed with the implicit assumption that the to-be-quantized data are permutations of samples from a zero-mean Gaussian, and this assumption is not true during QAPT. By enforcing the zero-mean Gaussian prior as a hard constraint, CEWT can suppress this detrimental noise. Empirical results on various combinations of SOTA quantizers and hypersphere optimizers suggest, that with a geomean increase of 4% in training time, CEWT can consistently reduce the pre-training perplexity (by an average of 2.5 and up to 21 points) of low-precision (down to 1-bit activations and weights and up to 610M parameters) LLaMA/GPT models without introducing any hyperparameters or storage overhead. Code is available at https://github.com/1733116199/cewt

Multivariate Time Series Forecasting (MTSF) models that mix information across channels assume that the past of one channel carries information about the future of another. Yet they are evaluated on a small fixed set of standard datasets whose cross-channel structure is rarely examined. We ask two questions: "How can we reliably measure lagged, non-linear, and joint coupling in MTSF datasets?" and "Do the standard datasets actually have such coupling?" To answer the first, we test four candidate measures on synthetic datasets with planted ground-truth coupling: Granger Causality (GC), Transfer Entropy (TE), lagged Mutual Information (MI), and the CD gain, a model-based measure we introduce that compares a channel-dependent (CD) model to its channel-independent (CI) variant. Only lagged MI and the CD gain recover every planted coupling. For the second question, the answer is a definite no, as the standard datasets have a median of only 23% lagged-coupled channel pairs and a median CD gain of -4.9%, compared to 78% and +4.6% on chaotic ODE systems. We therefore propose MixBench-TS, a benchmark of 10 real-world datasets with a median of 55.5% lagged-coupled pairs and a median CD gain of +1.7%. Across six state-of-the-art models tuned under one protocol, CI models win on 10/10 (MSE) and 8/10 (MAE) standard datasets, but on only 3/10 and 2/10 MixBench-TS datasets. We recommend using our benchmark for evaluating new CD models. Moreover, we propose profiling new datasets with lagged MI and the CD gain before using them to evaluate multivariate models. Code and data are available at https://anonymous.4open.science/r/mixbench-ts-B027.

Soft state representations assign states nonnegative class weights that sum to one, but agreement with reference labels does not guarantee accurate prediction of the next weight vector. We derive lower confidence bounds on the minimum population root-mean-square prediction error over linear predictors with a prescribed spectral-norm limit. The bounds compare within-class successor variation with retained within-class input variation, with an alternative based on reference-label deviation. They are estimated from independent transition pairs without fitting a prediction matrix to the evaluation data. A bound above a chosen error tolerance excludes that accuracy for every admissible predictor at the stated confidence level; a bound of zero does not establish whether an accurate predictor exists. For coordinates constructed from Kernel Affine Hull Machine reconstruction scores, we relate the reference class's affinity advantage over competitors to label agreement. The class-centered population bound equals the optimum in a four-state example. Using 117,000 previously generated evaluation datasets from fixed four-state designs, we compare the finite-sample certificates with an empirical-risk-minimization baseline. For one Van der Pol representation selected before evaluation, the label-based bound is positive in all 64 independent evaluation datasets at the fitted matrix's spectral norm. Separately, exact matrix propagation under deterministic dynamics yields a Koopman interpretation of the coordinate-function span. The certificates help identify forecasting targets that cannot be reached by refitting a fixed representation within a specified predictor class and evaluation distribution.

Graph neural networks typically rely on dense, floating-point node representations, which can impose substantial memory and computational costs. Binary graph hashing offers an alternative by encoding node information as compact bit strings. However, existing approaches either sacrifice global topological information for computational efficiency or incur substantial generation costs. We introduce an ultra-fast, entirely algebraic hashing method that constructs binary node representations directly from graph structure, without requiring node features or gradient-based training. Our method approximates a high-order structural transition matrix using randomized column sampling inspired by the Nyström method and combines it with an efficient label-safe semantic propagation mechanism. The resulting continuous representations are discretized through column-wise thresholding to obtain compact binary codes. Experiments on ten node classification datasets show that the proposed method consistently improves classification accuracy over existing feature-free binary baselines while requiring sub-second code generation on many datasets. The resulting binary representations are also naturally suited to event-driven computation, making them compatible with neuromorphic spiking neural networks and gradient-free learning rules. These results demonstrate that simple algebraic approximations can provide an efficient alternative to learned pipelines for discrete graph representation learning.

We present Schur-Neural KF (SN-KF), a learning-based correction to the extended Kalman filter (EKF) that preserves the probabilistic conditioning interpretation of the EKF. The method perturbs the predictive state-measurement cross-covariance and the Cholesky factor of the measurement noise covariance so that the resulting joint predictive covariance is always positive semidefinite. The positive semidefiniteness is ensured by a Schur complement-based parametrization. We instantiate the parametrization with a recurrent neural architecture whose matrix outputs are modulated by amplitude gates. We prove that incorporating a measurement does not increase the filter's state uncertainty, and show that no measurement can induce an arbitrarily large state correction relative to its statistical surprise. We also present a perturbative analysis suggesting SN-KF's structural strength in the data-scarce regime. We provide two numerical experiments to illustrate the practical benefits of SN-KF. In a two-radar experiment, enforcing Schur-consistency provides a much broader failure-free hyperparameter region and reduces RMSE for small training subsets, consistent with our theoretical analysis in the data-scarce regime. In the unicycle experiment, SN-KF achieves the best precision, recall, false alarm rate, and gated RMSE under innovation-based sensor-fault rejection.

Multiobjective Bayesian optimization (MOBO) with Gaussian process (GP) surrogates is a sample efficient approach to solving multiobjective optimization problems. In MOBO, a Bayesian decision theoretic acquisition function guides the adaptive selection of new candidate inputs, on which objectives and constraints are evaluated to update the surrogate model sequentially. Existing approaches maintain independent GP models for the objectives and constraints, with new observations evaluating all objectives and constraints in a coupled fashion. However, the objectives and constraints often contain inherent correlations which, if exploited, can enable ${decoupled}$ evaluations where only a subset of them are evaluated at each round. We present a new approach that leverages a multitask GP model to jointly learn all objectives and constraints, and propose a total correlation metric that enables identifying an ${optimal}$ subset of objectives and constraints to be evaluated at every round, even under uniform evaluation costs. Theoretically, we show that our acquisition policy is asymptotically consistent despite decoupling and that our proposed decoupled subset selection rule maximizes the expected posterior entropy reduction about unevaluated tasks under mild conditions. Empirically, we show that our approach outperforms coupled and decoupled baselines in the state of the art.

Large self-supervised vision models learn representations that support scene segmentation and the semantic decomposition of physical objects. We ask whether their representational geometry supports transfer to visual reasoning problems without any task-specific fine-tuning. We hypothesized that relational representations may bridge perception and abstract reasoning by encoding similarities and transformations among visual inputs such that an intuitive, implicit form of reasoning may be performed via latent vector arithmetic. Specifically, we examine DINOv3, MAE, and random pixel projections on abstract and naturalistic Bongard problems, ARC-AGI-1 and ARC-AGI-2, and novel ARC-GEN instances. On both Bongard benchmarks, the accuracy of a simple nearest-centroid readout of frozen visual embeddings is within four percentage points of the task-specific baselines reported with the original benchmarks. In ARC, latent difference vectors summarizing demonstration input-output transformations, which we refer to as intuition vectors, show greater alignment with test vectors from the same task, whereas those from unrelated tasks are near orthogonal. This latent geometry is operational: transporting a query along its intuition vector consistently improves exact-output retrieval, reaching 70.7 on ARC-AGI-2 evaluation. Across 397,000 ARC-GEN instances from 794 tasks, single-pair intuition vectors identify the generating task with approximately 87% leave-one-out accuracy. These findings suggest that latent vector arithmetic over frozen visual representations supports implicit rule inference across varied problem domains without a generative model component, indicating that inferring an abstract transformation and generating its instance-specific consequence may be separable capacities.

Training neural networks directly in a low-rank parameterization is an appealing route to reducing memory, compute, and storage simultaneously during both training and inference. Dynamic low-rank training (DLRT), which confines weights to a rank-$r$ manifold via the Galerkin projection of the gradient flow, is particularly attractive because it identifies efficient subnetworks on the fly without specialized initialization or post-factorization. However, DLRT fails to find trainable networks under high compression. In this paper, we derive the gradient flow of the best rank-$r$ approximation and point out that the offset of DLRT comes from a curvature-coupling term which is large and thus non-negligible under aggressive compression. Guided by this analysis, we propose a stable dynamic low-rank training method, named SDLRT, which maintains a lightweight compensation buffer that reinjects the top neglected singular directions. Additionally, we introduce a negative feedback on the truncation tolerance to stabilize each layer's rank. Experimentally, SDLRT reliably finds trainable subnetworks where DLRT collapses and as a PEFT adapter on DeBERTa-v3, it achieves the best average score on SuperGLUE at only $2.8\%$ parameter overhead over LoRA.

Graph neural networks lose much of their advantage on heterophilic graphs, where connected nodes often carry different labels. Graph rewiring is a popular remedy, but rewiring methods are usually evaluated with a single classifier, which makes it hard to tell whether the gains come from the new topology or from that particular pairing. We propose an affinity-guided rewiring method that estimates the graph and the node representation together. It alternates, in the spirit of expectation maximisation, between training a lightweight graph neural network on the current graph and re-weighting candidate edges under a modularity objective with a pseudo-label homophily term. Candidate edges come from a compact pool scored by a contrastively learned node similarity and a neighbourhood-distribution affinity. The method returns two classifier-independent outputs: a rewired graph and a node embedding learned on it. Across six heterophilic benchmarks and five downstream classifiers, it improves accuracy over the original graph with normalised features in 23 of 30 classifier-dataset combinations, with a mean gain of 5.8 points, and reduces the accuracy spread between classifiers about fourfold. A controlled ablation shows that the two outputs are each useful and play complementary roles: the embedding contributes most of the accuracy gain, while the rewired graph makes different classifiers agree. A fully unsupervised variant, which uses no labels during rewiring, retains most of the improvement. The rewired graphs are also more homophilic and improve label propagation and community detection.

Large language models represent context with a key-value (KV) cache. Caches are model-specific: for the same text, models with different architectures or weights produce incompatible representations. This makes it costly to switch models over a shared context: although the context has already been processed by one model, the incoming model must process it again to build its own cache. We introduce KV-Lingo, a method for translating the KV cache of a source model into one that can be read by a target model. KV-Lingo consists of a collection of linear maps, typically one per layer of the target model, that are applied independently on all tokens' key and value representations. We train these maps using distillation, minimising the divergence between the target model's predictions from its native cache and those from the translated cache. We consider several model pairs spanning multiple sizes and architectures, training one translator per pair on a generic text corpus. The resulting translators preserve strong downstream performance in both small-to-large and large-to-small transfers. Since a switch then costs a linear map and a single decoding step instead of a prefill, replacing re-prefill with cache translation reduces the time to first token after a model switch by 9.6x already on a 64-token prompt for Qwen models on an Apple M3 Ultra, and by up to 29x at 32k context length on an H100. These gains make KV-Lingo particularly useful for dynamic model routing: a context can be processed by one model and handed off to another only when needed, without re-prefilling the shared prefix. We finally show that KV-Lingo can be used for seamless model switching, staying close to re-prefill across repeated switches in our multi-turn evaluations.

Adversarial attacks optimized on a single open-weight LLM can transfer to and jailbreak architecturally different models, allowing an attacker with white-box access to one model to compromise independently deployed systems. This creates a shared vulnerability across models, yet existing defenses are not designed for this cross-model threat. We find that cross-model transfer aligns with shared internal representation geometry, making it a natural defense target. AnchorRep targets this geometry directly with a lightweight LoRA adapter that pushes the defended model's internal representations of harmful prompts away from those of a frozen anchor model on the same prompts. Training uses a small set of harmful prompts and no adversarial examples. Across five models and four architectural families, AnchorRep reduces cross-model attack success rate to <=1.1% on 2,000 transferred attacks (0% on two), including the largest drop on Mistral (36% -> 1.1%). Existing defenses can reduce transfer, but only at high cost either inducing up to 77% degenerate benign output or increasing over-refusal by up to 18%. Because such degenerate benign outputs are not captured by standard refusal-based metrics, we introduce the Benign Garble Rate to quantify them. Our results suggest that cross-model robustness can be achieved by shaping representation geometry, without requiring attack-specific training

Symbolic regression (SR) aims to discover compact and meaningful mathematical equations from data, but searching the vast combinatorial space of symbolic structures remains challenging. Existing methods typically guide this process using expression-level objectives, such as fitting error, which assess a candidate equation as a whole but provide little information about whether an individual term contributes genuinely new structure or is largely redundant with the rest of the expression. We introduce Sobolev Novelty, a term-level measure of structural independence for symbolic equations. For each term, we construct an empirical Sobolev signature from its function values and exact derivatives over the observed inputs, and quantify how much of this behavior cannot be reconstructed by the remaining terms. We further derive a theory-calibrated threshold, yielding a principled and tuning-free criterion for identifying structurally novel terms. Using this threshold, 92.6% of terms in benchmark ground-truth equations exhibit sufficient structural novelty, compared with only 38.3% on average for expressions produced by 15 SR methods, revealing a substantial gap between scientific equations and current SR solutions. As a lightweight plug-in, Sobolev Novelty can be incorporated into diverse SR paradigms to support term pruning, search guidance, LLM feedback, and data selection, yielding consistent performance gains and demonstrating broad applicability.

Current machine learning algorithms primarily rely on instantaneous signals such as loss, margin, and prediction confidence to characterize model behavior. These signals indicate how difficult a prediction is at the current optimization step, but they do not capture how long the model has remained incorrect. We study this temporal dimension of learning and introduce Age of Learning (AoL), a learning-state variable that measures the persistence of prediction errors over time. AoL increases while an error remains unresolved and resets when a correct prediction is achieved, thereby distinguishing persistent under-learning from transient mistakes. We develop AoL-based training strategies for both offline and streaming settings. In offline learning, sample-level AoL is accumulated over training and aggregated into class-level states that guide adaptive reweighting and resampling. In streaming learning, where full historical access is unavailable, we maintain lightweight class-level AoL states using current and buffered observations. Across long-tailed classification settings, AoL improves or matches standard training baselines, with larger benefits when learning difficulty persists over time. Multi-seed streaming experiments further show reproducible gains under temporally stable imbalance. Analysis of class frequency, loss, and margin shows that AoL is related to conventional difficulty measures but captures additional information about error duration. These results suggest that temporal persistence provides a useful complementary signal for characterizing and controlling learning dynamics in imbalanced and non-stationary environments.

Data-driven model predictive control (MPC) combines learned world models with online trajectory optimization, achieving strong performance in continuous control. However, the per-step cost of sampling and evaluating hundreds of candidate trajectories restricts deployment to control frequencies well below what real-time robotics demands. Motivated by the dual-process theory of human cognition, which distinguishes between fast, intuitive processing (System 1) and slower, deliberative reasoning (System 2), we ask whether every decision requires the same degree of computational deliberation. We propose Fast-TD-MPC, a lightweight framework that adaptively routes between fast policy execution and test-time planning, reserving costly deliberation for states where it is most needed. Fast-TD-MPC delivers competitive task performance across 103 continuous control tasks while achieving up to ~4x faster inference. Under external disturbances, Fast-TD-MPC selectively falls back to planning, maintaining robustness comparable to the original planner.

Mixture-of-Experts (MoE) architectures have become a standard way to scale model capacity while keeping computation sparse, yet routing remains a key determinant of MoE quality and training behavior. Prior empirical studies suggest that MoE routing reflects input semantics and upstream computation across depth, but standard routers do not explicitly use the routing distributions produced by preceding layers. We propose HERO-MoE, Historical Expert ROuting with Scale-Preserving Fusion, a routing framework that injects historical routing priors into MoE routers by reusing detached, dense routing distributions collected from preceding MoE layers. The key idea is simple: HERO-MoE preserves the original token-conditioned routing branch and adds a residual historical routing contribution before the standard softmax and top-$k$ dispatch. To stabilize this historical signal, HERO-MoE introduces a scale-preserving fusion mechanism that matches the magnitude of historical routing memory to the current hidden representation and accounts for the number of visible historical layers, without introducing an auxiliary routing loss or a fusion-specific tuning parameter. By reusing routing distributions already computed by preceding MoE layers, HERO-MoE improves training-loss reduction with modest end-to-end overhead. The resulting router remains compatible with standard sparse dispatch, including top-$k$ and group-limited routing, and can be inserted into existing MoE backbones with minimal architectural changes. Experiments on an approximately 8B-parameter MoE model with 0.5B active parameters, trained from scratch on 100B tokens, show that HERO-MoE reduces the final loss from 1.6393 to 1.6184, while peak memory and FLOPs increase by only 0.44% and 0.64%, respectively.

A linear projection can reduce the dimension of query and key vectors without updating the pretrained model, but it remains unclear which training objective best preserves model behavior. We ask whether preferences over keys and attention mass on the highest-weighted keys provide a better signal than matching the full attention distribution with KL divergence, especially in long-context settings. We introduce DimPO, which combines listwise preference optimization with a lightweight top-k cross-entropy term for head-fidelity. DimPO is trained offline from the attention patterns of a frozen language model, with one map per layer, shared by the query and the keys and trained separately from the other layers. Across LLaMA3.2-3B, LLaMA3.1-8B, Qwen2.5-7B, and Qwen3-4B Instruct models, pairwise preference objectives outperform the triplet baseline and retain 98% of the original score on short-context tasks when projecting to half the dimension on the last 40% of the layers. With more projected layers or on long-context RULER, they degrade rapidly. In contrast, KL and DimPO, which use every key during training, retain about 95% of the original RULER 4k score on the 8B model when projecting up to 50% of the layers. KL-based projections remain closer to the original attention distribution and attention output, yet DimPO achieves better downstream performance. Beyond 50% of projected layers, DimPO increasingly outperforms KL on tasks including SQuAD, common-word extraction, frequent-word extraction, and variable tracking. These results suggest that under dimensionality reduction, preserving the ordering and concentration of task-relevant attention can matter more than reproducing the full attention distribution.

Reinforcement learning (RL) for code agents often uses executable tests to provide binary rewards. With these rewards, Group Relative Policy Optimization (GRPO) assigns identical advantages to test-passing trajectories within each rollout group, overlooking differences in implementation quality and adherence to task requirements. This leaves the policy without a learning signal that favors clean, targeted implementations over those containing unnecessary or out-of-scope changes. We introduce GAGAR, a framework for quality-aware credit redistribution in code agent RL. Built on dynamic sampling that retains groups containing both passing and failing trajectories, GAGAR places all trajectories from each group in a shared workspace, where an SFT-trained agentic grader jointly inspects them and ranks the test-passing candidates. Based on this ranking, we downweight lower-ranked trajectories and proportionally rescale the advantages of all test-passing trajectories to restore their original sum. This sum-preserving redistribution retains the relative weights established by quality-based downweighting while shifting credit toward higher-quality implementations. We evaluate GAGAR at industrial scale using pre-RL SFT checkpoints of MiMo-V2.6-Flash (310B total parameters) and MiMo-V2.6-Pro (1.02T total parameters). Controlled code-only Flash experiments show improved code agent performance, reduced trajectory-length growth, and more stable training. We further apply GAGAR in large-scale mixed-task RL with both Flash and Pro. Our results support combining test-based verification with groupwise agentic grading to improve the quality and stability of code agent RL.

On-policy distillation (OPD) is a promising approach to language-model adaptation, aligning teacher supervision with the student's own generated trajectories. When adaptation prompts are distributed across clients, can this process benefit from federated collaboration? We study federated OPD and find that substantial collaboration gains can be obscured by learning-rate sensitivity: FedAvg can perform no better than independent local training at a small learning rate, yet recover a clear advantage at a larger rate. We explain this phenomenon through the student's dual role as learner and generator of future training data. An aggregation-induced optimization lag can delay access to useful teacher supervision, which in turn slows subsequent learning. Our theory establishes this aggregation–rollout feedback in a solvable model with a common optimum and stable updates, and identifies two coupled roles of learning rate: learning from current supervision and reaching future supervision. Guided by this analysis, we propose FedTOPS (Federated Teacher-guided On-Policy Scaling), which reuses teacher feedback on current trajectories to adapt the FedAvg update magnitude under clientwise predictive-change constraints. Across six mathematical reasoning benchmarks, FedTOPS improves macro Avg@8 over FedAvg by 4.56–14.57 percentage points across the evaluated student models and local learning rates.

Clustering algorithms that operate on nearest-neighbor graphs, such as FINCH (First Integer Neighbor Clustering Hierarchy), depend heavily on the quality of the embedding space they are given. However, pretrained vision and language model embeddings are not optimized for this purpose. We propose CAESAR, a method that reorganizes a pretrained embedding space: a reorganization network is trained to pull mutual nearest neighbors together and push non-neighbors apart, yielding a reorganized embedding space substantially better suited to clustering. CAESAR offers a second major advantage: it never requires the number of clusters $K$. This matters because in realistic unsupervised settings, $K$ is typically unknown and discovering it is often part of the problem, yet most strong clustering methods take it as an input. We therefore design the entire CAESAR pipeline to infer $K$ rather than assume it is known. Empirically, reorganizing the embeddings consistently improves clustering over the raw space on both text and image datasets. Since the few deep clustering methods that also infer $K$ do not release their code, we complement controlled comparisons with methods that infer $K$ on the same embedding space by comparisons with strong deep clustering methods that are given the true $K$, giving them a substantial oracle advantage. Even so, CAESAR outperforms all of them on text, achieves the best results on the most challenging image benchmark and remains competitive on the others. Reorganizing pretrained embeddings thus emerges as a simple and powerful route to clustering realistic data, where classes overlap and the number of clusters is unknown.

Intelligence is defined in many ways. One of these definitions defines intelligence as the pursuit of learnable novelty. However, learnable novelty can be meaningless without the ability to compose the learned structures to take action and achieve goals. Learnable novelty builds on epiplexity, which is a way to measure learnable structure in data through a bounded observer. In this paper, we investigate a closed-form spectral approximation to compute epiplexity. We use a fixed-trace constraint and find that the epiplexity objective prefers a more uniform distribution of spectral mass rather than concentrating it in a small number of directions. However, a representation may spread information across many directions without organizing that information into features useful for a particular task. To address this gap, we propose a compositional objective whose observer measures the relationships between the parts and interactions of an image. We compare it with the original spectral objective given only the masked parts. In our ImageNet training runs, the spectral objective with masked parts produces an almost maximally spread representation while achieving the strongest frozen-feature classification performance on most evaluations, more than doubling the linear-probe accuracy of the whole-image baseline. Across multiple image benchmarks, changing what the observer sees matters more than adding relation and interaction tokens. At the same time, our prediction-oriented compositional objective produces substantially better held-out observer prediction but relatively weaker classification, revealing that spectral diversity, predictability, and downstream utility are distinct properties. These results suggest that the usefulness of spectral spreading depends not only on how much structure is preserved, but on which relationships the observer makes available to the objective.

Based on an analysis of the role of language in thought and communication, this article proposes a new concept for designing corporate knowledge bases. The concept integrates the probabilistic vector space of a corporate vocabulary, reflect-ing industry specifics, subject focus, terminology, and culture, with traditional ontological modeling. This combination enables the efficient extraction of knowledge from accumulated corporate documents while strictly accounting for specific business processes. Consequently, this concept bridges statistical and semantic (cause-and-effect) methodologies. Furthermore, analyzing the projec-tions of probabilistic spaces and causal relationships can help identify bottlenecks in business logic. As a dynamic system, language functions as a separate, inde-pendent layer within the overall information architecture. Introducing a dynamic component into the probabilistic space of word distribution allows it to be mod-eled as a multidimensional solution space for various problem formulations. In this context, input data defining the problem conditions serve as control parame-ters for dynamic transformations.

Understanding the safety risks of vision-language-action (VLA) models is essential for their deployment in the physical world. Existing safety research has mainly considered persistent perturbations that are applied continuously to observations throughout an episode. However, momentary observation corruption, in which observations are severely perturbed only briefly within an episode, remains an underexplored safety threat. To address this gap, this work investigates robustness to one-step perturbations applied at a single time step per episode. Our experiments reveal that these perturbations substantially degrade VLA performance and that their impact depends on the action chunk execution length. Based on them, we propose CARE, which dynamically selects the execution length based on consistency with the previously predicted action chunk. CARE improves robustness with low computational overhead while preserving clean performance by selecting shorter execution lengths only under perturbations.

Large language models are commonly evaluated by generating a small number of stochastic responses for each prompt in a benchmark. Because inference budgets are limited, this shallow evaluation may fail to observe low-probability but operationally important failures. A prompt that produces no failures in a small sample may therefore appear reliable despite having a nonzero latent probability of failure under repeated inference. We formulate LLM reliability evaluation as a budget-constrained discovery problem in which each prompt is associated with an unknown per-generation failure probability. We propose a budgeted discovery framework that first performs shallow evaluation across the prompt set and then uses trial-level failure outcomes together with prompt-derived representations to learn a feature-based ranking of failure propensity. The resulting scores prioritize prompts with zero observed shallow failures for deeper evaluation, concentrating the deep-evaluation budget where hidden failures are more likely to be discovered. We evaluate this approach on AIRBench and StrongREJECT across multiple model and system-prompt conditions. The central empirical test asks whether models fit without access to deep-evaluation outcomes can rank prompts with zero observed shallow failures according to their likelihood of producing failures under deeper evaluation. On AIRBench, the highest-ranked 10% of unresolved prompts achieves 2.54x hidden-failure lift for Qwen 2.5 7B and 1.87x for Gemma 3n E4B, recovering 25.4% and 18.7% of subsequently observed hidden failures, respectively, compared with 10% expected under random allocation. Semantic-neighborhood and feature-ablation analyses further show that this predictive signal can be recovered from multiple representations of prompt content and relationships.

Large language models can generate several records within one autoregressive completion, making earlier answers available as context for later records. This paper shows that such shared-completion batching can distort relationships among variables in the resulting synthetic data, using controlled tests on synthetic survey respondents. In a matched experiment on 2,000 European Social Survey profiles, generating ten rather than one respondent per request increases mean absolute error in within-country correlations by 48-58% for Qwen3.8-27B and 114-127% for Llama-3.3-70B-Instruct across three seeds, holding profiles, examples, questions and decoding parameters fixed. The distortion primarily reflects exaggerated relationship strength, while retaining substantial agreement with the human ordering of correlations. Controlled interventions establish answer history as a causal channel: re-pairing the same preceding values, with profiles and marginal distributions fixed, changes correlations among subsequently generated responses. Hiding preceding answers reduces correlation error in the tested settings but worsens marginal accuracy. Exploratory corrections across social-attitude, health and economic data likewise show that lower correlation error can coexist with worse marginal distributions and regression estimates. Request construction is therefore part of the data-generating process, and synthetic-data validity must be evaluated against the analyses the generated data are intended to support.