Improving RL Exploration for LLM Reasoning through Retrospective Replay

Reinforcement learning (RL) has increasingly become a pivotal technique in the post-training of large language models (LLMs). The effective exploration of the output space is essential for the success of RL. We observe that for complex problems, during the early stages of training, the model exhibits strong exploratory capabilities and can identify promising solution ideas. However, its limited capability at this stage prevents it from successfully solving these problems. The early suppression of these potentially valuable solution ideas by the policy gradient hinders the model's ability to revisit and re-explore these ideas later. Consequently, although the LLM's capabilities improve in the later stages of training, it still struggles to effectively address these complex problems. To address this exploration issue, we propose a novel algorithm named Retrospective Replay-based Reinforcement Learning (RRL), which introduces a dynamic replay mechanism throughout the training process. RRL enables the model to revisit promising states identified in the early stages, thereby improving its efficiency and effectiveness in exploration. To evaluate the effectiveness of RRL, we conduct extensive experiments on complex reasoning tasks, including mathematical reasoning and code generation, and general dialogue tasks. The results indicate that RRL maintains high exploration efficiency throughout the training period, significantly enhancing the effectiveness of RL in optimizing LLMs for complicated reasoning tasks. Moreover, it also improves the performance of RLHF, making the model both safer and more helpful.

MAIN: Mutual Alignment Is Necessary for instruction tuning

Instruction tuning has enabled large language models (LLMs) to achieve remarkable performance, but its success heavily depends on the availability of large-scale, high-quality instruction-response pairs. However, current methods for scaling up data generation often overlook a crucial aspect: the alignment between instructions and responses. We hypothesize that high-quality instruction-response pairs are not defined by the individual quality of each component, but by the extent of their alignment with each other. To address this, we propose a Mutual Alignment Framework (MAIN) that ensures coherence between the instruction and response through mutual constraints. Experiments demonstrate that models such as LLaMA and Mistral, fine-tuned within this framework, outperform traditional methods across multiple benchmarks. This approach underscores the critical role of instruction-response alignment in enabling scalable and high-quality instruction tuning for LLMs.

Seedream 3.0 Technical Report

We present Seedream 3.0, a high-performance Chinese-English bilingual image generation foundation model. We develop several technical improvements to address existing challenges in Seedream 2.0, including alignment with complicated prompts, fine-grained typography generation, suboptimal visual aesthetics and fidelity, and limited image resolutions. Specifically, the advancements of Seedream 3.0 stem from improvements across the entire pipeline, from data construction to model deployment. At the data stratum, we double the dataset using a defect-aware training paradigm and a dual-axis collaborative data-sampling framework. Furthermore, we adopt several effective techniques such as mixed-resolution training, cross-modality RoPE, representation alignment loss, and resolution-aware timestep sampling in the pre-training phase. During the post-training stage, we utilize diversified aesthetic captions in SFT, and a VLM-based reward model with scaling, thereby achieving outputs that well align with human preferences. Furthermore, Seedream 3.0 pioneers a novel acceleration paradigm. By employing consistent noise expectation and importance-aware timestep sampling, we achieve a 4 to 8 times speedup while maintaining image quality. Seedream 3.0 demonstrates significant improvements over Seedream 2.0: it enhances overall capabilities, in particular for text-rendering in complicated Chinese characters which is important to professional typography generation. In addition, it provides native high-resolution output (up to 2K), allowing it to generate images with high visual quality.

LLM Can be a Dangerous Persuader: Empirical Study of Persuasion Safety in Large Language Models

Recent advancements in Large Language Models (LLMs) have enabled them to approach human-level persuasion capabilities. However, such potential also raises concerns about the safety risks of LLM-driven persuasion, particularly their potential for unethical influence through manipulation, deception, exploitation of vulnerabilities, and many other harmful tactics. In this work, we present a systematic investigation of LLM persuasion safety through two critical aspects: (1) whether LLMs appropriately reject unethical persuasion tasks and avoid unethical strategies during execution, including cases where the initial persuasion goal appears ethically neutral, and (2) how influencing factors like personality traits and external pressures affect their behavior. To this end, we introduce PersuSafety, the first comprehensive framework for the assessment of persuasion safety which consists of three stages, i.e., persuasion scene creation, persuasive conversation simulation, and persuasion safety assessment. PersuSafety covers 6 diverse unethical persuasion topics and 15 common unethical strategies. Through extensive experiments across 8 widely used LLMs, we observe significant safety concerns in most LLMs, including failing to identify harmful persuasion tasks and leveraging various unethical persuasion strategies. Our study calls for more attention to improve safety alignment in progressive and goal-driven conversations such as persuasion.

DynClean: Training Dynamics-based Label Cleaning for Distantly-Supervised Named Entity Recognition

Distantly Supervised Named Entity Recognition (DS-NER) has attracted attention due to its scalability and ability to automatically generate labeled data. However, distant annotation introduces many mislabeled instances, limiting its performance. Most of the existing work attempt to solve this problem by developing intricate models to learn from the noisy labels. An alternative approach is to attempt to clean the labeled data, thus increasing the quality of distant labels. This approach has received little attention for NER. In this paper, we propose a training dynamics-based label cleaning approach, which leverages the behavior of a model as training progresses to characterize the distantly annotated samples. We also introduce an automatic threshold estimation strategy to locate the errors in distant labels. Extensive experimental results demonstrate that: (1) models trained on our cleaned DS-NER datasets, which were refined by directly removing identified erroneous annotations, achieve significant improvements in F1-score, ranging from 3.18% to 8.95%; and (2) our method outperforms numerous advanced DS-NER approaches across four datasets.

Mitigating Object Hallucinations in MLLMs via Multi-Frequency Perturbations

Recently, multimodal large language models (MLLMs) have demonstrated remarkable performance in visual-language tasks. However, the authenticity of the responses generated by MLLMs is often compromised by object hallucinations. We identify that a key cause of these hallucinations is the model's over-susceptibility to specific image frequency features in detecting objects. In this paper, we introduce Multi-Frequency Perturbations (MFP), a simple, cost-effective, and pluggable method that leverages both low-frequency and high-frequency features of images to perturb visual feature representations and explicitly suppress redundant frequency-domain features during inference, thereby mitigating hallucinations. Experimental results demonstrate that our method significantly mitigates object hallucinations across various model architectures. Furthermore, as a training-time method, MFP can be combined with inference-time methods to achieve state-of-the-art performance on the CHAIR benchmark.

Deep Reinforcement Learning-based Video-Haptic Radio Resource Slicing in Tactile Internet

Enabling video-haptic radio resource slicing in the Tactile Internet requires a sophisticated strategy to meet the distinct requirements of video and haptic data, ensure their synchronized transmission, and address the stringent latency demands of haptic feedback. This paper introduces a Deep Reinforcement Learning-based radio resource slicing framework that addresses video-haptic teleoperation challenges by dynamically balancing radio resources between the video and haptic modalities. The proposed framework employs a refined reward function that considers latency, packet loss, data rate, and the synchronization requirements of both modalities to optimize resource allocation. By catering to the specific service requirements of video-haptic teleoperation, the proposed framework achieves up to a 25% increase in user satisfaction over existing methods, while maintaining effective resource slicing with execution intervals up to 50 ms.

Highly Efficient Direct Analytics on Semantic-aware Time Series Data Compression

Semantic communication has emerged as a promising paradigm to tackle the challenges of massive growing data traffic and sustainable data communication. It shifts the focus from data fidelity to goal-oriented or task-oriented semantic transmission. While deep learning-based methods are commonly used for semantic encoding and decoding, they struggle with the sequential nature of time series data and high computation cost, particularly in resource-constrained IoT environments. Data compression plays a crucial role in reducing transmission and storage costs, yet traditional data compression methods fall short of the demands of goal-oriented communication systems. In this paper, we propose a novel method for direct analytics on time series data compressed by the SHRINK compression algorithm. Through experimentation using outlier detection as a case study, we show that our method outperforms baselines running on uncompressed data in multiple cases, with merely 1% difference in the worst case. Additionally, it achieves four times lower runtime on average and accesses approximately 10% of the data volume, which enables edge analytics with limited storage and computation power. These results demonstrate that our approach offers reliable, high-speed outlier detection analytics for diverse IoT applications while extracting semantics from time-series data, achieving high compression, and reducing data transmission.

U2AD: Uncertainty-based Unsupervised Anomaly Detection Framework for Detecting T2 Hyperintensity in MRI Spinal Cord

T2 hyperintensities in spinal cord MR images are crucial biomarkers for conditions such as degenerative cervical myelopathy. However, current clinical diagnoses primarily rely on manual evaluation. Deep learning methods have shown promise in lesion detection, but most supervised approaches are heavily dependent on large, annotated datasets. Unsupervised anomaly detection (UAD) offers a compelling alternative by eliminating the need for abnormal data annotations. However, existing UAD methods rely on curated normal datasets and their performance frequently deteriorates when applied to clinical datasets due to domain shifts. We propose an Uncertainty-based Unsupervised Anomaly Detection framework, termed U2AD, to address these limitations. Unlike traditional methods, U2AD is designed to be trained and tested within the same clinical dataset, following a "mask-and-reconstruction" paradigm built on a Vision Transformer-based architecture. We introduce an uncertainty-guided masking strategy to resolve task conflicts between normal reconstruction and anomaly detection to achieve an optimal balance. Specifically, we employ a Monte-Carlo sampling technique to estimate reconstruction uncertainty mappings during training. By iteratively optimizing reconstruction training under the guidance of both epistemic and aleatoric uncertainty, U2AD reduces overall reconstruction variance while emphasizing regions. Experimental results demonstrate that U2AD outperforms existing supervised and unsupervised methods in patient-level identification and segment-level localization tasks. This framework establishes a new benchmark for incorporating uncertainty guidance into UAD, highlighting its clinical utility in addressing domain shifts and task conflicts in medical image anomaly detection. Our code is available: https://github.com/zhibaishouheilab/U2AD

Mitigating Ambiguities in 3D Classification with Gaussian Splatting

3D classification with point cloud input is a fundamental problem in 3D vision. However, due to the discrete nature and the insufficient material description of point cloud representations, there are ambiguities in distinguishing wire-like and flat surfaces, as well as transparent or reflective objects. To address these issues, we propose Gaussian Splatting (GS) point cloud-based 3D classification. We find that the scale and rotation coefficients in the GS point cloud help characterize surface types. Specifically, wire-like surfaces consist of multiple slender Gaussian ellipsoids, while flat surfaces are composed of a few flat Gaussian ellipsoids. Additionally, the opacity in the GS point cloud represents the transparency characteristics of objects. As a result, ambiguities in point cloud-based 3D classification can be mitigated utilizing GS point cloud as input. To verify the effectiveness of GS point cloud input, we construct the first real-world GS point cloud dataset in the community, which includes 20 categories with 200 objects in each category. Experiments not only validate the superiority of GS point cloud input, especially in distinguishing ambiguous objects, but also demonstrate the generalization ability across different classification methods.