Long-context LLM serving is bottlenecked by the cost of attending over ever-growing KV caches. Dynamic sparse attention promises relief by accessing only a small, query-dependent subset of the KV state per decoding step and extending the KV storage to CPU memory. In practice, however, these algorithmic savings rarely translate into end-to-end system-level gains because sparse methods typically operate at different granularities and thus rely on ad hoc, per-algorithm implementations. At the same time, hierarchical KV storage introduces a new systems bottleneck: retrieving fine-grained, irregular KV subsets across the GPU-CPU boundary can easily erase the benefits of sparsity. We present SPIN, a sparse-attention-aware inference framework that co-designs the execution pipeline with hierarchical KV storage through three techniques: (1) a unified partition abstraction that maps different sparsity granularities onto a shared page-based KV substrate; (2) a locality-aware KV cache manager that dynamically sizes per-request HBM budgets and uses a GPU-friendly bucketed LRU policy to cut PCIe round-trips; and (3) a two-level hierarchical metadata layout sized to the active working set rather than the worst-case address space. Built on vLLM with three representative sparse attention algorithms, SPIN delivers 1.66-5.66x higher end-to-end throughput and 7-9x lower TTFT than vLLM, and reduces TPOT by up to 58% over the original sparse-attention implementations. less

Adapting reasoning models to new tasks during post-training with only output-level supervision stalls under reinforcement learning from verifiable rewards (RLVR) when the initial success probability $p_0$ is small. Using the Tsallis $q$-logarithm, we define a loss family $J_Q$ that interpolates between RLVR (at $q{=}0$, the exploitation pole) and the log-marginal-likelihood over latent trajectories (at $q{=}1$, the density-estimation pole). All members share the same per-example gradient direction, differing only by a scalar amplification $P_{θ^{-q}}$ that reweights each instance independently of the learning rate. This amplification is the mechanism that addresses cold-start stalling: under gradient flow, the exploitation pole requires $Ω(\frac{1}{p_0})$ time to escape cold start, while the density-estimation pole escapes in $Θ\big(\log(\frac{1}{p_0})\big)$; intermediate $q$ trades escape speed against noise memorization. Because $P_θ$ is intractable, we derive two Monte Carlo estimators from the two factorizations of the gradient: Gradient-Amplified RL (GARL) samples from the prior and amplifies the RL gradient, and Posterior-Attenuated Fine-Tuning (PAFT) importance-resamples from the posterior and runs standard SFT. Both have bias $O\big(\frac{q}{M P_θ^{q+1}}\big)$; GARL has lower variance, PAFT has semantically coherent gradients. On FinQA, HotPotQA, and MuSiQue, GARL at $q{=}0.75$ substantially mitigates cold-start stalling, escaping cold start where GRPO fails entirely. In warm start, GARL at low $q$ dominates FinQA where training is stable; on HotPotQA and MuSiQue, GARL destabilizes during training, and PAFT at $q{=}0.75$ provides stable gradients (best overall on HotPotQA at 47.9 maj@16, $+14.4$ over GRPO). less

Imbalanced classification remains a pervasive challenge in machine learning, particularly when minority samples are too scarce to provide a robust discriminative boundary. In such extreme scenarios, conventional models often suffer from unstable decision boundaries and a lack of reliable error control. To bridge the gap between generative modeling and discriminative classification, we propose a two-stage framework \textbf{VAE-Inf} that integrates deep representation learning with statistically interpretable hypothesis testing. In the first stage, we adopt a one-class modeling perspective by training a variational autoencoder (VAE) exclusively on majority-class data to capture the underlying reference distribution. The resulting latent posteriors are aggregated via a Wasserstein barycenter to construct a global Gaussian reference model, providing a geometrically principled baseline for the majority class. In the second stage, we transform this generative foundation into a discriminative classifier by fine-tuning the encoder with limited minority samples. This is achieved through a novel distribution-aware loss that enforces probabilistic separation between classes based on variance-normalized projection statistics. For inference, we introduce a projection-based score that admits a natural hypothesis testing interpretation, allowing for a distribution-free calibration procedure. This approach yields exact finite-sample control of the Type-I error (false positive rate) without relying on restrictive parametric assumptions. Extensive experiments on diverse real-world benchmarks demonstrate that our framework achieves competitive performance against other approaches. The codes are available upon request. less

Block-sequential continual learning demands that a single model both protect prior solutions from catastrophic forgetting and efficiently infer at inference time which prior solution matches the current input without task labels. We present Functional Task Networks (FTN), a parameter-isolation method inspired by structural and dynamical motifs found in the mammalian neocortex. Similar to mixture-of-experts, this method uses a high dimensional, self-organizing binary mask over a large population of small but deep networks, inspired by dendritic models of pyramidal neurons. The mask is produced by a three-stage procedure: (1) gradient descent on a continuous mask identifies task-relevant neurons, (2) a smoothing kernel biases the result toward spatial contiguity, (3) and k-winner-take-all binarizes the resulting group at a fixed capacity budget. Like mixture-of-experts, each neuron is an independent deep network, so disjoint masks give exactly disjoint gradient updates, providing structural guarantees against catastrophic forgetting. This three-stage procedure recovers the sub-network of a previously-trained task in a single gradient step, providing unsupervised task segmentation at inference time. We test it on three continual-learning benchmarks: (1) a synthetic multi-task classification/regression generator, (2) MNIST with shuffled class labels (pure concept shift), and (3) Permuted MNIST (domain shift). On all three, FTN with fine grained smoothing (FTN-Slow) results in nearly zero forgetting. FTN with a large kernel and only 2 iterations of smoothing (FTN-Fast) trades off some retention for increased speed. We show that the spatial organization mechanism reduces the effective mask search from the combinatorial top-k subset problem in O(C(H,K)) to the complexity of a near-linear scan in O(H) over compact cortical neighborhoods, which is parallelized by the gradient-based update. less

Scientific publication compresses a branching, iterative research process into a linear narrative, discarding the majority of what was discovered along the way. This compilation imposes two structural costs: a Storytelling Tax, where failed experiments, rejected hypotheses, and the branching exploration process are discarded to fit a linear narrative; and an Engineering Tax, where the gap between reviewer-sufficient prose and agent-sufficient specification leaves critical implementation details unwritten. Tolerable for human readers, these costs become critical when AI agents must understand, reproduce, and extend published work. We introduce the Agent-Native Research Artifact (Ara), a protocol that replaces the narrative paper with a machine-executable research package structured around four layers: scientific logic, executable code with full specifications, an exploration graph that preserves the failures compilation discards, and evidence grounding every claim in raw outputs. Three mechanisms support the ecosystem: a Live Research Manager that captures decisions and dead ends during ordinary development; an Ara Compiler that translates legacy PDFs and repos into Aras; and an Ara-native review system that automates objective checks so human reviewers can focus on significance, novelty, and taste. On PaperBench and RE-Bench, Ara raises question-answering accuracy from 72.4% to 93.7% and reproduction success from 57.4% to 64.4%. On RE-Bench's five open-ended extension tasks, preserved failure traces in Ara accelerate progress, but can also constrain a capable agent from stepping outside the prior-run box depending on the agent's capabilities. less

A widely cited result by Dong et al. (2021) showed that Transformers built from self-attention alone, without skip connections or feed-forward layers, suffer from rapid rank collapse: all token representations converge to a single direction. The proposed remedy was the MLP. We show that this picture, while correct in the regime studied by Dong, is incomplete in ways that matter for architectural understanding. Three results are established. First, layer normalisation is precisely affine-rank-neutral: it preserves the affine rank of the token representation set exactly. The widespread claim that LN "plays no role" is imprecise; the correct statement is sharper. Second, residual connections generically obstruct rank collapse in real Transformers such as BERT-base, in a measure-theoretic sense, without contribution from the MLP. The MLP's irreplaceable function is different: generating feature directions outside the linear span of the original token embeddings, which no stack of attention layers can produce. Third, a phenomenon distinct from rank collapse is identified: head-channel non-identifiability. After multi-head attention sums per-head outputs through the output projection, individual contributions cannot be canonically attributed to a specific head; n(H-1)d_k degrees of freedom per layer remain ambiguous when recovering a single head from the mixed signal. The MLP cannot remedy this because it acts on the post-summation signal. A constructive partial remedy is proposed: a position-gated output projection (PG-OP) at parameter overhead below 1.6% of the standard output projection. The four collapse phenomena identified in the literature -- rank collapse in depth, in width, head-channel non-identifiability, and entropy collapse -- are unified under a symmetry-breaking framework, each corresponding to a distinct symmetry of the Transformer's forward pass. less

Hypernetwork-based methods such as Doc-to-LoRA internalize a document into an LLM's weights in a single forward pass, but they fail systematically on conflicts: when the document contradicts pretraining knowledge, accuracy collapses to 46.4% on the deepest facts. We show the failure is a magnitude problem rather than a representational one. The hypernetwork already targets the right layers, but its adapter margin is approximately constant across documents while the pretrained margin grows with training frequency, so deep conflicts lose by construction. The account predicts that failure should track prior strength: sorting 194 conflicts by the base model's log-probability on the contradicted fact, baseline accuracy falls from 68% on weak-prior questions to 16% on strong-prior ones, a 52 percentage-point gap. The cure is amplitude. Selective Layer Boosting scales the adapter at its top-norm layers, and Conflict-Aware Internalization triggers boosting only when the base model is confident. Both are training-free; together they raise deep-conflict accuracy from 46.4% to 71.0% on Gemma-2B and from 53.6% to 72.5% on Mistral-7B while preserving novel-knowledge recall, and beat vanilla retrieval-augmented generation on medium conflicts by 18 percentage points despite operating entirely in parameter space. We release KID-Bench, a 489-question benchmark that separates novel recall, cross-knowledge combination, and prior-graded conflicts. less

Fair machine learning (ML) methods help identify and mitigate the risk that algorithms encode or automate social injustices. Algorithmic approaches alone cannot resolve structural inequalities, but they can support socio-technical decision systems by surfacing discriminatory biases, clarifying trade-offs, and enabling governance. Although fairness is well studied in supervised learning, many real ML applications are online and sequential, with prior decisions informing future ones. Each decision is taken under uncertainty due to unobserved counterfactuals and finite samples, with dire consequences for under-represented groups, systematically under-observed due to historical exclusion and selective feedback. A bank cannot know whether a denied loan would have been repaid, and may have less data on marginalized populations. This paper introduces a taxonomy of uncertainty in sequential decision-making -- model, feedback, and prediction uncertainty -- providing shared vocabulary for assessing systems where uncertainty is unevenly distributed across groups. We formalize model and feedback uncertainty via counterfactual logic and reinforcement learning, and illustrate harms to decision makers (unrealized gains/losses) and subjects (compounding exclusion, reduced access) of policies that ignore the unobserved space. Algorithmic examples show it is possible to reduce outcome variance for disadvantaged groups while preserving institutional objectives (e.g. expected utility). Experiments on data simulated with varying bias show how unequal uncertainty and selective feedback produce disparities, and how uncertainty-aware exploration alters fairness metrics. The framework equips practitioners to diagnose, audit, and govern fairness risks. Where uncertainty drives unfairness rather than incidental noise, accounting for it is essential to fair and effective decision-making. less

Physics-informed neural networks (PINNs) have garnered significant interest for their potential in solving partial differential equations (PDEs) that govern a wide range of physical phenomena. By incorporating physical laws into the learning process, PINN models have demonstrated the ability to learn physical outcomes reasonably well. However, current PINN approaches struggle to predict or solve new PDEs effectively when there is a lack of training examples, indicating they do not generalize well to unseen problem instances. In this paper, we present a transferable learning approach for PINNs premised on a fast Pseudoinverse PINN framework (Pi-PINN). Pi-PINN learns a transferable physics-informed representation in a shared embedding space and enables rapid solving of both known and unknown PDE instances via closed-form head adaptation using a least-squares-optimal pseudoinverse under PDE constraints. We further investigate the synergies between data-driven multi-task learning loss and physics-informed loss, providing insights into the design of more performant PINNs. We demonstrate the effectiveness of Pi-PINN on various PDE problems, including Poisson's equation, Helmholtz equation, and Burgers' equation, achieving fast and accurate physics-informed solutions without requiring any data for unseen instances. Pi-PINN can produce predictions 100-1000 times faster than a typical PINN, while producing predictions with 10-100 times lower relative error than a typical data-driven model even with only two training samples. Overall, our findings highlight the potential of transferable representations with closed-form head adaptation to enhance the efficiency and generalization of PINNs across PDE families and scientific and engineering applications. less