MiCU: End-to-End Smart Home Command Understanding with Large Language Model

Command understanding systems in smart home ecosystems can automate device control and substantially improve user experience. However, while they perform well on precise utterances (e.g., "turn on the bedroom light"), they struggle with ambiguous or misaligned commands (e.g., "make the bedroom cozy"). Large language models (LLMs) generalize well across various domains and can outperform traditional rule-based systems on such tasks, but their effectiveness is often constrained by scarce domain-specific data, insufficient task-specific adaptation, and high computational costs. In this paper, we propose an automated training data synthesis workflow using user logs and LLMs; then we build MiCU, a domain-specific LLM that excels at command understanding. Specifically, we employ curriculum learning to inject domain knowledge into the base LLM, then we enhance its reasoning ability via cold-start training combined with reinforcement learning (RL) guided by domain-specific thinking rules. Additionally, we introduce a token compression technique that condenses device description into a single special token, substantially reducing inference overhead and enabling \model-fast, an efficient variant optimized for long inputs. Extensive experiments show that MiCU significantly outperforms baselines, with an average accuracy gain of 20.01% across all device categories. We have deployed MiCU in the Xiaomi Home app, receiving approximately 1.7 million page views per day. Production evaluations show that MiCU reduces user correction rate by 1.57% and increases human audited accuracy by 32.05%. Our data and code are available at https://github.com/xiaomi-research/iot_spec_llm

Unified Value Alignment for Generative Recommendation in Industrial Advertising

Generative Recommendation (GR) reformulates recommendation as a next-token generation problem and has shown promise in industrial applications. However, extending GR to industrial advertising is non-trivial because the system must optimize not only user interest but also commercial value. Existing GR pipelines remain largely semantics-centric, making it difficult to align value signals across tokenization, decoding, and online serving. To address this issue, we propose UniVA, a Unified Value Alignment framework for advertising recommendation. We first introduce a Commercial SID tokenizer that injects value-related attributes into SID construction, yielding value-discriminative item representations. We then develop a Generation-as-Ranking SID Decoder jointly optimized by supervised learning and eCPM-aware reinforcement learning, which fuses value scores into next-item SID generation to perform generation and ranking in one decoding process. Finally, we design a value-guided personalized beam search that reuses generation-as-ranking logits as online value guidance and applies a personalized trie tree to constrain decoding to request-valid SID paths. Experiments on the Tencent WeChat Channels advertising platform show that UniVA achieves a 37.04\% improvement in offline Hit Rate@100 over the baseline and a 1.5\% GMV lift in online A/B tests.

Efficient Serving for Dynamic Agent Workflows with Prediction-based KV-Cache Management

LLM-based workflows compose specialized agents to execute complex tasks, and these agents usually share substantial context, allowing KV-Cache reuse to save computation. Existing approaches either manage KV-Cache at agent level and fail to exploit the reuse opportunities within workflows, or manage cache at the workflow level but assume that each workflow calls a static sequence of agents. However, practical workflows are typically dynamic, where the sequence of invoked agents and thus induced cache reuse opportunities depend on the context of each task. To serve such dynamic workflows efficiently, we build a system dubbed PBKV (\textbf{P}rediction-\textbf{B}ased \textbf{KV}-Cache Management). For each workflow, PBKV predicts the agent invocations in several future steps by fusing the guidance from historical workflows and context of the target workflow. Based on the predictions, PBKV estimates the reuse potential of cache entries and keeps the high-potential entries in GPU memory. To be robust to prediction errors, PBKV utilizes the predictions conservatively during both cache eviction and prefetching. Experiments on three workflow benchmarks show that PBKV achieves up to $1.85\times$ speedup over LRU on dynamic workflows, and up to $1.26\times$ speedup over the SOTA baseline KVFlow on the static workflow.

MTServe: Efficient Serving for Generative Recommendation Models with Hierarchical Caches

Generative recommendation (GR) offers superior modeling capabilities but suffers from prohibitive inference costs due to the repeated encoding of long user histories. While cross-request Key-Value (KV) cache reuse presents a significant optimization opportunity, the massive scale of individual user states creates a storage explosion that far exceeds physical GPU limits. We propose MTServe, a hierarchical cache management system that virtualizes GPU memory by leveraging host RAM as a scalable backup store. To bridge the I/O gap between tiers, MTServe introduces a suite of system-level optimizations, including a hybrid storage layout, an asynchronous data transfer pipeline, and a locality-driven replacement policy. On both public and production datasets, MTServe delivers up to 3.1* speedup while maintaining near-perfect hit ratios (>98.5%).

Scheduling LLM Inference with Uncertainty-Aware Output Length Predictions

To schedule LLM inference, the \textit{shortest job first} (SJF) principle is favorable by prioritizing requests with short output lengths to avoid head-of-line (HOL) blocking. Existing methods usually predict a single output length for each request to facilitate scheduling. We argue that such a \textit{point estimate} does not match the \textit{stochastic} decoding process of LLM inference, where output length is \textit{uncertain} by nature and determined by when the end-of-sequence (EOS) token is sampled. Hence, the output length of each request should be fitted with a distribution rather than a single value. With an in-depth analysis of empirical data and the stochastic decoding process, we observe that output length follows a heavy-tailed distribution and can be fitted with the log-t distribution. On this basis, we propose a simple metric called Tail Inflated Expectation (TIE) to replace the output length in SJF scheduling, which adjusts the expectation of a log-t distribution with its tail probabilities to account for the risk that a request generates long outputs. To evaluate our TIE scheduler, we compare it with three strong baselines, and the results show that TIE reduces the per-token latency by $2.31\times$ for online inference and improves throughput by $1.42\times$ for offline data generation.

Visual Attention Drifts,but Anchors Hold:Mitigating Hallucination in Multimodal Large Language Models via Cross-Layer Visual Anchors

Multimodal Large Language Models often suffer from object hallucination. While existing research utilizes attention enhancement and visual retracing, we find these works lack sufficient interpretability regarding attention drift in final model stages. In this paper, we investigate the layer wise evolution of visual features and discover that hallucination stems from deep layer attention regressing toward initial visual noise from early layers. We observe that output reliability depends on acquiring visual anchors at intermediate layers rather than final layers. Based on these insights, we propose CLVA, which stands for Cross-Layer Visual Anchors, a training free method that reinforces critical mid layer features while suppressing regressive noise. This approach effectively pulls deep layer attention back to correct visual regions by utilizing essential anchors captured from attention dynamics. We evaluate our method across diverse architectures and benchmarks, demonstrating outstanding performance without significant increase in computational time and GPU memory.

AdaCultureSafe: Adaptive Cultural Safety Grounded by Cultural Knowledge in Large Language Models

With the widespread adoption of Large Language Models (LLMs), respecting indigenous cultures becomes essential for models' culturally safety and responsible global applications. Existing studies separately consider cultural safety and cultural knowledge and neglect that the former should be grounded by the latter. This severely prevents LLMs from yielding culture-specific respectful responses. Consequently, adaptive cultural safety remains a formidable task. In this work, we propose to jointly model cultural safety and knowledge. First and foremost, cultural-safety and knowledge-paired data serve as the key prerequisite to conduct this research. However, the cultural diversity across regions and the subtlety of cultural differences pose significant challenges to the creation of such paired evaluation data. To address this issue, we propose a novel framework that integrates authoritative cultural knowledge descriptions curation, LLM-automated query generation, and heavy manual verification. Accordingly, we obtain a dataset named AdaCultureSafe containing 4.8K manually decomposed fine-grained cultural descriptions and the corresponding 48K manually verified safety- and knowledge-oriented queries. Upon the constructed dataset, we evaluate three families of popular LLMs on their cultural safety and knowledge proficiency, via which we make a critical discovery: no significant correlation exists between their cultural safety and knowledge proficiency. We then delve into the utility-related neuron activations within LLMs to investigate the potential cause of the absence of correlation, which can be attributed to the difference of the objectives of pre-training and post-alignment. We finally present a knowledge-grounded method, which significantly enhances cultural safety by enforcing the integration of knowledge into the LLM response generation process.

Beyond Static Snapshots: Dynamic Modeling and Forecasting of Group-Level Value Evolution with Large Language Models

Social simulation is critical for mining complex social dynamics and supporting data-driven decision making. LLM-based methods have emerged as powerful tools for this task by leveraging human-like social questionnaire responses to model group behaviors. Existing LLM-based approaches predominantly focus on group-level values at discrete time points, treating them as static snapshots rather than dynamic processes. However, group-level values are not fixed but shaped by long-term social changes. Modeling their dynamics is thus crucial for accurate social evolution prediction--a key challenge in both data mining and social science. This problem remains underexplored due to limited longitudinal data, group heterogeneity, and intricate historical event impacts. To bridge this gap, we propose a novel framework for group-level dynamic social simulation by integrating historical value trajectories into LLM-based human response modeling. We select China and the U.S. as representative contexts, conducting stratified simulations across four core sociodemographic dimensions (gender, age, education, income). Using the World Values Survey, we construct a multi-wave, group-level longitudinal dataset to capture historical value evolution, and then propose the first event-based prediction method for this task, unifying social events, current value states, and group attributes into a single framework. Evaluations across five LLM families show substantial gains: a maximum 30.88\% improvement on seen questions and 33.97\% on unseen questions over the Vanilla baseline. We further find notable cross-group heterogeneity: U.S. groups are more volatile than Chinese groups, and younger groups in both countries are more sensitive to external changes. These findings advance LLM-based social simulation and provide new insights for social scientists to understand and predict social value changes.

End-to-End Semantic ID Generation for Generative Advertisement Recommendation

Generative Recommendation (GR) has excelled by framing recommendation as next-token prediction. This paradigm relies on Semantic IDs (SIDs) to tokenize large-scale items into discrete sequences. Existing GR approaches predominantly generate SIDs via Residual Quantization (RQ), where items are encoded into embeddings and then quantized to discrete SIDs. However, this paradigm suffers from inherent limitations: 1) Objective misalignment and semantic degradation stemming from the two-stage compression; 2) Error accumulation inherent in the structure of RQ. To address these limitations, we propose UniSID, a Unified SID generation framework for generative advertisement recommendation. Specifically, we jointly optimize embeddings and SIDs in an end-to-end manner from raw advertising data, enabling semantic information to flow directly into the SID space and thus addressing the inherent limitations of the two-stage cascading compression paradigm. To capture fine-grained semantics, a multi-granularity contrastive learning strategy is introduced to align distinct items across SID levels. Finally, a summary-based ad reconstruction mechanism is proposed to encourage SIDs to capture high-level semantic information that is not explicitly present in advertising contexts. Experiments demonstrate that UniSID consistently outperforms state-of-the-art SID generation methods, yielding up to a 4.62% improvement in Hit Rate metrics across downstream advertising scenarios compared to the strongest baseline.

Sparsity-Aware Unlearning for Large Language Models

Large Language Models (LLMs) inevitably memorize sensitive information during training, posing significant privacy risks. Machine unlearning has emerged as a promising solution to selectively remove such information without full retraining. However, existing methods are designed for dense models and overlook model sparsification-an essential technique for efficient LLM deployment. We find that unlearning effectiveness degrades substantially on sparse models. Through empirical analysis, we reveal that this degradation occurs because existing unlearning methods require updating all parameters, yet sparsification prunes substantial weights to zero, fundamentally limiting the model's forgetting capacity. To address this challenge, we propose Sparsity-Aware Unlearning (SAU), which decouples unlearning from sparsification objectives through gradient masking that redirects updates to surviving weights, combined with importance-aware redistribution to compensate for pruned parameters. Extensive experiments demonstrate that SAU significantly outperforms existing methods on sparse LLMs, achieving effective forgetting while preserving model utility.