Abstract:Reinforcement Learning with Verifiable Rewards (RLVR) is increasingly viewed as a tree pruning mechanism. However, we identify a systemic pathology termed Recursive Space Contraction (RSC), an irreversible collapse driven by the combined dynamics of positive sharpening and negative squeezing, where the sampling probability of valid alternatives vanishes. While Kullback-Leibler (KL) regularization aims to mitigate this, it imposes a rigid Shape Matching constraint that forces the policy to mimic the reference model's full density, creating a gradient conflict with the sharpening required for correctness. We propose Anchored Policy Optimization (APO), shifting the paradigm from global Shape Matching to Support Coverage. By defining a Safe Manifold based on the reference model's high-confidence support, APO permits aggressive sharpening for efficiency while selectively invoking a restorative force during error correction to prevent collapse. We theoretically derive that APO serves as a gradient-aligned mechanism to maximize support coverage, enabling an Elastic Recovery that re-inflates valid branches. Empirical evaluations on mathematical benchmarks demonstrate that APO breaks the accuracy-diversity trade-off, significantly improving Pass@1 while restoring the Pass@K diversity typically lost by standard policy gradient methods.
Abstract:A/B testing on platforms often faces challenges from network interference, where a unit's outcome depends not only on its own treatment but also on the treatments of its network neighbors. To address this, cluster-level randomization has become standard, enabling the use of network-aware estimators. These estimators typically trim the data to retain only a subset of informative units, achieving low bias under suitable conditions but often suffering from high variance. In this paper, we first demonstrate that the interior nodes - units whose neighbors all lie within the same cluster - constitute the vast majority of the post-trimming subpopulation. In light of this, we propose directly averaging over the interior nodes to construct the mean-in-interior (MII) estimator, which circumvents the delicate reweighting required by existing network-aware estimators and substantially reduces variance in classical settings. However, we show that interior nodes are often not representative of the full population, particularly in terms of network-dependent covariates, leading to notable bias. We then augment the MII estimator with a counterfactual predictor trained on the entire network, allowing us to adjust for covariate distribution shifts between the interior nodes and full population. By rearranging the expression, we reveal that our augmented MII estimator embodies an analytical form of the point estimator within prediction-powered inference framework. This insight motivates a semi-supervised lens, wherein interior nodes are treated as labeled data subject to selection bias. Extensive and challenging simulation studies demonstrate the outstanding performance of our augmented MII estimator across various settings.
Abstract:Achieving safe and stylized trajectory planning in complex real-world scenarios remains a critical challenge for autonomous driving systems. This paper proposes the SDD Planner, a diffusion-based framework designed to effectively reconcile safety constraints with driving styles in real time. The framework integrates two core modules: a Multi-Source Style-Aware Encoder, which employs distance-sensitive attention to fuse dynamic agent data and environmental contexts for heterogeneous safety-style perception; and a Style-Guided Dynamic Trajectory Generator, which adaptively modulates priority weights within the diffusion denoising process to generate user-preferred yet safe trajectories. Extensive experiments demonstrate that SDD Planner achieves state-of-the-art performance. On the StyleDrive benchmark, it improves the SM-PDMS metric by 3.9% over WoTE, the strongest baseline. Furthermore, on the NuPlan Test14 and Test14-hard benchmarks, SDD Planner ranks first with overall scores of 91.76 and 80.32, respectively, outperforming leading methods such as PLUTO. Real-vehicle closed-loop tests further confirm that SDD Planner maintains high safety standards while aligning with preset driving styles, validating its practical applicability for real-world deployment.
Abstract:The deployment of Large Language Models (LLMs) faces a critical bottleneck when handling lengthy inputs: the prohibitive memory footprint of the Key Value (KV) cache. To address this bottleneck, the token pruning paradigm leverages attention sparsity to selectively retain a small, critical subset of tokens. However, existing approaches fall short, with static methods risking irreversible information loss and dynamic strategies employing heuristics that insufficiently capture the query-dependent nature of token importance. We propose FASA, a novel framework that achieves query-aware token eviction by dynamically predicting token importance. FASA stems from a novel insight into RoPE: the discovery of functional sparsity at the frequency-chunk (FC) level. Our key finding is that a small, identifiable subset of "dominant" FCs consistently exhibits high contextual agreement with the full attention head. This provides a robust and computationally free proxy for identifying salient tokens. %making them a powerful and efficient proxy for token importance. Building on this insight, FASA first identifies a critical set of tokens using dominant FCs, and then performs focused attention computation solely on this pruned subset. % Since accessing only a small fraction of the KV cache, FASA drastically lowers memory bandwidth requirements and computational cost. Across a spectrum of long-context tasks, from sequence modeling to complex CoT reasoning, FASA consistently outperforms all token-eviction baselines and achieves near-oracle accuracy, demonstrating remarkable robustness even under constraint budgets. Notably, on LongBench-V1, FASA reaches nearly 100\% of full-KV performance when only keeping 256 tokens, and achieves 2.56$\times$ speedup using just 18.9\% of the cache on AIME24.
Abstract:Diffusion Large Language Models (DLLMs) have emerged as a powerful alternative to autoregressive models, enabling parallel token generation across multiple positions. However, preference alignment of DLLMs remains challenging due to high variance introduced by Evidence Lower Bound (ELBO)-based likelihood estimation. In this work, we propose AR-MAP, a novel transfer learning framework that leverages preference-aligned autoregressive LLMs (AR-LLMs) as implicit teachers for DLLM alignment. We reveal that DLLMs can effectively absorb alignment knowledge from AR-LLMs through simple weight scaling, exploiting the shared architectural structure between these divergent generation paradigms. Crucially, our approach circumvents the high variance and computational overhead of direct DLLM alignment and comprehensive experiments across diverse preference alignment tasks demonstrate that AR-MAP achieves competitive or superior performance compared to existing DLLM-specific alignment methods, achieving 69.08\% average score across all tasks and models. Our Code is available at https://github.com/AMAP-ML/AR-MAP.
Abstract:Multimodal reward models are crucial for aligning multimodal large language models with human preferences. Recent works have incorporated reasoning capabilities into these models, achieving promising results. However, training these models suffers from two critical challenges: (1) the inherent noise in preference datasets, which degrades model performance, and (2) the inefficiency of conventional training methods, which ignore the differences in sample difficulty. In this paper, we identify a strong correlation between response entropy and accuracy, indicating that entropy can serve as a reliable and unsupervised proxy for annotation noise and sample difficulty. Based on this insight, we propose a novel Entropy-Guided Training (EGT) approach for multimodal reasoning reward models, which combines two strategies: (1) entropy-guided data curation to mitigate the impact of unreliable samples, and (2) an entropy-guided training strategy that progressively introduces more complex examples. Extensive experiments across three benchmarks show that the EGT-trained model consistently outperforms state-of-the-art multimodal reward models.
Abstract:Unified Multimodal Models (UMMs) integrate both visual understanding and generation within a single framework. Their ultimate aspiration is to create a cycle where understanding and generation mutually reinforce each other. While recent post-training methods have successfully leveraged understanding to enhance generation, the reverse direction of utilizing generation to improve understanding remains largely unexplored. In this work, we propose UniMRG (Unified Multi-Representation Generation), a simple yet effective architecture-agnostic post-training method. UniMRG enhances the understanding capabilities of UMMs by incorporating auxiliary generation tasks. Specifically, we train UMMs to generate multiple intrinsic representations of input images, namely pixel (reconstruction), depth (geometry), and segmentation (structure), alongside standard visual understanding objectives. By synthesizing these diverse representations, UMMs capture complementary information regarding appearance, spatial relations, and structural layout. Consequently, UMMs develop a deeper and more comprehensive understanding of visual inputs. Extensive experiments across diverse UMM architectures demonstrate that our method notably enhances fine-grained perception, reduces hallucinations, and improves spatial understanding, while simultaneously boosting generation capabilities.
Abstract:Text-to-image (T2I) models have achieved remarkable success in generating high-fidelity images, but they often fail in handling complex spatial relationships, e.g., spatial perception, reasoning, or interaction. These critical aspects are largely overlooked by current benchmarks due to their short or information-sparse prompt design. In this paper, we introduce SpatialGenEval, a new benchmark designed to systematically evaluate the spatial intelligence of T2I models, covering two key aspects: (1) SpatialGenEval involves 1,230 long, information-dense prompts across 25 real-world scenes. Each prompt integrates 10 spatial sub-domains and corresponding 10 multi-choice question-answer pairs, ranging from object position and layout to occlusion and causality. Our extensive evaluation of 21 state-of-the-art models reveals that higher-order spatial reasoning remains a primary bottleneck. (2) To demonstrate that the utility of our information-dense design goes beyond simple evaluation, we also construct the SpatialT2I dataset. It contains 15,400 text-image pairs with rewritten prompts to ensure image consistency while preserving information density. Fine-tuned results on current foundation models (i.e., Stable Diffusion-XL, Uniworld-V1, OmniGen2) yield consistent performance gains (+4.2%, +5.7%, +4.4%) and more realistic effects in spatial relations, highlighting a data-centric paradigm to achieve spatial intelligence in T2I models.
Abstract:Hate Video Detection (HVD) is crucial for online ecosystems. Existing methods assume identical distributions between training (source) and inference (target) data. However, hateful content often evolves into irregular and ambiguous forms to evade censorship, resulting in substantial semantic drift and rendering previously trained models ineffective. Test-Time Adaptation (TTA) offers a solution by adapting models during inference to narrow the cross-domain gap, while conventional TTA methods target mild distribution shifts and struggle with the severe semantic drift in HVD. To tackle these challenges, we propose SCANNER, the first TTA framework tailored for HVD. Motivated by the insight that, despite the evolving nature of hateful manifestations, their underlying cores remain largely invariant (i.e., targeting is still based on characteristics like gender, race, etc), we leverage these stable cores as a bridge to connect the source and target domains. Specifically, SCANNER initially reveals the stable cores from the ambiguous layout in evolving hateful content via a principled centroid-guided alignment mechanism. To alleviate the impact of outlier-like samples that are weakly correlated with centroids during the alignment process, SCANNER enhances the prior by incorporating a sample-level adaptive centroid alignment strategy, promoting more stable adaptation. Furthermore, to mitigate semantic collapse from overly uniform outputs within clusters, SCANNER introduces an intra-cluster diversity regularization that encourages the cluster-wise semantic richness. Experiments show that SCANNER outperforms all baselines, with an average gain of 4.69% in Macro-F1 over the best.
Abstract:Reinforcement Learning with Verifiable Rewards (RLVR) offers a robust mechanism for enhancing mathematical reasoning in large models. However, we identify a systematic lack of emphasis on more challenging questions in existing methods from both algorithmic and data perspectives, despite their importance for refining underdeveloped capabilities. Algorithmically, widely used Group Relative Policy Optimization (GRPO) suffers from an implicit imbalance where the magnitude of policy updates is lower for harder questions. Data-wise, augmentation approaches primarily rephrase questions to enhance diversity without systematically increasing intrinsic difficulty. To address these issues, we propose a two-dual MathForge framework to improve mathematical reasoning by targeting harder questions from both perspectives, which comprises a Difficulty-Aware Group Policy Optimization (DGPO) algorithm and a Multi-Aspect Question Reformulation (MQR) strategy. Specifically, DGPO first rectifies the implicit imbalance in GRPO via difficulty-balanced group advantage estimation, and further prioritizes harder questions by difficulty-aware question-level weighting. Meanwhile, MQR reformulates questions across multiple aspects to increase difficulty while maintaining the original gold answer. Overall, MathForge forms a synergistic loop: MQR expands the data frontier, and DGPO effectively learns from the augmented data. Extensive experiments show that MathForge significantly outperforms existing methods on various mathematical reasoning tasks. The code and augmented data are all available at https://github.com/AMAP-ML/MathForge.