Huiming Wang

Xiao Zhu, Xinyu Zhou, Boyu Zhu et al.

Reinforcement Learning from Verifiable Rewards (RLVR) has driven recent progress in code large language models by leveraging execution-based feedback from unit tests, but its scalability is fundamentally constrained by the availability and reliability of high-quality test cases. We propose CodeScaler, a reward model designed to scale both reinforcement learning training and test-time inference for code generation. CodeScaler is trained on carefully curated preference data derived from verified code problems and incorporates syntax-aware code extraction and validity-preserving reward shaping to ensure stable and robust optimization. Across four coding benchmarks, CodeScaler consistently outperforms execution-based RL by +1.55 points on Qwen3-8B-Base and +4.23 points on Qwen3-14B-Base. By further scaling to 44K problems with additional synthetic data, CodeScaler yields +14.64 points improvement over the base model without requiring any test cases. At inference time, CodeScaler serves as an effective test-time scaling method, achieving performance comparable to unit test approaches while providing a 10-fold reduction in latency. Moreover, CodeScaler surpasses existing reward models on RM-Bench not only in the code domain (+3.3 points), but also in general and reasoning domains (+2.7 points on average).

Zaoyu Chen, Jianbo Dai, Boyu Zhu et al.

Large language models (LLMs) can generate code from natural language, but the extent to which they capture intended program behavior remains unclear. Executable behavioral specifications, defined via preconditions and postconditions, provide a concrete means to assess such understanding. However, existing work on specification generation is constrained in evaluation methodology, task settings, and specification expressiveness. We introduce CodeSpecBench, a benchmark for executable behavioral specification generation under an execution-based evaluation protocol. CodeSpecBench supports both function-level and repository-level tasks and encodes specifications as executable Python functions. Constructed from diverse real-world codebases, it enables a realistic assessment of both correctness (accepting valid behaviors) and completeness (rejecting invalid behaviors). Evaluating 15 state-of-the-art LLMs on CodeSpecBench, we observe a sharp performance degradation on repository-level tasks, where the best model attains only a 20.2% pass rate. We further find that specification generation is substantially more challenging than code generation, indicating that strong coding performance does not necessarily reflect deep understanding of intended program semantics. Our data and code are available at https://github.com/SparksofAGI/CodeSpecBench.

Huiming Wang, Zhaodonghui Li, Liying Cheng et al.

Recently, large language models (LLMs) have emerged as a groundbreaking technology and their unparalleled text generation capabilities have sparked interest in their application to the fundamental sentence representation learning task. Existing methods have explored utilizing LLMs as data annotators to generate synthesized data for training contrastive learning based sentence embedding models such as SimCSE. However, since contrastive learning models are sensitive to the quality of sentence pairs, the effectiveness of these methods is largely influenced by the content generated from LLMs, highlighting the need for more refined generation in the context of sentence representation learning. Building upon this premise, we propose MultiCSR, a multi-level contrastive sentence representation learning framework that decomposes the process of prompting LLMs to generate a corpus for training base sentence embedding models into three stages (i.e., sentence generation, sentence pair construction, in-batch training) and refines the generated content at these three distinct stages, ensuring only high-quality sentence pairs are utilized to train a base contrastive learning model. Our extensive experiments reveal that MultiCSR enables a less advanced LLM to surpass the performance of ChatGPT, while applying it to ChatGPT achieves better state-of-the-art results. Comprehensive analyses further underscore the potential of our framework in various application scenarios and achieving better sentence representation learning with LLMs.

Yinan Xia, Haotian Zhang, Huiming Wang

Large Reasoning Models (LRMs) have shown exceptional reasoning capabilities, but they also suffer from the issue of overthinking, often generating excessively long and redundant answers. For problems that exceed the model's capabilities, LRMs tend to exhibit the overconfidence phenomenon, generating overly short but incorrect answers, which may contribute to suboptimal performance. To address these issues, we propose Difficulty-Differentiated Policy Optimization (DDPO), an efficient reinforcement learning algorithm that optimizes simple and complex tasks separately based on the overconfidence phenomenon. Specifically, it reduces the output length for simple tasks without compromising accuracy, while for complex tasks, it expands the exploration space to improve performance. We further derive the theoretical conditions for maximizing expected accuracy, which require the length distribution to closely approximate the optimal length and be as concentrated as possible. Based on these conditions, we propose using the difficulty-level average as a well-founded reference for length optimization. Extensive experiments on both in-domain and out-of-domain benchmarks validate the superiority and effectiveness of DDPO. Compared to GRPO, DDPO reduces the average answer length by 12% while improving accuracy by 1.85% across multiple benchmarks, achieving a better trade-off between accuracy and length. The code is available at https://github.com/Yinan-Xia/DDPO.

Jingxuan Xu, Ken Deng, Weihao Li et al.

Evaluating large language models (LLMs) for software engineering has been limited by narrow task coverage, language bias, and insufficient alignment with real-world developer workflows. Existing benchmarks often focus on algorithmic problems or Python-centric bug fixing, leaving critical dimensions of software engineering underexplored. To address these gaps, we introduce SWE-Compass1, a comprehensive benchmark that unifies heterogeneous code-related evaluations into a structured and production-aligned framework. SWE-Compass spans 8 task types, 8 programming scenarios, and 10 programming languages, with 2000 high-quality instances curated from authentic GitHub pull requests and refined through systematic filtering and validation. We benchmark ten state-of-the-art LLMs under two agentic frameworks, SWE-Agent and Claude Code, revealing a clear hierarchy of difficulty across task types, languages, and scenarios. Moreover, by aligning evaluation with real-world developer practices, SWE-Compass provides a rigorous and reproducible foundation for diagnosing and advancing agentic coding capabilities in large language models.

Zizheng Zhan, Ken Deng, Huaixi Tang et al.

We present Kwaipilot-AutoThink (KAT), an open-source 40B large language model developed to address the overthinking problem in reasoning-intensive tasks, where an automatic thinking training paradigm is proposed to dynamically switch between reasoning and non-reasoning modes based on task complexity. Specifically, first, we construct the dual-regime dataset based on a novel tagging pipeline and a multi-agent synthesis strategy, and then we apply Multi-Token Prediction (MTP)-enhanced knowledge distillation, enabling efficient and fine-grained reasoning transfer with minimal pretraining cost. Besides, we implement a cold-start initialization strategy that introduces mode-selection priors using majority-vote signals and intent-aware prompting. Finally, we propose Step-SRPO, a reinforcement learning algorithm that incorporates intermediate supervision into the GRPO framework, offering structured guidance over both reasoning-mode selection and response accuracy. Extensive experiments across multiple benchmarks demonstrate that KAT consistently matches or even outperforms current state-of-the-art models, including DeepSeek-R1-0528 and Qwen3-235B-A22B, across a wide range of reasoning-intensive tasks while reducing token usage. Notably, KAT outperforms all open-source models and even surpasses o3-mini on the leakage-controlled LiveCodeBench Pro. Beyond academic evaluation, KAT has been successfully deployed in Kwaipilot (i.e., Kuaishou's internal coding assistant), where it improves real-world development workflows with high accuracy, efficiency, and controllable reasoning behaviors. Moreover, we are actively training a 200B Mixture-of-Experts (MoE) model with 40B active parameters, and early results already show significant gains, further demonstrating the scalability of the AutoThink paradigm.

Xizhong Yang, Yinan Xia, Huiming Wang et al.

Leveraging the model's internal information as the self-reward signal in Reinforcement Learning (RL) has received extensive attention due to its label-free nature. While prior works have made significant progress in applying the Test-Time Scaling (TTS) strategies to RL, the discrepancy in internal information between test and training remains inadequately addressed. Moreover, Test-Time Training based on voting-based TTS strategies often suffers from reward hacking problems. To address these issues, we propose DistriTTRL, which leverages the distribution prior of the model's confidence during RL to progressively optimize the reward signal, rather than relying solely on single-query rollouts. Additionally, we mitigate the phenomenon of consistent reward hacking caused by the voting-based TTS strategies through diversity-targeted penalties. Benefiting from this training mechanism where model capability and self-reward signals complement each other, and the mitigation of reward hacking, DistriTTRL has achieved significant performance improvements across multiple models and benchmarks.

Zizheng Zhan, Ken Deng, Jinghui Wang et al.

Recent advances in large language models (LLMs) have enabled progress in agentic coding, where models autonomously reason, plan, and act within interactive software development workflows. However, bridging the gap between static text-based training and dynamic real-world agentic execution remains a core challenge. In this technical report, we present KAT-Coder, a large-scale agentic code model trained through a multi-stage curriculum encompassing Mid-Term Training, Supervised Fine-Tuning (SFT), Reinforcement Fine-Tuning (RFT), and Reinforcement-to-Deployment Adaptation. The Mid-Term stage enhances reasoning, planning, and reflection capabilities through a corpus of real software engineering data and synthetic agentic interactions. The SFT stage constructs a million-sample dataset balancing twenty programming languages, ten development contexts, and ten task archetypes. The RFT stage introduces a novel multi-ground-truth reward formulation for stable and sample-efficient policy optimization. Finally, the Reinforcement-to-Deployment phase adapts the model to production-grade IDE environments using Error-Masked SFT and Tree-Structured Trajectory Training. In summary, these stages enable KAT-Coder to achieve robust tool-use reliability, instruction alignment, and long-context reasoning, forming a deployable foundation for real-world intelligent coding agents. Our KAT series 32B model, KAT-Dev, has been open-sourced on https://huggingface.co/Kwaipilot/KAT-Dev.

Xinyu Zhou, Boyu Zhu, Haotian Zhang et al.

Reinforcement learning (RL) plays a central role in improving the reasoning and alignment of large language models, yet its efficiency critically depends on how training data are selected. Existing online selection strategies predominantly rely on difficulty-based heuristics, favouring datapoints with intermediate success rates, implicitly equating difficulty with informativeness and neglecting epistemic uncertainty arising from limited evidence. We introduce InSight, an INformation-guided data SamplInG metHod for RL Training, grounded in a weighted mutual information objective. By modeling data outcomes with Bayesian latent success rates, we show that expected uncertainty reduction decomposes into complementary difficulty- and evidence-dependent components, revealing a fundamental limitation of difficulty-only selection. Leveraging this observation, InSight constructs a stable acquisition score based on the mean belief of datapoints' success rather than noisy sampled outcomes, and naturally extends to multi-rollout settings common in reinforcement learning with verifiable rewards (RLVR). Extensive experiments demonstrate that InSight consistently achieves state-of-the-art performance and improves training efficiency, including a +1.41 average gain on Planning & Mathmatics benchmarks, +1.01 improvement on general reasoning, and up to ~2.2x acceleration, with negligible additional computational overhead.

Xizhong Yang, Haotian Zhang, Huiming Wang et al.

Large Reasoning Models have demonstrated remarkable performance with the advancement of test-time scaling techniques, which enhances prediction accuracy by generating multiple candidate responses and selecting the most reliable answer. While prior work has analyzed that internal model signals like confidence scores can partly indicate response correctness and exhibit a distributional correlation with accuracy, such distributional information has not been fully utilized to guide answer selection. Motivated by this, we propose DistriVoting, which incorporates distributional priors as another signal alongside confidence during voting. Specifically, our method (1) first decomposes the mixed confidence distribution into positive and negative components using Gaussian Mixture Models, (2) then applies a reject filter based on positive/negative samples from them to mitigate overlap between the two distributions. Besides, to further alleviate the overlap from the perspective of distribution itself, we propose SelfStepConf, which uses step-level confidence to dynamically adjust inference process, increasing the separation between the two distributions to improve the reliability of confidences in voting. Experiments across 16 models and 5 benchmarks demonstrate that our method significantly outperforms state-of-the-art approaches.

Yiran Zhao, Wenxuan Zhang, Huiming Wang et al.

As an effective alternative to the direct fine-tuning on target tasks in specific languages, cross-lingual transfer addresses the challenges of limited training data by decoupling ''task ability'' and ''language ability'' by fine-tuning on the target task in the source language and another selected task in the target language, respectively. However, they fail to fully separate the task ability from the source language or the language ability from the chosen task. In this paper, we acknowledge the mutual reliance between task ability and language ability and direct our attention toward the gap between the target language and the source language on tasks. As the gap removes the impact of tasks, we assume that it remains consistent across tasks. Based on this assumption, we propose a new cross-lingual transfer method called $\texttt{AdaMergeX}$ that utilizes adaptive adapter merging. By introducing a reference task, we can determine that the divergence of adapters fine-tuned on the reference task in both languages follows the same distribution as the divergence of adapters fine-tuned on the target task in both languages. Hence, we can obtain target adapters by combining the other three adapters. Furthermore, we propose a structure-adaptive adapter merging method. Our empirical results demonstrate that our approach yields new and effective cross-lingual transfer, outperforming existing methods across all settings.

Zhaodonghui Li, Haitao Yuan, Huiming Wang et al.

Query rewrite, which aims to generate more efficient queries by altering a SQL query's structure without changing the query result, has been an important research problem. In order to maintain equivalence between the rewritten query and the original one during rewriting, traditional query rewrite methods always rewrite the queries following certain rewrite rules. However, some problems still remain. Firstly, existing methods of finding the optimal choice or sequence of rewrite rules are still limited and the process always costs a lot of resources. Methods involving discovering new rewrite rules typically require complicated proofs of structural logic or extensive user interactions. Secondly, current query rewrite methods usually rely highly on DBMS cost estimators which are often not accurate. In this paper, we address these problems by proposing a novel method of query rewrite named LLM-R2, adopting a large language model (LLM) to propose possible rewrite rules for a database rewrite system. To further improve the inference ability of LLM in recommending rewrite rules, we train a contrastive model by curriculum to learn query representations and select effective query demonstrations for the LLM. Experimental results have shown that our method can significantly improve the query execution efficiency and outperform the baseline methods. In addition, our method enjoys high robustness across different datasets.

Xizhong Yang, Huiming Wang, Ning Xu et al.

Distribution shifts between training and testing data are a critical bottleneck limiting the practical utility of models, especially in real-world test-time scenarios. To adapt models when the source domain is unknown and the target domain is unlabeled, previous works constructed pseudo-source domains via data generation and translation, then aligned the target domain with them. However, significant discrepancies exist between the pseudo-source and the original source domain, leading to potential divergence when correcting the target directly. From this perspective, we propose a Stepwise Semantic Alignment (SSA) method, viewing the pseudo-source as a semantic bridge connecting the source and target, rather than a direct substitute for the source. Specifically, we leverage easily accessible universal semantics to rectify the semantic features of the pseudo-source, and then align the target domain using the corrected pseudo-source semantics. Additionally, we introduce a Hierarchical Feature Aggregation (HFA) module and a Confidence-Aware Complementary Learning (CACL) strategy to enhance the semantic quality of the SSA process in the absence of source and ground truth of target domains. We evaluated our approach on tasks like semantic segmentation and image classification, achieving a 5.2% performance boost on GTA2Cityscapes over the state-of-the-art.

Linyue Cai, Yuyang Cheng, Xiaoding Shao et al.

As artificial intelligence advances toward artificial general intelligence (AGI), the need for robust and human-like memory systems has become increasingly evident. Current memory architectures often suffer from limited adaptability, insufficient multimodal integration, and an inability to support continuous learning. To address these limitations, we propose a scenario-driven methodology that extracts essential functional requirements from representative cognitive scenarios, leading to a unified set of design principles for next-generation AI memory systems. Based on this approach, we introduce the \textbf{COgnitive Layered Memory Architecture (COLMA)}, a novel framework that integrates cognitive scenarios, memory processes, and storage mechanisms into a cohesive design. COLMA provides a structured foundation for developing AI systems capable of lifelong learning and human-like reasoning, thereby contributing to the pragmatic development of AGI.

Longze Chen, Renke Shan, Huiming Wang et al.

Speculative decoding (SD) is a promising method for accelerating the decoding process of Large Language Models (LLMs). The efficiency of SD primarily hinges on the consistency between the draft model and the verify model. However, existing drafting approaches typically require additional modules to be trained, which can be challenging to implement and ensure compatibility across various LLMs. In this paper, we propose CLaSp, an in-context layer-skipping strategy for self-speculative decoding. Unlike prior methods, CLaSp does not require additional drafting modules or extra training. Instead, it employs a plug-and-play mechanism by skipping intermediate layers of the verify model to construct a compressed draft model. Specifically, we develop a dynamic programming algorithm that optimizes the layer-skipping process by leveraging the complete hidden states from the last verification stage as an objective. This enables CLaSp to dynamically adjust its layer-skipping strategy after each verification stage, without relying on pre-optimized sets of skipped layers. Experimental results across diverse downstream tasks demonstrate that CLaSp achieves a speedup of 1.3x ~ 1.7x on LLaMA3 series models without altering the original distribution of the generated text.

Jinghui Wang, Shaojie Wang, Yinghan Cui et al.

We introduce SeamlessFlow, a server based reinforcement learning (RL) framework that addresses two core challenges in industrial scale RL: (1) decoupling RL training from the complex execution flow of agents; (2) maximizing GPU utilization with minimal idle time while preserving the stability and scalability required for large-scale deployments. First, SeamlessFlow introduces a data plane that decouples the RL trainer from diverse, complex agent implementations while sustaining high throughput. A central trajectory manager maintains complete interaction histories and supports partial rollout, allowing rollout to pause for weight updates and resume seamlessly, keeping agents unaware of service interruptions. Second, we propose a tag driven scheduling paradigm that abstracts hardware into capability tagged resources, unifying colocated and disaggregated architectures. Based on this, SeamlessFlow introduces a spatiotemporal multiplexing pipeline that dynamically reassigns idle training nodes to rollout in a train rollout separated setup, eliminating pipeline bubbles and fully exploiting heterogeneous cluster resources. By combining these innovations, SeamlessFlow delivers both stability and high performance, making it well suited for multi agent, long horizon, and other complex RL tasks.

Huiming Wang, Liying Cheng, Wenxuan Zhang et al.

Recently, data augmentation (DA) methods have been proven to be effective for pre-trained language models (PLMs) in low-resource settings, including few-shot named entity recognition (NER). However, conventional NER DA methods are mostly aimed at sequence labeling models, i.e., token-level classification, and few are compatible with unified autoregressive generation frameworks, which can handle a wider range of NER tasks, such as nested NER. Furthermore, these generation frameworks have a strong assumption that the entities will appear in the target sequence with the same left-to-right order as the source sequence. In this paper, we claim that there is no need to keep this strict order, and more diversified but reasonable target entity sequences can be provided during the training stage as a novel DA method. Nevertheless, a naive mixture of augmented data can confuse the model since one source sequence will then be paired with different target sequences. Therefore, we propose a simple but effective Prompt Ordering based Data Augmentation (PODA) method to improve the training of unified autoregressive generation frameworks under few-shot NER scenarios. Experimental results on three public NER datasets and further analyses demonstrate the effectiveness of our approach.