arXiv subjects

Yaohua Tang

Publications and source records attributed to Yaohua Tang.

18 recordsLinked to original sources

arXiv2026

TCPO: Turn-Level Credit Policy Optimization

Verifier-guided reinforcement learning has become a powerful paradigm for improving LLM reasoning. In multi-turn settings, models receive a verifier score after each turn and iteratively refine their outputs. Although such scores provide dense feedback, they do not directly provide dense credit: a score measures the quality of the current output, while credit should measure how the current turn changes the refinement trajectory. We propose TCPO, a turn-level credit assignment method for verifier-guided multi-turn RL. TCPO casts credit assignment as score-to-credit conversion and constructs turn-level advantages through reference-based comparisons: retrospective credit captures immediate progress and regression relative to the best prior state; hindsight delayed credit identifies non-improving turns with later payoff; and selective fixed-history counterfactual estimation refines high-surprisal turns under the same history. Experiments on math reasoning, code generation, and AppWorld agent tasks show that TCPO improves or matches the strongest baselines across model scales, task domains, and verifier types. TCPO achieves the best or tied-best best-turn Pass@8 on Qwen3-4B and DeepSeek-R1-Distill-Llama-8B, reduces turns to success, and improves multi-turn agent performance. These results highlight score-to-credit conversion as a central ingredient for verifier-guided multi-turn policy optimization.

cs.AI↗

arXiv2026

LEAP: Lean Environment-Feedback via Adaptive Pruning for Code RL in GPU Kernel Generation

Post-training large language models (LLMs) via reinforcement learning (RL) has significantly advanced code generation capabilities. To bypass the heavy memory footprint of critic networks, current state-of-the-art frameworks leverage critic-free paradigms like Group Relative Policy Optimization (GRPO) tied to rule-based verification sandboxes. However, applying these frameworks to low-level systems programming, such as CUDA kernel generation-presents severe challenges: binary pass/fail rewards introduce severe signal sparsity, while multi-turn environmental feedback loops suffer from prohibitive compilation latencies and reward dilution across trajectories. In this work, we introduce LEAP (Lean Environment-Feedback via Adaptive Pruning), a scalable and computationally efficient multi-turn RL framework optimized for low-level hardware accelerator alignment. LEAP features Difficulty-Conditioned Pruning (DCP), a dynamic gating mechanism that adaptively cuts off simple and overly catastrophic tasks from multi-turn expansion, focusing resource-heavy compilation and hardware exploration exclusively on high-value, complex tasks. To fully operationalize these paths without manual hyperparameter engineering, we propose a Rank-Based Reward formulation. By deriving scale-free relative advantages from pairwise tournament outcomes within the GRPO rollout group, our method inherently penalizes token inefficiency on simple prompts while maximizing learning gradients on challenging distributions. Empirical evaluations show that LEAP achieves superior first-turn proficiency and robust multi-turn debugging resilience while converging faster than unpruned multi-turn baselines, establishing a practical paradigm for low-level code RL.

cs.LG↗

arXiv2026

Beyond the Best Teacher: Expanding and Compressing the Reasoning Solution Manifold

A single reinforcement-learning run can produce a strong reasoner yet an incomplete teacher: it often amplifies only a subset of the valid solution modes. We argue that reinforcement learning (RL)-trained policies should therefore be viewed as local probes of a multi-basin reasoning solution manifold, rather than as globally reliable supervisors. Based on this view, we propose an expand-then-compress framework that couples teacher construction with multi-teacher policy distillation. In the expansion stage, Residual Group Relative Policy Optimization (RGRPO) trains a sequence of teachers from a common initialization and redirects each later round toward examples not yet covered by the accumulated teacher union. In the compression stage, reliability-gated Teacher-Union On-policy Distillation (TU-OPD) lets the student learn from its own response prefixes. For each example, only reliable teachers contribute, and their sampled-token OPD losses are weighted by their per-example quality. We further introduce Consensus-Residual Decomposition, which preserves a winner teacher's excess token preferences over its reliable peers, preventing specialist behavior from being suppressed during teacher aggregation. Experiments on mathematical reasoning, code generation, and instruction following show that the resulting Qwen3-1.7B student consistently outperforms the strongest individual teacher across all three domains, yielding relative improvements of 2.0%, 8.3%, and 6.9%, respectively, while retaining single-model inference. These results establish a simple but powerful principle: stronger students can be obtained not by selecting a single better teacher, but by deliberately constructing and compressing a complementary teacher union.

cs.LG↗

arXiv2026

MusaCoder: Native GPU Kernel Generation with Full-Stack Training on Moore Threads GPU

Native GPU kernel generation turns high-level tensor programs into executable, efficient low-level code. Existing Large Language Models (LLMs) struggle with this task, while execution-based reinforcement learning suffers from sparse rewards, reward hacking, and training instability. We present MusaCoder, a full-stack training framework for native GPU kernel generation on CUDA and MUSA backends. MusaCoder combines progressive kernel-oriented data synthesis, diversity-preserving rejection fine-tuning, and execution-feedback Reinforcement Learning (RL) through MooreEval, a distributed verifier and reward environment. To stabilize RL, MusaCoder introduces PrimeEcho for first-turn-anchored multi-turn rewards, Buffered Dynamic Retry for recovering signals from all-failed hard samples, and MirrorPop for off-policy sequence filtering. Experiments on KernelBench and a MUSA-ported variant show that MusaCoder outperforms strong open-source and proprietary baselines in both correctness and empirical speedup, with the 9B model matching or exceeding frontier closed-source models and the 27B model establishing a new state of the art. These results demonstrate not only the effectiveness of full-stack execution-feedback training for native kernel generation, but also the capability of Moore Threads GPUs to support the complete LLM post-training stack, providing a practical foundation for large-model training and optimization on emerging accelerators.

cs.CV↗

arXiv2026

ReorgGS: Equivalent Distribution Reorganization for 3D Gaussian Splatting

A converged 3D Gaussian Splatting (3DGS) model may approximate the target scene while remaining poorly parameterized for further optimization. We identify this failure mode as \emph{parameterization degeneration}: high-opacity floaters attenuate gradients to true surfaces through alpha compositing, and redundant overlapping clusters create strongly coupled parameter blocks with nearly collinear Jacobian responses. These effects explain why continued optimization can plateau even when the model still contains removable artifacts. We propose ReorgGS, an equivalent distribution reorganization method for converged 3DGS models. ReorgGS treats the existing Gaussian set as an empirical probability field, resamples centers from it, estimates local anisotropic covariances with kNN, initializes low opacity, and continues optimization with the original 3DGS renderer and loss. Unlike opacity reset, which only rescales opacity on the old overlap graph, ReorgGS rebuilds centers, covariances, and visibility structure, thereby changing the graph itself. Our analysis shows that distributional equivalence is not optimization equivalence. The reorganized model preserves scene support while improving gradient accessibility under alpha compositing and reducing opacity-weighted overlap, thereby weakening local parameter coupling during subsequent optimization. Under the same additional optimization budget, ReorgGS improves fitting quality at a fixed Gaussian count, suppresses persistent floaters, and reduces rendering overhead from redundant overlap.

cs.CV↗

arXiv2026

LiteGUI: Distilling Compact GUI Agents with Reinforcement Learning

Developing lightweight, on-device vision-language GUI agents is essential for efficient cross-platform automated interaction. However, current on-device agents are constrained by limited model capacity, and further performance improvements remain urgently needed. Traditional Supervised Fine-Tuning (SFT) for small-scale models often leads to overfitting, catastrophic forgetting and policy rigidity, and thus fails to fully address these challenges. In this work, we propose a novel SFT-free training paradigm that significantly enhances the performance of small-scale models. We first present the initial systematic integration of generalized knowledge distillation into the GUI agent domain via Guided On-policy Distillation. By incorporating oracle reference trajectories together with a dynamic retrieval mechanism, our method reduces hallucinations and mitigates the cognitive misalignment inherent in multi-solution GUI tasks. Building on this foundation, we further introduce a Multi-solution Dual-level GRPO framework that jointly aligns macro-level subtask planning with micro-level execution matching, thereby improving exploration in long-horizon GUI agent scenarios. In addition, we construct an automated data generation pipeline to synthesize GUI task trajectories with rich multi-solution annotations. Extensive experiments show that our method achieves state-of-the-art performance among lightweight models while remaining competitive with substantially larger-scale models across all benchmarks. Ablation studies further demonstrate that structured on-policy distillation and multi-solution dual-level exploration can fully unlock the capabilities of 2B/3B scale agents, surpassing the performance limits of conventional imitation learning.

cs.AI↗

arXiv2026

UniSem: Generalizable Semantic 3D Reconstruction from Sparse Unposed Images

Semantic-aware 3D reconstruction from sparse, unposed images remains challenging for feed-forward 3D Gaussian Splatting (3DGS). Existing methods often predict an over-complete set of Gaussian primitives under sparse-view supervision, leading to unstable geometry and inferior depth quality. Meanwhile, they rely solely on 2D segmenter features for semantic lifting, which provides weak 3D-level and limited generalizable supervision, resulting in incomplete 3D semantics in novel scenes. To address these issues, we propose UniSem, a unified framework that jointly improves depth accuracy and semantic generalization via two key components. First, Error-aware Gaussian Dropout (EGD) performs error-guided capacity control by suppressing redundancy-prone Gaussians using rendering error cues, producing meaningful, geometrically stable Gaussian representations for improved depth estimation. Second, we introduce a Mix-training Curriculum (MTC) that progressively blends 2D segmenter-lifted semantics with the model's own emergent 3D semantic priors, implemented with object-level prototype alignment to enhance semantic coherence and completeness. Extensive experiments on ScanNet and Replica show that UniSem achieves superior performance in depth prediction and open-vocabulary 3D segmentation across varying numbers of input views. Notably, with 16-view inputs, UniSem reduces depth Rel by 15.2% and improves open-vocabulary segmentation mAcc by 3.7% over strong baselines.

cs.CV↗

arXiv2025

URPO: A Unified Reward & Policy Optimization Framework for Large Language Models

Large-scale alignment pipelines typically pair a policy model with a separately trained reward model whose parameters remain frozen during reinforcement learning (RL). This separation creates a complex, resource-intensive pipeline and suffers from a performance ceiling due to a static reward signal. We propose a novel framework, Unified Reward & Policy Optimization (URPO), that unifies instruction-following ("player") and reward modeling ("referee") within a single model and a single training phase. Our method recasts all alignment data-including preference pairs, verifiable reasoning, and open-ended instructions-into a unified generative format optimized by a single Group-Relative Policy Optimization (GRPO) loop. This enables the model to learn from ground-truth preferences and verifiable logic while simultaneously generating its own rewards for open-ended tasks. Experiments on the Qwen2.5-7B model demonstrate URPO's superiority. Our unified model significantly outperforms a strong baseline using a separate generative reward model, boosting the instruction-following score on AlpacaEval from 42.24 to 44.84 and the composite reasoning average from 32.66 to 35.66. Furthermore, URPO cultivates a superior internal evaluator as a byproduct of training, achieving a RewardBench score of 85.15 and surpassing the dedicated reward model it replaces (83.55). By eliminating the need for a separate reward model and fostering a co-evolutionary dynamic between generation and evaluation, URPO presents a simpler, more efficient, and more effective path towards robustly aligned language models.

cs.CV↗

arXiv2025

StableGS: A Floater-Free Framework for 3D Gaussian Splatting

3D Gaussian Splatting (3DGS) reconstructions are plagued by stubborn ``floater" artifacts that degrade their geometric and visual fidelity. We are the first to reveal the root cause: a fundamental conflict in the 3DGS optimization process where the opacity gradients of floaters vanish when their blended color reaches a pseudo-equilibrium of canceling errors against the background, trapping them in a spurious local minimum. To resolve this, we propose StableGS, a novel framework that decouples geometric regularization from final appearance rendering. Its core is a Dual Opacity architecture that creates two separate rendering paths: a ``Geometric Regularization Path" to bear strong depth-based constraints for structural correctness, and an ``Appearance Refinement Path" to generate high-fidelity details upon this stable foundation. We complement this with a synergistic set of geometric constraints: a self-supervised depth consistency loss and an external geometric prior enabled by our efficient global scale optimization algorithm. Experiments on multiple benchmarks show StableGS not only eliminates floaters but also resolves the common blur-artifact trade-off, achieving state-of-the-art geometric accuracy and visual quality.

cs.CV↗

arXiv2025

DeskVision: Large Scale Desktop Region Captioning for Advanced GUI Agents

The limitation of graphical user interface (GUI) data has been a significant barrier to the development of GUI agents today, especially for the desktop / computer use scenarios. To address this, we propose an automated GUI data generation pipeline, AutoCaptioner, which generates data with rich descriptions while minimizing human effort. Using AutoCaptioner, we created a novel large-scale desktop GUI dataset, DeskVision, along with the largest desktop test benchmark, DeskVision-Eval, which reflects daily usage and covers diverse systems and UI elements, each with rich descriptions. With DeskVision, we train a new GUI understanding model, GUIExplorer. Results show that GUIExplorer achieves state-of-the-art (SOTA) performance in understanding/grounding visual elements without the need for complex architectural designs. We further validated the effectiveness of the DeskVision dataset through ablation studies on various large visual language models (LVLMs). We believe that AutoCaptioner and DeskVision will significantly advance the development of GUI agents, and will open-source them for the community.

cs.CL↗

arXiv2025

LiteGS: A High-performance Framework to Train 3DGS in Subminutes via System and Algorithm Codesign

3D Gaussian Splatting (3DGS) has emerged as promising alternative in 3D representation. However, it still suffers from high training cost. This paper introduces LiteGS, a high performance framework that systematically optimizes the 3DGS training pipeline from multiple aspects. At the low-level computation layer, we design a ``warp-based raster'' associated with two hardware-aware optimizations to significantly reduce gradient reduction overhead. At the mid-level data management layer, we introduce dynamic spatial sorting based on Morton coding to enable a performant ``Cluster-Cull-Compact'' pipeline and improve data locality, therefore reducing cache misses. At the top-level algorithm layer, we establish a new robust densification criterion based on the variance of the opacity gradient, paired with a more stable opacity control mechanism, to achieve more precise parameter growth. Experimental results demonstrate that LiteGS accelerates the original 3DGS training by up to 13.4x with comparable or superior quality and surpasses the current SOTA in lightweight models by up to 1.4x speedup. For high-quality reconstruction tasks, LiteGS sets a new accuracy record and decreases the training time by an order of magnitude.

cs.CV↗

arXiv2025

SEKI: Self-Evolution and Knowledge Inspiration based Neural Architecture Search via Large Language Models

We introduce SEKI, a novel large language model (LLM)-based neural architecture search (NAS) method. Inspired by the chain-of-thought (CoT) paradigm in modern LLMs, SEKI operates in two key stages: self-evolution and knowledge distillation. In the self-evolution stage, LLMs initially lack sufficient reference examples, so we implement an iterative refinement mechanism that enhances architectures based on performance feedback. Over time, this process accumulates a repository of high-performance architectures. In the knowledge distillation stage, LLMs analyze common patterns among these architectures to generate new, optimized designs. Combining these two stages, SEKI greatly leverages the capacity of LLMs on NAS and without requiring any domain-specific data. Experimental results show that SEKI achieves state-of-the-art (SOTA) performance across various datasets and search spaces while requiring only 0.05 GPU-days, outperforming existing methods in both efficiency and accuracy. Furthermore, SEKI demonstrates strong generalization capabilities, achieving SOTA-competitive results across multiple tasks.

cs.CL↗

arXiv2025

Round Attention: A Novel Round-Level Attention Mechanism to Accelerate LLM Inference

The increasing context window size in large language models (LLMs) has improved their ability to handle complex, long-text tasks. However, as the conversation rounds continue, it is required to store a large amount of KV cache in GPU memory, which significantly affects the efficiency and even availability of the model serving systems. This paper analyzes dialogue data from real users on the granularity of round and discovers that the LLM inference manifests a watershed layer, after which the distribution of round-level attention shows notable similarity. Based on this, we propose Round Attention - a novel round-level attention mechanism that selectively processes the KV cache of top-k relevant rounds, where k is dynamically determined through the attention matrix in the watershed layer. Theoretical analysis demonstrates that our method reduces memory usage by 54\% to 82\%, while experimental results confirm that loading sparse critical-round KV cache maintains answer accuracy without performance degradation.

cs.CL↗

arXiv2024

TurboRAG: Accelerating Retrieval-Augmented Generation with Precomputed KV Caches for Chunked Text

Current Retrieval-Augmented Generation (RAG) systems concatenate and process numerous retrieved document chunks for prefill which requires a large volume of computation, therefore leading to significant latency in time-to-first-token (TTFT). To reduce the computation overhead as well as TTFT, we introduce TurboRAG, a novel RAG system that redesigns the inference paradigm of the current RAG system by first pre-computing and storing the key-value (KV) caches of documents offline, and then directly retrieving the saved KV cache for prefill. Hence, online computation of KV caches is eliminated during inference. In addition, we provide a number of insights into the mask matrix and positional embedding mechanisms, plus fine-tune a pretrained language model to maintain model accuracy of TurboRAG. Our approach is applicable to most existing large language models and their applications without any requirement in modification of models and inference systems. Experimental results across a suite of RAG benchmarks demonstrate that TurboRAG reduces TTFT by up to 9.4x compared to the conventional RAG systems (on an average of 8.6x), but reserving comparable performance to the standard RAG systems.

cs.CV↗

arXiv2024

A Full-duplex Speech Dialogue Scheme Based On Large Language Models

We present a generative dialogue system capable of operating in a full-duplex manner, allowing for seamless interaction. It is based on a large language model (LLM) carefully aligned to be aware of a perception module, a motor function module, and the concept of a simple finite state machine (called neural FSM) with two states. The perception and motor function modules operate in tandem, allowing the system to speak and listen to the user simultaneously. The LLM generates textual tokens for inquiry responses and makes autonomous decisions to start responding to, wait for, or interrupt the user by emitting control tokens to the neural FSM. All these tasks of the LLM are carried out as next token prediction on a serialized view of the dialogue in real-time. In automatic quality evaluations simulating real-life interaction, the proposed system reduces the average conversation response latency by more than threefold compared with LLM-based half-duplex dialogue systems while responding within less than 500 milliseconds in more than 50% of evaluated interactions. Running an LLM with only 8 billion parameters, our system exhibits an 8% higher interruption precision rate than the best available commercial LLM for voice-based dialogue.

cs.CL↗

arXiv2021

CNN-based Realized Covariance Matrix Forecasting

It is well known that modeling and forecasting realized covariance matrices of asset returns play a crucial role in the field of finance. The availability of high frequency intraday data enables the modeling of the realized covariance matrices directly. However, most of the models available in the literature depend on strong structural assumptions and they often suffer from the curse of dimensionality. We propose an end-to-end trainable model built on the CNN and Convolutional LSTM (ConvLSTM) which does not require to make any distributional or structural assumption but could handle high-dimensional realized covariance matrices consistently. The proposed model focuses on local structures and spatiotemporal correlations. It learns a nonlinear mapping that connect the historical realized covariance matrices to the future one. Our empirical studies on synthetic and real-world datasets demonstrate its excellent forecasting ability compared with several advanced volatility models.

cs.CE↗

arXiv2016

Neural Machine Translation with External Phrase Memory

In this paper, we propose phraseNet, a neural machine translator with a phrase memory which stores phrase pairs in symbolic form, mined from corpus or specified by human experts. For any given source sentence, phraseNet scans the phrase memory to determine the candidate phrase pairs and integrates tagging information in the representation of source sentence accordingly. The decoder utilizes a mixture of word-generating component and phrase-generating component, with a specifically designed strategy to generate a sequence of multiple words all at once. The phraseNet not only approaches one step towards incorporating external knowledge into neural machine translation, but also makes an effort to extend the word-by-word generation mechanism of recurrent neural network. Our empirical study on Chinese-to-English translation shows that, with carefully-chosen phrase table in memory, phraseNet yields 3.45 BLEU improvement over the generic neural machine translator.

cs.CL↗

arXiv2013

Approximating Soft-Capacitated Facility Location Problem With Uncertainty

We first show that a better analysis of the algorithm for The Two-Sage Stochastic Facility Location Problem from Srinivasan \cite{sri07} and the algorithm for The Robust Fault Tolerant Facility Location Problem from Byrka et al \cite{bgs10} can render improved approximation factors of 2.206 and α+4 where αis the maximum number an adversary can close, respectively, and which are the best ratios so far. We then present new models for the soft-capacitated facility location problem with uncertainty and design constant factor approximation algorithms to solve them. We devise the stochastic and robust approaches to handle the uncertainty incorporated into the original model. Explicitly, in this paper we propose two new problem, named The 2-Stage Soft-Capacitated Facility Location Problem and The Robust Soft-Capacitated Facility Location Problem respectively, and present constant factor approximation algorithms for them both. Our method uses reductions between facility location problems and linear-cost models, the randomized thresholding technique of Srinivasan \cite{sri07} and the filtering and clustering technique of Byrka et al \cite{bgs10}.

cs.DS↗