Heterogeneous Multi-Expert Reinforcement Learning for Long-Horizon Multi-Goal Tasks in Autonomous Forklifts

Autonomous mobile manipulation in unstructured warehouses requires a balance between efficient large-scale navigation and high-precision object interaction. Traditional end-to-end learning approaches often struggle to handle the conflicting demands of these distinct phases. Navigation relies on robust decision-making over large spaces, while manipulation needs high sensitivity to fine local details. Forcing a single network to learn these different objectives simultaneously often causes optimization interference, where improving one task degrades the other. To address these limitations, we propose a Heterogeneous Multi-Expert Reinforcement Learning (HMER) framework tailored for autonomous forklifts. HMER decomposes long-horizon tasks into specialized sub-policies controlled by a Semantic Task Planner. This structure separates macro-level navigation from micro-level manipulation, allowing each expert to focus on its specific action space without interference. The planner coordinates the sequential execution of these experts, bridging the gap between task planning and continuous control. Furthermore, to solve the problem of sparse exploration, we introduce a Hybrid Imitation-Reinforcement Training Strategy. This method uses expert demonstrations to initialize the policy and Reinforcement Learning for fine-tuning. Experiments in Gazebo simulations show that HMER significantly outperforms sequential and end-to-end baselines. Our method achieves a task success rate of 94.2\% (compared to 62.5\% for baselines), reduces operation time by 21.4\%, and maintains placement error within 1.5 cm, validating its efficacy for precise material handling.

Infer As You Train: A Symmetric Paradigm of Masked Generative for Click-Through Rate Prediction

Generative models are increasingly being explored in click-through rate (CTR) prediction field to overcome the limitations of the conventional discriminative paradigm, which rely on a simple binary classification objective. However, existing generative models typically confine the generative paradigm to the training phase, primarily for representation learning. During online inference, they revert to a standard discriminative paradigm, failing to leverage their powerful generative capabilities to further improve prediction accuracy. This fundamental asymmetry between the training and inference phases prevents the generative paradigm from realizing its full potential. To address this limitation, we propose the Symmetric Masked Generative Paradigm for CTR prediction (SGCTR), a novel framework that establishes symmetry between the training and inference phases. Specifically, after acquiring generative capabilities by learning feature dependencies during training, SGCTR applies the generative capabilities during online inference to iteratively redefine the features of input samples, which mitigates the impact of noisy features and enhances prediction accuracy. Extensive experiments validate the superiority of SGCTR, demonstrating that applying the generative paradigm symmetrically across both training and inference significantly unlocks its power in CTR prediction.

DGenCTR: Towards a Universal Generative Paradigm for Click-Through Rate Prediction via Discrete Diffusion

Recent advances in generative models have inspired the field of recommender systems to explore generative approaches, but most existing research focuses on sequence generation, a paradigm ill-suited for click-through rate (CTR) prediction. CTR models critically depend on a large number of cross-features between the target item and the user to estimate the probability of clicking on the item, and discarding these cross-features will significantly impair model performance. Therefore, to harness the ability of generative models to understand data distributions and thereby alleviate the constraints of traditional discriminative models in label-scarce space, diverging from the item-generation paradigm of sequence generation methods, we propose a novel sample-level generation paradigm specifically designed for the CTR task: a two-stage Discrete Diffusion-Based Generative CTR training framework (DGenCTR). This two-stage framework comprises a diffusion-based generative pre-training stage and a CTR-targeted supervised fine-tuning stage for CTR. Finally, extensive offline experiments and online A/B testing conclusively validate the effectiveness of our framework.

Unveiling Over-Memorization in Finetuning LLMs for Reasoning Tasks

The pretrained large language models (LLMs) are finetuned with labeled data for better instruction following ability and alignment with human values. In this paper, we study the learning dynamics of LLM finetuning on reasoning tasks and reveal the uncovered over-memorization phenomenon during a specific stage of LLM finetuning. At this stage, the LLMs have excessively memorized training data and exhibit high test perplexity while maintaining good test accuracy. We investigate the conditions that lead to LLM over-memorization and find that training epochs and large learning rates contribute to this issue. Although models with over-memorization demonstrate comparable test accuracy to normal models, they suffer from reduced robustness, poor out-of-distribution generalization, and decreased generation diversity. Our experiments unveil the over-memorization to be broadly applicable across different tasks, models, and finetuning methods. Our research highlights that overparameterized, extensively finetuned LLMs exhibit unique learning dynamics distinct from traditional machine learning models. Based on our observations of over-memorization, we provide recommendations on checkpoint and learning rate selection during finetuning.

Towards Bridging the Reward-Generation Gap in Direct Alignment Algorithms

Direct Alignment Algorithms (DAAs), such as Direct Preference Optimization (DPO) and Simple Preference Optimization (SimPO), have emerged as efficient alternatives to Reinforcement Learning from Human Feedback (RLHF) algorithms for aligning large language models (LLMs) with human preferences. However, DAAs suffer from a fundamental limitation we identify as the "reward-generation gap" -- a misalignment between optimization objectives during training and actual generation performance during inference. In this paper, we find a contributor to the reward-generation gap is the mismatch between the inherent importance of prefix tokens during the LLM generation process and how this importance is reflected in the implicit reward functions of DAAs. To bridge the gap, we introduce a simple yet effective approach called Prefix-Oriented Equal-length Training (POET), which truncates both preferred and dispreferred responses to match the shorter one's length. Training with POET, where both responses in each sample are truncated to equal length, resulting in diverse truncated lengths across samples, the optimization of DAAs objective is implicitly constrained to converge across all positions, thus paying more attention to prefix tokens than the standard DAAs. We conduct experiments with DPO and SimPO, two representative DAAs, demonstrating that POET improves over their standard implementations, achieving up to 15.6 points in AlpacaEval 2 and overall improvements across downstream tasks. Our results highlight the importance of addressing the misalignment between reward optimization and generation performance in DAAs.

Automatic Robustness Stress Testing of LLMs as Mathematical Problem Solvers

Large language models (LLMs) have achieved distinguished performance on various reasoning-intensive tasks. However, LLMs might still face the challenges of robustness issues and fail unexpectedly in some simple reasoning tasks. Previous works evaluate the LLM robustness with hand-crafted templates or a limited set of perturbation rules, indicating potential data contamination in pre-training or fine-tuning datasets. In this work, inspired by stress testing in software engineering, we propose a novel framework, Automatic Robustness Checker (AR-Checker), to generate mathematical problem variants that maintain the semantic meanings of the original one but might fail the LLMs. The AR-Checker framework generates mathematical problem variants through multi-round parallel streams of LLM-based rewriting and verification. Our framework can generate benchmark variants dynamically for each LLM, thus minimizing the risk of data contamination. Experiments on GSM8K and MATH-500 demonstrate the strong performance of AR-Checker on mathematical tasks. We also evaluate AR-Checker on benchmarks beyond mathematics, including MMLU, MMLU-Pro, and CommonsenseQA, where it also achieves strong performance, further proving the effectiveness of AR-Checker.

AudioTurbo: Fast Text-to-Audio Generation with Rectified Diffusion

Diffusion models have significantly improved the quality and diversity of audio generation but are hindered by slow inference speed. Rectified flow enhances inference speed by learning straight-line ordinary differential equation (ODE) paths. However, this approach requires training a flow-matching model from scratch and tends to perform suboptimally, or even poorly, at low step counts. To address the limitations of rectified flow while leveraging the advantages of advanced pre-trained diffusion models, this study integrates pre-trained models with the rectified diffusion method to improve the efficiency of text-to-audio (TTA) generation. Specifically, we propose AudioTurbo, which learns first-order ODE paths from deterministic noise sample pairs generated by a pre-trained TTA model. Experiments on the AudioCaps dataset demonstrate that our model, with only 10 sampling steps, outperforms prior models and reduces inference to 3 steps compared to a flow-matching-based acceleration model.

TAG-INSTRUCT: Controlled Instruction Complexity Enhancement through Structure-based Augmentation

High-quality instruction data is crucial for developing large language models (LLMs), yet existing approaches struggle to effectively control instruction complexity. We present TAG-INSTRUCT, a novel framework that enhances instruction complexity through structured semantic compression and controlled difficulty augmentation. Unlike previous prompt-based methods operating on raw text, TAG-INSTRUCT compresses instructions into a compact tag space and systematically enhances complexity through RL-guided tag expansion. Through extensive experiments, we show that TAG-INSTRUCT outperforms existing instruction complexity augmentation approaches. Our analysis reveals that operating in tag space provides superior controllability and stability across different instruction synthesis frameworks.

ImPart: Importance-Aware Delta-Sparsification for Improved Model Compression and Merging in LLMs

With the proliferation of task-specific large language models, delta compression has emerged as a method to mitigate the resource challenges of deploying numerous such models by effectively compressing the delta model parameters. Previous delta-sparsification methods either remove parameters randomly or truncate singular vectors directly after singular value decomposition (SVD). However, these methods either disregard parameter importance entirely or evaluate it with too coarse a granularity. In this work, we introduce ImPart, a novel importance-aware delta sparsification approach. Leveraging SVD, it dynamically adjusts sparsity ratios of different singular vectors based on their importance, effectively retaining crucial task-specific knowledge even at high sparsity ratios. Experiments show that ImPart achieves state-of-the-art delta sparsification performance, demonstrating $2\times$ higher compression ratio than baselines at the same performance level. When integrated with existing methods, ImPart sets a new state-of-the-art on delta quantization and model merging.

Fin-R1: A Large Language Model for Financial Reasoning through Reinforcement Learning

Reasoning large language models are rapidly evolving across various domains. However, their capabilities in handling complex financial tasks still require in-depth exploration. In this paper, we introduce Fin-R1, a reasoning large language model specifically designed for the financial sector. Fin-R1 is built using a two-stage architecture, leveraging a financial reasoning dataset distilled and processed based on DeepSeek-R1. Through supervised fine-tuning (SFT) and reinforcement learning (RL) training, it demonstrates performance close to DeepSeek-R1 with a parameter size of 7 billion across a range of financial reasoning tasks. It achieves the state-of-the-art (SOTA) in the FinQA and ConvFinQA tasks between those LLMs in our evaluation, surpassing larger models in other tasks as well. Fin-R1 showcases strong reasoning and decision-making capabilities, providing solutions to various problems encountered in the financial domain. Our code is available at https://github.com/SUFE-AIFLM-Lab/Fin-R1.