Step-DeepResearch Technical Report

As LLMs shift toward autonomous agents, Deep Research has emerged as a pivotal metric. However, existing academic benchmarks like BrowseComp often fail to meet real-world demands for open-ended research, which requires robust skills in intent recognition, long-horizon decision-making, and cross-source verification. To address this, we introduce Step-DeepResearch, a cost-effective, end-to-end agent. We propose a Data Synthesis Strategy Based on Atomic Capabilities to reinforce planning and report writing, combined with a progressive training path from agentic mid-training to SFT and RL. Enhanced by a Checklist-style Judger, this approach significantly improves robustness. Furthermore, to bridge the evaluation gap in the Chinese domain, we establish ADR-Bench for realistic deep research scenarios. Experimental results show that Step-DeepResearch (32B) scores 61.4% on Scale AI Research Rubrics. On ADR-Bench, it significantly outperforms comparable models and rivals SOTA closed-source models like OpenAI and Gemini DeepResearch. These findings prove that refined training enables medium-sized models to achieve expert-level capabilities at industry-leading cost-efficiency.

FlakyGuard: Automatically Fixing Flaky Tests at Industry Scale

Flaky tests that non-deterministically pass or fail waste developer time and slow release cycles. While large language models (LLMs) show promise for automatically repairing flaky tests, existing approaches like FlakyDoctor fail in industrial settings due to the context problem: providing either too little context (missing critical production code) or too much context (overwhelming the LLM with irrelevant information). We present FlakyGuard, which addresses this problem by treating code as a graph structure and using selective graph exploration to find only the most relevant context. Evaluation on real-world flaky tests from industrial repositories shows that FlakyGuard repairs 47.6 % of reproducible flaky tests with 51.8 % of the fixes accepted by developers. Besides it outperforms state-of-the-art approaches by at least 22 % in repair success rate. Developer surveys confirm that 100 % find FlakyGuard's root cause explanations useful.

An Integrated Fusion Framework for Ensemble Learning Leveraging Gradient Boosting and Fuzzy Rule-Based Models

The integration of different learning paradigms has long been a focus of machine learning research, aimed at overcoming the inherent limitations of individual methods. Fuzzy rule-based models excel in interpretability and have seen widespread application across diverse fields. However, they face challenges such as complex design specifications and scalability issues with large datasets. The fusion of different techniques and strategies, particularly Gradient Boosting, with Fuzzy Rule-Based Models offers a robust solution to these challenges. This paper proposes an Integrated Fusion Framework that merges the strengths of both paradigms to enhance model performance and interpretability. At each iteration, a Fuzzy Rule-Based Model is constructed and controlled by a dynamic factor to optimize its contribution to the overall ensemble. This control factor serves multiple purposes: it prevents model dominance, encourages diversity, acts as a regularization parameter, and provides a mechanism for dynamic tuning based on model performance, thus mitigating the risk of overfitting. Additionally, the framework incorporates a sample-based correction mechanism that allows for adaptive adjustments based on feedback from a validation set. Experimental results substantiate the efficacy of the presented gradient boosting framework for fuzzy rule-based models, demonstrating performance enhancement, especially in terms of mitigating overfitting and complexity typically associated with many rules. By leveraging an optimal factor to govern the contribution of each model, the framework improves performance, maintains interpretability, and simplifies the maintenance and update of the models.

Towards Adaptive Humanoid Control via Multi-Behavior Distillation and Reinforced Fine-Tuning

Humanoid robots are promising to learn a diverse set of human-like locomotion behaviors, including standing up, walking, running, and jumping. However, existing methods predominantly require training independent policies for each skill, yielding behavior-specific controllers that exhibit limited generalization and brittle performance when deployed on irregular terrains and in diverse situations. To address this challenge, we propose Adaptive Humanoid Control (AHC) that adopts a two-stage framework to learn an adaptive humanoid locomotion controller across different skills and terrains. Specifically, we first train several primary locomotion policies and perform a multi-behavior distillation process to obtain a basic multi-behavior controller, facilitating adaptive behavior switching based on the environment. Then, we perform reinforced fine-tuning by collecting online feedback in performing adaptive behaviors on more diverse terrains, enhancing terrain adaptability for the controller. We conduct experiments in both simulation and real-world experiments in Unitree G1 robots. The results show that our method exhibits strong adaptability across various situations and terrains. Project website: https://ahc-humanoid.github.io.

Implicit Federated In-context Learning For Task-Specific LLM Fine-Tuning

As large language models continue to develop and expand, the extensive public data they rely on faces the risk of depletion. Consequently, leveraging private data within organizations to enhance the performance of large models has emerged as a key challenge. The federated learning paradigm, combined with model fine-tuning techniques, effectively reduces the number of trainable parameters. However,the necessity to process high-dimensional feature spaces results in substantial overall computational overhead. To address this issue, we propose the Implicit Federated In-Context Learning (IFed-ICL) framework. IFed-ICL draws inspiration from federated learning to establish a novel distributed collaborative paradigm, by converting client local context examples into implicit vector representations, it enables distributed collaborative computation during the inference phase and injects model residual streams to enhance model performance. Experiments demonstrate that our proposed method achieves outstanding performance across multiple text classification tasks. Compared to traditional methods, IFed-ICL avoids the extensive parameter updates required by conventional fine-tuning methods while reducing data transmission and local computation at the client level in federated learning. This enables efficient distributed context learning using local private-domain data, significantly improving model performance on specific tasks.

HaluMem: Evaluating Hallucinations in Memory Systems of Agents

Memory systems are key components that enable AI systems such as LLMs and AI agents to achieve long-term learning and sustained interaction. However, during memory storage and retrieval, these systems frequently exhibit memory hallucinations, including fabrication, errors, conflicts, and omissions. Existing evaluations of memory hallucinations are primarily end-to-end question answering, which makes it difficult to localize the operational stage within the memory system where hallucinations arise. To address this, we introduce the Hallucination in Memory Benchmark (HaluMem), the first operation level hallucination evaluation benchmark tailored to memory systems. HaluMem defines three evaluation tasks (memory extraction, memory updating, and memory question answering) to comprehensively reveal hallucination behaviors across different operational stages of interaction. To support evaluation, we construct user-centric, multi-turn human-AI interaction datasets, HaluMem-Medium and HaluMem-Long. Both include about 15k memory points and 3.5k multi-type questions. The average dialogue length per user reaches 1.5k and 2.6k turns, with context lengths exceeding 1M tokens, enabling evaluation of hallucinations across different context scales and task complexities. Empirical studies based on HaluMem show that existing memory systems tend to generate and accumulate hallucinations during the extraction and updating stages, which subsequently propagate errors to the question answering stage. Future research should focus on developing interpretable and constrained memory operation mechanisms that systematically suppress hallucinations and improve memory reliability.

PIVOTS: Aligning unseen Structures using Preoperative to Intraoperative Volume-To-Surface Registration for Liver Navigation

Non-rigid registration is essential for Augmented Reality guided laparoscopic liver surgery by fusing preoperative information, such as tumor location and vascular structures, into the limited intraoperative view, thereby enhancing surgical navigation. A prerequisite is the accurate prediction of intraoperative liver deformation which remains highly challenging due to factors such as large deformation caused by pneumoperitoneum, respiration and tool interaction as well as noisy intraoperative data, and limited field of view due to occlusion and constrained camera movement. To address these challenges, we introduce PIVOTS, a Preoperative to Intraoperative VOlume-To-Surface registration neural network that directly takes point clouds as input for deformation prediction. The geometric feature extraction encoder allows multi-resolution feature extraction, and the decoder, comprising novel deformation aware cross attention modules, enables pre- and intraoperative information interaction and accurate multi-level displacement prediction. We train the neural network on synthetic data simulated from a biomechanical simulation pipeline and validate its performance on both synthetic and real datasets. Results demonstrate superior registration performance of our method compared to baseline methods, exhibiting strong robustness against high amounts of noise, large deformation, and various levels of intraoperative visibility. We publish the training and test sets as evaluation benchmarks and call for a fair comparison of liver registration methods with volume-to-surface data. Code and datasets are available here https://github.com/pengliu-nct/PIVOTS.

Step-Audio 2 Technical Report

This paper presents Step-Audio 2, an end-to-end multi-modal large language model designed for industry-strength audio understanding and speech conversation. By integrating a latent audio encoder and reasoning-centric reinforcement learning (RL), Step-Audio 2 achieves promising performance in automatic speech recognition (ASR) and audio understanding. To facilitate genuine end-to-end speech conversation, Step-Audio 2 incorporates the generation of discrete audio tokens into language modeling, significantly enhancing its responsiveness to paralinguistic information such as speaking styles and emotions. To effectively leverage the rich textual and acoustic knowledge in real-world data, Step-Audio 2 integrates retrieval-augmented generation (RAG) and is able to call external tools such as web search to mitigate hallucination and audio search to switch timbres. Trained on millions of hours of speech and audio data, Step-Audio 2 delivers intelligence and expressiveness across diverse conversational scenarios. Evaluation results demonstrate that Step-Audio 2 achieves state-of-the-art performance on various audio understanding and conversational benchmarks compared to other open-source and commercial solutions. Please visit https://github.com/stepfun-ai/Step-Audio2 for more information.

Residual Prior-driven Frequency-aware Network for Image Fusion

Image fusion aims to integrate complementary information across modalities to generate high-quality fused images, thereby enhancing the performance of high-level vision tasks. While global spatial modeling mechanisms show promising results, constructing long-range feature dependencies in the spatial domain incurs substantial computational costs. Additionally, the absence of ground-truth exacerbates the difficulty of capturing complementary features effectively. To tackle these challenges, we propose a Residual Prior-driven Frequency-aware Network, termed as RPFNet. Specifically, RPFNet employs a dual-branch feature extraction framework: the Residual Prior Module (RPM) extracts modality-specific difference information from residual maps, thereby providing complementary priors for fusion; the Frequency Domain Fusion Module (FDFM) achieves efficient global feature modeling and integration through frequency-domain convolution. Additionally, the Cross Promotion Module (CPM) enhances the synergistic perception of local details and global structures through bidirectional feature interaction. During training, we incorporate an auxiliary decoder and saliency structure loss to strengthen the model's sensitivity to modality-specific differences. Furthermore, a combination of adaptive weight-based frequency contrastive loss and SSIM loss effectively constrains the solution space, facilitating the joint capture of local details and global features while ensuring the retention of complementary information. Extensive experiments validate the fusion performance of RPFNet, which effectively integrates discriminative features, enhances texture details and salient objects, and can effectively facilitate the deployment of the high-level vision task.

High-Layer Attention Pruning with Rescaling

Pruning is a highly effective approach for compressing large language models (LLMs), significantly reducing inference latency. However, conventional training-free structured pruning methods often employ a heuristic metric that indiscriminately removes some attention heads across all pruning layers, without considering their positions within the network architecture. In this work, we propose a novel pruning algorithm that strategically prunes attention heads in the model's higher layers. Since the removal of attention heads can alter the magnitude of token representations, we introduce an adaptive rescaling parameter that calibrates the representation scale post-pruning to counteract this effect. We conduct comprehensive experiments on a wide range of LLMs, including LLaMA3.1-8B, Mistral-7B-v0.3, Qwen2-7B, and Gemma2-9B. Our evaluation includes both generation and discriminative tasks across 27 datasets. The results consistently demonstrate that our method outperforms existing structured pruning methods. This improvement is particularly notable in generation tasks, where our approach significantly outperforms existing baselines.