Searcharxiv⌕ Search

arXiv subjects

Kaiyuan Li

Publications and source records attributed to Kaiyuan Li.

At least 19 recordsLinked to original sources
arXiv2026

TWICE: Two-Clock, Two-Window Learning for Long-Horizon Conversion Prediction in Online Advertising

Long-horizon conversion prediction under delayed feedback creates a two-clock, two-window learning problem in online advertising. A short base observation window releases recent clicks on the click clock before their outcomes mature, whereas conversions continue to arrive on the conversion clock throughout a longer target conversion window. The click clock provides timely but partially observed status supervision. The conversion clock reveals long-tail delays, but the delay composition within an arrival-time slice is weighted by historical click cohorts with different traffic volumes and target-window conversion rates. We present TWICE, a framework that factorizes long-horizon post-click conversion rate (CVR) into a target-window conversion probability and a grouped elapsed-delay cumulative distribution function (CDF). The two clocks provide complementary supervision. Click-clock records train the target-window CVR head through a current-status likelihood over the base observation window. Newly arrived conversions train the delay model on the conversion clock. To account for the cohort mixture, TWICE uses fixed click-time predicted CVR (pCVR) mass as cohort exposure in an arrival-conditioned likelihood. This accounts for differences in cohort traffic and conversion propensity. The resulting aggregate records are self-contained. A single learned CDF produces monotone predictions for all requested horizons up to the target conversion window. Serving requires neither historical lookup nor convolution. Experiments on a public benchmark and an industrial advertising dataset demonstrate the effectiveness of TWICE. In an online A/B test in Kwai's advertising system, TWICE increased expected revenue, revenue, and conversions by 2.486%, 1.858%, and 2.061%, respectively. It was subsequently deployed to full traffic.

cs.LG↗
arXiv2026

The Point, the Vision and the Text: Does Point Cloud Boost Spatial Reasoning of Large Language Models? A Bias-Controlled Study

3D Large Language Models (LLMs) leveraging spatial information in point clouds for 3D spatial reasoning attract great attention. Despite some promising results, the advantages of point clouds over other modalities remain unclear. Moreover, existing 3D benchmarks are insufficient for fairly evaluating the ability of multimodal LLMs to comprehend spatial concepts. To address these challenges, we introduce ScanReQA, a 3D spatial reasoning benchmark encompassing text, vision, and point cloud modalities. We then evaluate the performance of text, 2D, and 3D LLMs on the benchmark to compare the effectiveness of different modalities in understanding spatial concepts. Furthermore, we analyze the reasoning mechanisms behind 3D LLMs using point clouds. Our findings reveal that: 1) binary spatial reasoning remains challenging for current 3D LLMs, 2) MLLMs based on point cloud and visual modalities demonstrate stronger spatial reasoning capabilities than LLMs, and 3) 3D LLMs exhibit the attention sink phenomenon similar to that in 2D LLMs, impairing spatial reasoning. We think these conclusions can help the next step of 3D LLMs and also offer insights for foundation models in other modalities. We release datasets and codes in the project page: https://github.com/EmbodiedCity/ScanReQA.code.

cs.CV↗
arXiv2026

EventPrune: Cascaded Event-Assisted Token Pruning for Efficient First-Person Dynamic Spatial Reasoning

First-person dynamic spatial reasoning requires models to track continuous motion and precise geometric structure, but the quadratic attention cost of Transformer-based Video-LLMs makes dense visual tokens computationally expensive. Existing token pruning paradigms predominantly rely on discrete static snapshots, failing to preserve the motion and geometric cues essential for reasoning. We propose Event Cascade Pruning (ECP), to our knowledge the first training-free framework that leverages the high-frequency motion cues from event cameras as a continuous event-guided motion prior to guide token selection. ECP combines three stages: Event-Triggered Causal Sampling to anchor motion-informative keyframes, Event-guided Motion Saliency Filtering to suppress event-inactive visual tokens, and Event-Attention Ranking Fusion to calibrate spatial attention with motion-salient dynamics. With 80% visual token reduction, ECP outperforms the full-token baseline (37.62% vs. 36.31%) while achieving 1.89x inference speedup and 52% GFLOPs reduction. We further introduce ESR-Real, the first real-world RGB-event benchmark for first-person spatial reasoning, where ECP improves accuracy by 2.68 percentage points over full-token baselines.

cs.CV↗
arXiv2026

Counterfactual Multi-task Learning for Delayed Conversion Modeling in E-commerce Sales Pre-Promotion

Sales promotions, as short-term incentives to stimulate product purchases, play a pivotal role in modern e-commerce marketing strategies. During promotional events, user behavior patterns exhibit distinct characteristics compared to regular periods. In the pre-promotion phase, users typically engage in product search and browsing without immediate purchases, adding items to carts in anticipation of promotional discounts. This behavior leads to delayed conversions, resulting in significantly lower conversion rates (CVR) before the promotion day. Although existing research has made progress in CVR prediction for promotion days using historical data, it largely overlooks the critical pre-promotion period. And delayed feedback modeling has been extensively studied, current approaches fail to account for the unique distribution shifts in conversion behavior before promotional events, where delayed conversions predominantly occur on the promotion day rather than over continuous time windows. To address these limitations, we propose the Counterfactual Multi-task Delayed Conversion Model (CM-DCM), which leverages historical pre-promotion data to enhance CVR prediction for both delayed and direct conversions. Our model incorporates three key innovations: (i) A multi-task architecture that jointly models direct and delayed conversions using historical pre-promotion data; (ii) A personalized user behavior gating module to mitigate data sparsity issues during brief pre-promotion periods; (iii) A counterfactual causal approach to model the transition probability from add-to-cart (ATC) to delayed conversion. Extensive experiments demonstrate that CM-DCM outperforms baselines in pre-promotion scenarios. Online A/B tests during major promotional events showed significant improvements in advertising revenue, delayed conversion GMV, and overall GMV, validating the effectiveness of our approach.

cs.IR↗
arXiv2026

VLGOR: Visual-Language Knowledge Guided Offline Reinforcement Learning for Generalizable Agents

Combining Large Language Models (LLMs) with Reinforcement Learning (RL) enables agents to interpret language instructions more effectively for task execution. However, LLMs typically lack direct perception of the physical environment, which limits their understanding of environmental dynamics and their ability to generalize to unseen tasks. To address this limitation, we propose Visual-Language Knowledge-Guided Offline Reinforcement Learning (VLGOR), a framework that integrates visual and language knowledge to generate imaginary rollouts, thereby enriching the interaction data. The core premise of VLGOR is to fine-tune a vision-language model to predict future states and actions conditioned on an initial visual observation and high-level instructions, ensuring that the generated rollouts remain temporally coherent and spatially plausible. Furthermore, we employ counterfactual prompts to produce more diverse rollouts for offline RL training, enabling the agent to acquire knowledge that facilitates following language instructions while grounding in environments based on visual cues. Experiments on robotic manipulation benchmarks demonstrate that VLGOR significantly improves performance on unseen tasks requiring novel optimal policies, achieving a success rate over 24% higher than the baseline methods.

cs.LG↗
arXiv2026

RLVR Training of LLMs Does Not Improve Thinking Ability for General QA: Evaluation Method and a Simple Solution

Reinforcement learning from verifiable rewards (RLVR) stimulates the thinking processes of large language models (LLMs), substantially enhancing their reasoning abilities on verifiable tasks. It is often assumed that similar gains should transfer to general question answering (GQA), but this assumption has not been thoroughly validated. To assess whether RLVR automatically improves LLM performance on GQA, we propose a Cross-Generation evaluation framework that measures the quality of intermediate reasoning by feeding the generated thinking context into LLMs of varying capabilities. Our evaluation leads to a discouraging finding: the efficacy of the thinking process on GQA tasks is markedly lower than on verifiable tasks, suggesting that explicit training on GQA remains necessary in addition to training on verifiable tasks. We further observe that direct RL training on GQA is less effective than RLVR. Our hypothesis is that, whereas verifiable tasks demand robust logical chains to obtain high rewards, GQA tasks often admit shortcuts to high rewards without cultivating high-quality thinking. To avoid possible shortcuts, we introduce a simple method, Separated Thinking And Response Training (START), which first trains only the thinking process, using rewards defined on the final answer. We show that START improves both the quality of thinking and the final answer across several GQA benchmarks and RL algorithms.

cs.CL↗
arXiv2026

MALLOC: Benchmarking the Memory-aware Long Sequence Compression for Large Sequential Recommendation

The scaling law, which indicates that model performance improves with increasing dataset and model capacity, has fueled a growing trend in expanding recommendation models in both industry and academia. However, the advent of large-scale recommenders also brings significantly higher computational costs, particularly under the long-sequence dependencies inherent in the user intent of recommendation systems. Current approaches often rely on pre-storing the intermediate states of the past behavior for each user, thereby reducing the quadratic re-computation cost for the following requests. Despite their effectiveness, these methods often treat memory merely as a medium for acceleration, without adequately considering the space overhead it introduces. This presents a critical challenge in real-world recommendation systems with billions of users, each of whom might initiate thousands of interactions and require massive memory for state storage. Fortunately, there have been several memory management strategies examined for compression in LLM, while most have not been evaluated on the recommendation task. To mitigate this gap, we introduce MALLOC, a comprehensive benchmark for memory-aware long sequence compression. MALLOC presents a comprehensive investigation and systematic classification of memory management techniques applicable to large sequential recommendations. These techniques are integrated into state-of-the-art recommenders, enabling a reproducible and accessible evaluation platform. Through extensive experiments across accuracy, efficiency, and complexity, we demonstrate the holistic reliability of MALLOC in advancing large-scale recommendation. Code is available at https://anonymous.4open.science/r/MALLOC.

cs.IR↗
arXiv2026

Aggregate and Broadcast: Scalable and Efficient Feature Interaction for Recommender Systems

Feature interaction is a core ingredient in ranking models for large-scale recommender systems, yet making it both expressive and efficiently scalable remains challenging. Exhaustive pairwise interaction is powerful but incurs quadratic complexity in the number of tokens/features, while many efficient alternatives rely on restrictive structures that limit information exchange. We further identify two common bottlenecks in practice: (1) early aggregation of behavior sequences compresses fine-grained signals, making it difficult for deeper layers to reuse item-level details; and (2) late fusion injects task signals only at the end, preventing task objectives from directly guiding the interaction process. To address these issues, we propose the Information Flow Network (INFNet), a lightweight architecture that enables scalable, task-aware feature interaction with linear complexity. INFNet represents categorical features, behavior sequences, and task identifiers as tokens, and introduces a small set of hub tokens for each group to serve as communication hubs. Interaction is realized through an efficient aggregate-and-broadcast information flow: hub tokens aggregate global context across groups via cross-attention, and a lightweight gated broadcast unit injects the refined context back to update the categorical, sequence, and task tokens. This design supports width-preserving stacking that preserves item-level signals in sequence and enables task-guided interaction throughout the network, while reducing interaction cost from quadratic to linear in the number of feature tokens. Experiments on a public benchmark and a large-scale industrial dataset demonstrate that INFNet consistently outperforms strong baselines and exhibits strong scaling behavior. In a commercial online advertising system, deploying INFNet improves revenue by +1.587% and click-through rate by +1.155%.

cs.IR↗
arXiv2026

CropCraft: Complete Structural Characterization of Crop Plants From Images

The ability to automatically build 3D digital twins of plants from images has countless applications in agriculture, environmental science, robotics, and other fields. However, current 3D reconstruction methods fail to recover complete shapes of plants due to heavy occlusion and complex geometries. In this work, we present a novel method for 3D modeling of agricultural crops based on optimizing a parametric model of plant morphology via inverse procedural modeling. Our method first estimates depth maps by fitting a neural radiance field and then optimizes a specialized loss to estimate morphological parameters that result in consistent depth renderings. The resulting 3D model is complete and biologically plausible. We validate our method on a dataset of real images of agricultural fields, and demonstrate that the reconstructed canopies can be used for a variety of monitoring and simulation applications.

cs.CV↗
arXiv2025

Progressive Supernet Training for Efficient Visual Autoregressive Modeling

Visual Auto-Regressive (VAR) models significantly reduce inference steps through the "next-scale" prediction paradigm. However, progressive multi-scale generation incurs substantial memory overhead due to cumulative KV caching, limiting practical deployment. We observe a scale-depth asymmetric dependency in VAR: early scales exhibit extreme sensitivity to network depth, while later scales remain robust to depth reduction. Inspired by this, we propose VARiant: by equidistant sampling, we select multiple subnets ranging from 16 to 2 layers from the original 30-layer VAR-d30 network. Early scales are processed by the full network, while later scales utilize subnet. Subnet and the full network share weights, enabling flexible depth adjustment within a single model. However, weight sharing between subnet and the entire network can lead to optimization conflicts. To address this, we propose a progressive training strategy that breaks through the Pareto frontier of generation quality for both subnets and the full network under fixed-ratio training, achieving joint optimality. Experiments on ImageNet demonstrate that, compared to the pretrained VAR-d30 (FID 1.95), VARiant-d16 and VARiant-d8 achieve nearly equivalent quality (FID 2.05/2.12) while reducing memory consumption by 40-65%. VARiant-d2 achieves 3.5 times speedup and 80% memory reduction at moderate quality cost (FID 2.97). In terms of deployment, VARiant's single-model architecture supports zero-cost runtime depth switching and provides flexible deployment options from high quality to extreme efficiency, catering to diverse application scenarios.

cs.CV↗
arXiv2025

Balanced Token Pruning: Accelerating Vision Language Models Beyond Local Optimization

Large Vision-Language Models (LVLMs) have shown impressive performance across multi-modal tasks by encoding images into thousands of tokens. However, the large number of image tokens results in significant computational overhead, and the use of dynamic high-resolution inputs further increases this burden. Previous approaches have attempted to reduce the number of image tokens through token pruning, typically by selecting tokens based on attention scores or image token diversity. Through empirical studies, we observe that existing methods often overlook the joint impact of pruning on both the current layer's output (local) and the outputs of subsequent layers (global), leading to suboptimal pruning decisions. To address this challenge, we propose Balanced Token Pruning (BTP), a plug-and-play method for pruning vision tokens. Specifically, our method utilizes a small calibration set to divide the pruning process into multiple stages. In the early stages, our method emphasizes the impact of pruning on subsequent layers, whereas in the deeper stages, the focus shifts toward preserving the consistency of local outputs. Extensive experiments across various LVLMs demonstrate the broad effectiveness of our approach on multiple benchmarks. Our method achieves a 78% compression rate while preserving 96.7% of the original models' performance on average. Our code is available at https://github.com/EmbodiedCity/NeurIPS2025-Balanced-Token-Pruning.

cs.CV↗
arXiv2025

VQL: An End-to-End Context-Aware Vector Quantization Attention for Ultra-Long User Behavior Modeling

In large-scale recommender systems, ultra-long user behavior sequences encode rich signals of evolving interests. Extending sequence length generally improves accuracy, but directly modeling such sequences in production is infeasible due to latency and memory constraints. Existing solutions fall into two categories: (1) top-k retrieval, which truncates the sequence and may discard most attention mass when L >> k; and (2) encoder-based compression, which preserves coverage but often over-compresses and fails to incorporate key context such as temporal gaps or target-aware signals. Neither class achieves a good balance of low-loss compression, context awareness, and efficiency. We propose VQL, a context-aware Vector Quantization Attention framework for ultra-long behavior modeling, with three innovations. (1) Key-only quantization: only attention keys are quantized, while values remain intact; we prove that softmax normalization yields an error bound independent of sequence length, and a codebook loss directly supervises quantization quality. This also enables L-free inference via offline caches. (2) Multi-scale quantization: attention heads are partitioned into groups, each with its own small codebook, which reduces quantization error while keeping cache size fixed. (3) Efficient context injection: static features (e.g., item category, modality) are directly integrated, and relative position is modeled via a separable temporal kernel. All context is injected without enlarging the codebook, so cached representations remain query-independent. Experiments on three large-scale datasets (KuaiRand-1K, KuaiRec, TMALL) show that VQL consistently outperforms strong baselines, achieving higher accuracy while reducing inference latency, establishing a new state of the art in balancing accuracy and efficiency for ultra-long sequence recommendation.

cs.IR↗
arXiv2025

EBaReT: Expert-guided Bag Reward Transformer for Auto Bidding

Reinforcement learning has been widely applied in automated bidding. Traditional approaches model bidding as a Markov Decision Process (MDP). Recently, some studies have explored using generative reinforcement learning methods to address long-term dependency issues in bidding environments. Although effective, these methods typically rely on supervised learning approaches, which are vulnerable to low data quality due to the amount of sub-optimal bids and low probability rewards resulting from the low click and conversion rates. Unfortunately, few studies have addressed these challenges. In this paper, we formalize the automated bidding as a sequence decision-making problem and propose a novel Expert-guided Bag Reward Transformer (EBaReT) to address concerns related to data quality and uncertainty rewards. Specifically, to tackle data quality issues, we generate a set of expert trajectories to serve as supplementary data in the training process and employ a Positive-Unlabeled (PU) learning-based discriminator to identify expert transitions. To ensure the decision also meets the expert level, we further design a novel expert-guided inference strategy. Moreover, to mitigate the uncertainty of rewards, we consider the transitions within a certain period as a "bag" and carefully design a reward function that leads to a smoother acquisition of rewards. Extensive experiments demonstrate that our model achieves superior performance compared to state-of-the-art bidding methods.

cs.LG↗
arXiv2025

EgoPrune: Efficient Token Pruning for Egomotion Video Reasoning in Embodied Agent

Egomotion videos are first-person recordings where the view changes continuously due to the agent's movement. As they serve as the primary visual input for embodied AI agents, making egomotion video reasoning more efficient is therefore essential for real-world deployment. Recent advances in vision-language models have enabled strong multimodal reasoning capabilities, but their computational cost remains prohibitive for long, redundant video inputs. Existing token pruning methods, typically designed for third-person videos, fail to leverage the spatiotemporal continuity and motion constraints inherent in egomotion settings. To address this, we propose EgoPrune, a training-free token pruning method tailored for egomotion video reasoning. EgoPrune comprises three components: a keyframe selector adapted from EmbodiedR for temporally efficient sampling; Perspective-Aware Redundancy Filtering (PARF), which aligns visual tokens using perspective transformations and removes redundant tokens; and a Maximal Marginal Relevance (MMR)-based token selector that jointly considers visual-text relevance and intra-frame diversity. Experiments on two egomotion video benchmarks show that EgoPrune consistently outperforms prior training-free methods across various pruning ratios while significantly reducing FLOPs, memory usage, and latency. Moreover, we deploy EgoPrune on an embodied agent equipped with a Jetson Orin NX 16GB edge device, demonstrating its real-world efficiency and suitability for on-device egomotion video reasoning.

cs.CV↗
arXiv2025

Physically interpretable diffractive optical networks for high-dimensional vortex mode sorting

Despite the significant progress achieved by diffractive optical networks in diverse computing tasks, such as mode multiplexing and demultiplexing, investigations into the physical meanings behind complex diffractive networks at the layer level have been quite limited. Here, for highdimensional vortex mode sorting tasks, we show how various physical transformation rules for each layer within trained diffractive networks can be revealed under properly defined input/output mode relations. An intriguing physical transformation division phenomenon, associated with the saturated sorting performance of the system, has been observed with an increasing number of masks. In addition, we have also demonstrated the use of physical interpretation for efficiently designing parameter-varying networks with high performance. These physically interpretable optical networks resolve the contradiction between rigorous physical theorems and operationally vague network structures, paving the way for designing and understanding systems for various mode conversion tasks, and inspiring further interpretation of diffractive networks in advanced tasks and other network structures.

physics.optics↗
arXiv2025

ImagineBench: Evaluating Reinforcement Learning with Large Language Model Rollouts

A central challenge in reinforcement learning (RL) is its dependence on extensive real-world interaction data to learn task-specific policies. While recent work demonstrates that large language models (LLMs) can mitigate this limitation by generating synthetic experience (noted as imaginary rollouts) for mastering novel tasks, progress in this emerging field is hindered due to the lack of a standard benchmark. To bridge this gap, we introduce ImagineBench, the first comprehensive benchmark for evaluating offline RL algorithms that leverage both real rollouts and LLM-imaginary rollouts. The key features of ImagineBench include: (1) datasets comprising environment-collected and LLM-imaginary rollouts; (2) diverse domains of environments covering locomotion, robotic manipulation, and navigation tasks; and (3) natural language task instructions with varying complexity levels to facilitate language-conditioned policy learning. Through systematic evaluation of state-of-the-art offline RL algorithms, we observe that simply applying existing offline RL algorithms leads to suboptimal performance on unseen tasks, achieving 35.44% success rate in hard tasks in contrast to 64.37% of method training on real rollouts for hard tasks. This result highlights the need for algorithm advancements to better leverage LLM-imaginary rollouts. Additionally, we identify key opportunities for future research: including better utilization of imaginary rollouts, fast online adaptation and continual learning, and extension to multi-modal tasks. Our code is publicly available at https://github.com/LAMDA-RL/ImagineBench.

cs.LG↗
arXiv2025

BBQRec: Behavior-Bind Quantization for Multi-Modal Sequential Recommendation

Multi-modal sequential recommendation systems leverage auxiliary signals (e.g., text, images) to alleviate data sparsity in user-item interactions. While recent methods exploit large language models to encode modalities into discrete semantic IDs for autoregressive prediction, we identify two critical limitations: (1) Existing approaches adopt fragmented quantization, where modalities are independently mapped to semantic spaces misaligned with behavioral objectives, and (2) Over-reliance on semantic IDs disrupts inter-modal semantic coherence, thereby weakening the expressive power of multi-modal representations for modeling diverse user preferences. To address these challenges, we propose a Behavior-Bind multi-modal Quantization for Sequential Recommendation (BBQRec for short) featuring dual-aligned quantization and semantics-aware sequence modeling. First, our behavior-semantic alignment module disentangles modality-agnostic behavioral patterns from noisy modality-specific features through contrastive codebook learning, ensuring semantic IDs are inherently tied to recommendation tasks. Second, we design a discretized similarity reweighting mechanism that dynamically adjusts self-attention scores using quantized semantic relationships, preserving multi-modal synergies while avoiding invasive modifications to the sequence modeling architecture. Extensive evaluations across four real-world benchmarks demonstrate BBQRec's superiority over the state-of-the-art baselines.

cs.IR↗
arXiv2025

CHIME: A Compressive Framework for Holistic Interest Modeling

Modeling holistic user interests is important for improving recommendation systems but is challenged by high computational cost and difficulty in handling diverse information with full behavior context. Existing search-based methods might lose critical signals during behavior selection. To overcome these limitations, we propose CHIME: A Compressive Framework for Holistic Interest Modeling. It uses adapted large language models to encode complete user behaviors with heterogeneous inputs. We introduce multi-granular contrastive learning objectives to capture both persistent and transient interest patterns and apply residual vector quantization to generate compact embeddings. CHIME demonstrates superior ranking performance across diverse datasets, establishing a robust solution for scalable holistic interest modeling in recommendation systems.

cs.IR↗