Papers

Source-Free Universal Domain Adaptation (SF-UniDA) extends Universal Domain Adaptation by removing access to source data at adaptation time while still handling label-set mismatches between domains. Despite growing interest in this setting for image data, no benchmark exists for time series, which are more challenging. We present the first SF-UniDA benchmark on time series. In addition, we provide the first study of pretrained foundation models as feature extractors for time series domain adaptation. In this context, we identify a critical and previously underexplored limitation of all existing SF-UniDA methods: the inference threshold for unknown-sample rejection is highly sensitive. We address this by proposing a plug-in auto-thresholding module that can be integrated into any SF-UniDA method. Experiments on three well-known time series datasets confirm the suitability of this module. They also highlight that foundation models do not systematically outperform classical backbones and that SF-UniDA tailored for time series is yet to be developed.

Lie detection probes aim to predict from a language model's internal states whether its output is truthful or dishonest. However, role-play complicates what "truth" means for an LLM: language models can adopt a wide range of personas that take very different claims to be true, including personas whose beliefs clearly contradict reality, such as a conspiracy theorist. In this work, we investigate whether lie detection probes reliably flag falsehoods generated under such an anti-factual persona or whether they instead follow the persona's beliefs. We introduce a dataset of 8,916 human-reviewed, on-policy responses from three LLMs adopting anti-factual personas. Evaluating eight probes from prior work, we find that many fail in this setting, particularly when correct and incorrect answers are evaluated under the same persona prompt. To investigate why, we construct three novel confounder datasets in which truth is anti-correlated with a potential confounding concept. Our experiments reveal that many existing probes strongly track concepts that are spuriously correlated with truth in their training data, such as instruction compliance or response likelihood. Based on these findings, we introduce a simple linear probe that achieves the strongest overall performance on both the persona and confounder stress tests. Our results suggest that current lie detection probes are far from reliable and highlight the need for training data in which truth is decorrelated from confounding concepts.

Post-training quantization (PTQ) has become a widely adopted technique for reducing the memory footprint and inference cost of large language models (LLMs). However, recent studies reveal that when applied to reasoning models, PTQ not only degrades reasoning performance but also exacerbates overthinking, leading to longer reasoning trajectories. These issues may offset the efficiency gains expected from lower-precision inference. Existing approaches mainly rely on complex optimization procedures. More recent lightweight inference strategies instead use predefined overthinking markers, limiting their adaptability across quantized models. To address these issues, we propose Reasoning Analysis and Token-level Inference Optimization (RATIO), a framework that identifies model-specific overthinking tokens and assigns each a tailored penalty. RATIO first introduces Quantization-aware Reasoning Behavior Analysis (QRBA) to identify overthinking tokens by analyzing discrepancies between full-precision and quantized models. It then adopts Token-Specific Penalty Determination (TSPD), which leverages full-precision guidance to derive token-specific penalties without additional training. Extensive experiments show that RATIO achieves a better accuracy-efficiency trade-off than existing token-level interventions. Specifically, RATIO achieves up to 9.8 points accuracy improvement and reduces chain-of-thought (CoT) length by up to 51.3% compared with quantized baselines. The code will be available at https://github.com/steven-bao1/RATIO.

We analyse allocation, admission and post-write retention in finite-horizon linear-Gaussian noisy recurrent memories. At every horizon, the directional Fisher memory $M_n$ satisfies $\operatorname{tr}M_n=N$: non-normality redistributes information but cannot raise its spherical average, while normal carriers satisfy $M_n=I$. For bi-power-bounded carriers, we derive uniform $1/n$ lag bounds, identify the limit of $M_n$ with the inverse of the classical Cesàro asymptotic limit of $W^\top$, and give finite-horizon error bounds. A time-varying coupling defines an end-to-end store operator. The writer-optimal direction need not be store-optimal. After writing ends, an invertible hold preserves the full stored Fisher matrix. Additive contamination bounded by $\alpha$ times the closure covariance retains at least $1/(1+\alpha)$ of that matrix; a covariance-aware decoder attains the corresponding accuracy. With recurrent carriers held fixed, training input masks and linear readouts approached the task-specific optimum in 160 runs, with median normalized Rayleigh efficiency above $0.998$. Binary accuracy matched the Gaussian prediction to mean absolute error below $0.002$ over more than four orders of magnitude in $J$. In a separate pre-specified study of 320 runs, trained masks followed the designated input-time objective in both carrier types, in 16 of 16 draws. These studies used development-seen carriers and are pre-specified validations, not blind holdouts. The same fixed design reproduced the objective-specific result in 16 of 16 draws on carriers unused before run commitment. Exact isolation preserved information, while a decoder fixed at its training horizon fell to chance; inverse-adjoint transport restored its sampled decisions to numerical precision.

Building on a probabilistic perspective in which adversarial examples arise from the overlap between a distance-based distribution $p_{\mathrm{dis}}$ and a victim-classifier-induced distribution $p_{\mathrm{vic}}$, we start from a simple intuition: adversarial examples become harder to generate when these two distributions are pushed apart, as their overlap becomes smaller, thereby increasing robustness. This intuition naturally motivates a KL-based robustness objective. We then prove that $\mathrm{KL}(p_{\mathrm{dis}}\|p_{\mathrm{vic}})-\log Z_{\mathrm{vic}}$ is a lower bound on probabilistic robustness (PR), where $Z_{\mathrm{vic}}$ denotes the normalizing constant of $p_{\mathrm{vic}}$. Since PR is generally intractable to compute directly, maximizing this KL-based lower bound provides a tractable surrogate objective for improving PR. We further show that this objective recovers a scaled form of adversarial training, offering a probabilistic interpretation of adversarial training and a principled route to robustness improvement. We call the resulting method probabilistic adversarial training. Experiments show that it consistently improves PR, and ablation studies demonstrate that the induced scaling factor can even enhance the PR of non-probabilistic adversarial training methods.

Distinguishing periodic from chaotic dynamics in a time series is a fundamental challenge in both physics and engineering. Yet, end-to-end learned architectures must discover both a representation and a decision boundary from data, at substantial cost. We introduce TopTimeNet, which decouples these tasks: a fixed, non-learned stage extracts a $42$-dimensional geometric and topological descriptor from Takens delay embeddings and persistent homology, and a lightweight learnable stage performs classification. On a benchmark of $49$ nonlinear dynamical systems, a $1{,}638$-parameter configuration matches the mean accuracy of one with $33\times$ more trainable parameters. Additionally, this approach delivers mean accuracy comparable to convolutional neural networks and surpasses the average performance of converged Transformer models, while requiring three to four orders of magnitude fewer trainable parameters. Robustness also depends sharply on where noise is introduced: TopTimeNet degrades gracefully under perturbations to its precomputed features, but degrades sharply when noise is introduced into the raw signal and the full feature-extraction pipeline is recomputed, showing that robustness to perturbations of the precomputed features does not imply robustness of the complete raw-signal-to-prediction pipeline. These results show that decoupling fixed geometric and topological feature construction from a lightweight discriminative stage can achieve comparable classification accuracy with substantially fewer trainable parameters.

Real-world time series forecasting systems operate under non-stationary data streams, where forecasting performance may degrade over time. Although retraining can recover the performance, it incurs non-trivial computational and operational costs. Under limited deployment resources, the key challenge is therefore not only how to retrain but also when to retrain. While existing retraining policies often rely on indirect indicators such as drift alarms or model staleness, we instead use realized forecast errors as direct deployment feedback. In this paper, we propose PILOT (Pseudo-label-Informed Learned Online Trigger), an online retraining framework that learns when to retrain from forecast-error dynamics. Since ground-truth retraining labels are unavailable, PILOT constructs a pseudo-label from future increases in forecast error and trains a lightweight scorer to predict it from observed error states. At deployment, PILOT uses only completed forecast errors and serves as a plug-in module for arbitrary forecasting backbones without architectural modification. We evaluate PILOT under standard multivariate forecasting settings across eight benchmarks with three representative backbones—DLinear, iTransformer, and TimesNet. Across all three backbones, PILOT achieves state-of-the-art average-rank performance among retraining policies while maintaining a favorable performance–efficiency trade-off.

Latent communication enables multi-agent systems to exchange information directly in internal representation space, reducing the token, computation, and latency overhead of text-based communication. To this end, lightweight trainable links are introduced to map the sender's representations into the receiver's input space. In this work, we show that even benign link training can increase harmful compliance relative to text-based communication while the underlying safety-aligned agents remain unchanged. An attacker can amplify this effect by optimizing the links on harmful query–response pairs or poisoning otherwise benign training data. We further develop a reinforcement-learning attack that rewards harmful compliance alongside benign task performance without requiring harmful target responses. Across three communication topologies and four safety benchmarks, this attack raises the mean harmful-compliance score from 27.9 with benignly trained links to 76.9. Compared with direct supervised optimization, it also achieves higher average accuracy on two benign utility benchmarks. Adapting the rewards toward safer behavior also enables repair of compromised links, substantially reducing harmful compliance across all evaluated attacks without updating the agents. Overall, our results show that safety alignment requires considering the multi-agent system as a whole. Code: https://github.com/Muhammad-Huzaifaa/latent-safety

Can heterogeneous physical degradation systems benefit from joint pretraining and move beyond system-specific prognostics toward reusable cross-system representation learning? CORD combines type-specific observation interfaces with a shared degradation backbone. Its two self-supervised objectives learn at complementary scales: Intra-Observation Structure Modeling (ISM) captures structure within observations, while Inter-Observation Dynamics Modeling (IDM) captures latent degradation evolution across observation histories. We evaluate CORD under two transfer boundaries: Pretraining-Included System Types, where downstream datasets and held-out units are unseen but their system types are represented during source pretraining, and Pretraining-Excluded System Types, where the entire turbofan-engine type is absent from pretraining. Across bearings, batteries, and cutting tools, CORD (Multi-domain) consistently improves over CORD (Single-domain) under Frozen adaptation, provides further gains under Full FT in most settings, and remains competitive with representative external baselines. Source-pretrained initialization also improves low-label adaptation to the pretraining-excluded engine type. Frozen-representation analysis further shows improved cross-unit lifecycle consistency after multi-domain pretraining. Joint pretraining across heterogeneous physical systems thus produces degradation representations reusable across devices, datasets, and system types.

LLM-based multi-agent systems coordinate specialized reasoning through aggregation, interaction, and adaptive control, yet their potential for graph learning remains unexplored. Graph learning is a natural setting for such systems because useful evidence may arise from heterogeneous local, long-range, global structural, and semantic perspectives whose relevance varies across instances. Existing LLM-based graph learning approaches primarily rely on single-agent reasoning, while multi-agent coordination has been studied mainly in general reasoning settings. Consequently, it remains unclear whether multiple specialized agents can improve graph learning and how coordination strategies should be designed and evaluated. To address this gap, we introduce GraphMAS, a systematic benchmark of multi-agent coordination for graph learning. GraphMAS builds a shared pool of graph reasoning specialists and organizes coordination along two dimensions, inter-agent interaction and runtime adaptivity, yielding four paradigms and seven representative coordination methods. Under a unified protocol, we evaluate these methods across seven text-attributed graphs, three domains, and two graph learning tasks. We find that heterogeneous graph perspectives are complementary, and that coordinating specialists improves over individual specialists and single-agent graph reasoning, with gains from decomposing reasoning across specialists rather than from broader evidence access alone. However, richer inter-agent interaction does not reliably help, whereas instance-adaptive specialist selection yields the strongest accuracy-efficiency trade-off. We further show that coordination can be learned over a fixed specialist pool and transfers to held-out graphs. GraphMAS therefore provides a controlled evaluation framework and empirical principles for understanding when and how multi-agent coordination benefits graph learning.

Crystal structure governs material properties, making crystal structure prediction (CSP) a fundamental problem in materials science. Generative models are a promising approach for solving this problem, but the prevalence of polymorphism, coupled with large unit cells and complex packing geometry, makes the molecular CSP task challenging for existing models. To address this, we introduce Coarse-Grained Open Materials Generation (CG-OMatG), an equivariant Riemannian flow-based generative model. CG-OMatG predicts molecular crystal structures via a coarse-grained, hierarchical representation. CG-OMatG treats molecules as rigid bodies—performing both inter- and intra-molecular message passing to construct a geometric representation for molecular packings—and learns to reconstruct molecule centroid positions, orientations, and lattice parameters, conditioned on chemical species and conformer geometry. We train the model on subsets of the Open Molecular Crystals (OMC25) and Cambridge Structural Database (CSD) datasets. Further, we fine-tune the model via policy gradient reinforcement learning to steer the model towards generating low-energy candidate structures. We validate the generated structures on the CSP blind test benchmark, assessing agreement with experimentally determined crystals using COMPACK packing-similarity analysis. CG-OMatG exhibits strong performance for generative molecular crystal structure prediction, paving the way for accelerated polymorph screening and organic solid-state materials discovery.

Certifying a deployed neural network raises decision problems that the verification literature has not classified: whether the model carries a backdoor planted in its training data, whether a fault in its stored parameters can drive it into an unsafe state, whether its output leaks a private part of its input. We formalise eight such problems and classify what we can. The organising observation is a logical one. The function computed by a piecewise linear network, together with all its node values, is definable by a quantifier-free formula of real addition of size linear in the network, so a property of the network is a quantifier-alternation sentence, which Sontag's 1985 theorem places in the polynomial hierarchy at the level of its prefix. Membership results are thus corollaries, and the argument makes plain what they need: that the quantified objects are inputs rather than the network's own parameters. Non-interference, monotonicity and counterfactual fairness have exactly the complexity of network equivalence and of interval verification, all co-NP- complete over ReLU. Detection of backdoor triggers from a quantised alphabet is Sigma_2^P-complete, one level above robustness certification, so it does not reduce to polynomially many robustness queries unless the hierarchy collapses. Inversion resistance is co-NP-complete for every l_p metric, p a fixed positive integer. Quantifying over parameters instead of inputs - the fault model of bit-flip attacks, radiation upsets and analog accelerators - makes verification exists-R-complete already for networks of identity nodes, for which every previously studied problem is in P, and it stays so when each parameter is confined to a box of inverse-polynomial width; the corresponding safety question is forall-R-complete for ReLU.

The scale and expense of pre-training language models make efficient hyperparameter tuning essential, yet a principled guidance is still missing. In this work, we analyze language model pre-training dynamics from a local landscape geometry perspective. Our study reveals two distinct phases. In Phase I, sharpness of the local landscape is initially high, leading to instability and loss plateaus under large learning rates (LRs). The landscape shifts from sharp to flatter regions early in training. This dynamic explains the necessity of LR warmup and further suggests that larger peak LRs require proportionally longer warmup periods. In Phase II, the local landscape is governed by the gradient noise scale. Our theory identifies a depth flatness trade-off: high noise from smaller batches widens the loss basin, whereas reduced noise from larger batches deepens it. This theory motivates a dynamic batch-size (BS) scheduler that begins with a small BS and increases it late in training. Together, we provide a unified view of loss landscape evolution, which translates into actionable tuning strategies for large-scale pre-training.

Black-box distillation is a practical route for transferring capabilities from API-accessible large language models that expose only text outputs into smaller student models. Recent on-policy adversarial methods such as GAD improve over SeqKD by forming an adversarial loop between a critic and a student, where the critic provides rewards for GRPO-based student policy optimization over the student's sampled responses. However, GRPO computes advantages from the within-group relative rewards of student samples for the same prompt, whereas the critic is trained primarily to distinguish teacher responses from student responses. This objective mismatch can produce reward groups with collapsed scale or fragile margins, leading to brittle grouped optimization signals. We propose Groupwise Reward Geometry Conditioning (GRGC), a two-stage framework that improves advantage construction by shaping student-side reward groups during both critic training and policy optimization. To improve critic-side conditioning, Gaussian groupwise Optimal Transport calibration regularizes the critic during training to produce reward groups with non-collapsed spread and smooth rank-wise gaps by matching sorted prompt-wise rewards to group-centered Gaussian quantiles. Building on this conditioned reward geometry, policy-side group power modulation reshapes the prompt-wise reward groups before they are converted into advantages, preserving the critic-induced ordering while increasing optimization-relevant margin separability. Extensive experiments across diverse teachers, student model families and scales, and training datasets demonstrate the effectiveness of GRGC on both in-distribution and out-of-distribution evaluations, while introducing negligible overhead over GAD. The code is available at https://github.com/2018cx/GRGC.

Routing and switch placement are fundamental combinatorial optimization problems in chip design, requiring the joint optimization of routing topology and physical placement under strict structural, geometric and logical constraints. Existing approaches typically rely on carefully engineered heuristics that incorporate strong problem-specific biases to navigate the enormous space of possible designs. In this work, we introduce a hierarchical reinforcement learning framework for joint routing and switch placement at the level of logical communication routes. Starting from a minimal routing graph, our method progressively constructs increasingly expressive solutions through three coupled operations: switch expansion, switch placement, and route refinement. These operations preserve routing validity by construction, restricting exploration to feasible configurations where every communicating initiator-target pair has one assigned loop-free route. We explore the induced solution space using Gumbel Monte Carlo Tree Search, showing that neural-guided search substantially improves solution quality over non-learning optimization methods. Furthermore, pretraining across floorplans provides a strong initialization for fine-tuning on unseen instances.

Large forecasting applications often combine statistical, machine-learning, and neural models. These families solve the same problem but differ in fitted state, training procedures, and how they parallelize work. Forecasting software must therefore either hide these differences behind a single estimator interface, or keep the families in separate packages, forcing users to rewrite data preparation and evaluation for every package. We present the Nixtlaverse, an ecosystem of open-source Python libraries for time series forecasting, as a case study of a third design: all libraries share the same long-format panel data and keyed forecast outputs, while every model family keeps its own specialized implementation. We demonstrate this design through three use cases on the public M5 competition data. First, we evaluate statistical, machine-learning, and neural models, and an external engine from a separate ecosystem, in a single rolling-origin evaluation with per-series and hierarchy-weighted metrics. Second, we profile runtime and peak memory from 100 to 30,490 series and locate each family's bottleneck: statistical fitting scales approximately linearly in the number of series, feature construction dominates machine-learning memory, and neural training time is nearly independent of panel size under a fixed training budget. Third, we reconcile the forecasts of multiple engines, including the external one, over all 42,840 series of the M5 hierarchy, with sparse reconciliation where dense implementations exhausted memory. These use cases establish the costs, boundaries, and utility of shared data and output contracts. The Nixtlaverse has seen substantial public distribution, scholarly reuse, and adoption through other forecasting frameworks, and is released under permissive open-source licenses with public datasets, reproducible examples, and verifiable benchmark artifacts.

Graph generative models increasingly rely on Graph Transformers (GT) to capture complex dependencies among nodes and edges. While deeper architectures should provide greater expressive capacity and a broader receptive field, their effectiveness can decline with depth: repeated self-attention progressively contracts node representations, impeding information flow and gradient propagation. We analyse this phenomenon from a dynamical systems perspective, focusing on how the denoiser's spectral dynamics affect graph generation. We show that standard GT denoisers become increasingly dissipative as depth grows, leading to vanishing gradients and representation collapse. To isolate the effect of these dynamics, we construct a permutation-equivariant GT with inherently stable, non-dissipative transport. We also introduce a damping mechanism that continuously interpolates between non-dissipative and increasingly contractive regimes, enabling a direct assessment of how dissipation influences generation. Experiments on synthetic and molecular graph generation benchmarks show that the gap between these regimes widens with depth: non-dissipative dynamics preserve representation diversity and gradient flow, sustaining strong generative performance, whereas greater contraction progressively impairs it. These findings identify the denoiser's dynamical regime as a key design factor for deep graph generative models.

Learning to generate machining process plans and toolpaths from B-rep CAD requires coupling discrete operation decisions with continuous tool motion as the workpiece evolves. Correctly predicting an operation sequence does not by itself ensure correct material removal, because each toolpath acts on the stock left by preceding cuts. We formulate this problem around persistent manufacturing objects: object identity determines the target of an operation, while the evolving stock state conditions the generation of its toolpath. Based on this formulation, we propose CNCGEN, a dataset and learning framework for three-axis machining. CNCGEN-Dataset contains approximately 50k geometrically verified synthetic machining flows and 800 held-out real CNC records. Each flow aligns B-rep geometry with object-referenced operations, parameterized toolpaths, intermediate stock states, and verification outcomes, enabling supervision of the correspondence between planning decisions and their geometric effects. CNCGEN generates operations and toolpaths for selected objects step by step, updating a compact machining state to guide subsequent predictions. During training, a learned surrogate verifier provides material-removal feedback that links local predictions to their geometric consequences. Experiments on synthetic and held-out real CNC records show that CNCGEN improves the resulting workpiece geometry and reduces residual material and overcut compared with adapted CNC generation baselines.

The two-dimensional sphere embedded in three-dimensional Euclidean space S2, plays a central role in a variety of scientific and engineering domains, including geophysics, planetary science, geodesy, atmospheric physics, quantum chemistry, cosmology, and virtual reality, among many others. As machine learning increasingly permeates these fields, the demand grows for robust tools that process and model functions on the sphere, while respecting the inherent topological and symmetry properties of the domain. We present torch-harmonics, a comprehensive library that offers efficient, differentiable implementations of advanced signal processing and machine learning (ML) methods for spherical data. These include the spherical harmonic transform (SHT), the spherical analogue of the Fourier transform, vector spherical harmonics, discrete-continuous and spectral convolutions, as well as both global and neighborhood spherical attention mechanisms. Beyond traditional representations, torch-harmonics provides the building blocks for state-of-the-art spherical ML architectures such as spherical transformers in order to enable scalable, rotationally-aware learning and inference in modern scientific and engineering applications.

Adapting a language model to a specialized corpus means choosing which instruction-tuning tasks to train on under a fixed budget, and testing one choice costs a fine-tuning run. Common heuristics add more source tasks or pick sources similar to the target. The first assumes transfer is never negative; the second, that it is symmetric. We show that both assumptions fail: task $A$ can help task $B$ while $B$ hurts $A$, so helpfulness is a signed property of ordered source–target pairs. We introduce the transfer map, a signed estimate of how much each source helps or hurts each held-out target. We fit the map in hundreds of fine-tuning runs on Qwen3 and Mistral models from 0.6B to 32B parameters, with all sources drawn from one corpus and no training examples from the target. The map predicts a held-out target's accuracy on unseen mixtures: recorded before those runs, its predictions have less than half the error of a mixture-agnostic baseline. The map is specific to its target and corpus but transfers across model scale: a mixture selected in advance at one size beats training on all source tasks at every other size we tested. Transfer is thus a property of the data. The map selects the tasks that help and drops the one that interferes: accuracy on the reasoning targets (causal explanation, multi-hop questions and methodological critique) rises by up to 14 percentage points over training on all source tasks.

Value steering should change an LLM's normative priorities while preserving the scenario, facts, and task constraints underlying its answer. Conventional activation edits often change both. We introduce an editable semantic-value interface on frozen residual states, with a one-way semantic-to-value pathway that grounds value recognition in context. Stop-gradient blocks feedback through this pathway; swap consistency, topic de-confounding, and decorrelation encourage selective codes. At inference, editing the value code produces a residual delta while holding the semantic code fixed. On two instruction-tuned backbones, this interface improves semantic preservation and reduces benign refusals at comparable value alignment. A matched mixing-by-gating ablation separates representation learning from selective edit activation, and dimension-matched probes establish improved code selectivity. Against validation-selected prompting on LLaMA-3.1-8B, the method achieves comparable alignment (0.750 vs. 0.748), higher BERTScore (0.938 vs. 0.923), and fewer contradictions (5.1% vs. 7.6%). Human ratings and cross-taxonomy controls provide complementary evidence for low-damage value steering.

On-policy distillation (OPD) has demonstrated two important capabilities in language models: compressing large teachers into smaller students and merging expert models into a single model. Existing diffusion OPD, however, mostly focus on the latter, with teachers and students sharing the same backbone and scale. We investigate large-to-small diffusion opd from large teachers to a small student and find that the standard recipe fails. To find the underlying cause, we propose Fixed-State KL, an effective and fair way to measure the distribution gap between student and teacher during OPD training for diffusion models. We are the first to clarify why large-to-small OPD is challenging for diffusion models: a smaller student struggles to perfectly match the distribution of a larger teacher, while classifier-free guidance can accumulate and amplify the distributional discrepancies between the student's conditional and unconditional branches and those of the teacher. To solve this problem, we propose GFD-OPD, a simple yet effective method that reduces the student-teacher gap while avoiding the error amplification of the CFG composition. Across numerous experiments, GFD outperforms previous baselines in both training efficiency and final performance, achieving state-of-the-art results on all benchmarks.

On-policy self-distillation (OPSD) trains mathematical reasoning models using a privileged teacher that sees a reference solution and supervises student-sampled prefixes. Standard OPSD uses one fixed parameter setting at every state, but nearby settings may offer additional supervision. We find that local parameter perturbations reveal complementary reference-aligned corrections under the same reference context. Different experts supply these corrections at different reference positions. Their pool covers more such positions than the unperturbed privileged teacher. We introduce Neighborhood OPSD (N-OPSD) to turn these corrections into supervision at student-visited states. Offline, greedy selection builds a compact pool of frozen experts by rewarding filtered reference-token gains beyond the pool's current best at each position. The highest-peak expert need not provide the best training target. Online routing therefore separates the anchor direction from its level of support. MaxPeak selects the anchor token, and quantile selection chooses among experts whose top token matches it. The student learns from the chosen expert's full next-token distribution through the clipped forward-KL objective inherited from OPSD. We evaluate on AIME 2024, AIME 2025, and HMMT February 2025. Across three independent runs per method, Neighborhood OPSD improves the three-benchmark Average@12 over OPSD by 2.75, 1.67, and 1.94 points on Qwen3-1.7B, 4B, and 8B, respectively. Student-prefix continuations support using the pool beyond the reference trajectories used for selection. Matched ablations support filtered reference-token gains as a selection criterion. Accounting for overlap within the pool and routing by state further improve student accuracy. Inference uses only the distilled student.

Privacy has become an increasingly important concern in the Document Analysis community, to the extent that in many environments such as archives, governmental institutions, and local businesses, the adoption of automation is restricted by legal and policy constraints. While federated learning has often been regarded as a “necessary evil”, implying an unavoidable performance trade-off in exchange for decentralization and privacy, many prior works overlook its potential to improve robustness to out-of-distribution data. In this paper, we present Unapologetically Distributed, the first comprehensive study evaluating distributed learning in Document Analysis along three key axes simultaneously: the tasks addressed, the architectures employed, and the fine-tuning strategies applied. Specifically, we demonstrate how various distributed training approaches enhance generalization capabilities across diverse tasks such as Table Recognition, handwriting recognition, and Word Spotting, particularly during transfer learning stages. Our results provide strong evidence that decentralization is not merely a constraint, but a valuable opportunity to improve model robustness and adaptability in real-world Document Analysis scenarios.

Audio deepfake detection (ADD) must remain effective when new spoofing attacks emerge after deployment. Emerging audio language model (ALM)-based ADD methods are built on predefined supervision from ground-truth labels or verified forensic rationales. However, this paradigm overlooks an ALM's own mistakes, which indicate where targeted supervision is most needed. To this end, we first introduce evolving spoofing environments for ALM-based ADD, where a new attack becomes dominant while previously observed attacks persist. Motivated by the above learning-from-mistakes perspective, we further propose SE-ADD, a self-evolving framework that iteratively adapts an ALM via low-rank adaptation (LoRA) using mistake-driven supervision built from its verdicts and self-generated forensic cues. All training samples receive direct authenticity supervision, while misclassified ones receive additional cue-augmented supervision. As verdicts and cues are regenerated by the updated ALM, the resulting supervision evolves accordingly. Experiments on two ALMs demonstrate the effectiveness of SE-ADD in generalizing to unseen attacks, reducing the equal error rate (EER) from $36.72\%$ to $7.52\%$ for Qwen2-Audio and from $19.93\%$ to $3.97\%$ for MOSS-Audio.

Can a pretrained graph model replace training and tuning a separate predictor for each dataset? Answering this requires evaluating prediction quality alongside computational cost. We present NodeGround, a node classification benchmark that puts graph foundation models (GFMs) and dataset-specific supervised learning under a common evaluation framework. The benchmark spans 51 datasets and evaluates six GFMs alongside 15 supervised methods under two label-availability regimes. Shared data partitions, validation-only model selection, controlled hyperparameter searches, and multiple predictive metrics make comparisons systematic, while workflow measurements account for adaptation, training, tuning, and inference. The results favor carefully tuned graph neural networks overall. GraphPFN reaches third place by Elo when more labels are available, yet its relative strengths vary substantially with dataset properties. Efficiency comparisons further qualify the benefits of pretrained reuse: GVT and GraphPFN appear on the Pareto frontiers when supervised methods are represented by their default and fully tuned configurations. Adding intermediate tuning budgets removes this advantage for GVT and leaves GraphPFN extending the estimated frontier in the label-rich setting alone. Thus, reusing pretrained parameters does not yet provide a broadly reliable route to either stronger predictions or cheaper workflows. We release the evaluation pipeline, run-level records, and an open leaderboard at https://github.com/nums-ai/nodeground.

Diffusion models generate data by reversing a forward corruption process that typically approaches a simple Gaussian prior. Recent work has extended this framework to signals supported on fixed graphs, e.g., road-network traffic and sensor-network measurements. Many graph signals have nonuniform spectral energy, whereas isotropic corruption adds the same conditional noise variance to every graph-frequency mode. Driving all modes to near-zero terminal signal-to-noise ratio (SNR) requires strong corruption, which increases the noise range that must be covered under a fixed sampling budget. We introduce Graph Residual Conjugate Diffusion (GRCD), which replaces the shared clock of graph heat diffusion with a mode-dependent clock that gives every graph-Fourier mode the same conditional SNR. GRCD fits a zero-mean graph-spectral Gaussian reference on the training split and stops at a finite terminal SNR at which the propagated reference still carries the fitted spectral variances. The Gaussian component has an exact modewise propagator in the probability-flow ODE, so sampling advances it analytically and integrates only the learned residual score numerically. We evaluate GRCD on five settings (METR-LA traffic, Molene weather, and three stochastic block models) against seven comparators under a matched protocol: Graph-Aware Diffusion (GAD), EDM (graph backbone), two adaptations of Whitened Score Diffusion (WSD), and three preconditioning controls. At four function evaluations (NFEs), GRCD lowers averaged maximum mean discrepancy (aMMD) by 22 to 36 times over the best comparator on all five settings, reaching 0.054 on METR-LA, where it clears an aMMD 0.1 target with 87% less sampling wall-clock time than the cheapest comparator that reaches it. Fitting the terminal reference reduces aMMD by 2.7 to 7.3 times at finite terminal SNR, while the factors shrink to 1.00 to 1.01 near zero.

Diffusion models are typically viewed as stochastic processes that transform noise into data. We take a complementary perspective: a diffusion model defines a family of deterministic dynamical systems indexed by noise scale. At each fixed scale $\sigma$, we treat the denoiser as a self-map and study its dynamics. For an exact denoiser, fixed points correspond to critical points of the smoothed data density, while attractors correspond to its modes; as $\sigma$ increases, sample-level modes merge into progressively coarser ones. This suggests a geometric view of memorization: examples that receive excess probability mass due to duplication or overfitting, as well as outliers, should remain distinguishable under stronger smoothing than ordinary examples. We quantify this persistence by the critical scale $\sigma_c$, the largest noise scale at which an example is retained by the fixed-scale dynamics. In conditional models, the same construction extends naturally to image–caption pairs. Experiments in controlled settings and on large-scale models show that $\sigma_c$ tracks memorization arising from duplication, overfitting, and outliers, and identifies both memorized and partially memorized examples in Stable Diffusion. Moreover, $\sigma_c$ yields interpretable measures of the image spatial distribution and caption dependence of memorization.

Federated learning faces severe communication bottlenecks when clients upload high-dimensional model updates. Existing methods often compress these updates uniformly across all layers. This uniform approach ignores the heterogeneous value of different parameter blocks and wastes limited bandwidth on insensitive layers. To address this issue, we propose Layer-wise Budgeted Adaptive Transmission (LBAT). LBAT reframes federated communication under extreme uplink budgets as a resource allocation problem. Our framework dynamically estimates the transmission value of different layers utilising local training signals. It then employs an exact byte dynamic programming allocator to determine optimal rank and bit configurations under strict budgets. We validate LBAT on highly heterogeneous federated tabular prediction and data generation tasks. Extensive experiments demonstrate that LBAT consistently outperforms uniform rank, uniform quantisation, and fixed compression baselines across various extreme budget regimes. Furthermore, it achieves significantly better communication and utility tradeoffs while preserving essential distributional fidelity.

Multi-agent collaboration lets large language models (LLMs) improve question answering through deliberation and feedback. Yet shared discussion couples correction with exposure to the same mistakes, which can erode the diversity needed for voting. Self-consistency offers sampling diversity without feedback, while single-pair Actor-Critic collaboration refines only one candidate. We introduce SEPAL, which assigns three private Actor-Critic teams to direct reasoning, evidence grounding, and verification. Role-specific training gives the teams different reasoning objectives beyond sampling variation. Each Critic guides revisions within its own team, preventing feedback from carrying errors across candidates. Once revision ends, majority voting combines only the final answers, keeping the reasoning histories separate until the decision. Across five open-weight backbones and five question-answering benchmarks, SEPAL improves mean accuracy by 1.81 percentage points over a matched single Actor-Critic pair, with improvements across all five backbones. Code is available at https://github.com/zhansan114514/SEPAL.