In local-life service platforms, query suggestion reduces user effort by generating candidate queries from input prefixes. Traditional multi-stage systems rely heavily on historical popular queries, limiting their ability to capture long-tail and emerging demand. Although LLMs provide strong semantic generalization, their deployment in local-life services faces three challenges: insufficient city-preference awareness, exposure bias in preference optimization, and strict online latency constraints. We propose LocalSUG, an LLM-based query suggestion framework for local-life services. LocalSUG mines city-preference-enhanced candidates from term co-occurrence and injects them into prompts as dynamic references rather than fusing them into model parameters. This allows the model to adapt to changing city preferences, such as merchant openings or closures, while reducing stale or locally invalid suggestions. We further introduce a beam-search-driven GRPO algorithm to align training with inference-time decoding and optimize relevance together with business-oriented rewards. Finally, quality-aware beam acceleration and vocabulary pruning reduce online latency while preserving generation quality. Offline evaluations and large-scale online A/B testing show that LocalSUG improves CTR by +0.35

Troubleshooting performance problems of large model training (LMT) is immensely challenging, due to unprecedented scales of modern GPU clusters, the complexity of software-hardware interactions, and the data intensity of the training process. Existing troubleshooting approaches designed for traditional distributed systems or datacenter networks fall short and can hardly apply to real-world training systems. In this paper, we present EROICA, the first online troubleshooting system that provides both fine-grained observation based on profiling, and coverage of all machines in GPU clusters, to diagnose performance issues in production, including both hardware and software problems (or the mixture of both). EROICA effectively summarizes runtime behavior patterns of LMT function executions via online profiling, and leverages differential observability to localize the root cause with minimal production impact. EROICA has been deployed as a production service for large-scale GPU clusters of 100,000 GPUs for 1.5 years. It has diagnosed a variety of difficult performance issues with 97.5

The integration of Large Language Model (LLM) agents is transforming recommender systems from simple query-item matching towards deeply personalized and interactive recommendations. Reinforcement Learning (RL) provides an essential framework for the optimization of these agents in recommendation tasks. However, current methodologies remain limited by a reliance on single dimensional outcome-based rewards that focus exclusively on final user interactions, overlooking critical intermediate capabilities, such as instruction following and complex intent understanding. Despite the necessity for designing multi-dimensional reward, the field lacks a standardized benchmark to facilitate this development. To bridge this gap, we introduce RecRM-Bench, the largest and most comprehensive benchmark to date for agentic recommender systems. It comprises over 1 million structured entries across four core evaluation dimensions: instruction following, factual consistency, query-item relevance, and fine-grained user behavior prediction. By supporting comprehensive assessment from syntactic compliance to complex intent grounding and preference modeling, RecRM-Bench provides a foundational dataset for training sophisticated reward models. Furthermore, we propose a systematic framework for the construction of multi-dimensional reward models and the integration of a hybrid reward function, establishing a robust foundation for developing reliable and highly capable agentic recommender systems. The complete RecRM-Bench dataset is publicly available at https://huggingface.co/datasets/wwzeng/RecRM-Bench.

Compositional e-commerce queries express multiple requirements that must hold jointly, yet existing rerankers collapse these constraints into aggregate relevance and often promote topical near misses over feasible products. In this paper, we introduce REAlign, a novel requirement-evidence-aligned reranking framework that explicitly connects typed query requirements with visible evidence. REAlign distinguishes satisfied, violated, and unsupported conditions, constructs requirement-targeted contrasts that expose failure modes, and optimizes duplicate-free partial rankings through Requirement-Aware Group-Relative Policy Optimization. Its list utility preserves relevance while incorporating requirement satisfaction, evidence support, material violations, and output validity. Experiments on two fixed-pool e-commerce benchmarks show consistent improvements over strong supervised and policy-optimization baselines under matched training budgets, with fewer violations among top-ranked candidates and larger gains at shallow ranks. Controlled ablations confirm the complementary value of requirement modeling, evidence grounding, and decomposed optimization.

Large reasoning models (LRMs) have recently achieved significant progress in complex reasoning tasks, aided by reinforcement learning with verifiable rewards. However, LRMs often suffer from overthinking, expending excessive computation on simple problems and reducing efficiency. Existing efficient reasoning methods typically require accurate task assessment to preset token budgets or select reasoning modes, which limits their flexibility and reliability. In this work, we revisit the essence of overthinking and identify that encouraging effective steps while penalizing ineffective ones is key to its solution. To this end, we propose a novel rule-based verifiable stepwise reward mechanism (VSRM), which assigns rewards based on the performance of intermediate states in the reasoning trajectory. This approach is intuitive and naturally fits the step-by-step nature of reasoning tasks. We conduct extensive experiments on standard mathematical reasoning benchmarks, including AIME24 and AIME25, by integrating VSRM with PPO and Reinforce++. Results show that our method achieves substantial output length reduction while maintaining original reasoning performance, striking an optimal balance between efficiency and accuracy. Further analysis of overthinking frequency and pass@k score before and after training demonstrates that our approach indeed effectively suppresses ineffective steps and encourages effective reasoning, fundamentally alleviating the overthinking problem.

Logical consistency, the requirement that statements remain non-contradictory under logical rules, is fundamental for trustworthy reasoning, yet current LLMs often fail to maintain it even on simple inference tasks. Existing benchmarks for LLM logical consistency are not scalable, not diverse, and not challenging, with state-of-the-art models already surpassing 95% accuracy. LogiConBench is the first benchmark that (1) generates unlimited logical rule combinations with precise labels, (2) provides controllable-depth graphs with explicit reasoning paths, and (3) remains challenging for state-of-the-art LLMs. To achieve this, LogiConBench automatically generates logical graphs where nodes represent symbolic propositions and edges denote reasoning relations. From these graphs, it samples lists of propositions, extracts reasoning paths, determines all consistent label lists, and translates them into diverse natural language expressions. While we release a 280K-sample corpus in this work, the framework can be scaled to generate unlimited data. To strengthen its evaluative significance, we evaluate 14 frontier LLMs on two tasks with varying difficulty levels, and find that the Enumerative task remains extremely challenging, with the best exact accuracy as only 34%. Our code and data are available at https://anonymous.4open.science/r/LogiConBench-11D1/.

Generative Recommendation (GR) formulates recommendation as autoregressive generation over discrete semantic identifiers (IDs). Although recent multimodal GR methods improve semantic ID construction with visual and textual information, they typically require item-level paired observations, restricting tokenization to the intersection of modality availability. Moreover, incorporating unpaired observations is nontrivial because small representation shifts may cross quantization boundaries and produce incompatible identifier sequences. To address this challenge, we propose Unpaired Modality-Agnostic Generative Recommendation (UnpairGR), which learns a unified semantic-ID space from paired, image-only, and text-only observations. UnpairGR confines modality-specific processing to lightweight input projections while sharing the subsequent Transformer and residual codebooks across all observation conditions. Paired observations establish a reliability-guided cross-modal consensus, whereas unimodal observations directly refine the same representations and codes. The learned tokenizer is then fixed to provide stationary targets for a single autoregressive recommender, without feature imputation, modality-specific codebooks, or fallback mappings. Extensive experiments on three benchmark datasets demonstrate that UnpairGR consistently improves recommendation performance under both fully observed and incomplete-observation settings.

Large Language Models (LLMs) are prone to logical hallucinations and stochastic drifts during long-chain reasoning. While Classifier-Free Guidance (CFG) can improve instruction adherence, standard static implementations often cause semantic dilution and linguistic degradation. We propose SPREG (Structured Plan-guided Real-time Entropy Gating), a lightweight inference-time framework for surgical error rectification. SPREG employs an adaptive dual-threshold mechanism to monitor real-time entropy, identifying sudden “entropy spikes” as reliable indicators of logical failure. Upon detection, it triggers a dynamic repair by replacing uninformative null-priors with reference distributions synthesized from historical high-confidence states. By modulating guidance intensity according to structured reasoning stages (e.g., Action, Observation), SPREG steers the model back to a stable manifold without compromising fluency. Our experiments demonstrate significant gains, notably a 20.0

Recommendation is crucial for both user experience and company revenue in Meituan as a leading lifestyle company, and generative recommendation models (GRMs) are shown to produce quality recommendations recently. However, existing systems are limited by insufficient functionality support and inefficient implementations for training GRMs in industrial scenarios. As such, we introduce MTGenRec as an efficient and scalable system for GRM training. Specifically, to handle real-time insertions/deletions of sparse embeddings, MTGenRec employs dynamic hash tables to replace static ones. To improve training efficiency, MTGenRec conducts dynamic sequence balancing to address the computation load imbalances among GPUs and adopts feature ID deduplication alongside automatic table merging to accelerate embedding lookup. Extensive experiments show that MTGenRec improves training throughput by 1.6x - 2.4x while achieving good scalability when running over 100 GPUs. MTGenRec has been deployed for many applications in Meituan and is now handling hundreds of millions of requests on a daily basis. On the delivery platform, we observe a 1.22% growth in user order volume and a 1.31% enhancement in online PV_CTR.

Reinforcement Learning with Verifiable Rewards (RLVR) is an effective paradigm for improving the reasoning capabilities of large language models. However, existing RLVR methods utilize rollouts in an indiscriminate and short-horizon manner: responses of heterogeneous quality within each prompt are treated uniformly, and historical rollouts are discarded after a single use. This leads to noisy supervision, poor sample efficiency, and suboptimal policy updates. We address these issues by formulating rollout scheduling in RLVR as a contextual bandit problem and proposing a unified neural scheduling framework that adaptively selects high-value rollouts throughout training. Each rollout is treated as an arm whose reward is defined by the induced performance gain between consecutive optimization steps. The resulting scheduler supports both noise-aware intra-group selection and adaptive global reuse of historical rollouts within a single principled framework. We provide theoretical justification by deriving sublinear regret bounds and showing that enlarging the rollout buffer improves the achievable performance upper bound. Experiments on six mathematical reasoning benchmarks demonstrate consistent gains in performance and training efficiency across multiple RLVR optimization methods.

Exploration is not an optional behavior in natural intelligence; it is an evolutionary principle underlying the emergence and adaptation of intelligence. Curiosity, play, and deliberate probing emerge as evolved responses to uncertainty, enabling organisms to construct internal models, expand competence, and preserve adaptability in changing environments. We argue that this evolutionary logic is equally indispensable for artificial general intelligence (AGI): exploration is not a heuristic appended to learning, but the mechanism through which generality becomes possible. We develop a unified view of Epistemic Exploration for agentic systems: the capacity of an agent to actively acquire information that reduces uncertainty about the world, seek experiences at the boundary of its current capabilities and convert them into durable capability improvement, and preserve epistemic reachability as the readiness and ability to adapt when the world changes. This view yields three criteria: Information Gain, Value Improvement, and Epistemic Reachability. We then introduce an exploration-centered five-level trajectory toward AGI, in which each level is characterised by a distinct exploration capacity and exploration serves as the transition mechanism among levels: • Responder: minimal or no explicit epistemic exploration; the system mainly relies on learned input-output mappings and local token-level variation. • Reasoner: reasoning-space exploration enables hypothesis search, deliberate reasoning trajectories, branching, backtracking, and self-verification beyond reactive response generation. • Agent: interaction-space exploration extends internal deliberation into embodied perception, tool use, memory, and closed-loop action under partial observability. • Prospector: imagination-space exploration uses learned world models to simulate counterfactual futures, reduce the cost and risk of real interaction, and support long-horizon policy improvement. • Ecosystem: coordination-space exploration enables collectives of heterogeneous agents to co-evolve roles, shared representations, communications, and collaborative strategies beyond the limits of any single agent. Finally, we conclude with evaluation principles and open challenges for building exploration-centric agents that continually reduce uncertainty, improve their own capabilities, and maintain readiness to adapt beyond predefined tasks.

Reinforcement Learning from Verifier Rewards (RLVR) has emerged as a widely used approach for post-training large language models on reasoning tasks, with group-based methods such as GRPO and its variants gaining broad adoption. These methods rely on group-relative advantage estimation to avoid learned critics, yet its theoretical properties remain poorly understood. In this work, we uncover a fundamental issue of group-based RL: the group-relative advantage estimator is inherently biased relative to the true (expected) advantage. We provide the first theoretical analysis showing that it systematically underestimates advantages for hard prompts and overestimates them for easy prompts, leading to imbalanced exploration and exploitation. To address this issue, we propose History-Aware Adaptive Difficulty Weighting (HA-DW), an adaptive reweighting scheme that adjusts advantage estimates based on an evolving difficulty anchor and training dynamics. Both theoretical analysis and experiments on five mathematical reasoning benchmarks demonstrate that HA-DW consistently improves performance when integrated into GRPO and its variants. Our results suggest that correcting biased advantage estimation is critical for robust and efficient RLVR training.

Recent advancements in Vision-Language Models (VLMs) have revolutionized general visual understanding. However, their application in the food domain remains constrained by benchmarks that rely on coarse-grained categories, single-view imagery, and inaccurate metadata. To bridge this gap, we introduce DiningBench, a hierarchical, multi-view benchmark designed to evaluate VLMs across three levels of cognitive complexity: Fine-Grained Classification, Nutrition Estimation, and Visual Question Answering. Unlike previous datasets, DiningBench comprises 3,021 distinct dishes with an average of 5.27 images per entry, incorporating fine-grained "hard" negatives from identical menus and rigorous, verification-based nutritional data. We conduct an extensive evaluation of 29 state-of-the-art open-source and proprietary models. Our experiments reveal that while current VLMs excel at general reasoning, they struggle significantly with fine-grained visual discrimination and precise nutritional reasoning. Furthermore, we systematically investigate the impact of multi-view inputs and Chain-of-Thought reasoning, identifying five primary failure modes. DiningBench serves as a challenging testbed to drive the next generation of food-centric VLM research. All codes are released in https://github.com/meituan/DiningBench.

Generative retrieval (GR) has gained significant attention as an effective paradigm that integrates the capabilities of large language models (LLMs). It generally consists of two stages: constructing discrete semantic identifiers (IDs) for documents and retrieving documents by autoregressively generating ID tokens.The core challenge in GR is how to construct document IDs (DocIDS) with strong representational power. Good IDs should exhibit two key properties: similar documents should have more similar IDs, and each document should maintain a distinct and unique ID.However, most existing methods ignore native category information, which is common and critical in E-commerce. Therefore, we propose a novel ID learning method, CAtegory-Tree Integrated Document IDentifier (CAT-ID^2), incorporating prior category information into the semantic IDs.CAT-ID^2 includes three key modules: a Hierarchical Class Constraint Loss to integrate category information layer by layer during quantization, a Cluster Scale Constraint Loss for uniform ID token distribution, and a Dispersion Loss to improve the distinction of reconstructed documents. These components enable CAT-ID^2 to generate IDs that make similar documents more alike while preserving the uniqueness of different documents' representations.Extensive offline and online experiments confirm the effectiveness of our method, with online A/B tests showing a 0.33

The limited capacity for fine-grained visual perception presents a critical bottleneck for Vision-Language Models (VLMs) in real-world applications. Addressing this is challenging due to the scarcity of high-quality data and the limitations of existing methods: supervised fine-tuning (SFT) often compromises general capabilities, while reinforcement fine-tuning (RFT) prioritizes textual reasoning over visual perception. To bridge this gap, we propose a novel two-stage task that structures visual perception learning as a coarse-to-fine progressive process. Based on this task formulation, we develop ViPER, a self-bootstrapping framework specifically designed to enable iterative evolution through self-critiquing and self-prediction. By synergistically integrating image-level and instance-level reconstruction with a two-stage reinforcement learning strategy, ViPER establishes a closed-loop training paradigm, where internally synthesized data directly fuel the enhancement of perceptual ability. Applied to the Qwen2.5-VL family, ViPER produces the Qwen-Viper series. With an average gain of 1.7\% on seven comprehensive benchmarks spanning various tasks and up to 6.0\% on fine-grained perception, Qwen-Viper consistently demonstrates superior performance across different vision-language scenarios while maintaining generalizability. Beyond enabling self-improvement in perceptual capabilities, ViPER provides concrete evidence for the reciprocal relationship between generation and understanding, a breakthrough to developing more autonomous and capable VLMs.

Reinforcement Learning with Verifiable Rewards (RLVR) enhances reasoning of Large Language Models (LLMs) but usually exhibits limited generation diversity due to the over-incentivization of positive rewards. Although methods like Negative Sample Reinforcement (NSR) mitigate this issue by upweighting penalty from negative samples, they may suppress the semantic distributions shared between positive and negative responses. To boost reasoning ability without losing diversity, this paper proposes negative sample projection Residual Reinforcement Learning (ResRL) that decouples similar semantic distributions among positive and negative responses. We theoretically link Lazy Likelihood Displacement (LLD) to negative-positive head-gradient interference and derive a single-forward proxy that upper-bounds representation alignment to guide conservative advantage reweighting. ResRL then projects negative-token hidden representations onto an SVD-based low-rank positive subspace and uses projection residuals to modulate negative gradients, improving reasoning while preserving diversity and outperforming strong baselines on average across twelve benchmarks spanning Mathematics, Code, Agent Tasks, and Function Calling. Notably, ResRL surpasses NSR on mathematical reasoning by 9.4\% in Avg@16 and 7.0\% in Pass@128. Code is available at https://anonymous.4open.science/r/ResRL.

Terminal agents extend Large Language Models with the ability to execute tasks directly in command-line environments, but their progress is bottlenecked by the scarcity of high-quality training data. Existing approaches bootstrap from partial sources such as human-defined seeds or GitHub repositories to instantiate one component and then complete the rest, producing tasks confined to narrow seed distributions, environments misaligned with task semantics, and inefficient trajectories from unguided exploration. To address these limitations, we introduce Terminal-World, a fully automated pipeline that uses agent skills as the central synthesis primitive, which jointly encode what to accomplish, when to apply (preconditions and environment state), and how to execute, enabling task instructions, environments, and teacher trajectories to be co-derived. To further broaden the synthesis space, Terminal-World composes skills into skill teams and skill graphs for multi-role and cross-domain task synthesis. Using this pipeline, we construct 5,723 training environments and train Terminal-World-8B/14B/32B, evaluated across 6 benchmarks where the Terminal-World series consistently outperforms terminal-agent baselines. Notably, using the same teacher model and only 1.2