Search arXiv⌕ Search

arXiv subjects

Yuhang Zhou

Publications and source records attributed to Yuhang Zhou.

At least 19 recordsLinked to original sources
arXiv2026

UniPolicy: Unified Objective-Specific Policies for Generative Search Advertising

Search advertising connects user intent with commercial content and plays a critical role in platform monetization. Recent systems typically align pretrained generative models with a single business reward, such as eCPM, or use naive reward fusion for preliminary multi-objective alignment. However, an ideal search advertising system must jointly account for heterogeneous objectives, including relevance, click propensity, and commercial value, to balance user experience and business value while mitigating globally suboptimal performance caused by gradient competition. We propose UniPolicy, an objective-aware multi-policy alignment framework. UniPolicy combines objective-specific prefix tokens, sparse MoE-LoRA routing, and objective-specific residual FFNs to hierarchically decouple parameters within a shared backbone, providing differentiated parameter and policy-expression spaces for different business objectives. It further constructs pairwise preferences from multi-stage behavioral feedback, supplementing the relative preference information in exposed-but-unclicked samples and strengthening the relative advantage of clicked candidates in the generation distribution. At inference, UniPolicy supports parallel, business-customizable multi-policy beam search, flexibly allocating candidate quotas across objectives under a fixed retrieval budget. Large-scale offline experiments show that UniPolicy delivers balanced improvements across multiple metrics while preserving retrieval quality, outperforming single-objective reinforcement learning and naive reward-fusion baselines. In a 7-day online A/B test on a real search advertising system, UniPolicy improves CTR by 0.71%, RPS by 1.58%, and advertising revenue by 1.32%, while maintaining stable serving latency.

cs.CL↗
arXiv2026

CORAL: An LLM-Native Harness for Production Recommender Systems

Production recommender systems shape what billions of people see, and sustaining their performance requires continual optimization: as content, user behavior, and upstream models shift, the choices governing retrieval, ranking, and serving must be revisited. Traditionally, human engineers test such changes through online experiments--a slow, reactive process limited by engineering effort, leaving parts of the system unrevised as conditions change. Although large language models have been applied to ranking, user modeling, and offline model development, few systems place an agent in a continual closed loop that acts on a live recommender and learns from the measured effects of its decisions. We present CORAL (Constraint-Optimized Recommender via an Agentic Loop), an LLM-native harness that closes this loop: each cycle, the agent observes operating signals, reasons over a memory of past decisions and outcomes, and invokes tools--including a numerical optimizer that keeps changes within a fixed operating budget--to reconfigure the recommender, with measured outcomes informing the next cycle. We formulate this as a partially observed, non-stationary, constrained optimization problem in which the policy improves in context, without parameter updates, from its prior actions. Across two large-scale social platforms, evaluated with A/B experiments, the same harness improves engagement at no additional serving cost on one and reduces serving cost without degrading engagement on the other, spanning the engagement-efficiency frontier. Performance improves as the loop iterates, suggesting that a single agentic loop can automate continual optimization work traditionally performed by human algorithm engineers under explicit guardrails.

cs.CL↗
arXiv2026

Uniform non-homogeneous bundles on quadrics

Let $X$ be an $n$-dimensional generalized Grassmannian not isomorphic to $\mathbb{P}^n$. We prove that $k(X)\le n-1$, where $k(X)$ denotes the maximal integer such that every uniform bundle on $X$ of rank at most $k(X)$ is homogeneous. In particular, for smooth quadrics $\mathbb{Q}^n$, we have $k(\mathbb{Q}^n)=n-1$ for odd $n$, and $n-2\le k(\mathbb{Q}^n)\le n-1$ for even $n$. We classify uniform rank $n$ bundles on $\mathbb{Q}^{n}$ for $n=3$, $5$. Furthermore, we characterize projective spaces among generalized Grassmannians in terms of uniform bundles.

math.AG↗
arXiv2026

CAR: Query-Guided Confidence-Aware Reranking for Retrieval-Augmented Generation

Retrieval-augmented generation (RAG) relies on evidence ranking to determine what information is exposed to the generator, yet existing retrieval and reranking methods primarily estimate query--document relevance. Relevance, however, is not equivalent to generator-side usefulness: a relevant passage may introduce ambiguity or distraction, whereas a lower-ranked passage may stabilize the generator's answer. We present CAR (Confidence-Aware Reranking), a training-free rank-correction framework that uses query-only answer stability as a control and measures each candidate by the change it induces in sampled-answer semantic stability. This controlled contrast estimates a document's marginal contribution to generator behavior without treating semantic stability as relevance or calibrated correctness. CAR converts these confidence changes into coarse precedence constraints and returns the feasible ranking with minimum Kendall distance from the baseline, preserving existing pairwise preferences unless generator-side evidence supports reversing them. Experiments on NQ, HotpotQA and FEVER across sparse and dense retrievers, seven ranking methods and three generator families show robust improvements. In the BM25-centered main analysis, CAR achieves a \textbf{+5.53\% mean relative NDCG@5 gain}; on the fixed NQ-answerable downstream evaluation, it improves token-level F1 by \textbf{+0.43 points}, with ranking and generation gains strongly aligned across rankers ($ρ= 0.93$). These results position CAR as a deployment-friendly, generator-aware correction layer that complements relevance while preserving informative prior rankings. CAR requires neither task-specific training nor access to model internals such as logits or hidden states, making it applicable to black-box LLMs through generated outputs alone.

cs.CL↗
arXiv2026

Adaptive Matrix Multiplication for Dynamic Shapes on Ascend NPUs

Matrix Multiplication (MatMul) faces a "generalization crisis" driven by highly dynamic tensor shapes. This crisis is particularly acute on Ascend NPUs, where explicitly controlled architectures and strict physical constraints render existing GPU-centric optimizations ineffective. To resolve this, we propose AdaptCore, an adaptive framework for universally high-performance MatMul on Ascend NPUs. AdaptCore systematically decouples operator optimization into spatial tiling and instruction orchestration. It first maps dynamic shapes into a hardware-aware 2D tiling taxonomy to balance on-chip capacity limits and multi-core parallelism. Furthermore, it integrates a composable optimization library with a deterministic analytical performance model. By mathematically evaluating hardware state mutations, AdaptCore proactively selects and caches optimal implementations, enabling O(1) overhead runtime dispatching. Evaluations demonstrate that AdaptCore delivers a remarkable 1.85x mean speedup across 80,000 input shapes, and achieves up to a 1.48x acceleration in representative end-to-end models over the highly-tuned native vendor library (ACLNN).

cs.AR↗
arXiv2026

MoE Proxy Models for Low-Cost Failure Reproduction and Diagnosis in LLM RL Post-Training

Reinforcement learning (RL) post-training of large language models (LLMs) is computationally intensive and involves complex system pipelines with substantial debugging overhead. In practice, factors such as framework adaptation, numerical precision, and operator implementation can cause failures, including gradient overflow and loss divergence. Reproducing such failures directly on large models requires considerable time and computational resources. This paper systematically analyzes failures encountered during large-scale RL training on the Huawei Ascend platform, summarizes representative failure types, and identifies three model-side factors relevant to fault reproduction. Based on these factors, we propose a proxy-model construction method for low-cost fault investigation and auxiliary diagnosis. It employs structure-preserving, clustering-based expert pruning to select representative experts while retaining the model's backbone architecture, routing mechanism, and basic task capabilities. Our experimental results show that the proxy models reduce accelerator requirements by 50%-87.5% and achieve up to a 33.3x reduction in per-step NPU-hour cost, while preserving major training dynamics and reproducing fault responses consistent with the original models. Overall, the proxy models can serve as low-cost surrogates for fault reproduction, targeted validation, and auxiliary diagnosis in RL post-training.

cs.LG↗
arXiv2026

SmartSwap: Swap-Based Memory Optimization for LLM Training under Varying Operator Sequences

The increasing size of large language models (LLMs) has led to a surge in memory requirements during training, often exceeding the capacity of high-bandwidth memory (HBM). Swap-based memory optimization incurs neither accuracy loss nor additional end-to-end overhead when effectively overlapped, thus being an attractive solution. However, existing swap methods assume consistent operator sequences, which is impractical in Eager Mode, where operator sequences can vary during change. We propose Chameleon, which redesigns the end-to-end process of swap-based memory optimization and is the first work to consider varying operator sequences in Eager Mode. Chameleon (i) introduces a lightweight online profiler to enable continuous profiling for monitoring operator sequences, (ii) generates effective swap policies with limited operator information, and (iii) optimizes the policy execution module for accurate policy application and better performance. Experimental results demonstrate that Chameleon reduces profiling overhead by 84.25%, enables training models up to 4x larger than hardware memory while adapting to changes in operator sequences, improves performance by up to 38.94% compared to recomputation or high-degree parallelism.

cs.DC↗
arXiv2026

Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents

In long-horizon tasks, decision-relevant state is often scattered across an expanding trajectory, while the action agent must surface it and act. As trajectories grow, task requirements, environment facts, prior attempts, diagnoses, and open subgoals can be buried in the context window or pushed beyond it, failing to influence decisions when needed. We call this failure mode "behavioral state decay". We study memory as an active intervention mechanism rather than passive retrieval. A separate memory agent runs alongside an unmodified action agent, updating a structured memory bank from the recent trajectory and deciding whether to inject a memory-grounded reminder or remain silent. The module is plug-and-play with frontier action agents and existing agent harnesses. Across Terminal-Bench 2.0 and $τ^2$-Bench, it improves pass@1 for both weaker and stronger action agents, with gains of +8.3 pp on Terminal-Bench and +6.8 pp on $τ^2$-Bench. Ablations show that selective intervention outperforms passive bank exposure, always-on injection, advisor-only guidance, and general retrieval. As an early step toward open-weight memory policies, we train Qwen3.5-27B on SETA using SFT and GRPO, improving validation reward and achieving partial transfer to Terminal-Bench.

cs.AI↗
arXiv2026

TurnOPD: Making On-Policy Distillation Turn-Aware for Efficient Long-Horizon Agent Training

On-policy distillation (OPD) trains a student policy by matching a stronger teacher on the student's own trajectories, offering a promising framework for language agent training. However, its application to long-horizon agentic tasks remains insufficiently explored. We identify two key inefficiencies in vanilla agent OPD: (1) full-horizon rollouts often waste wall-clock resources on tail turns that provide weak and noisy KL supervision, and (2) trajectory-level KL objectives concentrate most of the loss on shallow tokens, leaving deeper decision turns under-trained once initial behaviors are aligned. To address these challenges, we propose TurnOPD, a turn-level budgeting strategy for efficient on-policy distillation of long-horizon agents. TurnOPD consists of two budget controllers: adaptive rollout-depth budgeting, which uses probe-based turn statistics to determine rollout length, and progressive turn-normalized loss budgeting, which gradually shifts KL weighting from token-level to turn-balanced supervision. Experiments on ALFWorld, WebShop, and Multi-Hop Search with task-specialized teacher models show that TurnOPD achieves superior validation accuracy under equal wall-clock training budgets and advances the accuracy--time frontier beyond vanilla OPD.

cs.AI↗
arXiv2026

COMFYCLAW: Self-Evolving Skill Harnesses for Image Generation Workflows

Agents are increasingly used to construct workflows and assist humans in completing recurring tasks more efficiently. As these workflows become repeated and domain-specific, agent memory and reusable skills become increasingly important: agents should be able to recall workflow patterns, execution constraints, and user preferences from previous runs. We study this problem in workflow-based image generation and introduce COMFYCLAW, an agentic skill evolution harness for controlling ComfyUI workflows. COMFYCLAW formulates workflow construction as typed graph editing, exposes tools organized by construction stage, automatically reverts invalid edits, and uses a region-level vision-language model (VLM) verifier to translate visual failures into actionable repair suggestions. The framework further evolves a progressively disclosed skill library, where trajectories, execution errors, and verifier feedback from previous runs are distilled into reusable Agent Skills. Across four benchmark splits, three agent models, and two image backbones, COMFYCLAW achieves the best average image-generation evaluation score across all six agent configurations, outperforming a verifier-only baseline without skill evolution. Human annotations further show that annotators prefer COMFYCLAW over variants without skill evolution. Our results suggest that skill evolution is an effective mechanism for improving agent reliability and performance in recurring visual workflow construction.

cs.AI↗
arXiv2026

SAGE-OPD: Selective Agent-Guided Intervention for Multi-Turn On-Policy Distillation

On-policy distillation (OPD) improves student models by training them on trajectories induced by their own policy, making it a promising approach for mitigating exposure bias in agent training. However, most OPD studies focus on single-turn settings, while realistic LLM agents interact with environments over multiple turns. In this regime, early errors can alter future observations and compound across the trajectory, and standard dense token-level OPD becomes brittle, as it may over-penalize semantically valid alternatives, reinforce local degeneracies such as repeated actions, and propagate unreliable teacher supervision on off-distribution histories. We propose SAGE-OPD, a verifier-free selective intervention framework specifically designed for multi-turn OPD. Instead of applying teacher supervision uniformly across all turns, SAGE-OPD first observes environment feedback and uses teacher judgment to decide whether each student response should be skipped or intervened on. To further address compounding errors, SAGE-OPD weights token-level distillation by teacher confidence, reducing the influence of uncertain teacher distributions on corrupted or ambiguous histories. Finally, SAGE-OPD applies loss normalization to preserve the overall loss scale of standard OPD while retaining selective turn-level weighting. Experiments on agent tasks show that SAGE-OPD consistently improves over baselines, achieving up to a 13.3% relative improvement in ALFWorld unseen success rate over standard OPD. Ablation studies further demonstrate that turn-level intervention, teacher confidence weighting, and loss normalization provide complementary benefits. Our results suggest that effective multi-turn OPD should remain on-policy, but teacher supervision should be selectively allocated to turns where intervention is necessary and reliable.

cs.CL↗
arXiv2026

Token-Level LLM Collaboration via FusionRoute

Large language models (LLMs) exhibit strengths across diverse domains. However, achieving strong performance across these domains with a single general-purpose model typically requires scaling to sizes that are prohibitively expensive to train and deploy. On the other hand, while smaller domain-specialized models are much more efficient, they struggle to generalize beyond their training distributions. To address this dilemma, we propose FusionRoute, a robust and effective token-level multi-LLM collaboration framework in which a lightweight router simultaneously (i) selects the most suitable expert at each decoding step and (ii) contributes a complementary logit that refines or corrects the selected expert's next-token distribution via logit addition. Unlike existing token-level collaboration methods that rely solely on fixed expert outputs, we provide a theoretical analysis showing that pure expert-only routing is fundamentally limited: unless strong global coverage assumptions hold, it cannot in general realize the optimal decoding policy. By augmenting expert selection with a trainable complementary generator, FusionRoute expands the effective policy class and enables recovery of optimal value functions under mild conditions. Empirically, across both Llama-3 and Gemma-2 families and diverse benchmarks spanning mathematical reasoning, code generation, and instruction following, FusionRoute outperforms both sequence- and token-level collaboration, model merging, and direct fine-tuning, while remaining competitive with domain experts on their respective tasks.

cs.AI↗
arXiv2026

OmniOPD: Logit-Free On-Policy Distillation via Speculative Verification

On-Policy Distillation (OPD) trains a student model on its own generative trajectories under dense token-level feedback from a stronger teacher, mitigating both the off-policy distribution shift of Supervised Fine-Tuning (SFT) and the sparse credit assignment of Reinforcement Learning (RL). However, standard OPD faces two coupled limitations. First, it requires direct access to the teacher's token-level logits, excluding a broad class of capable proprietary models from serving as teachers. Second, the token-level logit signal itself is brittle, depending on a narrow overlap of plausible next tokens between teacher and student, and prone to amplifying degenerate patterns such as repetition loops. In this paper, we introduce OmniOPD, a novel framework that addresses both limitations through a logit-free, chunk-level supervision signal. OmniOPD replaces deterministic logit matching with Monte Carlo rollouts that approximate the teacher's local preferences through a continuous semantic similarity metric over multi-token chunks, and concentrates this supervision via a peak-entropy scheduler that audits the student only at its high-uncertainty reasoning forks. A Dirichlet-Multinomial Bayesian prior and a base-model KL anchor further bound the variance of discrete sampling and prevent policy collapse across unaudited tokens. Across competitive benchmarks, OmniOPD surpasses the standard OPD approach by up to +28.64% on math, confirming that chunk-level semantic verification extracts a more reliable learning signal than token-level logit matching, whose high information density is offset by significant noise and brittleness. Furthermore, when paired with stronger black-box teachers such as Claude-4.5-Haiku and Gemini-2.5-Flash, OmniOPD achieves an additional +9.54% relative on math over its open-weight teacher counterpart, advancing the student past the performance of self-exploratory RL.

cs.LG↗
arXiv2026

From Correctness to Utility: Gain-Based Prefix Evaluation for LLM Reasoning

Reasoning prefixes shape the future trajectory of LLM problem solving, yet existing process reward models usually evaluate them through local step correctness. We argue that correctness is a useful but indirect proxy for the effect we ultimately care about: whether a prefix increases the probability of successful completion. We define this effect as prefix gain, the solve-rate improvement induced by conditioning lightweight student model group on a prefix, and use it to train a Prefix Utility Model (PUM) with a simple pairwise ranking objective. PUM learns outcome-grounded prefix utility and can score both complete trajectories and partial reasoning prefixes. Across Best-of-$N$ selection, beam search, and reinforcement learning on mathematical reasoning, PUM provides a strong prefix-level supervision signal, especially when candidate pools are large, search budgets increase, or rule-based rewards are sparse. We release all data, models, and code at https://zhiqix.github.io/pum-project-page.

cs.CL↗
arXiv2026

SSV: Sparse Speculative Verification for Efficient LLM Inference

Speculative decoding and dynamic sparse attention are two complementary approaches for accelerating long-context LLM inference: the former amortizes target-model execution across multiple verifier queries, while the latter reduces each query's KV-cache working set. Directly combining them, however, exposes a structural mismatch: speculative verification relies on cross-query commonality, whereas dynamic sparse attention assigns query-specific sparse layouts. This mismatch limits KV-block reuse, amplifies NSA's branch-wise overheads, and makes verification strategy selection input- and regime-dependent. We present SSV, a sparse speculative-verification framework that turns dynamic sparse attention into a verification-oriented workload. SSV combines overlap-aware grouped-query execution, refresh/reuse-based NSA kernel fusion, and profile-guided prompt-adaptive orchestration to improve cross-query reuse, reduce selected-index and branch-fusion overheads, and select effective draft-verification strategies under user-specified precision classes. Experiments on NVIDIA H100 GPUs show that SSV achieves up to 3.49x end-to-end throughput over autoregressive NSA decoding and up to 6.86x kernel speedups for sparse speculative verification.

cs.OS↗
arXiv2026

Agentic Recommender System with Hierarchical Belief-State Memory

Memory-augmented LLM agents have advanced personalized recommendation, yet existing approaches universally adopt flat memory representations that conflate ephemeral signals with stable preferences, and none provides a complete lifecycle governing how memory should evolve. We propose MARS (Memory-Augmented Agentic Recommender System), a framework that treats recommendation as a partially observable problem and maintains a structured belief state that progressively abstracts noisy behavioral observations into a compact estimate of user preferences. MARS organizes this belief state into three tiers: event memory buffers raw signals, preference memory maintains fine-grained mutable chunks with explicit strength and evidence tracking, and profile memory distills all preferences into a coherent natural language narrative. A complete lifecycle of six operations -- extraction, reinforcement, weakening, consolidation, forgetting, and resynthesis -- is adaptively scheduled by an LLM-based planner rather than fixed-interval heuristics. Experiments on four InstructRec benchmark domains show that MARS achieves state-of-the-art performance with average improvements of 26.4% in HR@1 and 10.3% in NDCG@10 over the strongest baselines with further gains from agentic scheduling in evolving settings.

cs.CL↗
arXiv2026

Deep Reprogramming Distillation for Medical Foundation Models

Medical foundation models pre-trained on large-scale datasets have shown powerful versatile performance. However, when adapting medical foundation models for specific medical scenarios, it remains the inevitable challenge due to the gap induced by the discrepancy between pre-training and downstream tasks, the real-world computation, and speed constraints. Relevant techniques that probably handle this challenge more or less suffer from some intrinsic limitations. For example, knowledge distillation (KD) assumes that teacher and student models share the same task, training strategy, and model structure family, while prevalent parameter-efficient fine-tuning (PEFT) fails to achieve personalized and lightweight deployment. Even the combination of PEFT and KD still struggles to resolve model structures and training strategies inconsistencies between teacher and student models, leading to inefficient knowledge transfer. In this study, we propose a novel framework called Deep Reprogramming Distillation (DRD) to combat the general adaptation challenge. Specifically, DRD introduces the novel reprogramming module that on the one side overcomes the domain and task discrepancy between pretraining and downstream scenarios, and on the other side builds the student-friendly efficient distillation from foundation models to lightweight downstream models. Furthermore, to mitigate variability under different training conditions, we design a centered kernel alignment (CKA) distillation method to promote robust knowledge transfer. Empirical results show that DRD surpasses previous PEFT and KD methods across 18 medical downstream tasks under different foundation models, covering various scenarios including 2D/3D classification and 2D/3D segmentation.

cs.CV↗
arXiv2026

Chameleon: Adaptive Fault Tolerance for Distributed Training via Real-time Policy Selection

Training large language models faces frequent interruptions due to various faults, demanding robust fault-tolerance. Existing backup-free methods, such as redundant computation, dynamic parallelism, and data rerouting, each incur performance penalties, whether from ongoing overhead, lengthy reconfigurations, or post-recovery inefficiencies. We propose Chameleon, an adaptive fault-tolerant system that intelligently selects optimal recovery strategies when a failure occurs. Chameleon achieves this through a unified performance model, expedient execution plan search, accurate performance estimation, and efficient communication optimizations. Experiments on a 32-card cluster show that Chameleon maintains a performance gap of within 11.00% between post-recovery and failure-free training, while preserving model convergence and efficient memory usage. Compared to state-of-the-art methods, Chameleon achieves up to 1.229x and 1.355x higher average throughput than Oobleck and Recycle, respectively.

cs.DC↗