Short-video recommenders such as Douyin must exploit extremely long user histories without breaking latency or cost budgets. We present an end-to-end system that scales long-sequence modeling to 10k -length histories in production. First, we introduce Stacked Target-to-History Cross Attention (STCA), which replaces history self-attention with stacked cross-attention from the target to the history, reducing complexity from quadratic to linear in sequence length and enabling efficient end-to-end training. Second, we propose Request Level Batching (RLB), a user-centric batching scheme that aggregates multiple targets for the same user/request to share the user-side encoding, substantially lowering sequence-related storage, communication, and compute without changing the learning objective. Third, we design a length-extrapolative training strategy---train on shorter windows, infer on much longer ones---so the model generalizes to 10k histories without additional training cost. Across offline and online experiments, we observe predictable, monotonic gains as we scale history length and model capacity, mirroring the scaling law behavior observed in large language models. Deployed at full traffic on Douyin, our system delivers significant improvements on key engagement metrics while meeting production latency, demonstrating a practical path to scaling end-to-end long-sequence recommendation to the 10k regime.
论文检索
输入标题、作者或关键词，从 165 篇学术成果中精准定位
Towards Offline Imitation Learning: Strictly Batch Settings, Generalization, and Variability in Expertise
PDF ↗My doctoral research develops a unified framework for offline imitation learning (IL) that tackles three central challenges: achieving sample efficiency in strictly batch settings, ensuring robustness and generalization under dynamics shifts, and learning from demonstrations of varying quality. At the core of this work is a new paradigm for strictly offline IL based on enforcing the Markov Balance Equation (MBE), a fundamental structural property of trajectory data. Using advanced conditional density estimation, I developed two algorithms, CKIL and MBIL, which achieve state-of-the-art performance in high-dimensional continuous-control tasks. Building upon this foundation, I developed the first Distributionally Robust Offline IL framework under a stationarity constraint, enabling robustness to transition-model mismatch without requiring any additional interaction. I am now extending this direction through Robust Behavior Foundation Models (RBFMs), which aim to generalize across dynamics shifts for a wide range of tasks. Finally, I propose a variational approach for learning from crowdsourced demonstrations by inferring and accounting for demonstrator expertise. Together, these contributions yield principled and practical IL algorithms with strong performance and robustness, broadening the applicability of IL to real-world domains such as robotics, healthcare, and autonomous systems.
We study streaming data with categorical features where the vocabulary of categorical feature values is changing and can even grow unboundedly over time. Feature hashing is commonly used as a pre-processing step to map these categorical values into a feature space of fixed size before learning their embeddings. While these methods have been developed and evaluated for offline or batch settings, in this paper we consider online settings. We show that deterministic embeddings are sensitive to the arrival order of categories and suffer from forgetting in online learning, leading to performance deterioration. To mitigate this issue, we propose a probabilistic hash embedding (PHE) model that treats hash embeddings as stochastic and applies Bayesian online learning to learn incrementally from data. Based on the structure of PHE, we derive a scalable inference algorithm to learn model parameters and infer/update the posteriors of hash embeddings and other latent variables. Our algorithm (i) can handle an evolving vocabulary of categorical items, (ii) is adaptive to new items without forgetting old items, (iii) is implementable with a bounded set of parameters that does not grow with the number of distinct observed values on the stream, and (iv) is invariant to the item arrival order. Experiments in classification, sequence modeling, and recommendation systems in online learning setups demonstrate the superior performance of PHE while maintaining high memory efficiency (consumes as low as 2→4% memory of a one-hot embedding table).
Neural textures have emerged as pivotal assets in next-generation neural rendering pipelines. However, hardware limitations and programming interface constraints lead to suboptimal performance in multi-instance real-time rendering scenarios. This bottleneck becomes particularly acute for texture-intensive tasks such as font rendering. To address this, we propose Neural Outline Cache (NOC), a novel neural font texture supporting real-time anti-aliased rendering and procedural editing within modern neural graphics pipelines. NOC's lightweight network leverages multi-resolution hash encoding to cache spline-derived SDFs, delivering anti-aliased rendering via standard graphics pipelines. For massive-instance scalability, our cache buffer layout (CBL) and batch-fused inference (BFI), tailored for NOC, mitigate neural texture streaming bottlenecks. We constructed an evaluation dataset using five font styles. In offline rendering, our proposed method achieves overall average results of 57.35 dB PSNR, 0.998 SSIM, and 1.1584e-3 pixel RMSE, while maintaining approximately 0.5ms frame latency with 500 real-time instances. To demonstrate its versatility, we integrated a procedural editor for visual effects editing of NOC textures. These results all prove that NOC is a reliable, production-ready neural asset.
Turning Pre-Trained Vision Transformers into End-to-End Histopathology Whole Slide Image Models for Survival Prediction
PDF ↗Conventional whole slide image (WSI) analysis pipelines follow a two-stage process. First, an image encoder, such as a vision transformer (ViT), is used to perform batched offline feature extraction on a series of tiles cropped from the WSI. Second, a multiple instance learning (MIL) model is trained with slide-level labels to obtain task-specific slide embeddings. However, several limitations exist: strong reliance on pre-trained weights of the tile encoder, the absence of receptive fields from the original image, and a lack of task-independent WSI representations. An ideal improvement would be to develop an end-to-end pre-trained WSI model, but training it from scratch will face challenges such as high training costs and computational complexity. In this work, we deconstruct the key steps of ViT-based pathology image representation and propose a conversion strategy called E2E-ViT, which transforms a vanilla ViT into an end-to-end pre-trained WSI model without introducing additional parameters. E2E-ViT directly inputs the entire tissue region in WSIs to efficiently feed image sequences into the transformer backbone, achieving information interaction from the original receptive fields and generating slide features. Through multiple survival prediction tasks, we demonstrate that transformed pre-trained ViTs outperform two-stage MIL models and slide foundation models (SFM). Our work presents a new end-to-end learning paradigm that provides a promising direction for the next generation of computational pathology models.
What data should a vision-language model be trained on? To answer this question, many data curation efforts center on the quality of a dataset. However, most of these existing methods are (i) offline, i.e. they produce a static dataset from a set of predetermined filtering criteria, and (ii) concept-agnostic, i.e. they use model-based filters which induce additional dataset bias. In this work, we go beyond such offline, concept-agnostic methods and advocate for more flexible, task-adaptive online concept-based curation. Our first contribution is DataConcept, a collection of 128M web-crawled image-text pairs annotated with fine-grained details about their concept composition. Building on DataConcept, we introduce CABS (Concept-Aware Batch Sampling), a simple yet effective batch-sampling framework that flexibly constructs batches on-the-fly based on specific target distributions. We propose two variants: (i) Diversity Maximization, to curate batches with the broadest coverage of available concepts, and (ii) CABS-FM (Frequency Maximization), to curate batches with maximal object multiplicity. Through extensive evaluations with four visual backbones and a suite of 28 benchmarks, we demonstrate that CABS significantly benefits Language-Image Pretraining (LIP) and yields highly performant models on long-tailed evaluations. Overall, CABS represents a strong open-source alternative to proprietary online curation algorithms, enabling practitioners to define custom concept distributions that optimize for specific downstream tasks. Both DataConcept and the source code for CABS will be made public.
Breaking the Autoregressive Chain: Hyper-Parallel Decoding for Efficient LLM-Based Attribute Value Extraction
PDF ↗Some text generation tasks, such as Attribute Value Extraction (AVE), require decoding multiple independent sequences from the same document context. While standard autoregressive decoding is slow due to its sequential nature, the independence between output sequences offers an opportunity for parallelism. We present Hyper-Parallel Decoding, a novel decoding algorithm that accelerates offline decoding by leveraging both shared memory and computation across batches. HPD enables out-of-order token generation through position ID manipulation, significantly improving efficiency. Experiments on AVE show that attribute-value pairs are conditionally independent, enabling us to parallelize value generation within each prompt. By further stacking multiple documents within a single prompt, we can decode in parallel up to 96 tokens per prompt. HPD works with all LLMs, and reduces both inference costs and total inference time by up to 13.8X without compromising output quality, potentially saving hundreds of thousands of dollars on industry AVE tasks. Although designed for attribute extraction, HPD makes no assumptions unique to the AVE domain and can in theory be applied to other scenarios with independent output structures.
Topology-Enhanced Alignment for Large Language Models: Trajectory Topology Loss and Topological Preference Optimization
PDF ↗Alignment of large language models (LLMs) typically relies on supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF), or more recently direct preference optimization (DPO). However, existing objectives largely ignore the global geometry and topology of the representation space: they operate on local token-level likelihoods or scalar preference scores, and do not explicitly constrain how hidden states move from a user prompt to an answer.We view generation as tracing a semantic trajectory in hidden space, and propose a topology-enhanced alignment framework that regularizes these trajectories using 0-dimensional persistent homology. First, at the SFT stage, we introduce a Trajectory Topology Loss (TTL). For each batch, we treat mean-pooled embeddings of prompts and gold answers as a mixed point cloud, run a Union-Find-based 0D persistent homology algorithm, and extract ”prompt–answer bridge” edges that connect previously disconnected components. TTL encourages the model’s actual update direction from prompt to answer to align with these topologically derived bridges, rather than with arbitrary or per-example directions.Second, at the RLHF/DPO stage, we propose Topological Preference Optimization (TPO). TPO constructs topic-specific semantic preference vectors from an offline pipeline and aligns the semantic improvement direction between rejected and chosen responses with these vectors in an intermediate hidden layer. We further introduce an exponential-moving-average-based dynamic weighting scheme to balance DPO and TPO losses, and also explore a fully topological variant that applies persistent homology on the chosen/rejected embedding cloud.We instantiate our methods on Qwen2.5-7B-Instruct and evaluate on UltraChat and Anthropic HH-RLHF. Across both SFT and DPO training, topology-enhanced objectives consistently outperform strong non-topological baselines (including per-example, nearest-neighbor, and random direction regularizers) on automatic preference metrics and LLM-judge evaluations, while maintaining or slightly improving toxicity. These results suggest that incorporating persistent homology and trajectory geometry is a promising and practical direction for more controllable LLM alignment.
The expanding long-context capabilities of large language models are constrained by a significant memory bottleneck: the key-value (KV) cache required for autoregressive generation. This bottleneck is substantial; for instance, a Llama-3.1-8B model processing a 32K-token prompt at a batch size of 4 requires approximately 16 GB for its KV cache, exceeding the model’s weights. While KV-cache compression via low-rank projection is promising, existing methods rely on a static, offline-learned subspace that performs poorly under distribution shifts. To overcome these limitations, we introduce OjaKV, a novel framework integrating a hybrid storage policy with online subspace adaptation. OjaKV preserves crucial tokens in full rank as high-fidelity anchors, while applying low-rank compression to intermediate tokens by adapting the projection basis using Oja’s algorithm for online PCA. This adaptation involves a comprehensive update during prefilling and lightweight periodic updates during decoding, ensuring the subspace remains aligned with evolving context. Our framework is fully compatible with FlashAttention. Experiments demonstrate that OjaKV maintains or improves zero-shot accuracy at high compression ratios, achieving the strongest gains on long-context benchmarks requiring complex reasoning. Furthermore, our approach combines with token-selection methods for compounded memory savings, establishing a practical, plug-and-play solution for memory-efficient long-context inference without fine-tuning.
PipeSD: An Efficient Cloud-Edge Collaborative Pipeline Inference Framework with Speculative Decoding
Deep Learning · Algorithms
Speculative decoding can significantly accelerate LLM inference, especially given that its cloud-edge collaborative deployment offers cloud workload offloading, offline robustness, and privacy enhancement. However, existing collaborative inference frameworks with speculative decoding are constrained by (i) sequential token generation and communication with low resource utilization, and (ii) inflexible cloud non-autoregressive verification (NAV) triggering that induces premature verification or costly rollbacks. In this paper, we propose PipeSD, an efficient cloud-edge collaborative pipeline inference framework with speculative decoding. PipeSD overlaps token generation and communication by a token-batch pipeline scheduling mechanism optimized by dynamic programming, and improves verification flexibility through a dual-threshold NAV triggering mechanism with a lightweight Bayesian optimization autotuner. We implement PipeSD using llama-cpp-python, PyTorch, and FastAPI, and evaluate it on a real-world cloud-edge testbed with two draft-target model pairs across four scenarios. Results show that PipeSD consistently outperforms state-of-the-art baselines, achieving $1.16\times$-$2.16\times$ speedup and reducing energy consumption by $14.3$%-$25.3$%. Our code is available at https://anonymous.4open.science/r/PipeSD.
Theory · Learning Theory
On-policy preference learning algorithms for language model alignment such as online direct policy optimization (DPO) can significantly outperform their offline counterparts. We provide a theoretical explanation for this phenomenon by analyzing how the sampling policy's coverage evolves throughout on-policy training. We propose and rigorously justify the \emph{coverage improvement principle}: with sufficient batch size, each update moves into a region around the target where coverage is uniformly better, making subsequent data increasingly informative and enabling rapid convergence. In the contextual bandit setting with Bradley-Terry preferences and linear softmax policy class, we show that on-policy DPO converges exponentially in the number of iterations for batch size exceeding a generalized coverage threshold. In contrast, any learner restricted to offline samples from the initial policy suffers a slower minimax rate, leading to a sharp separation in total sample complexity. Motivated by this analysis, we further propose a simple hybrid sampler based on a novel \emph{preferential} G-optimal design, which removes dependence on coverage and guarantees convergence in just two rounds. Finally, we develop principled on-policy schemes for reward distillation in the general function class setting, and show faster noiseless rates under an alternative deviation-based notion of coverage.
Reinforcement Learning · Batch/Offline
In offline Reinforcement Learning, immediate rewards in logged batch data are often unobserved due to sparse or irregular record-keeping, or censored beyond certain reward values. This issue frequently arises in practical settings, including health care and marketing. We investigate off-policy evaluation (OPE) in finite-horizon Markov decision processes when rewards are missing not at random (MNAR), which breaks ignorability and induces selection bias even after conditioning on states and actions. To address this, we formalize a reward-dependent propensity model and use future states as shadow variables to identify the full-data conditional mean reward. We further introduce a bridge function that recovers the conditional mean reward without explicitly modeling the MNAR mechanism, and estimate it via a min-max procedure to avoid double sampling. Building upon these identification results, we propose an Fitted-Q-Evaluation-style estimator that propagates the recovered rewards while allowing target policies to depend on past missingness indicators. Finally, we establish consistency and finite-sample error bounds for our OPE estimator, and show through simulations the strong performance of our method compared to existing benchmarks.
Reinforcement Learning · Batch/Offline
Current online reinforcement learning (RL) algorithms like GRPO share a key limitation in LLM reasoning: they cannot learn from problems that are "unsolvable" to the model. In other words, they can only improve performance on problems where the model is capable of exploring the correct answer. If a problem is too difficult -- such that even hundreds of attempts never produce a correct solution -- the model cannot learn from it. Consequently, the model's "upper limit" remains unchanged after RL training, even though the likelihood of solving easier, solvable problems may increase. These hard, unsolvable samples -- though potentially rich in learning signal -- cannot contribute to training, as no rollouts yield rewards and thus no gradients are produced. To unlock learning from these hard samples, we propose NuRL, a "nudging" method that aims to push the upper bound of LLM reasoning using self-generated hints, i.e., abstract cues that help reduce the problem difficulty for the model. Given a question and its gold answer, the model generates a Chain-of-Thought (CoT) and then produces a hint containing the core knowledge needed to solve the problem. During online RL training, we generate G rollouts from the base policy and use the pass rate to decide whether the hint should be injected. For hard samples with a 0% pass rate, we inject the offline-generated hint and regenerate a new batch of trajectories. This yields two benefits: (1) the hint boosts pass rates (from 0% to non-zero), thereby introducing training signals for previously unsolvable samples, and (2) the hints are self-generated (conditioned on the gold answer), avoiding distributional shift and do not rely on external models. Compared to standard GRPO, NuRL achieves consistent improvements across six diverse benchmarks and three models, while remaining complementary to test-time scaling. Notably, NuRL can raise the model's upper limit, whereas GRPO leaves pass@1024 unchanged from the base model. Furthermore, we present a systematic study of what makes an effective hint and when hints are most useful. Interestingly, the best hints are abstract and high-level -- as revealing gold answers actually hurt performance -- and are most beneficial when applied necessarily and after GRPO has converged.
General Machine Learning · Everything Else
We introduce k-NLPmeans and k-LLMmeans, text-clustering variants of k-means that periodically replace numeric centroids with textual summaries. The key idea—summary-as-centroid—retains k-means assignments in embedding space while producing human-readable, auditable cluster prototypes. The method is LLM-optional: k-NLPmeans uses lightweight, deterministic summarizers, enabling offline, low-cost, and stable operation; k-LLMmeans is a drop-in upgrade that uses an LLM for summaries under a fixed per-iteration budget whose cost does not grow with dataset size. We also present a mini-batch extension for real-time clustering of streaming text. Across diverse datasets, embedding models, and summarization strategies, our approach consistently outperforms classical baselines and approaches the accuracy of recent LLM-based clustering without extensive LLM calls. Finally, we provide a case study on sequential text streams and release a StackExchange-derived benchmark for evaluating streaming text clustering.
Social Aspects · Trustworthy Machine Learning
Trajectory Optimization (TO) models have achieved remarkable success in offline reinforcement learning (offline RL). However, their vulnerability to backdoor attacks remains largely unexplored. We find that existing backdoor attacks in RL, which typically rely on reward manipulation throughout training, are largely ineffective against TO models due to their inherent sequence modeling nature and large network size. Moreover, the complexities introduced by high-dimensional continuous action further compound the challenge of injecting effective backdoors. To address these gaps, we propose TrojanTO, the first action-level backdoor attack against TO models. TrojanTO is a post-training attack and employs alternating training to forge a strong connection between triggers and target actions, ensuring high attack effectiveness. To maintain attack stealthiness, it utilizes trajectory filtering to preserve the benign performance and batch poisoning for trigger consistency. Extensive evaluations demonstrate that TrojanTO effectively implants backdoors across diverse tasks and attack objectives with a low attack budget (0.3\% of trajectories). Furthermore, TrojanTO exhibits broad applicability to DT, GDT, and DC, underscoring its scalability across diverse TO model architectures.
The ability to learn from large batches of autonomously collected data for policy improvement---a paradigm we refer to as batch online reinforcement learning---holds the promise of enabling truly scalable robot learning by significantly reducing the need for human effort of data collection while getting benefits from self-improvement. Yet, despite the promise of this paradigm, it remains challenging to achieve due to algorithms not being able to learn effectively from the autonomous data. For example, prior works have applied imitation learning and filtered imitation learning methods to the batch online RL problem, but these algorithms often fail to efficiently improve from the autonomously collected data or converge quickly to a suboptimal point. This raises the question of what matters for effective batch online reinforcement learning in robotics. Motivated by this question, we perform a systematic empirical study of three axes---(i) algorithm class, (ii) policy extraction methods, and (iii) policy expressivity---and analyze how these axes affect performance and scaling with the amount of autonomously collected data. Through our analysis, we make several observations. First, we observe that the use of Q-functions to guide batch online RL significantly improves performance over imitation-based methods. Building on this, we show that an implicit method of policy extraction---via choosing the best action in the distribution of the policy---is necessary over traditional explicit policy extraction methods from offline RL. Next, we show that an expressive policy class is preferred over less expressive policy classes. Based on this analysis, we propose a general recipe for effective batch online RL. We then show a simple addition to the recipe, namely using temporally-correlated noise to obtain more diversity, results in further performance gains. Our recipe obtains significantly better performance and scaling compared to prior methods.
Large-scale recommendation systems mainly rely on sparse ID features, struggling with data sparsity. It's important to use multimodal information to assist ID learning for better performance. However, there exists two challenges: (1) distribution discrepancy between multimodal and ID makes direct integration prone to user-item mismatch; (2) slower convergence of multimodal representations compared to ID, causing optimization imbalance under a unified objective, which limits the potential of multimodal representations. In this paper, we comprehensively investigate the two problems and proposes a framework named AB-Rec to align and balance ID and multimodal representations learning for recommendation. We design three alignment tasks to fine-tune a pre-trained multimodal large language model (MLLM), which is then utilized to generate a unified multimodal representation for each item. AB-Rec aligns the distributions of ID and multimodal representations by minimizing the in-batch Wasserstein distance, and maximizes the distance between the two types of representations for the same item to avoid representation collapse. To solve the optimization imbalance, we propose a gradient modulation method that adaptively controls the optimization process by monitoring the contribution differences between ID and multimodal representations. Finally, we conduct extensive offline experiments on four datasets and an A/B test on an online video platform, demonstrating the effectiveness and scalability of our proposed method.
Hyperparameter optimization is crucial for maximizing machine learning model performance, yet most existing algorithms are designed for batch or offline scenarios and assume static data distributions. Such assumptions fall short in data stream settings, where models must adapt to evolving inputs in real time. To address these limitations, we propose the Bayesian Stream Tuner (BST), a novel framework for online hyperparameter optimization in non-stationary data streams. BST maintains a dynamic set of candidate hyperparameter configurations and periodically refines them using an incremental Bayesian model, which estimates configuration performance based on recent data statistics and hyperparameter values. This systematic exploration and refinement strategy allows BST to detect and respond to concept drift by resetting its adaptation mechanisms whenever necessary, ensuring strong performance under changing distributions. Our theoretical analysis establishes sublinear regret bounds for BST in dynamic environments, and extensive experiments on classification and regression tasks demonstrate that BST consistently outperforms state-of-the-art online hyperparameter optimization methods in both predictive accuracy and adaptability, making it a powerful solution for real-time hyperparameter tuning in evolving data streams.
Offline safe reinforcement learning (OSRL) involves learning a decision-making policy to maximize rewards from a fixed batch of training data to satisfy pre-defined safety constraints. However, adapting to varying safety constraints during deployment without retraining remains an under-explored challenge. To address this challenge, we introduce constraint-adaptive policy switching (CAPS), a wrapper framework around existing offline RL algorithms. During training, CAPS uses offline data to learn multiple policies with a shared representation that optimize different reward and cost trade-offs. During testing, CAPS switches between those policies by selecting at each state the policy that maximizes future rewards among those that satisfy the current cost constraint. Our experiments on 38 tasks from the DSRL benchmark demonstrate that CAPS consistently outperforms existing methods, establishing a strong wrapper-based baseline for OSRL.
Imitation learning (IL) is notably effective for robotic tasks where directly programming behaviors or defining optimal control costs is challenging. In this work, we address a scenario where the imitator relies solely on observed behavior and cannot make environmental interactions during learning. It does not have additional supplementary datasets beyond the expert's dataset nor any information about the transition dynamics. Unlike state-of-the-art (SOTA) IL methods, this approach tackles the limitations of conventional IL by operating in a more constrained and realistic setting. Our method uses the Markov balance equation and introduces a novel conditional density estimation-based imitation learning framework. It employs conditional normalizing flows for transition dynamics estimation and aims at satisfying a balance equation for the environment. Through a series of numerical experiments on Classic Control and MuJoCo environments, we demonstrate consistently superior empirical performance compared to many SOTA IL algorithms.