The paper studies the personalization–collaboration trade-off in federated learning via Generalized Total Variation Minimization (GTVMin) with least-squares models on an arbitrary communication graph. Using random matrix theory in the proportional high-dimensional limit, it derives exact asymptotic test-error formulas under Gaussian-mixture data and proves that the GTVMin solution concentrates at rate $1/\sqrt{p}$, independently of the graph structure or regularization strength $\alpha$. If the modeling assumptions hold, the results provide a principled way to tune $\alpha$ from training data alone.

Long-context LLM inference is bottlenecked by expanding key-value caches. This paper proposes treating the KV cache as a learnable layer and optimizing it per document via test-time training while keeping model weights frozen. The two-stage procedure first matches teacher attention statistics and then distills teacher predictions, yielding large accuracy gains on synthetic and natural-document benchmarks under aggressive compression.

This paper proposes SKG, a geometric reasoning engine that inverts the usual data-driven paradigm: physical laws, symmetries and conservation constraints are treated as hard, a priori structures, while data are used only to calibrate a small task-specific energy term. The core object is an octuple $\mathcal{S}=(M,g,\omega,H,S,G,\Delta,\Pi)$ where inference becomes geometric evolution on a manifold, and the metric $g$ is pulled back from invariants via a Fisher-determined kernel bandwidth. The authors claim this yields provable out-of-distribution bounds, decouples learning from the metric, and reduces induction to estimating only the task energy $S_{\text{task}}$.

InstructPro2 tackles de novo protein design by translating open-ended natural language instructions into amino acid sequences while incorporating biological context such as target ligands or antigens. The framework augments a large protein language model with text and modality encoders, using lightweight gated cross-attention adapters to condition generation on both semantic and molecular signals. If the approach holds up experimentally, it could replace narrow task-specific models with a single generalist system that designers steer through plain text.

Mobile manipulation demands precise coordination between a mobile base and a robotic arm, making real-world policy fine-tuning difficult because of high-dimensional, tightly coupled action spaces. This paper proposes HiRe-MoMa, a framework that freezes a pre-trained base policy and learns a small, hierarchical residual: one head corrects the base motion while the other corrects the arm conditioned on the base correction. It also guides learning with sparse human teleoperation interventions via an auxiliary critic-ranking loss, reporting large success-rate gains on a TIAGo++ robot after roughly 30 minutes of online interaction.

Wearable activity recognition suffers from sensor heterogeneity and subtle, scale-specific motion patterns that models often confuse with participant-specific artifacts. W-JEPA tackles this by replacing raw temporal patches with fixed wavelet tokens—indexed by physical scale, sensor axis, and time—as the units of masked latent prediction in a Joint-Embedding Predictive Architecture. The key payoff is improved generalization to unseen participants across six benchmarks.

DyG-WM tackles dynamic graph world modelling by predicting future latent node and relation states from temporal context, rather than fitting specific supervised tasks or reconstructing raw edges. The core idea is a dual-level joint-embedding predictive architecture that combines historical node-relation context encoders with path-signature encoders to capture ordered structural changes. This matters because it aims to generalize across both link prediction and node classification without expensive task-specific labels.

SPARS addresses catastrophic forgetting in replay-free continual learning by partitioning network parameters into three importance-aware regions—winning, neighboring, and non-important—and treating each with freezing, soft regularization, or unrestricted updates based on either the Fisher Information Matrix or a novel label-free Confidence Score. To further mitigate classifier bias in class-incremental learning without stored exemplars, the authors replace the standard fully connected classifier with Quadratic Discriminant Analysis at inference time.

While large language models concentrate attention on initial tokens to prevent over-mixing, state-of-the-art graph transformers rarely exhibit this 'attention sink' behavior. This paper identifies the attention scoring function—not graph symmetry or locality—as the decisive factor, showing that replacing GRIT's gated logit with a standard dot product raises sink rates dramatically. The finding reframes attention sinks as an artifact of dot-product scoring rather than an inherent property of deep transformers.

TopicMerge tackles the scalability bottleneck of LLM-based topic modeling by combining efficient clustering with targeted LLM reasoning. Its core idea is to first over-segment a corpus into fine-grained semantic clusters, then use a lightweight LLM to hierarchically merge them with contextual neighborhood guidance. The authors claim this slashes LLM prompts by over 92% while matching or exceeding state-of-the-art clustering alignment on standard benchmarks.

SGQN introduces selective growth and lazy pruning into decoupled Q-learning for continuous control, aiming to concentrate action-grid resolution where the value function peaks. On eight DMControl tasks it reports competitive mean returns, but its central—and most valuable—finding is that the adaptive grid is not responsible for the performance: growth behaves like uniform expansion, pruning changes returns by at most 1.1%, and what actually matters is starting from a coarse grid and using $\varepsilon$-greedy exploration. The paper is therefore best read as a rigorous self-critical analysis of which ingredients in growing action-grid methods matter and which do not.

When multiple LLM retrieval actions are generated jointly, downstream systems typically return a single reward for the merged result set, making it impossible to tell which action helped or hurt. CASI closes this information gap through snapshot-grounded set intervention: it caches per-action retrievals, evaluates all alternative subsets under a frozen training snapshot, and aggregates the resulting marginals into Shapley-based credit. This matters because broadcasting the same scalar reward ignores redundancy and competition among actions, which misleads policy optimization in capacity-limited regimes such as sponsored search.

Continuum Attention replaces the discrete key-value lookup in Transformers with exact integration over a continuous, piecewise-linear memory trajectory. Each query induces an exponential relevance density along the curve, and the operator admits closed-form forward and backward passes via low-order exponential moments. On small-scale byte-level language modeling it improves over tuned softmax, though the empirical story is limited to a single small dataset and model size.

Goal misgeneralization in RL looks uniform from the outside—agents competently pursue proxies instead of intended goals—but this paper argues it is not a single failure. By tracing goal information through networks in three Procgen games and five simulated robot tasks, the authors identify three distinct alignment failures—functional target substitution, policy shortcuts, and cue competition—and show that the depth at which the goal is lost determines which test-time repair works. The result is a principled diagnostic framework that connects internal representations to intervention strategies, moving beyond purely behavioral descriptions of misalignment.

Scaling transformer deployments is hampered by the fact that model families are released in only a handful of fixed sizes. This paper tackles elastic inference: letting a single pre-trained transformer adapt its compute budget on-the-fly without retraining. The core idea is a differentiable soft-pruning formulation that jointly learns a global ranking of every individual MLP neuron and attention head across multiple sparsity targets, keeping the original weights frozen. It matters because it promises to replace rigid model selection with a smooth accuracy-compute trade-off, and the authors back this with strong vision results and measured GPU speedups.

On-policy distillation (OPD) trains a student on its own rollouts using token-level supervision from a teacher, but its effectiveness relies on the teacher's ability to recover from any student prefix. The authors show that this recoverability collapses on incorrect rollouts: the teacher follows the student's errors rather than correcting them, even though it can still recognize them. Their method, CriticOPD, restores recoverability by having the teacher critique an incorrect rollout so the student can rewrite its solution from the last error; the teacher scores the rewrite with the critique as privileged information, while the student trains on the revision without it. On six mathematical reasoning benchmarks, CriticOPD improves standard OPD by roughly 4–6 points across two student scales, with the largest gains on competition-level problems, and the student needs no feedback at inference.

Potato leaf diseases such as early and late blight reduce yields in Ethiopia, where farmers often lack access to laboratory diagnosis. This study develops a lightweight convolutional neural network to classify healthy, early-blight, and late-blight leaves, training on 2,215 PlantVillage images and testing on 247 smartphone photos collected from local farms. The central claim is that a 1.70 MB model can achieve high accuracy while remaining efficient enough for deployment on low-resource hardware.

Fairness-weighted federated learning deliberately inflates updates from high-loss clients, but under an update-only interface a server cannot tell legitimate amplification from adversarial scaling. This paper identifies that $q$-fair loss scaling preserves the class-mixture structure of honest updates while warping their coefficient constraint from a unit-sum simplex to a radial nonnegative cone. Building on this geometry, FUSE uses a small trusted labeled set to recover latent coefficients, screens submissions jointly for radial scale and cone consistency, and renormalizes the aggregate to protect both robustness and fairness.

Training deep networks without normalization remains difficult, and relating their trainability to the exact geometry of the loss surface is rarely attempted. This paper proposes a projective polynomial activation $f(z)=z+\lambda z^3$ paired with either an $\ell_2$ projection onto the unit sphere or an unnormalized identity path, and studies it at two scales: networks that train to depth $100$--$300$ on MNIST, and tiny algebraic versions whose square-loss critical points are solved exactly via Gröbner bases and numerical irreducible decomposition. The point is to derive concrete design, initialization, and diagnostic rules from the small model's geometry and validate them at depth.

Multimodal code intelligence demands programs that are correct, well-structured, and visually faithful, yet existing training signals are either too coarse or confined to single tasks. This paper introduces PrismaForge, which synthesizes per-instruction evaluation rubrics that separately assess code-level and visual-level criteria, then uses them to filter training data and reward reinforcement learning. The resulting dataset, PrismaCoder-2M, and the trained 9B and 35B models advance the open state of the art and compete with proprietary systems on charts, webpages, and scientific graphics.