Searcharxiv⌕ Search

arXiv subjects

Jinluan Yang

Publications and source records attributed to Jinluan Yang.

At least 19 recordsLinked to original sources
arXiv2026

Evolving Together, Failing Apart: The Skill-Policy Co-evolution Trap in Self-Improving Agents

Skill-policy co-evolution, in which an agent's policy model and its skill bank improve each other during training, is a promising paradigm for self-improving agents. Current evaluations score the co-evolved policy on the final bank it trained with (the delivered bank). After deployment, however, skills may be rewritten, added or removed more frequently than the policy model is retrained, leaving the policy to operate on a drifted skill bank that neither the training objective nor existing evaluation protocols address. In this paper, we first establish, through theoretical and empirical analysis, that the fragility hidden by this blind spot is not incidental. The acceptance rule these methods share (cooperative acceptance) ties a policy's competence to the bank it co-evolved with, leaving it fragile once skills drift, regardless of the implementation; empirically, sensitivity to skill drift grows as co-evolution training proceeds. We therefore introduce the CO-evolution Drift Audit (CODA), a deployment-oriented evaluation protocol that simulates skill drift and, alongside success rate under worst-case drift, introduces two new metrics: a consistency rate under meaning-preserving edits and a resistance rate to corrupted skills. Across four co-evolution methods and five benchmarks (six settings), worst-case success is 5-13 points below the peak measured for each delivered pair, erasing 52-93% of the delivered bank's gain over the no-skill baseline. CODA thus reframes the target of co-evolution, from peak performance on the co-evolved bank to the ability to harness skills however they evolve, and provides a shared standard against which future methods can be audited.

cs.AI↗
arXiv2026

TopoCurate:Modeling Interaction Topology for Tool-Use Agent Training

Training tool-use agents typically relies on outcome-based filtering: Supervised Fine-Tuning (SFT) on successful trajectories and Reinforcement Learning (RL) on pass-rate-selected tasks. However, this paradigm ignores interaction dynamics: successful trajectories may lack error recovery or exhibit redundancy, while pass rates fail to distinguish structurally informative tasks from trivial ones. We propose \textbf{TopoCurate}, an interaction-aware framework that projects multi-trial rollouts from the same task into a unified semantic quotient topology. By merging equivalent action-observation states, this projection transforms scattered linear trajectories into a structured manifold that explicitly captures how tool invocations and environmental responses drive the divergence between effective strategies and failure modes. Leveraging this representation, we introduce a dual-selection mechanism: for SFT, we prioritize trajectories demonstrating reflective recovery, semantic efficiency, and strategic diversity to mitigate covariate shift and mode collapse; for RL, we select tasks with high error branch ratios and strategic heterogeneity, maximizing gradient Signal-to-Noise Ratio to address vanishing signals in sparse-reward settings. Evaluations on BFCLv3 and Tau2 Bench show that TopoCurate achieves consistent gains of 4.2\% (SFT) and 6.9\% (RL) over state-of-the-art baselines. We will release the code and data soon for further investigations.

cs.LG↗
arXiv2026

SkillScaffold: Scaffolding Agentic Reinforcement Learning with Policy-Adaptive Skill Credit Assignment and Selective Injection

Agentic reinforcement learning (RL) enables large language models (LLMs) to learn from multi-turn interactions with verifiable outcomes. Textual skills can improve exploration by distilling prior interactions into reusable guidance. However, existing skill-based RL approaches often either assume skill injection is beneficial or rely on indirect feedback or rollout-intensive evaluation to assess its utility. Our pilot studies show that rollout-based utility estimation for skill credit assignment is costly and unstable under limited sampling, while skill benefits vary across candidates and depend on the current policy's competence on each query. To address these challenges, we introduce SkillScaffold, a policy-adaptive training scaffold that combines low-cost skill-utility estimation with selective skill injection. By reusing outcome-verified training trajectories, SkillScaffold generates candidate skills and estimates their utility through outcome-aware likelihood comparisons, without candidate-specific environment rollouts. The resulting utility guides whether and which skill to inject, retaining skill-free execution when no candidate is estimated to help. Joint training enables the Generator to produce skills tailored to the evolving Executor. Scaffolded RL encourages skill internalization for skill-free inference. Experiments on ALFWorld, WebShop, ScienceWorld, and $τ^2$-Bench demonstrate gains over strong agentic-RL and skill-based RL baselines across both short- and long-horizon tasks. Ablations and further analyses support the importance of utility-based selection and abstention. Our code is available at https://github.com/Vaidurya00/SkillScaffold.

cs.LG↗
arXiv2026

SOAR: Source-Aware Reinforcement Learning with Precise Credit Assignment for Search Agents

Reinforcement learning (RL) for search agents often suffers from evidence fabrication, a failure mode where agents bypass external information ($E_{ext}$) and fabricate an incorrect answer. We argue that this fabrication arises from an over-reliance on unreliable internal knowledge ($E_{int}$) as shortcuts due to the inherent opacity of reasoning trajectories. To address this, we propose SOAR (SOurce-Aware Reinforcement learning), a paradigm that transforms monolithic reasoning into source-explicit trajectories. SOAR (i) disentangles $E_{ext}$ and $E_{int}$ into parallel, observable streams via a Source-Aware Workflow, and (ii) incorporates Source-Consistent Reward to guide the policy evolution. By making information-source attribution explicit, SOAR enables precise credit assignment between different knowledge streams. Evaluations show that SOAR outperforms baselines on 3B and 7B backbones by 7.3% and 5.4%, while reducing evidence fabrication by 74.8%. These confirm that SOAR significantly improves performance on knowledge-intensive benchmarks while effectively mitigating evidence fabrication, fostering more faithful and grounded reasoning with enhanced robustness.

cs.CL↗
arXiv2026

Towards Advanced Mathematical Reasoning for LLMs via First-Order Logic Theorem Proving

Large language models (LLMs) have shown promising first-order logic (FOL) reasoning capabilities with applications in various areas. However, their effectiveness in complex mathematical reasoning involving multi-step FOL deductions is still under-researched. While LLMs perform competitively on established mathematical reasoning benchmarks, they struggle with multi-step FOL tasks, as demonstrated by Deepseek-Prover-V2-7B's low accuracy (4.2%) on our proposed theorem proving dataset. This issue arises from the limited exploration of diverse proof strategies and the potential for early reasoning mistakes to undermine entire proofs. To address these issues, we propose DREAM, a self-adaptive solution that enhances the Diversity and REAsonability of LLMs' generation strategies. DREAM incorporates an Axiom-Driven Strategy Diversification mechanism to promote varied strategic outcomes and a Sub-Proposition Error Feedback to help LLMs reflect on and correct their proofs. Our contributions include pioneering advancements in LLMs' mathematical reasoning through FOL theorem proving, introducing a novel inference stage solution that improves performance by 0.6% to 6.4%, and providing a curated dataset of 447 mathematical theorems in Lean 4 format for evaluation.

cs.AI↗
arXiv2026

Pushing the Boundaries of Natural Reasoning: Interleaved Bonus from Formal-Logic Verification

Large Language Models (LLMs) show remarkable capabilities, yet their stochastic next-token prediction creates logical inconsistencies and reward hacking that formal symbolic systems avoid. To bridge this gap, we introduce a formal logic verification-guided framework that dynamically interleaves formal symbolic verification with the natural language generation process, providing real-time feedback to detect and rectify errors as they occur. Distinguished from previous neuro-symbolic methods limited by passive post-hoc validation, our approach actively penalizes intermediate fallacies during the reasoning chain. We operationalize this framework via a novel two-stage training pipeline that synergizes formal logic verification-guided supervised fine-tuning and policy optimization. Extensive evaluation on six benchmarks spanning mathematical, logical, and general reasoning demonstrates that our 7B and 14B models outperform state-of-the-art baselines by average margins of 10.4% and 14.2%, respectively. These results validate that formal verification can serve as a scalable mechanism to significantly push the performance boundaries of advanced LLM reasoning.

cs.LG↗
arXiv2026

ExpertWeaver: Unlocking the Inherent MoE in Dense LLMs with GLU Activation Patterns

Mixture-of-Experts (MoE) effectively scales model capacity while preserving computational efficiency through sparse expert activation. However, training high-quality MoEs from scratch is prohibitively expensive. A promising alternative is to convert pretrained dense models into sparse MoEs. Existing dense-to-MoE methods fall into two categories: \textbf{dynamic structural pruning} that converts dense models into MoE architectures with moderate sparsity to balance performance and inference efficiency, and \textbf{downcycling} approaches that use pretrained dense models to initialize highly sparse MoE architectures. However, existing methods break the intrinsic activation patterns within dense models, leading to suboptimal expert construction. In this work, we argue that the Gated Linear Unit (GLU) mechanism provides a natural blueprint for dense-to-MoE conversion. We show that the fine-grained neural-wise activation patterns of GLU reveal a coarse-grained structure, uncovering an inherent MoE architecture composed of consistently activated universal neurons and dynamically activated specialized neurons. Leveraging this discovery, we introduce ExpertWeaver, a training-free framework that partitions neurons according to their activation patterns and constructs shared experts and specialized routed experts with layer-adaptive configurations. Our experiments demonstrate that ExpertWeaver significantly outperforms existing methods, both as a training-free dynamic structural pruning technique and as a downcycling strategy for superior MoE initialization.

cs.CL↗
arXiv2026

SIGHT: Reinforcement Learning with Self-Evidence and Information-Gain Diverse Branching for Search Agent

Reinforcement Learning (RL) has empowered Large Language Models (LLMs) to master autonomous search for complex question answering. However, particularly within multi-turn search scenarios, this interaction introduces a critical challenge: search results often suffer from high redundancy and low signal-to-noise ratios. Consequently, agents easily fall into "Tunnel Vision," where the forced interpretation of early noisy retrievals leads to irreversible error accumulation. To address these challenges, we propose SIGHT, a framework that enhances search-based reasoning through Self-Evidence Support (SES) and Information-Gain Driven Diverse Branching. SIGHT distills search results into high-fidelity evidence via SES and calculates an Information Gain score to pinpoint pivotal states where observations maximally reduce uncertainty. This score guides Dynamic Prompting Interventions - including de-duplication, reflection, or adaptive branching - to spawn new branches with SES. Finally, by integrating SES and correctness rewards via Group Relative Policy Optimization, SIGHT internalizes robust exploration strategies without external verifiers. Experiments on single-hop and multi-hop QA benchmarks demonstrate that SIGHT significantly outperforms existing approaches, particularly in complex reasoning scenarios, using fewer search steps.

cs.CL↗
arXiv2026

LongCat-Flash-Thinking-2601 Technical Report

We introduce LongCat-Flash-Thinking-2601, a 560-billion-parameter open-source Mixture-of-Experts (MoE) reasoning model with superior agentic reasoning capability. LongCat-Flash-Thinking-2601 achieves state-of-the-art performance among open-source models on a wide range of agentic benchmarks, including agentic search, agentic tool use, and tool-integrated reasoning. Beyond benchmark performance, the model demonstrates strong generalization to complex tool interactions and robust behavior under noisy real-world environments. Its advanced capability stems from a unified training framework that combines domain-parallel expert training with subsequent fusion, together with an end-to-end co-design of data construction, environments, algorithms, and infrastructure spanning from pre-training to post-training. In particular, the model's strong generalization capability in complex tool-use are driven by our in-depth exploration of environment scaling and principled task construction. To optimize long-tailed, skewed generation and multi-turn agentic interactions, and to enable stable training across over 10,000 environments spanning more than 20 domains, we systematically extend our asynchronous reinforcement learning framework, DORA, for stable and efficient large-scale multi-environment training. Furthermore, recognizing that real-world tasks are inherently noisy, we conduct a systematic analysis and decomposition of real-world noise patterns, and design targeted training procedures to explicitly incorporate such imperfections into the training process, resulting in improved robustness for real-world applications. To further enhance performance on complex reasoning tasks, we introduce a Heavy Thinking mode that enables effective test-time scaling by jointly expanding reasoning depth and width through intensive parallel thinking.

cs.AI↗
arXiv2026

Mix Data or Merge Models? Balancing the Helpfulness, Honesty, and Harmlessness of Large Language Model via Model Merging

Achieving balanced alignment of large language models (LLMs) in terms of Helpfulness, Honesty, and Harmlessness (3H optimization) constitutes a cornerstone of responsible AI. Existing methods like data mixture strategies face limitations, including heavy reliance on expert knowledge and conflicting optimization signals. While model merging offers parameter-level conflict-resolution strategies through integrating specialized models' parameters, its potential for 3H optimization remains underexplored. This paper systematically compares the effectiveness of model merging and data mixture methods in constructing 3H-aligned LLMs for the first time, revealing previously overlooked collaborative and conflict relationships among the 3H dimensions and discussing the advantages and drawbacks of data mixture (\textit{data-level}) and model merging (\textit{parameter-level}) methods in mitigating the conflict for balanced 3H optimization. Specially, we propose a novel \textbf{R}eweighting \textbf{E}nhanced task \textbf{S}ingular \textbf{M}erging method, \textbf{RESM}, through outlier weighting and sparsity-aware rank selection strategies to address the challenges of preference noise accumulation and layer sparsity adaptation inherent in 3H-aligned LLM merging. Extensive evaluations can verify the effectiveness and robustness of RESM compared to previous data mixture (2\%-5\% gain) and model merging (1\%-3\% gain) methods in achieving balanced LLM alignment. We release our models through \href{https://huggingface.co/Jinluan}{3H\_Merging} for further investigations.

cs.CL↗
arXiv2026

LRAS: Advanced Legal Reasoning with Agentic Search

While Large Reasoning Models (LRMs) have demonstrated exceptional logical capabilities in mathematical domains, their application to the legal field remains hindered by the strict requirements for procedural rigor and adherence to legal logic. Existing legal LLMs, which rely on "closed-loop reasoning" derived solely from internal parametric knowledge, frequently suffer from lack of self-awareness regarding their knowledge boundaries, leading to confident yet incorrect conclusions. To address this challenge, we present Legal Reasoning with Agentic Search (LRAS), the first framework designed to transition legal LLMs from static and parametric "closed-loop thinking" to dynamic and interactive "Active Inquiry". By integrating Introspective Imitation Learning and Difficulty-aware Reinforcement Learning, LRAS enables LRMs to identify knowledge boundaries and handle legal reasoning complexity. Empirical results demonstrate that LRAS outperforms state-of-the-art baselines by 8.2-32\%, with the most substantial gains observed in tasks requiring deep reasoning with reliable knowledge. We will release our data and models for further exploration soon.

cs.AI↗
arXiv2026

Reinforcement Learning for Tool-Integrated Interleaved Thinking towards Cross-Domain Generalization

Recent advances in large language models (LLMs) have demonstrated remarkable capabilities in reasoning and tool utilization. However, the generalization of tool-augmented reinforcement learning (RL) across diverse domains remains a significant challenge. Standard paradigms often treat tool usage as a linear or isolated event, which becomes brittle when transferring skills from restricted domains (e.g., mathematics) to open-ended tasks. In this work, we investigate the cross-domain generalization of an LLM agent trained exclusively on mathematical problem-solving. To facilitate robust skill transfer, we propose a {\textbf{R}einforcement Learning for \textbf{I}nterleaved \textbf{T}ool \textbf{E}xecution (RITE)}. Unlike traditional methods, RITE enforces a continuous ``Plan-Action-Reflection'' cycle, allowing the model to ground its reasoning in intermediate tool outputs and self-correct during long-horizon tasks. To effectively train this complex interleaved policy, we introduce {Dr. GRPO}, a robust optimization objective that utilizes token-level loss aggregation with importance sampling to mitigate reward sparsity and high-variance credit assignment. Furthermore, we employ a dual-component reward system and dynamic curriculum via online rollout filtering to ensure structural integrity and sample efficiency. Extensive experiments reveal that our approach, despite being trained solely on math tasks, achieves state-of-the-art performance across diverse reasoning domains, demonstrating high token efficiency and strong generalization capabilities.

cs.LG↗
arXiv2025

Scaling and Transferability of Annealing Strategies in Large Language Model Training

Learning rate scheduling is crucial for training large language models, yet understanding the optimal annealing strategies across different model configurations remains challenging. In this work, we investigate the transferability of annealing dynamics in large language model training and refine a generalized predictive framework for optimizing annealing strategies under the Warmup-Steady-Decay (WSD) scheduler. Our improved framework incorporates training steps, maximum learning rate, and annealing behavior, enabling more efficient optimization of learning rate schedules. Our work provides a practical guidance for selecting optimal annealing strategies without exhaustive hyperparameter searches, demonstrating that smaller models can serve as reliable proxies for optimizing the training dynamics of larger models. We validate our findings on extensive experiments using both Dense and Mixture-of-Experts (MoE) models, demonstrating that optimal annealing ratios follow consistent patterns and can be transferred across different training configurations.

cs.LG↗
arXiv2025

Noise Projection: Closing the Prompt-Agnostic Gap Behind Text-to-Image Misalignment in Diffusion Models

In text-to-image generation, different initial noises induce distinct denoising paths with a pretrained Stable Diffusion (SD) model. While this pattern could output diverse images, some of them may fail to align well with the prompt. Existing methods alleviate this issue either by altering the denoising dynamics or by drawing multiple noises and conducting post-selection. In this paper, we attribute the misalignment to a training-inference mismatch: during training, prompt-conditioned noises lie in a prompt-specific subset of the latent space, whereas at inference the noise is drawn from a prompt-agnostic Gaussian prior. To close this gap, we propose a noise projector that applies text-conditioned refinement to the initial noise before denoising. Conditioned on the prompt embedding, it maps the noise to a prompt-aware counterpart that better matches the distribution observed during SD training, without modifying the SD model. Our framework consists of these steps: we first sample some noises and obtain token-level feedback for their corresponding images from a vision-language model (VLM), then distill these signals into a reward model, and finally optimize the noise projector via a quasi-direct preference optimization. Our design has two benefits: (i) it requires no reference images or handcrafted priors, and (ii) it incurs small inference cost, replacing multi-sample selection with a single forward pass. Extensive experiments further show that our prompt-aware noise projection improves text-image alignment across diverse prompts.

cs.CV↗
arXiv2025

Unifying Adversarial Perturbation for Graph Neural Networks

This paper studies the vulnerability of Graph Neural Networks (GNNs) to adversarial attacks on node features and graph structure. Various methods have implemented adversarial training to augment graph data, aiming to bolster the robustness and generalization of GNNs. These methods typically involve applying perturbations to the node feature, weights, or graph structure and subsequently minimizing the loss by learning more robust graph model parameters under the adversarial perturbations. Despite the effectiveness of adversarial training in enhancing GNNs' robustness and generalization abilities, its application has been largely confined to specific datasets and GNN types. In this paper, we propose a novel method, PerturbEmbedding, that integrates adversarial perturbation and training, enhancing GNNs' resilience to such attacks and improving their generalization ability. PerturbEmbedding performs perturbation operations directly on every hidden embedding of GNNs and provides a unified framework for most existing perturbation strategies/methods. We also offer a unified perspective on the forms of perturbations, namely random and adversarial perturbations. Through experiments on various datasets using different backbone models, we demonstrate that PerturbEmbedding significantly improves both the robustness and generalization abilities of GNNs, outperforming existing methods. The rejection of both random (non-targeted) and adversarial (targeted) perturbations further enhances the backbone model's performance.

cs.LG↗
arXiv2025

Discovering Invariant Neighborhood Patterns for Heterophilic Graphs

This paper studies the problem of distribution shifts on non-homophilous graphs Mosting existing graph neural network methods rely on the homophilous assumption that nodes from the same class are more likely to be linked. However, such assumptions of homophily do not always hold in real-world graphs, which leads to more complex distribution shifts unaccounted for in previous methods. The distribution shifts of neighborhood patterns are much more diverse on non-homophilous graphs. We propose a novel Invariant Neighborhood Pattern Learning (INPL) to alleviate the distribution shifts problem on non-homophilous graphs. Specifically, we propose the Adaptive Neighborhood Propagation (ANP) module to capture the adaptive neighborhood information, which could alleviate the neighborhood pattern distribution shifts problem on non-homophilous graphs. We propose Invariant Non-Homophilous Graph Learning (INHGL) module to constrain the ANP and learn invariant graph representation on non-homophilous graphs. Extensive experimental results on real-world non-homophilous graphs show that INPL could achieve state-of-the-art performance for learning on large non-homophilous graphs.

cs.LG↗
arXiv2025

Each Rank Could be an Expert: Single-Ranked Mixture of Experts LoRA for Multi-Task Learning

Low-Rank Adaptation (LoRA) is widely used for adapting large language models (LLMs) to specific domains due to its efficiency and modularity. Meanwhile, vanilla LoRA struggles with task conflicts in multi-task scenarios. Recent works adopt Mixture of Experts (MoE) by treating each LoRA module as an expert, thereby mitigating task interference through multiple specialized LoRA modules. While effective, these methods often isolate knowledge within individual tasks, failing to fully exploit the shared knowledge across related tasks. In this paper, we establish a connection between single LoRA and multi-LoRA MoE, integrating them into a unified framework. We demonstrate that the dynamic routing of multiple LoRAs is functionally equivalent to rank partitioning and block-level activation within a single LoRA. We further empirically demonstrate that finer-grained LoRA partitioning, within the same total and activated parameter constraints, leads to better performance gains across heterogeneous tasks. Building on these findings, we propose Single-ranked Mixture of Experts LoRA (\textbf{SMoRA}), which embeds MoE into LoRA by \textit{treating each rank as an independent expert}. With a \textit{dynamic rank-wise activation} mechanism, SMoRA promotes finer-grained knowledge sharing while mitigating task conflicts. Experiments demonstrate that SMoRA activates fewer parameters yet achieves better performance in multi-task scenarios.

cs.LG↗
arXiv2025

Ada-R1: Hybrid-CoT via Bi-Level Adaptive Reasoning Optimization

Recently, long-thought reasoning models achieve strong performance on complex reasoning tasks, but often incur substantial inference overhead, making efficiency a critical concern. Our empirical analysis reveals that the benefit of using Long-CoT varies across problems: while some problems require elaborate reasoning, others show no improvement, or even degraded accuracy. This motivates adaptive reasoning strategies that tailor reasoning depth to the input. However, prior work primarily reduces redundancy within long reasoning paths, limiting exploration of more efficient strategies beyond the Long-CoT paradigm. To address this, we propose a novel two-stage framework for adaptive and efficient reasoning. First, we construct a hybrid reasoning model by merging long and short CoT models to enable diverse reasoning styles. Second, we apply bi-level preference training to guide the model to select suitable reasoning styles (group-level), and prefer concise and correct reasoning within each style group (instance-level). Experiments demonstrate that our method (Ada-R1) significantly reduces inference costs compared to other baseline approaches, while maintaining performance. Notably, on five mathematical datasets, the average length of reasoning is reduced by more than 50%, highlighting the potential of adaptive strategies to optimize reasoning efficiency in large language models. Our code is coming soon at https://github.com/StarDewXXX/AdaR1

cs.AI↗