arXiv subjects

Wayne Xin Zhao

Publications and source records attributed to Wayne Xin Zhao.

At least 55 records · Page 3Linked to original sources

arXiv2026

Controlled LLM Training on Spectral Sphere

Scaling large models requires optimization strategies that ensure rapid convergence grounded in stability. Maximal Update Parametrization ($\boldsymbolμ$P) provides a theoretical safeguard for width-invariant $Θ(1)$ activation control, whereas emerging optimizers like Muon are only ``half-aligned'' with these constraints: they control updates but allow weights to drift. To address this limitation, we introduce the \textbf{Spectral Sphere Optimizer (SSO)}, which enforces strict module-wise spectral constraints on both weights and their updates. By deriving the steepest descent direction on the spectral sphere, SSO realizes a fully $\boldsymbolμ$P-aligned optimization process. To enable large-scale training, we implement SSO as an efficient parallel algorithm within Megatron. Through extensive pretraining on diverse architectures, including Dense 1.7B, MoE 8B-A1B, and 200-layer DeepNet models, SSO consistently outperforms AdamW and Muon. Furthermore, we observe significant practical stability benefits, including improved MoE router load balancing, suppressed outliers, and strictly bounded activations.

cs.LG↗

arXiv2026

SWE-Master: Unleashing the Potential of Software Engineering Agents via Post-Training

In this technical report, we present SWE-Master, an open-source and fully reproducible post-training framework for building effective software engineering agents. SWE-Master systematically explores the complete agent development pipeline, including teacher-trajectory synthesis and data curation, long-horizon SFT, RL with real execution feedback, and inference framework design. Starting from an open-source base model with limited initial SWE capability, SWE-Master demonstrates how systematical optimization method can elicit strong long-horizon SWE task solving abilities. We evaluate SWE-Master on SWE-bench Verified, a standard benchmark for realistic software engineering tasks. Under identical experimental settings, our approach achieves a resolve rate of 61.4\% with Qwen2.5-Coder-32B, substantially outperforming existing open-source baselines. By further incorporating test-time scaling~(TTS) with LLM-based environment feedback, SWE-Master reaches 70.8\% at TTS@8, demonstrating a strong performance potential. SWE-Master provides a practical and transparent foundation for advancing reproducible research on software engineering agents. The code is available at https://github.com/RUCAIBox/SWE-Master.

cs.SE↗

arXiv2026

Improving LLM-based Recommendation with Self-Hard Negatives from Intermediate Layers

Large language models (LLMs) have shown great promise in recommender systems, where supervised fine-tuning (SFT) is commonly used for adaptation. Subsequent studies further introduce preference learning to incorporate negative samples into the training process. However, existing methods rely on sequence-level, offline-generated negatives, making them less discriminative and informative when adapting LLMs to recommendation tasks with large negative item spaces. To address these challenges, we propose ILRec, a novel preference fine-tuning framework for LLM-based recommendation, leveraging self-hard negative signals extracted from intermediate layers to improve preference learning. Specifically, we identify self-hard negative tokens from intermediate layers as fine-grained negative supervision that dynamically reflects the model's preference learning process. To effectively integrate these signals into training, we design a two-stage framework comprising cross-layer preference optimization and cross-layer preference distillation, enabling the model to jointly discriminate informative negatives and enhance the quality of negative signals from intermediate layers. In addition, we introduce a lightweight collaborative filtering model to assign token-level rewards for negative signals, mitigating the risk of over-penalizing false negatives. Extensive experiments on three datasets demonstrate ILRec's effectiveness in enhancing the performance of LLM-based recommender systems.

cs.IR↗

arXiv2026

RV-Syn: Rational and Verifiable Mathematical Reasoning Data Synthesis based on Structured Function Library

The advancement of reasoning capabilities in Large Language Models (LLMs) requires substantial amounts of high-quality reasoning data, particularly in mathematics. Existing data synthesis methods, such as data augmentation from annotated training sets or direct question generation based on relevant knowledge points and documents, have expanded datasets but face challenges in mastering the inner logic of the problem during generation and ensuring the verifiability of the solutions. To address these issues, we propose RV-Syn, a novel Rational and Verifiable mathematical Synthesis approach. RV-Syn constructs a structured mathematical operation function library based on initial seed problems and generates computational graphs as solutions by combining Python-formatted functions from this library. These graphs are then back-translated into complex problems. Based on the constructed computation graph, we achieve solution-guided logic-aware problem generation. Furthermore, the executability of the computational graph ensures the verifiability of the solving process. Experimental results show that RV-Syn surpasses existing synthesis methods, including those involving human-generated problems, achieving greater efficient data scaling. This approach provides a scalable framework for generating high-quality reasoning datasets.

cs.AI↗

arXiv2026

ReForm: Reflective Autoformalization with Prospective Bounded Sequence Optimization

Autoformalization, which translates natural language mathematics into machine-verifiable formal statements, is critical for using formal mathematical reasoning to solve math problems stated in natural language. While Large Language Models can generate syntactically correct formal statements, they often fail to preserve the original problem's semantic intent. This limitation arises from the LLM approaches' treating autoformalization as a simplistic translation task which lacks mechanisms for self-reflection and iterative refinement that human experts naturally employ. To address these issues, we propose ReForm, a Reflective Autoformalization method that tightly integrates semantic consistency evaluation into the autoformalization process. This enables the model to iteratively generate formal statements, assess its semantic fidelity, and self-correct identified errors through progressive refinement. To effectively train this reflective model, we introduce Prospective Bounded Sequence Optimization (PBSO), which employs different rewards at different sequence positions to ensure that the model develops both accurate autoformalization and correct semantic validations, preventing superficial critiques that would undermine the purpose of reflection. Extensive experiments across four autoformalization benchmarks demonstrate that ReForm achieves an average improvement of 22.6 percentage points over the strongest baselines. To further ensure evaluation reliability, we introduce ConsistencyCheck, a benchmark of 859 expert-annotated items that not only validates LLMs as judges but also reveals that autoformalization is inherently difficult: even human experts produce semantic errors in up to 38.5% of cases.

cs.CL↗

arXiv2026

SWE-World: Building Software Engineering Agents in Docker-Free Environments

Recent advances in large language models (LLMs) have enabled software engineering agents to tackle complex code modification tasks. Most existing approaches rely on execution feedback from containerized environments, which require dependency-complete setup and physical execution of programs and tests. While effective, this paradigm is resource-intensive and difficult to maintain, substantially complicating agent training and limiting scalability. We propose SWE-World, a Docker-free framework that replaces physical execution environments with a learned surrogate for training and evaluating software engineering agents. SWE-World leverages LLM-based models trained on real agent-environment interaction data to predict intermediate execution outcomes and final test feedback, enabling agents to learn without interacting with physical containerized environments. This design preserves the standard agent-environment interaction loop while eliminating the need for costly environment construction and maintenance during agent optimization and evaluation. Furthermore, because SWE-World can simulate the final evaluation outcomes of candidate trajectories without real submission, it enables selecting the best solution among multiple test-time attempts, thereby facilitating effective test-time scaling (TTS) in software engineering tasks. Experiments on SWE-bench Verified demonstrate that SWE-World raises Qwen2.5-Coder-32B from 6.2\% to 52.0\% via Docker-free SFT, 55.0\% with Docker-free RL, and 68.2\% with further TTS. The code is available at https://github.com/RUCAIBox/SWE-World

cs.SE↗

arXiv2026

MARS: Co-evolving Dual-System Deep Research via Multi-Agent Reinforcement Learning

Large Reasoning Models (LRMs) face two fundamental limitations: excessive token consumption when overanalyzing simple information processing tasks, and inability to access up-to-date knowledge beyond their training data. We introduce MARS (Multi-Agent System for Deep ReSearch), a novel co-evolution framework that jointly optimizes dual cognitive systems through multi-agent reinforcement learning. Unlike prior approaches that employ fixed or independently-trained summarizers, MARS enables System 1 (fast, intuitive processing) and System 2 (deliberate reasoning) to co-adapt through shared trajectory rewards, developing complementary strategies where System 1 learns to distill information specifically useful for System 2's reasoning. We extend Group Relative Policy Optimization (GRPO) for multi-agent settings with three key innovations: (1) decoupled gradient computation ensuring proper credit assignment despite shared rewards, (2) bin-packing optimization for efficient parallel information processing, and (3) advantage-weighted balanced sampling preventing training imbalance. Extensive experiments demonstrate that MARS (8B), trained under a challenging Zero RL setting without any supervised fine-tuning, achieves 8.17% on HLE -- outperforming WebThinker (32B with SFT, 6.87%) and narrowing the gap with proprietary models like Claude 3.7 Sonnet (7.89%) -- while achieving an average gain of 8.9% across 7 knowledge-intensive tasks.

cs.AI↗

arXiv2026

IterResearch: Rethinking Long-Horizon Agents with Interaction Scaling

Recent advances in deep-research agents have shown promise for autonomous knowledge construction through dynamic reasoning over external sources. However, existing approaches rely on a mono-contextual paradigm that accumulates all information in a single, expanding context window, leading to context suffocation and noise contamination that limit their effectiveness on long-horizon tasks. We introduce \textbf{IterResearch}, a novel iterative deep-research paradigm that revisits long-horizon research through the lens of Interaction Scaling. Instead of relying on linear context accumulation, we adopt an MDP-inspired architecture with strategic workspace reconstruction. By maintaining an evolving report as memory and periodically synthesizing insights, our approach preserves consistent reasoning capacity across arbitrary exploration depths. To effectively train this paradigm, we employ Efficiency-Aware Policy Optimization (EAPO), a training strategy that adapts geometric reward discounting to incentivize efficient exploration and utilizes adaptive downsampling for stable distributed training. Extensive experiments demonstrate that IterResearch achieves substantial improvements over existing open-source agents with average +14.5pp across six benchmarks and narrows the gap with frontier proprietary systems. Remarkably, our paradigm exhibits unprecedented interaction scaling, extending to 2048 interactions with dramatic performance gains (from 3.5\% to 42.5\%), and serves as an effective prompting strategy, improving frontier models by up to 19.2pp over ReAct on long-horizon tasks. These findings position IterResearch as a versatile solution for long-horizon reasoning, effective both as a trained agent and as a prompting paradigm for frontier models.

cs.AI↗

arXiv2026

Adaptive Ability Decomposing for Unlocking Large Reasoning Model Effective Reinforcement Learning

Reinforcement learning with verifiable rewards (RLVR) has shown great potential to enhance the reasoning ability of large language models (LLMs). However, due to the limited amount of information provided during the RLVR process, the model can only engage in largely blind exploration, which often results in failure on challenging problems. To provide additional information for the RLVR process without relying on a teacher model, we propose A$^2$D, an Adaptive Ability Decomposing method for enhancing the effectiveness of RLVR. Specifically, we first train a decomposer via RLVR without distillation, enabling it to decompose complex questions into a set of simpler sub-questions. Next, we use this decomposer to annotate sub-questions for each question in the training dataset, and then train the reasoner under RLVR with sub-question guidance. To better understand A$^2$D, we first compare its performance with competitive baselines, showing its effectiveness. Next, we observe that our method functions as a plug-and-play module that can be applied to different RLVR algorithms. Furthermore, we conduct an analysis of the decomposer, revealing how the RLVR process affects its performance and behavior, and which type of guidance is better suited for enhancing the reasoner's exploration and exploitation abilities.

cs.CL↗

arXiv2026

RecNet: Self-Evolving Preference Propagation for Agentic Recommender Systems

Agentic recommender systems leverage Large Language Models (LLMs) to model complex user behaviors and support personalized decision-making. However, existing methods primarily model preference changes based on explicit user-item interactions, which are sparse, noisy, and unable to reflect the real-time, mutual influences among users and items. To address these limitations, we propose RecNet, a self-evolving preference propagation framework that proactively propagates real-time preference updates across related users and items. RecNet consists of two complementary phases. In the forward phase, the centralized preference routing mechanism leverages router agents to integrate preference updates and dynamically propagate them to the most relevant agents. To ensure accurate and personalized integration of propagated preferences, we further introduce a personalized preference reception mechanism, which combines a message buffer for temporary caching and an optimizable, rule-based filter memory to guide selective preference assimilation based on past experience and interests. In the backward phase, the feedback-driven propagation optimization mechanism simulates a multi-agent reinforcement learning framework, using LLMs for credit assignment, gradient analysis, and module-level optimization, enabling continuous self-evolution of propagation strategies. Extensive experiments on various scenarios demonstrate the effectiveness of RecNet in modeling preference propagation for recommender systems.

cs.AI↗

arXiv2026

GenCI: Generative Modeling of User Interest Shift via Cohort-based Intent Learning for CTR Prediction

Click-through rate (CTR) prediction plays a pivotal role in online advertising and recommender systems. Despite notable progress in modeling user preferences from historical behaviors, two key challenges persist. First, exsiting discriminative paradigms focus on matching candidates to user history, often overfitting to historically dominant features and failing to adapt to rapid interest shifts. Second, a critical information chasm emerges from the point-wise ranking paradigm. By scoring each candidate in isolation, CTR models discard the rich contextual signal implied by the recalled set as a whole, leading to a misalignment where long-term preferences often override the user's immediate, evolving intent. To address these issues, we propose GenCI, a generative user intent framework that leverages semantic interest cohorts to model dynamic user preferences for CTR prediction. The framework first employs a generative model, trained with a next-item prediction (NTP) objective, to proactively produce candidate interest cohorts. These cohorts serve as explicit, candidate-agnostic representations of a user's immediate intent. A hierarchical candidate-aware network then injects this rich contextual signal into the ranking stage, refining them with cross-attention to align with both user history and the target item. The entire model is trained end-to-end, creating a more aligned and effective CTR prediction pipeline. Extensive experiments on three widely used datasets demonstrate the effectiveness of our approach.

cs.IR↗

arXiv2026

MergeMix: Optimizing Mid-Training Data Mixtures via Learnable Model Merging

Optimizing data mixtures is essential for unlocking the full potential of large language models (LLMs), yet identifying the optimal composition remains computationally prohibitive due to reliance on heuristic trials or expensive proxy training. To address this, we introduce \textbf{MergeMix}, a novel approach that efficiently determines optimal data mixing ratios by repurposing model merging weights as a high-fidelity, low-cost performance proxy. By training domain-specific experts on minimal tokens and optimizing their merging weights against downstream benchmarks, MergeMix effectively optimizes the performance of data mixtures without incurring the cost of full-scale training. Extensive experiments on models with 8B and 16B parameters validate that MergeMix achieves performance comparable to or surpassing exhaustive manual tuning while drastically reducing search costs. Furthermore, MergeMix exhibits high rank consistency (Spearman $ρ> 0.9$) and strong cross-scale transferability, offering a scalable, automated solution for data mixture optimization.

cs.LG↗

arXiv2026

Beyond the Last Frame: Process-aware Evaluation for Generative Video Reasoning

Recent breakthroughs in video generation have demonstrated an emerging capability termed Chain-of-Frames (CoF) reasoning, where models resolve complex tasks through the generation of continuous frames. While these models show promise for Generative Video Reasoning (GVR), existing evaluation frameworks often rely on single-frame assessments, which can lead to outcome-hacking, where a model reaches a correct conclusion through an erroneous process. To address this, we propose a process-aware evaluation paradigm. We introduce VIPER, a comprehensive benchmark spanning 16 tasks across temporal, structural, symbolic, spatial, physics, and planning reasoning. Furthermore, we propose Process-outcome Consistency (POC@r), a new metric that utilizes VLM-as-Judge with a hierarchical rubric to evaluate both the validity of the intermediate steps and the final result. Our experiments reveal that state-of-the-art video models achieve POC@1.0 only about 20% and exhibit a significant outcome-hacking. We further explore the impact of test-time scaling and sampling robustness, highlighting a substantial gap between current video generation and true generalized visual reasoning. Our benchmark are released at https://github.com/RUCAIBox/VIPER.

cs.CV↗

arXiv2025

Entropy-Guided Token Dropout: Training Autoregressive Language Models with Limited Domain Data

As access to high-quality, domain-specific data grows increasingly scarce, multi-epoch training has become a practical strategy for adapting large language models (LLMs). However, autoregressive models often suffer from performance degradation under repeated data exposure, where overfitting leads to a marked decline in model capability. Through empirical analysis, we trace this degradation to an imbalance in learning dynamics: predictable, low-entropy tokens are learned quickly and come to dominate optimization, while the model's ability to generalize on high-entropy tokens deteriorates with continued training. To address this, we introduce EntroDrop, an entropy-guided token dropout method that functions as structured data regularization. EntroDrop selectively masks low-entropy tokens during training and employs a curriculum schedule to adjust regularization strength in alignment with training progress. Experiments across model scales from 0.6B to 8B parameters show that EntroDrop consistently outperforms standard regularization baselines and maintains robust performance throughout extended multi-epoch training. These findings underscore the importance of aligning regularization with token-level learning dynamics when training on limited data. Our approach offers a promising pathway toward more effective adaptation of LLMs in data-constrained domains.

cs.CL↗

arXiv2025

Rethinking Sample Polarity in Reinforcement Learning with Verifiable Rewards

Large reasoning models (LRMs) are typically trained using reinforcement learning with verifiable reward (RLVR) to enhance their reasoning abilities. In this paradigm, policies are updated using both positive and negative self-generated rollouts, which correspond to distinct sample polarities. In this paper, we provide a systematic investigation into how these sample polarities affect RLVR training dynamics and behaviors. We find that positive samples sharpen existing correct reasoning patterns, while negative samples encourage exploration of new reasoning paths. We further explore how adjusting the advantage values of positive and negative samples at both the sample level and the token level affects RLVR training. Based on these insights, we propose an Adaptive and Asymmetric token-level Advantage shaping method for Policy Optimization, namely A3PO, that more precisely allocates advantage signals to key tokens across different polarities. Experiments across five reasoning benchmarks demonstrate the effectiveness of our approach.

cs.CL↗

arXiv2025

Scaling Laws for Code: Every Programming Language Matters

Code large language models (Code LLMs) are powerful but costly to train, with scaling laws predicting performance from model size, data, and compute. However, different programming languages (PLs) have varying impacts during pre-training that significantly affect base model performance, leading to inaccurate performance prediction. Besides, existing works focus on language-agnostic settings, neglecting the inherently multilingual nature of modern software development. Therefore, it is first necessary to investigate the scaling laws of different PLs, and then consider their mutual influences to arrive at the final multilingual scaling law. In this paper, we present the first systematic exploration of scaling laws for multilingual code pre-training, conducting over 1000+ experiments (Equivalent to 336,000+ H800 hours) across multiple PLs, model sizes (0.2B to 14B parameters), and dataset sizes (1T tokens). We establish comprehensive scaling laws for code LLMs across multiple PLs, revealing that interpreted languages (e.g., Python) benefit more from increased model size and data than compiled languages (e.g., Rust). The study demonstrates that multilingual pre-training provides synergistic benefits, particularly between syntactically similar PLs. Further, the pre-training strategy of the parallel pairing (concatenating code snippets with their translations) significantly enhances cross-lingual abilities with favorable scaling properties. Finally, a proportion-dependent multilingual scaling law is proposed to optimally allocate training tokens by prioritizing high-utility PLs (e.g., Python), balancing high-synergy pairs (e.g., JavaScript-TypeScript), and reducing allocation to fast-saturating languages (Rust), achieving superior average performance across all PLs compared to uniform distribution under the same compute budget.

cs.CL↗

arXiv2025

ToolMind Technical Report: A Large-Scale, Reasoning-Enhanced Tool-Use Dataset

Large Language Model (LLM) agents have developed rapidly in recent years to solve complex real-world problems using external tools. However, the scarcity of high-quality trajectories still hinders the development of stronger LLM agents. Most existing works on multi-turn dialogue synthesis validate correctness only at the trajectory level, which may overlook turn-level errors that can propagate during training and degrade model performance. To address these limitations, we introduce ToolMind, a large-scale, high-quality tool-agentic dataset with 160k synthetic data instances generated using over 20k tools and 200k augmented open-source data instances. Our data synthesis pipeline first constructs a function graph based on parameter correlations and then uses a multi-agent framework to simulate realistic user-assistant-tool interactions. Beyond trajectory-level validation, we employ fine-grained turn-level filtering to remove erroneous or suboptimal steps, ensuring that only high-quality reasoning traces are retained. This approach mitigates error amplification during training while preserving self-corrective reasoning signals essential for robust tool-use learning. Models fine-tuned on ToolMind show significant improvements over baselines on several benchmarks.

cs.AI↗

arXiv2025

Revisiting the Necessity of Lengthy Chain-of-Thought in Vision-centric Reasoning Generalization

We study how different Chain-of-Thought (CoT) designs affect the acquisition of the generalizable visual reasoning ability in vision-language models (VLMs). While CoT data, especially long or visual CoT such as "think with image", has been widely used to supervise intermediate reasoning, it remains unclear why specific CoT designs help and which ones truly support generalizable reasoning. To systematically evaluate this, we focus on a controlled maze-solving benchmark where reasoning rules are fully visual, difficulty can be tuned by grid size, and all the intermediate steps can be automatically generated. Using Qwen2.5-VL-7B under a standard SFT-then-RL pipeline, we compare three representative CoT formats: Language CoT, Grounding CoT (with spatial coordinate trajectories), and Visual CoT (with image manipulations). Our experiments reveal that visual and longer CoT mainly accelerate convergence but do not lift the final performance ceiling; concise CoT containing only essential grounding steps outperforms longer traces; and, strikingly, CoT retaining only the minimal grounding results generalizes best across different maze sizes. We further validate these insights on other vision-centric tasks. These findings highlight a "short is long" effect and provide practical guidance for constructing more generalizable SFT datasets for visual reasoning.

cs.CV↗