论文库

新论文 · v1arXiv:2611.13560

更新于 2026-11-10

Are Dense Rewards Worth Their Tokens in GRPO?

Vikash Singh, Weicong Chen, Debargha Ganguly, Yanyan Zhang, Mohsen Hariri, Sreehari Sankar, Wang Yang, Chaoda Song 等 19 位作者

原文摘要

RLVR trains language models to reason by rewarding correct final answers. A widely used algorithm, group-relative policy optimization (GRPO), gives every token of a solution the same credit, so dense process rewards have been proposed to credit intermediate steps, for example with a Monte-Carlo reward that scores each step by sampling continuations from it. Comparing such a reward with GRPO per training update, or vetting it with a gradient test on a single batch, leaves out what it costs: in our setting, the continuations alone take 91% of all generated tokens. We introduce a matched-compute protocol that gives every training intervention the same budget, measured in generated tokens and training GPU-hours, and tests it in preregistered studies with controls for side effects. Across 1,668 runs on six models, matching compute reverses the ranking. On Qwen2.5-1.5B, the Monte-Carlo reward leads standard GRPO per update, yet ends 14.3 percentage points lower at equal tokens and 7.8 points lower at equal training time, and none of four dense-reward variants outperforms GRPO even with as much training time. A per-token gradient test orders these rewards as training does but overestimates each by one to two points, as much as the gains it is meant to detect. In contrast, training on the prompts whose predicted success is nearest 1/2 improves accuracy at matched compute, by 1.7 points on Qwen3-0.6B and 0.8 on Qwen2.5-1.5B. These gains replicate on fresh seeds and unseen problems, and a simple rule, that the candidate prompts are predicted to fail more often than not, predicted where they appear in five of six preregistered tests. Both gains come early in training, and standard GRPO with twice the initial learning rate comes within half a point of them. We argue that training interventions should be validated at matched compute, which overturns some apparent gains and confirms others.

cs.CL · cs.AI

版本更新 · v2arXiv:2508.14948

更新于 2026-11-09

LFM4Ads: Large Foundation Model for Ads Recommendation

Jie Jiang, Shangyu Zhang, Shijie Quan, Zhongren Wang, Tianqu Zhuang, Junwei Pan, Bo Fu, Yilong Sun 等 35 位作者

原文摘要

Recent work leverages large-scale pretrained foundation models (FMs) to transfer generalizable representations to downstream advertising tasks. Most existing approaches transfer user representations (URs) and item representations (IRs), while underutilizing the user-item cross representations (CRs) generated by deeper interaction layers. Although a recent work transfers target-aware embeddings (TAEs), it requires real-time upstream FM inference alongside the downstream, introducing additional engineering complexity and overhead. To address these limitations, we propose Large Foundation Model for Ads Recommendation (LFM4Ads), an efficient pretrain-then-transfer paradigm for cross-domain ad recommendation. LFM4Ads jointly exploits URs, IRs, and CRs from the upstream FM. It extracts sample-level CRs from pretrained interaction layers and aggregates them into user-level CRs using a time-aware exponential moving average, thereby decoupling upstream inference from downstream online serving. To use CRs efficiently, we propose three downstream usages at different levels of granularity: feature-level gating to fuse CRs into downstream embeddings, module-level fine-tuning to yield downstream CRs, and model-level retrieval by matching CRs with IRs. LFM4Ads has been deployed on Tencent's advertising platform since Q4 2024. It has supported 10+ launches across major scenarios, yielding a platform-wide gross merchandise value (GMV) lift of 2.45% and annual revenue gains in the hundreds of millions of dollars. Offline experiments on KuaiRand-1K further demonstrate our effectiveness. Our code is available at https://github.com/ZhuangQu/LFM4Ads.

cs.LG

新论文 · v1arXiv:2611.11189

更新于 2026-11-08

TEAM: Topology-Aware, Semantic-Sensitive, General Transformer for Multimodal-Attributed Graphs

Yushuo Li, Xunkai Li, Zhengyu Wu, Ronghua Li, Guoren Wang

原文摘要

Multimodal-Attributed Graphs (MAGs) organize heterogeneous data in graphs whose nodes are characterized by modalities such as text and images rather than traditional unimodal attributes. However, many representative MAG methods organize propagation around node-level states and topology-defined neighborhoods, which can obscure fine-grained modality contributions: a bounded neighborhood may contain semantically weak evidence, while compatible context need not be confined to it. To address both limitations, we propose TEAM, a modality-centric framework that constructs and encodes graph context with node-modality tokens rather than treating each node as a single interaction unit. Concretely, TEAM unfolds the local topology into a Modality Encoding Tree and performs tree-anchored semantic evidence acquisition, organizing local structural context while expanding the accessible evidence space outside the constructed local Tree. We further introduce the Modality-Source Context Transformer (MSCT), which encodes the selected node-modality tokens and coordinates interactions among modality-source groups. By retaining textual and visual attributes as separate interaction units until readout, it supports fine-grained cross-node cross-modal exchange without first collapsing them into whole-node features. Across the evaluated baselines, TEAM attains the highest reported mean on all graph-learning metrics, all three G2Text metrics, and both G2Image metrics. On Toys, it surpasses the strongest baseline by 3.37 and 2.93 percentage points in Accuracy and Macro-F1, respectively.

cs.LG

新论文 · v1arXiv:2611.08631

更新于 2026-11-06

BEVCast: Path-Adaptive BEV for Efficient End-to-End Autonomous Driving

Qiren Wang, Mingyu Guo, Zixuan Wang, Weisong Shi

原文摘要

Camera-based driving models build a dense bird's-eye-view (BEV) grid and spend the same computation on every cell of it, although the vehicle crosses only a few of those cells within the planning horizon. We present BEVCast, a two-stage BEV whose resolution adapts to the predicted driving path. A coarse pass encodes the whole scene on a grid of 4.096 m cells. A heat head trained on the recorded path assigns a heat to every coarse cell and the cells with the highest heat are refined into 1.024 m cells under a fixed budget. The remaining cells keep their coarse resolution, so the trajectory head reads a mixed-resolution BEV that still covers the entire scene and needs neither an HD map nor a navigation command. Against a full grid of 1.024 m cells with the same architecture and training recipe, BEVCast at a 10% refinement budget encodes 28% of the BEV queries and needs 66% of the BEV encoder time at batch size 1. On nuScenes its mean displacement error is 1.078 m against 1.149 m and its drivable area violation rate is 1.46% against 2.29%. On NAVSIM navtest its mean displacement error is 0.577 m against 0.598 m and its extended predictive driver model score is 82.51 against 81.28. Next to published planners BEVCast reaches a comparable error on both benchmarks without a navigation command. The resolution of a driving BEV can therefore follow the task instead of the grid.

cs.CV

新论文 · v1arXiv:2611.02332

更新于 2026-11-02

LEANGUARD: Lean-Verified Guardrails for Tool-Using Agents

Sreehari Sankar, Debargha Ganguly, Vikash Singh, Mohsen Hariri, Weicong Chen, Yanyan Zhang, Shouren Wang, Wang Yang 等 20 位作者

原文摘要

Tool-using language-model agents often execute actions that their domain policies forbid. We present LeanGuard, a runtime that checks every proposed tool call against a policy written in Lean4 before the call reaches the environment. Policies combine state predicates with scoped past-time temporal operators over the execution history, evaluation is fail-closed, and the decision procedure is proved equivalent to a LeanLTL semantics. A denied call can be corrected in two ways. With feedback, the agent receives the violated rule and a hint; with proof-guided repair, Lean metaprogramming searches for a nearby call that preserves the user's intent and proves that the policy admits it, without the model. On Tau-Bench with five Qwen models (0.8B-27B), unguarded agents execute at least 33 calls that the Lean policy forbids, while LeanGuard executes none; with feedback, task success does not change significantly. On DTap, LeanGuard reduces direct prompt-injection success from 57.9% to 5.9%, and in telecom, where benign tasks need no writes, benign success is unchanged. Feedback reduces repeated blocked calls and raises task success for four of five models over a bare denial. On replayed denials, repair proves an admissible edit for 88 of 252 denials caused by argument values, in a median of 2,ms.

cs.AI · cs.CL

新论文 · v1arXiv:2610.40331

更新于 2026-10-29

多源协变量偏移下高维非参数回归的最优神经估计

Optimal Neural Estimation for High-Dimensional Nonparametric Regression under Multisource Covariate Shift

Chenghao Xu, Haotian Xu, Shivam Kumar, Daren Wang

AI 速读基于摘要

研究多源协变量偏移下高维非参数回归，提出oracle加权神经估计并给出minimax最优收敛率，同时给出无需已知噪声方差与偏移参数的自适应方法。

math.ST · stat.ME · stat.ML

查看原文摘要

We study high-dimensional nonparametric regression under multisource covariate shift, where $\mathrm J$ sources of training data share a common regression function but differ in their covariate distributions, sample sizes, and noise variances. Our goal is to combine these sources to estimate the regression function accurately under a target covariate distribution. We propose an oracle weighted neural estimator and establish its minimax optimality up to logarithmic factors through matching upper and lower bounds. The resulting rate is $\mathsf n_{\rm eff}^{-2p/(2p+K)}$, where $p$ and $K$ denote the effective smoothness and dimension of the regression function, and the effective sample size is \begin{align} \mathsf n_{\rm eff} =\sum_{j=1}^{\mathrm J} \frac{n_j}{(1+σ_{ρ,j})σ_{ε,j}^{2}}. \nonumber \end{align} Here, $n_j$ is the sample size of source $j$, $σ_{ε,j}^{2}$ is its noise variance, and $σ_{ρ,j}$ controls the deviation between the covariate distribution of source $j$ and the target covariate distribution. We also develop an adaptive procedure that requires no knowledge of source noise variances or covariate-shift parameters. Under suitable conditions, this procedure attains the oracle minimax rate up to logarithmic factors.

版本更新 · v2arXiv:2604.07794

更新于 2026-10-28

Density Decomposition on Hypergraphs

Xiaoyu Leng, Hongchao Qin, Rong-Hua Li, Guoren Wang

原文摘要

Hypergraphs capture higher-order interactions involving multiple entities. Existing decomposition models, such as $k$-core and nbr-$k$-core, rely on degree or neighborhood constraints and may overlook fine-grained density variations. We propose a $(k,δ)$-dense subhypergraph model based on capped hyperedge contributions. For each parameter pair, the unique dense set is the inclusionwise smallest maximizer of a penalized capped hyperedge contribution. These dense sets form a nested hierarchy characterized by egalitarian $δ$-orientations and hypergraph reachability. We develop three algorithms for dense subhypergraph search and two for complete density decomposition. Both decomposition algorithms use integral flow balancing: Flow-Decomp extracts density levels through strongly connected component condensation, while Reverse-Decomp uses direct reverse propagation. Once an egalitarian orientation is obtained, both recover all density levels in linear time. We further derive sandwich bounds that restrict computation to equivalent residual subproblems and develop incremental algorithms for maintaining the hierarchy under hyperedge insertions and deletions. Experiments on nine real-world hypergraphs, including datasets with over 15 million vertices, compare our methods with four established decomposition models. The results show that Reverse-Decomp improves decomposition efficiency, incremental maintenance outperforms recomputation in the tested settings, and the proposed model captures density structures overlooked by degree-based decompositions.

cs.SI · cs.DS

新论文 · v1arXiv:2610.38091

更新于 2026-10-28

Jailbreak Connectivity: Towards Diverse, Transferable, and Universal Multimodal Large Language Model Jailbreak

Xu Zhang, Ziqing Hu, Shuo Han, Ren Wang

原文摘要

Multimodal large language models (MLLMs) are vulnerable to visual jailbreak attacks that bypass safety safeguards. Existing methods typically optimize isolated adversarial examples, limiting attack diversity, transferability, and generalization. We propose Jailbreak Connectivity (JC), a framework based on the observation that effective adversarial inputs can be connected through low-loss regions in image space. JC optimizes continuous adversarial paths: path sampling yields diverse jailbreak images, lightweight surrogate classifiers guide cross-model transfer, and a distribution-level objective enables universal attacks across harmful queries. On MM-SafetyBench, JC achieves average attack success rates of 84.01%, 85.77%, and 97.95% on Qwen3-VL-8B-Instruct, InternVL3.5-8B, and Gemma-3-12B-IT, respectively, exceeding the strongest baselines by up to 68.36 percentage points. In black-box evaluations, JC improves attack success rates over FigStep by 24.69 and 27.34 percentage points on GPT-5.6 and Gemini 3.1 Pro, respectively. These findings reveal structured vulnerabilities in MLLM safety alignment and inform more comprehensive security evaluation.

cs.LG · cs.AI · cs.CR

新论文 · v1arXiv:2610.36986

更新于 2026-10-27

LearnPrunerV2: Learnable Progressive Token Pruning for Vision-Language Models

Rinyoichi Takezoe, Yaqian Li, Weiren Wang, Zihao Bo, Huihong Shi, Anzhou Hou, Mo Guang, Jiale Yan 等 9 位作者

原文摘要

Vision-Language Models (VLMs) have demonstrated remarkable multimodal capabilities but suffer from prohibitive computational overhead due to the massive number of visual tokens. While token pruning is a promising solution, existing methods still suffer from two major drawbacks: 1) unreliable pruning criteria that fail to accurately localize task-specific regions due to inherent attention biases or query-agnostic designs, and 2) suboptimal pruning strategies that employ rigid, non-adaptive schedules, overlooking the dynamic evolution of token redundancy across different model depths. To address these challenges, we propose LearnPrunerV2, which extends the previous two-stage heuristic framework (LearnPruner) into a fully learnable architecture. Specifically, we integrate lightweight learnable modules into every layer of the VLM backbone, which estimate token importance by fusing intrinsic visual saliency with extrinsic query relevance captured via a shadow token. By uncovering a consistent staircase redundancy pattern across VLM layers, we further introduce Budget-Deterministic Pruning (BDP) to establish a hardware-friendly schedule with fixed per-layer budgets. Extensive experimental results across various benchmarks and VLMs demonstrate that LearnPrunerV2 consistently outperform previous state-of-the-art methods. Notably, LearnPrunerV2 preserves 99.8% of the original performance when pruning 77.8% of visual tokens on LLaVA-v1.5, achieving a superior accuracy-efficiency trade-off.

cs.CV · cs.AI

首次观测 · 记录 #1ACL Anthology · ACL 2026:acl:2026.acl-long.201ACL Anthology · ACL 2026

来源日期 2026-07-01（仅到月份）

CodeFlowBench: A Multi-turn, Iterative Benchmark for Complex Code Generation

Sizhe Wang, Zhengren Wang Peking University, Dongsheng Ma, Yongan Yu Mila - Quebec Artificial Intelligence Institute and McGill University, McGill University, Rui Ling, Zhiyu li, Feiyu Xiong MemTensor (Shanghai) Technology Co., Ltd., Wentao Zhang

原文摘要

Modern software development demands code that is maintainable, testable, and scalable by organizing the implementation into modular components with iterative reuse of existing codes. We formalize this iterative, multi-turn paradigm as codeflow and introduce CodeFlowBench, the first benchmark designed to comprehensively evaluate LLMs’ ability to perform codeflow - implementing new functionality by reusing existing functions over multiple turns. CodeFlowBench comprises two complementary components: CodeFlowBench-Comp, a core collection of 5,000+ competitive programming problems from Codeforces updated via an automated pipeline and CodeFlowBench-Repo, which is sourced from GitHub repositories to better reflect real-world scenarios. Furthermore, a novel evaluation framework featured dual assessment protocol and structural metrics derived from dependency trees is introduced. Extensive experiments reveal significant performance degradation in multi-turn codeflow scenarios. Furthermore, our in-depth analysis illustrates that model performance inversely correlates with dependency complexity. These findings not only highlight the critical challenges for supporting real-world workflows, but also establish CodeFlowBench as an essential tool for advancing code generation research.

cs.CL

Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表

首次观测 · 记录 #1ACL Anthology · ACL 2026:acl:2026.acl-long.259ACL Anthology · ACL 2026

来源日期 2026-07-01（仅到月份）

Demystifying Uncertainty in LLMs: Active Calibration between Concepts and Human Evaluations

Pengqi Li, Lizhong Ding Beijing Institute of Technology, Zhehao Zhou, Chunhui Zhang, Jiarun Fu, Hao Li Beijing Institute of Technology, Ye Yuan Beijing Institute of Technology, Guoren Wang Beijing Institute of Technology

原文摘要

Hallucinations arise when large language models (LLMs) guess rather than acknowledge their underlying uncertainty. Existing static strategies for mitigating hallucinations have been only partially successful, largely because they do not explicitly model the information gain from interacting with the external environment. Researchers need a general method to proactively steer users toward informative clarifications, thereby unlocking the model’s effective capacity under underspecified inputs. We model the uncertainty of LLMs in interactive settings and uncover the mechanism of active calibration between model concepts and human evaluations, improving reliability. We show that calibration error in LLMs density estimation admits a non-vanishing lower bound under non-interactive learning, while interaction empirically reduces it. We further characterize that calibration error identifies informative queries and that calibration can be accelerated by shifting query distributions from imbalanced to balanced regimes. Guided by these insights, we propose a calibration-driven Interactive Learning Strategy (ILS) that selects clarification queries by optimizing calibration error, providing both theoretical guarantees and empirical gains for reliability. Code and data are available at https://github.com/zhouyeah215/Demystifying_Uncertainty.

cs.CL

Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表

首次观测 · 记录 #1ACL Anthology · ACL 2026:acl:2026.acl-long.605ACL Anthology · ACL 2026

来源日期 2026-07-01（仅到月份）

DORA: A Dual-Objective Reinforcement Learning Framework for Effective and Efficient Multimodal Agentic Search

Guangming Qin, Yuhao Deng, Yukun Zhao, Zhenyang Li Baidu, Junfeng Wang, Dawei Yin Baidu, Ye Yuan Beijing Institute of Technology, Guoren Wang 等 11 位作者

原文摘要

The most recent research uses reinforcement learning (RL) to post-train Multi-modal Large Language Models (MLLMs) such that these models are able to iteratively call search engines to dynamically access external knowledge when handling complex Visual Question Answering (VQA) tasks. However, existing methods face two major limitations in effectiveness and efficiency: i) For effectiveness, the objective of these methods, which only considers the correctness of the generated final response, overlooks the quality of intermediate search results, thus leading to suboptimal search strategies. ii) For efficiency, existing methods often unnecessarily invoke search calls during reasoning, making the inference inefficient. To address these issues, we propose , a customized dual-objective reinforcement learning framework to improve the search strategies of MLLMs, enhancing their search quality yet minimizing search frequency. The key ideas include (1) a reward function that promotes correct reasoning trajectories with fewer search calls; and (2) dual optimization objectives that jointly optimize search quality and answer correctness. Extensive experiments on 3 real-world datasets demonstrate that DORA outperforms state-of-the-art methods, achieving up to 8.4% higher accuracy while reducing the number of search calls by 9.7%.

cs.CL · cs.AI · cs.CV

Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表

首次观测 · 记录 #1ACL Anthology · ACL 2026:acl:2026.acl-long.857ACL Anthology · ACL 2026

来源日期 2026-07-01（仅到月份）

Closing the Modality Reasoning Gap for Speech Large Language Models

Chaoren Wang, Heng Lu, Xueyao Zhang The Chinese University of Hong Kong, Shenzhen, Shujie Liu, Yan Lu Microsoft Research Asia, Jinyu Li Microsoft, Zhizheng Wu Chinese University of Hong Kong, Shenzhen

原文摘要

Although Speech Large Language Models have achieved notable progress, a substantial modality reasoning gap remains: their reasoning performance on speech inputs is markedly weaker than on text. This gap could be associated with representational drift across Transformer layers and behavior deviations in long-chain reasoning. To address this issue, we introduce TARS, a reinforcement-learning framework that aligns text-conditioned and speech-conditioned trajectories through an asymmetric reward design. The framework employs two dense and complementary signals: representation alignment, which measures layer-wise hidden-state similarity between speech- and text-conditioned trajectories, and behavior alignment, which evaluates semantic consistency between generated outputs and reference text completions. Experiments on challenging reasoning benchmarks, including MMSU and OBQA, show that our approach significantly narrows the modality reasoning gap and achieves state-of-the-art performance among 7B-scale Speech LLMs.

cs.CL

Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表