Heavy-Tailed Flow Matching via Random Clocks

Heavy-tailed data arise in many domains where rare events carry disproportionate importance, such as imbalanced image datasets, financial returns, and weather extremes. Standard diffusion and flow-matching models typically begin from Gaussian noise or Gaussian source distributions, which yield tractable training targets but provide a poor inductive match for heavy-tailed data. We propose Heavy-Tailed Flow Matching via Random Clocks (HTFM), a framework that portrays heavy-tailed sources as mixtures of clock-conditioned Gaussian sources. Conditioning on a given clock path, the source distribution and flow are Gaussian; marginalizing over the clock gives a Gaussian scale mixture covering Gaussian, $α$-stable, and Student-t families. To make the clock-conditioned vector field practical, we encode the path-valued clock using truncated logsignature features, allowing the velocity field to adapt to the realized conditional space with negligible overhead. Empirically, on 2D imbalanced $α$-stable mixtures, CIFAR10-LT, and HRRR weather fields, HTFM improves mode coverage, sample quality, and tail-statistic recovery over Gaussian flow matching and competitive heavy-tailed baselines, while retaining the low-NFE sampling advantage of flow matching. Moreover, the random-clock formulation further provides a practical tail-control interface: by varying only the clock law or tail parameter, the same architecture can calibrate the ``heaviness'' of generated tails across different distribution families.

When Implausible Tokens Get Reinforced: Tail-Aware Credit Calibration for LLM Reinforcement Learning

Reinforcement learning (RL) has achieved remarkable success in enhancing the reasoning capabilities of large language models (LLMs). However, widely used critic-free RL methods rely on uniform credit assignment, broadcasting the same advantage to all tokens regardless of their differences. We identify a critical failure mode of this design, which we refer to as Positive-Credit Contamination: low-probability tail tokens that are contextually erroneous receive identical positive credit to plausible ones within the same trajectory, resulting in the indiscriminate reinforcement of flawed reasoning behavior. To mitigate this issue, we propose Tail-Aware Credit calibratiOn (TACO), a method that calibrates uniform credit assignment to suppress undesirable positive updates. TACO first computes a tail-risk score that incorporates the local generation context to assess each token's risk of falling into the unreliable tail, distinguishing unexpected rarity from uncertainty-driven exploration. TACO then uses this score to tune positive credit for risky tokens without removing their gradients entirely, so that recurring useful rare patterns can accumulate reinforcement while incidental noise is progressively dampened. Experimental results across three LLMs and eight benchmarks show that TACO consistently outperforms GRPO-style baselines. Notably, TACO improves training stability, supporting sustained performance gains in long-horizon RL. The source code is available at: https://github.com/xiuyilou/TACO.

DynamicMem: A Long-Horizon Memory Benchmark in Real-World Settings

LLM agents increasingly act as personal assistants that must remember a user's profile over months: who they are (attributes), what they routinely do (habits), and what they prefer (preferences), and keep it updated as jobs, routines, and tastes drift. Existing benchmarks evaluate this "memory" ability through short, simplified interactions, missing three core properties of real behavior: the profile is heterogeneous, with attributes, habits, and preferences evolving on different timelines; changes are driven by external context such as seasons and life events; and evidence is rarely stated explicitly, instead scattered across many small actions in different apps that a memory system must infer from. We introduce DynamicMem, a synthetic benchmark that constructs 15 months of activity per user, providing long-term multi-app data that real users' privacy keeps out of reach. It provides user-consistent trajectories averaging 2.2M tokens and 1,772 grounded events per user across 16 applications such as e-commerce, fitness, and social platforms. The profile evolves over this period and is never given explicitly: each attribute, habit, or preference must be inferred from small signals scattered across apps. We evaluate at five quarterly checkpoints to track how systems scale as history grows. Benchmarking five representative systems exposes problems a single accuracy score hides: (i) profile reconstruction degrades with history length while service-task accuracy stays flat, despite both drawing on the same memory; (ii) no system both keeps facts that stay true and replaces facts that change, with errors clustering on preferences and on naming the exact referent; and (iii) over 93% of failures trace to what the memory retrieves, not to the model writing the answer, so the largest room for improvement lies in memory itself. Code: https://wenyaxie023.github.io/DynamicMem/

Learning at the Right Pace: Adaptive Data Scheduling Improves LLM Reinforcement Learning

Large Language Models (LLMs) achieve remarkable reasoning capabilities through reinforcement learning (RL) post-training. However, existing RL post-training commonly relies on uniform data sampling, which ignores the semantic structure of the training data and the changing capability of the training policy. To address these limitations, we propose Adaptive Data Scheduling (ADS), a dual-level data scheduling framework for pacing RL post-training that replaces uniform sampling with an adaptive distribution over semantic clusters and policy-boundary sample selection. At the cluster level, ADS organizes samples according to semantic patterns and maintains an adaptive inter-cluster distribution to solidify current training progress. At the sample level, ADS performs intra-cluster scheduling to continuously sample policy-boundary samples, which provides informative relative advantages. Experimental results across three LLMs and seven reasoning benchmarks demonstrate that ADS improves average accuracy by 5.2% over Group Relative Policy Optimization (GRPO). Notably, ADS consistently improves RL methods with different objective designs, highlighting its potential as a general data scheduling strategy for LLM RL post-training. The source code is available at: https://github.com/Richard-zrx/ADS.

Online Learning with Recency: Algorithms for Sliding-window Streaming Multi-armed Bandits

Motivated by the recency effect in online learning, we study algorithms for single-pass *sliding-window streaming multi-armed bandits (MABs)* in this paper. In this setting, we are given $n$ arms with unknown sub-Gaussian reward distributions and a parameter $W$. The arms arrive in a single-pass stream, and only the most recent $W$ arms are considered valid. The algorithm is required to perform pure exploration and regret minimization with limited memory, defined as the number of stored arms. The model is a natural extension of the streaming multi-armed bandits model (without the sliding window) that has been extensively studied in recent years. We provide a comprehensive analysis of both the pure exploration and regret minimization problems with the model. For pure exploration, we prove that finding the best arm is hard with sublinear memory while finding an approximate best arm admits an efficient algorithm. For regret minimization, we explore a new notion of regret and give sharp memory-regret trade-offs for any single-pass algorithm. We complement our theoretical results with experiments, demonstrating the trade-offs between sample, regret, and memory.

Demystifying OPD: Length Inflation and Stabilization Strategies for Large Language Models

On-policy distillation (OPD) trains student models under their own induced distribution while leveraging supervision from stronger teachers. We identify a failure mode of OPD: as training progresses, on-policy rollouts can undergo abrupt length inflation, causing truncated trajectories to dominate the training data. This truncation collapse coincides with abrupt repetition saturation and induces biased gradient signals, leading to severe training instability and sharp degradation in validation performance. We attribute this problem to the interaction between student-induced data collection and the distillation objective, which implicitly favors long and repetitive rollouts. To address this issue, we propose StableOPD, a stabilized OPD framework that combines a reference-based divergence constraint with rollout mixture distillation. These together mitigate repetition-induced length inflation and further stabilize OPD training. Across multiple math reasoning datasets, our approach prevents truncation collapse, stabilizes training dynamics, and improves performance by 7.2% on average.

High-Dimensional Robust Mean Estimation with Untrusted Batches

We study high-dimensional mean estimation in a collaborative setting where data is contributed by $N$ users in batches of size $n$. In this environment, a learner seeks to recover the mean $μ$ of a true distribution $P$ from a collection of sources that are both statistically heterogeneous and potentially malicious. We formalize this challenge through a double corruption landscape: an $\varepsilon$-fraction of users are entirely adversarial, while the remaining ``good'' users provide data from distributions that are related to $P$, but deviate by a proximity parameter $α$. Unlike existing work on the untrusted batch model, which typically measures this deviation via total variation distance in discrete settings, we address the continuous, high-dimensional regime under two natural variants for deviation: (1) good batches are drawn from distributions with a mean-shift of $\sqrtα$, or (2) an $α$-fraction of samples within each good batch are adversarially corrupted. In particular, the second model presents significant new challenges: in high dimensions, unlike discrete settings, even a small fraction of sample-level corruption can shift empirical means and covariances arbitrarily. We provide two Sum-of-Squares (SoS) based algorithms to navigate this tiered corruption. Our algorithms achieve the minimax-optimal error rate $O(\sqrt{\varepsilon/n} + \sqrt{d/nN} + \sqrtα)$, demonstrating that while heterogeneity $α$ represents an inherent statistical difficulty, the influence of adversarial users is suppressed by a factor of $1/\sqrt{n}$ due to the internal averaging afforded by the batch structure.

Online Learning with Limited Information in the Sliding Window Model

Motivated by recent work on the experts problem in the streaming model, we consider the experts problem in the sliding window model. The sliding window model is a well-studied model that captures applications such as traffic monitoring, epidemic tracking, and automated trading, where recent information is more valuable than older data. Formally, we have $n$ experts, $T$ days, the ability to query the predictions of $q$ experts on each day, a limited amount of memory, and should achieve the (near-)optimal regret $\sqrt{nW}\text{polylog}(nT)$ regret over any window of the last $W$ days. While it is impossible to achieve such regret with $1$ query, we show that with $2$ queries we can achieve such regret and with only $\text{polylog}(nT)$ bits of memory. Not only are our algorithms optimal for sliding windows, but we also show for every interval $\mathcal{I}$ of days that we achieve $\sqrt{n|\mathcal{I}|}\text{polylog}(nT)$ regret with $2$ queries and only $\text{polylog}(nT)$ bits of memory, providing an exponential improvement on the memory of previous interval regret algorithms. Building upon these techniques, we address the bandit problem in data streams, where $q=1$, achieving $n T^{2/3}\text{polylog}(T)$ regret with $\text{polylog}(nT)$ memory, which is the first sublinear regret in the streaming model in the bandit setting with polylogarithmic memory; this can be further improved to the optimal $\mathcal{O}(\sqrt{nT})$ regret if the best expert's losses are in a random order.

Support Basis: Fast Attention Beyond Bounded Entries

The quadratic complexity of softmax attention remains a central bottleneck in scaling large language models (LLMs). [Alman and Song, NeurIPS 2023] proposed a sub-quadratic attention approximation algorithm, but it works only under the restrictive bounded-entry assumption. Since this assumption rarely holds in practice, its applicability to modern LLMs is limited. In this paper, we introduce support-basis decomposition, a new framework for efficient attention approximation beyond bounded entries. We empirically demonstrate that the entries of the query and key matrices exhibit sub-Gaussian behavior. Our approach uses this property to split large and small entries, enabling exact computation on sparse components and polynomial approximation on dense components. We establish rigorous theoretical guarantees, proving a sub-quadratic runtime, and extend the method to a multi-threshold setting that eliminates all distributional assumptions. Furthermore, we provide the first theoretical justification for the empirical success of polynomial attention [Kacham, Mirrokni, and Zhong, ICML 2024], showing that softmax attention can be closely approximated by a combination of multiple polynomial attentions with sketching.

Learning-Augmented Hierarchical Clustering

Hierarchical clustering (HC) is an important data analysis technique in which the goal is to recursively partition a dataset into a tree-like structure while grouping together similar data points at each level of granularity. Unfortunately, for many of the proposed HC objectives, there exist strong barriers to approximation algorithms with the hardness of approximation. Thus, we consider the problem of hierarchical clustering given auxiliary information from natural oracles. Specifically, we focus on a *splitting oracle* which, when provided with a triplet of vertices $(u,v,w)$, answers (possibly erroneously) the pairs of vertices whose lowest common ancestor includes all three vertices in an optimal tree, i.e., identifying which vertex ``splits away'' from the others. Using such an oracle, we obtain the following results: - A polynomial-time algorithm that outputs a hierarchical clustering tree with $O(1)$-approximation to the Dasgupta objective (Dasgupta [STOC'16]). - A near-linear time algorithm that outputs a hierarchical clustering tree with $(1-o(1))$-approximation to the Moseley-Wang objective (Moseley and Wang [NeurIPS'17]). Under the plausible Small Set Expansion Hypothesis, no polynomial-time algorithm can achieve any constant approximation for Dasgupta's objective or $(1-C)$-approximation for the Moseley-Wang objective for some constant $C>0$. As such, our results demonstrate that the splitting oracle enables algorithms to outperform standard HC approaches and overcome hardness constraints. Furthermore, our approaches extend to sublinear settings, in which we show new streaming and PRAM algorithms for HC with improved guarantees.