Large language models (LLMs) are increasingly recognized for their remarkable capabilities; federated learning (FL) offers a promising framework for collaboratively fine-tuning LLMs on edge devices while achieving data privacy. However, edge devices are inherently heterogeneous: clients may exhibit diverse data distributions and resource constraints, including limited memory and computation budgets. Recent federated Mixture-of-Experts (MoE) approaches address these challenges by using sparse expert activation or partial expert loading. However, existing partial-expert methods rely on heuristic full-pool observations or prioritize system-wide load balance, without guarantees for recovering client-specific relevant experts from partial observations. We propose $S^3$-MoE, a server-side sparse expert assignment framework that estimates client-specific expert relevance from partial observations and refines assignments through an exploit--explore mechanism. We also establish relevance-estimation error and exact Top-$K$ recovery guarantees. Experiments using a CNN-based MoE on CIFAR-100 show that $S^3$-MoE recovers task-dependent client--expert preferences and reduces conflicts caused by correlated expert selection across clients. Compared with Full-20 FedAvg, which gives each client access to all 20 experts for local Top-5 routing, $S^3$-MoE achieves 1.31 percentage points higher mean final accuracy while loading only $5$ experts per client.
Symbolic models offer a powerful account of cognition, but their scope has historically been limited by the need for hand-engineering and computational tractability. We show how advances in artificial intelligence (AI) can scale symbolic modeling and help overcome these engineering limitations. We introduce Behavior-Conditioned Symbolic Synthesis (BCSS)---an approach for automated search of interpretable, deterministic symbolic programs that verifiably reproduce behavior---and apply it to model both human and machine behavior across hundreds of challenging reasoning problems. Analyzing these programs reveals that both people and machines tend to oversimplify, adopt rules that are partial explanations for their observations, and misinfer dependencies between problem features. These shared signatures are modulated by available resources, with reasoning effort predicting the extent to which machines simplify and explain their observations. Together, these results point to underlying mechanisms reminiscent of satisficing in both people and machines, and illustrate how BCSS offers a general approach for interpreting and comparing rule-based reasoning of humans and black-box AI systems.
Multi-perspective-camera systems are widely adopted since they provide broad visual coverage for autonomous driving, motivating the synthesis of a unified wide view from spatially separated observations. In this work, we formulate wide view synthesis (WVS), which aims to synthesize a geometrically consistent wide view at a specified virtual viewpoint from multiple perspective cameras. Unlike conventional image stitching, WVS requires reasoning about 3D geometry and visibility to handle depth-dependent parallax and occlusions arising from spatially separated camera centers. While vanilla 3D Gaussian Splatting (3DGS) enables high-fidelity view synthesis, its reliance on per-scene optimization limits its applicability to driving scenes. Furthermore, WVS introduces additional challenges, *e.g.*, peripheral stretching under perspective projection and photometric inconsistencies across cameras. To address these challenges, we propose **Wide-View Gaussian Splatting (WV-GS)**, a feed-forward Gaussian-splatting framework that synthesizes wide views without per-scene optimization. Specifically, we introduce Piecewise Perspective Rendering (PPR) to mitigate peripheral stretching through local perspective projections sharing a common camera center, and Spatial Affine Photometric Modeling (SAPM) to account for camera-dependent photometric variations during training. We further introduce **WideDrive**, a synthetic benchmark providing dense ground-truth wide views and adapt a LiDAR-projected sparse-reference protocol to real-world datasets. Extensive experiments on multiple datasets demonstrate the effectiveness of WV-GS for WVS. Dataset and code will be publicly released.
Optimistic Hedge, a multiplicative weight update with one step of extrapolation, is one of the most natural regret-minimizing dynamics for games, and characterizing its regret has been a longstanding open question. Daskalakis, Fishelson and Golowich proved that if all $m$ players in a finite general-sum game follow Optimistic Hedge with a fixed learning rate, each player's regret is polylogarithmic in the horizon $T$. Subsequently, Ha showed that an appropriate choice of learning rate makes each player's regret bounded independently of $T$. This bound, however, is non-constructive and its dependence on the number of actions $n$ is not explicit, leaving open whether Optimistic Hedge can achieve a constant regret bound which is polynomial in the size of the game. We show that the answer hinges on the number of players. In two-player games, there is a learning rate for which Optimistic Hedge guarantees $O(n\log n)$ regret independent of the horizon. In three-player games, however, no such bound is possible: for any learning rate up to an absolute constant and for sufficiently large horizon $T$, there is a game with $n$ actions per player in which some player's regret is $2^{\Omega(n)}$, and more generally, for every $m\ge3$ we construct $m$-player games in which some player's regret is $2^{\Omega(mn)}$; for learning rates above $2^{-\Omega(mn)}$, that player's regret even grows linearly in $T$. Our upper bound is based on a temporal approximation framework for the analysis of Optimistic Hedge. Starting from the variance-based RVU bound of Daskalakis et al., we approximate the losses over time-localized windows by well-behaved functions, whose total variation is a small fraction of their cumulative variance. In $m$-player games, the same framework gives a significantly simplified proof of the recent $O(\sqrt{m}\log n\log T)$ regret bound due to Ha.
Automated feedback on classroom discourse improves instruction in professional and online teaching contexts, and recent work suggests that delivering similar feedback to learners may strengthen these effects by sharing responsibility for participation. Whether this holds in small-group volunteer tutoring has not been tested. We present results from a randomized controlled trial across two four-week SAT mathematics bootcamps on the Schoolhouse.world platform, involving 1,030 volunteer tutors, 7,443 learners and 6,508 recorded sessions. Tutors were randomly assigned to receive no feedback, automated feedback on their own discourse, or that feedback together with participation feedback delivered to their learners under one of two motivational framings. Feedback to tutors increased eliciting of learner ideas, uptake of learner contributions, and learner reasoning by 8-15%, and the pattern replicated across both bootcamps. However, tutor talk time, which averaged 91.5% in the control group, did not change in any condition, and learner-facing feedback did not further change discourse. Interviews with 17 tutors indicate that they adopted the moves the feedback recommended but were limited by learners who did not respond. Interviews with 13 learners suggest why the learner-facing feedback had little effect: it reported a group-level rather than a learner-level statistic, some learners preferred to listen, and learner-to-learner interaction was rare in these sessions. Automated feedback can change how tutors invite learner thinking, but shifting how much learners talk likely requires changes beyond feedback.
We propose Mechanistic Shaping, a unified framework for shaping models through their internal representations. Mechanistic Shaping is applied during training so that models follow specifications (rules) about which concepts they express, measured via their outputs or internal representations (e.g., no harmful content), and also which concepts they rely on, i.e., what they use to make their decisions (e.g., not a candidate's gender when screening a r\'esum\'e). Across multiple domains (vision, medical imaging, language), scales (up to 32B parameters), and training paradigms (classification, SFT, DPO, GRPO), we find that Mechanistic Shaping substantially reduces targeted behaviors while preserving task performance. Since training against a representation can also modify it, we ask whether shaped concepts remain monitorable, a question that depends on the class of monitors considered. We therefore introduce a hierarchy of such classes, which separates a concept that escapes one probe from a concept that is truly gone. We characterize three geometric mechanisms of circumvention: the concept can drift away from the monitor, be erased from the layer, or persist in higher-order statistics as a nonlinear encoding, which we explain through a product-code construction. Although other mechanisms may arise in different settings or at larger scales, these three account for all of our observations, and each can be detected by a matching monitor. In practice, we find that shaped concepts largely remain monitorable: refitted linear probes recover the concept in most settings, and the original probe remains reliable whenever gradients do not flow through it. Taken together, these results point to the potential of using interpretability as a tool for training, and call for a correspondingly rigorous science of interpretability-based training.
Reinforcement learning (RL) in Partially Observable Markov Decision Processes (POMDPs) is computationally challenging due to the curse of long-range dependence on history. Fortunately, recent studies have shown that a range of POMDPs that satisfy the filter stability assumption can be well approximated by a superstate Markov Decision Process (MDP) based on a truncated history in a finite-memory window, with a value gap that decays with the window length. However, existing finite-time convergence analyses on such superstate MDP surrogates are either limited to over-simplified linear function approximation or rely on an i.i.d. sampling oracle that is violated by POMDP trajectories, both of which severely limit the applicability of these theoretical results in practice. This motivates us to bridge the gap by investigating RL in POMDPs with nonlinear neural value function approximation. Our main contributions are threefold: i) We first establish a new finite-state MDP point-wise approximation error bound, which lays the foundation for our RL algorithm design and analysis in POMDPs; ii) By applying Temporal-Difference (TD) learning to POMDP-generated superstate transitions and targeting the Bellman fixed point of the superstate MDP, we prove a finite-time critic error bound with nonlinear neural value function approximation; and iii) By combining the neural critic function approximation error bound with the policy mirror descent framework, we obtain a finite-time sublinear regret against the optimal POMDP value. To the best of our knowledge, such a finite-time sublinear regret based on neural function approximation is the first in the literature for finite-memory POMDPs. Our numerical experiments also verify our theoretical results.
This paper develops a stochastic dynamical-systems perspective on three classes of generative AI models. First, we formulate large language models (LLMs) as high-dimensional nonlinear autoregressive models with attention-based dependencies. This formulation connects pretraining by conditional maximum likelihood to sequential token generation and provides a probabilistic interpretation of LLM operation. Second, we consider networks of interacting LLMs and characterize their collective behavior through mean-field ordinary differential equations. These dynamics provide a framework for understanding emergent preferential attachment and persistent structural biases, including type-dependent glass-ceiling effects. Third, we consider neural stochastic differential equations (SDEs) as continuous-time generative models for stochastic trajectories, including those observed at irregular time intervals. We discuss their training with particular emphasis on Wiener-space cubature methods based on rough-path analysis as an efficient alternative to Monte Carlo estimation. Together, the three settings address probabilistic prediction, information exchange and emergent behavior in interacting networks, and path-space generative modeling.
Few-shot adaptation of vision-language models must balance two competing goals: learning fine-grained visual distinctions from very limited supervision and preserving the efficiency that makes frozen CLIP-like models attractive for deployment. Existing lightweight adaptation methods, including prompt tuning, LoRA, and cache-based adapters, typically supervise global image embeddings and therefore underuse token-level structure. Conversely, structure-aware methods exploit local evidence but often introduce additional inference-time computation. We propose SD-LoRA, a structural distillation framework that uses token-level graph reasoning only during training and transfers its effects into a lightweight LoRA-adapted CLIP model. During training, a heterogeneous graph teacher models patch-patch, global-local, and visual-text relations over high-resolution ViT tokens. Rather than retaining this graph at test time, SD-LoRA transfers its structural bias into the shared CLIP-LoRA backbone through implicit distillation and Prototype Predictive Alignment, which aligns graph-conditioned representations with support-derived class prototypes. At inference, the graph teacher is discarded; prediction requires only a single CLIP-LoRA forward pass and a cache-based classifier. Across 11 few-shot benchmarks, SD-LoRA improves over a strong CLIP-LoRA baseline by 2.4/1.8/3.0/2.4/2.0 percentage points in the 1/2/4/8/16-shot settings, respectively, without additional graph computation at test time. Results on base-to-novel generalization, cross-dataset transfer, out-of-distribution robustness, and calibration further show that training-time structural supervision improves both accuracy and transfer reliability. Code is available at https://github.com/MR-Sherif/SD-LoRA.git.
Skin lesion segmentation without annotated masks requires the target to be specified through other cues. Existing approaches rely on dermoscopy-specific rules, annotation-derived model priors, or user-selected regions. A general text-to-image diffusion transformer offers a text interface to the target, but its concept-word attention is unreliable on dermoscopy images: in the model's attention-output space, the word-derived skin-to-lesion direction is poorly aligned with the direction separating real skin from real lesions. We show that the direction derived from generated healthy-skin and lesion images is substantially better aligned. Based on this finding, we propose GENIQ, a training-free method that constructs initial skin and lesion prototypes from generated images and refines them using the most skin-like and lesion-like patches from unlabelled dermoscopy images. It scores test patches along the refined direction and applies a gated threshold to suppress contextual spill-over around small lesions. Without inference-time annotations or additional models, GENIQ achieves Dice scores of 0.85-0.88 on three dermoscopy test sets. It outperforms zero-shot and word-based attention methods on all three sets, and a hand-designed dermoscopy pipeline in Dice and IoU on two.
Flow matching (FM) generators often face a quality-coverage tradeoff. Inference-time techniques such as classifier-free guidance raise quality, but at the cost of coverage. We hypothesize this limitation stems from a basic property of the objective: an intermediate noisy image could correspond to many clean images, yet the model can only predict a single velocity, at best averaging over the multiple possibilities. To mitigate this, we introduce competition among predictions during training. We achieve this by augmenting the predictor with a randomly drawn low-dimensional conditioning code $z$. For each training example, we sample $K$ values of $z$ to produce $K$ candidate velocities that compete to explain the target, and apply the loss only to the best-matching candidate. This winner-take-all objective allows different modes to shape different predictions, rather than being averaged into a single compromise. Furthermore, this term naturally enables a new form of guidance, auxiliary conditioning scaling, which complements traditional guidance methods. Together, these two guidance mechanisms show that our method substantially improves the quality-coverage frontier. We validate on ImageNet-256, using two different testbeds, SiT and RAEv2. Given less than 3% of the original SiT pretraining compute for post-training, at the 67% Recall point, our method reduces $\mathrm{FD}_{r,6}$ from 19.22 to 5.91. On RAEv2, at the 70% Recall point, it reduces $\mathrm{FD}_{r,6}$ from 16.83 to 3.18.
Genetic variants are routinely reclassified in ClinVar after a patient is tested, sometimes from uncertain to pathogenic, yet patient recontact is inconsistent, with no clear duty and responsibility diffused among labs, providers, and patients. Many affected people discuss their results in public online health communities that clinical recontact often does not reach. We built PRISM, an agentic large-language-model pipeline that screens these posts, checks whether the variant was later reclassified, and drafts a disclaimer-bearing outreach message, in a no-contact dry run. Across 3.9 million posts from 32 communities, agent-predicted reinterpretation-relevant discussion is rare (0.288%) but concentrated in specialist genetics communities (6.30% versus 0.10% in general medical-advice communities) and growing fivefold since 2014; PRISM surfaces 123 candidate classification differences, 5 of them from an uncertain submission to pathogenic or likely pathogenic. On a held-out set of 30 real-post drafts, a clinician judged 27 factually grounded and 15 fit to send: factual accuracy is necessary but not sufficient for clinically acceptable outreach. PRISM contributes a population-scale reinterpretation-need map with code and data to be released upon publication.
Once an agent is live, teams want to know whether it is doing its job, not just whether it throws errors. Yet most evaluation is one-off: a benchmark run, a look at traces, a single score. Its conclusions neither check themselves nor reach the agent. We propose closed-loop evaluation, an evaluation method for deployed agents. It models evaluation as performance estimation under biased raters and censored outcomes, and it requires every level's conclusion to trace back to atomic data and to be computed by versioned rules. The method covers four levels of analysis (span, trace, conversation and engagement), measures along a spine of Execution, Substance, Trajectory and Results, treats Conduct as a constraint throughout, and uses Provenance and Adjudication components to certify the evidence and the evaluators. Evaluator agents generated by an Agent Factory carry these components, each with a domain knowledge pack and an adapter harness. Two loops close the evaluation: an inner loop audits the evidence, rules and metrics themselves, and an outer loop uses per-component regression checks to drive agent improvement. We also give a pipeline that runs this process automatically. Using the business-analysis agent SharpInsight as a case and porting the method to public conversation data, we show how problems differ across levels and how evidence-and-rule judgment is more reliable than holistic scoring (ICC 0.86 vs. 0.81; kappa 0.81 vs. 0.63).
Hybrid Mesh-Gaussian rendering combines compact textured surfaces with flexible 3D Gaussian primitives, offering an efficient alternative to purely Gaussian scene representations for structured environments. Existing scene-reconstruction hybrids commonly follow an additive formulation, where Gaussian residuals compensate for errors left by the mesh. However, not every Mesh-Gaussian conflict is caused by missing content: residual geometry errors, occlusion mismatch, texture-projection artifacts, and view-dependent appearance can make the same mesh content consistent in one view but inconsistent in another. We therefore formulate hybrid rendering as a multi-view inconsistency problem and present Gaussolve, a subtract-and-refill framework that selectively dissolves inconsistent mesh content before residual Gaussians refill the revealed regions. We introduce a lightweight Neural Dissolve Field, parameterized by scalar spherical-harmonic Gaussians, to learn where and from which viewing directions mesh content should be dissolved. A continuous transmittance fusion rule progressively exposes residual Gaussians behind the dissolved mesh, preserving their reconstruction gradients and avoiding hard visibility discontinuities. Experiments on structured and semi-structured scenes demonstrate higher visual fidelity with fewer Gaussian primitives and faster rendering. Project page: https://ooosn.github.io/gaussians-can-subtract/
Policy-based decision benchmarks evaluate whether systems apply written rules to financial and operational cases. Yet a rule-following response can be marked wrong when the benchmark misstates the required decision. We audit nine public policy-based decision benchmarks through source tracing, program execution, and controlled checks. We find unsupported targets, reference errors that survive replay or propagate as solved steps, and evaluators that reject acceptable answers or credit incomplete decisions. To measure the scoring impact, we rescore unchanged model responses. On the 96 LegalBench SARA numeric cases, correcting the scorer raises acceptance by more than 80 percentage points for each of three models, without changing references or responses. In RuleArena Tax, restoring a required zero floor changes 27 of 300 references and narrows an observed model gap. A controlled federal-exemption task with explicit ownership and eligibility assumptions also shows that errors in supplied intermediate answers can affect final decisions. We present BAR (Benchmark Audit and Revision), a versioned collection of audit evidence and revisions for reevaluation.
Autoregressive decoding relies on the KV cache, whose memory footprint grows linearly with context length and caps batch size and throughput. Existing post-training methods exploit redundancy across tokens, layers, and feature dimensions, or lower numerical precision. We identify a complementary axis: within each attention head, values are largely predictable from their corresponding keys. We exploit this with value cache amortisation, which fits a lightweight per-head predictor at test time and stores residuals only for the tokens it reconstructs worst, giving $4\times$ near-lossless value compression. The predictor can be initialised by meta-learning for substantially reduced prefill overhead, or fitted from scratch in settings that allow a one-off cost, such as multi-query inference. We pair value amortisation with cross-layer low-rank key compression to produce LowRAM, an asymmetric compression pipeline that stays close to lossless across three model families. By reconstructing only a selected subset of KV pairs at each decoding step, LowRAM can keep decoding overhead minimal and achieve up to $5\times$ higher peak decode throughput at $4\times$ compression. LowRAM further composes with quantisation and eviction, reaching $26.9\times$ compression at $94.5\%$ of baseline performance on long contexts. These results show that treating keys and values as structurally distinct enables efficient, aggressive, and composable KV cache compression.
Transductive adaptation improves vision-language models at test time by refining predictions on an unlabeled batch, but current methods face a key trade-off. Aggressive approaches exploit batch structure but can drift when batches are sparse, have prior shift, or cover little of the label space. Conservative methods better preserve CLIP text semantics, but fixed semantic references can under-adapt even when the batch provides strong corrective evidence. We propose JPTA (Joint Prior and Transport Adaptation), a prior-aware framework that casts transductive VLM adaptation as batch-conditioned semantic reference estimation. Instead of treating CLIP text prototypes as fixed priors or freely replaceable, JPTA views them as semantic references that shift toward image-side batch structure only under calibrated evidence. It estimates batch priors and soft image-side prototypes to form transported semantic references whose impact is scaled by batch reliability. Across standard benchmarks, low effective-class regimes, all-class evaluation, and online streams, JPTA consistently outperforms TransCLIP and StatA, with the largest gains on sparse and prior-shifted batches. Anchor-drift, active-support, and absent-class-mass diagnostics show that JPTA fixes effective-class mismatch without unconstrained transductive drift, indicating that robust transductive VLM adaptation should recalibrate semantic references using reliable batch evidence rather than choose between unrestricted transduction and fixed prototypes. Code is available at https://github.com/MR-Sherif/JPTA.git.
Data Streams are characterised as a potentially infinite source of data subject to non-stationarity. This non-stationarity, also called Concept Drift, often results in deployed predictive models having to be perpetually retrained as the underlying data generating process of the stream changes over time. Drift Detectors are a class of methods used to identify these changes, triggering these retraining procedures. Historically, Stream Learning has operated under the assumption that instances are only seen once, and cannot be stored for later use. Contemporary work relaxes this assumption, yet drift detection and simple resetting procedures are still ubiquitous. Building on related work which has shown that traditional Batch Learning algorithms often outperform Stream Learning algorithms when instances can be stored, we introduce Time Neutralising Trees (TNT), a Decision Tree architecture that enables robust stream classification in settings with Concept Drift. During training, TNT will filter out ("neutralise") old training samples that are no longer relevant given the current context of the stream. If those instances become relevant again, TNT will then re-include them. We evaluate TNT, and related algorithms on 24 real and semi-real classification data streams from the USP DS repository. Results show strong evidence that TNT achieves best-in-class performance on a wide-range of data streams with varying concept drift types.
Data arising in scientific applications often exhibits complex nonlinear structure. Identifying and parameterizing said nonlinearities can enable better modeling, enhance predictive analysis, data interpolation, denoising, and anomaly detection. In this paper, we develop theoretical and computational methods to detect and parametrize Euclidean point cloud data lying near a wedge sum of tori. Specifically, we use persistent cohomology and cup products to identify the distinct toroidal components, and on each component, we optimize over low-energy multivariate trigonometric polynomials for best data fit. The ensuing optimization problem reduces to generalized ridge regression, which can be solved efficiently via standard techniques. We demonstrate the utility of the proposed framework in both synthetic and real datasets, including interpolation of low frame-rate recurrent videos, attractor recovery in quasiperiodic dynamical systems, and the parametrization of gene-space cycle data for parasites and humans.
Forecasting foundation models are increasingly used as frozen feature extractors for time series classification, but it remains unclear whether their pretrained representations offer an advantage over strong classifiers that do not require pretraining. This raises an important question: do these representations provide useful information beyond what the model architecture alone can capture, particularly when labelled data are scarce? We present a controlled benchmarking of two forecasting models, Reverso and Chronos-Bolt, on six large datasets from the MONSTER benchmark, using 1%, 10%, and 100% of the available training labels. We compare their frozen representations against four established time series classification methods and random-weight versions of the same forecasting models. Our results show that pretraining improves classification accuracy over random-weight counterparts on five of the six datasets, with gains of up to 23 percentage points. However, the frozen representations do not reliably outperform the strongest non-pretrained baseline in any setting: they perform worse in 15 of 17 settings and tie in the remaining two. Further experiments show that restoring amplitude information recovers part of the performance gap, while fine-tuning and forecast-error features do not close it. These findings highlight the importance of strong task specific baselines and random weight controls when evaluating forecasting foundation models for classification. Our code is available at https://github.com/Rachet2006/frozen-forecasters-monster.