Can RL Improve Generalization of LLM Agents? An Empirical Study

Reinforcement fine-tuning (RFT) has shown promise for training LLM agents to perform multi-turn decision-making based on environment feedback. However, most existing evaluations remain largely in-domain: training and testing are conducted in the same environment or even on the same tasks. In real-world deployment, agents may operate in unseen environments with different background knowledge, observation spaces, and action interfaces. To characterize the generalization profile of RFT under such shifts, we conduct a systematic study along three axes: (1) within-environment generalization across task difficulty, (2) cross-environment transfer to unseen environments, and (3) sequential multi-environment training to quantify transfer and forgetting. Our results show that RFT generalizes well across task difficulty within an environment, but exhibits weaker transfer to unseen environments, which correlates with shifts in both semantic priors and observation/action interfaces. In contrast, sequential training yields promising downstream gains with minimal upstream forgetting, and mixture training across environments improves the overall balance. We further provide detailed analyses and deeper insights, and hope our work helps the community develop and deploy generalizable LLM agents.

UniG2U-Bench: Do Unified Models Advance Multimodal Understanding?

Unified multimodal models have recently demonstrated strong generative capabilities, yet whether and when generation improves understanding remains unclear. Existing benchmarks lack a systematic exploration of the specific tasks where generation facilitates understanding. To this end, we introduce UniG2U-Bench, a comprehensive benchmark categorizing generation-to-understanding (G2U) evaluation into 7 regimes and 30 subtasks, requiring varying degrees of implicit or explicit visual transformations. Extensive evaluation of over 30 models reveals three core findings: 1) Unified models generally underperform their base Vision-Language Models (VLMs), and Generate-then-Answer (GtA) inference typically degrades performance relative to direct inference. 2) Consistent enhancements emerge in spatial intelligence, visual illusions, or multi-round reasoning subtasks, where enhanced spatial and shape perception, as well as multi-step intermediate image states, prove beneficial. 3) Tasks with similar reasoning structures and models sharing architectures exhibit correlated behaviors, suggesting that generation-understanding coupling induces class-consistent inductive biases over tasks, pretraining data, and model architectures. These findings highlight the necessity for more diverse training data and novel paradigms to fully unlock the potential of unified multimodal modeling.

SAE as a Crystal Ball: Interpretable Features Predict Cross-domain Transferability of LLMs without Training

In recent years, pre-trained large language models have achieved remarkable success across diverse tasks. Besides the pivotal role of self-supervised pre-training, their effectiveness in downstream applications also depends critically on the post-training process, which adapts models to task-specific data and objectives. However, this process inevitably introduces model shifts that can influence performance in different domains, and how such shifts transfer remains poorly understood. To open up the black box, we propose the SAE-based Transferability Score (STS), a new metric that leverages sparse autoencoders (SAEs) to forecast post-training transferability. Taking supervised fine-tuning as an example, STS identifies shifted dimensions in SAE representations and calculates their correlations with downstream domains, enabling reliable estimation of transferability \textit{before} fine-tuning. Extensive experiments across multiple models and domains show that STS accurately predicts the transferability of supervised fine-tuning, achieving Pearson correlation coefficients above 0.7 with actual performance changes. Beyond this, we take an initial step toward extending STS to reinforcement learning. We believe that STS can serve as an {\color{black} interpretable} tool for guiding post-training strategies in LLMs. Code is available at https://github.com/PKU-ML/STS.

Controllable Generative Sandbox for Causal Inference

Method validation and study design in causal inference rely on synthetic data with known counterfactuals. Existing simulators trade off distributional realism, the ability to capture mixed-type and multimodal tabular data, against causal controllability, including explicit control over overlap, unmeasured confounding, and treatment effect heterogeneity. We introduce CausalMix, a variational generative framework that closes this gap by coupling a mixture of Gaussian latent priors with data-type-specific decoders for continuous, binary, and categorical variables. The model incorporates explicit causal controls: an overlap regularizer shaping propensity-score distributions, alongside direct parameterizations of confounding strength and effect heterogeneity. This unified objective preserves fidelity to the observed data while enabling factorial manipulation of causal mechanisms, allowing overlap, confounding strength, and treatment effect heterogeneity to be varied independently at design time. Across benchmarks, CausalMix achieves state-of-the-art distributional metrics on mixed-type tables while providing stable, fine-grained causal control. We demonstrate practical utility in a comparative safety study of metastatic castration-resistant prostate cancer treatments, using CausalMix to compare estimators under calibrated data-generating processes, tune hyperparameters, and conduct simulation-based power analyses under targeted treatment effect heterogeneity scenarios.

Turning Semantics into Topology: LLM-Driven Attribute Augmentation for Collaborative Filtering

Large Language Models (LLMs) have shown great potential for enhancing recommender systems through their extensive world knowledge and reasoning capabilities. However, effectively translating these semantic signals into traditional collaborative embeddings remains an open challenge. Existing approaches typically fall into two extremes: direct inference methods are computationally prohibitive for large-scale retrieval, while embedding-based methods primarily focus on unilateral feature augmentation rather than holistic collaborative signal enhancement. To bridge this gap, we propose Topology-Augmented Graph Collaborative Filtering (TAGCF), a novel framework that transforms semantic knowledge into topological connectivity. Unlike existing approaches that depend on textual features or direct interaction synthesis, TAGCF employs LLMs to infer interaction intents and underlying causal relationships from user-item pairs, representing these insights as intermediate attribute nodes within an enriched User-Attribute-Item (U-A-I) graph. Furthermore, to effectively model the heterogeneous relations in this augmented structure, we propose Adaptive Relation-weighted Graph Convolution (ARGC), which employs relation-specific prediction networks to dynamically estimate the importance of each relation type. Extensive experiments across multiple benchmark datasets and CF backbones demonstrate consistent improvements, with comprehensive evaluations including cold-start scenarios validating the effectiveness and robustness of our framework. All code will be made publicly available. For anonymous review, our code is available at the following anonymous link: https://anonymous.4open.science/r/AGCF-2441353190/.

MagicAgent: Towards Generalized Agent Planning

The evolution of Large Language Models (LLMs) from passive text processors to autonomous agents has established planning as a core component of modern intelligence. However, achieving generalized planning remains elusive, not only by the scarcity of high-quality interaction data but also by inherent conflicts across heterogeneous planning tasks. These challenges result in models that excel at isolated tasks yet struggle to generalize, while existing multi-task training attempts suffer from gradient interference. In this paper, we present \textbf{MagicAgent}, a series of foundation models specifically designed for generalized agent planning. We introduce a lightweight and scalable synthetic data framework that generates high-quality trajectories across diverse planning tasks, including hierarchical task decomposition, tool-augmented planning, multi-constraint scheduling, procedural logic orchestration, and long-horizon tool execution. To mitigate training conflicts, we propose a two-stage training paradigm comprising supervised fine-tuning followed by multi-objective reinforcement learning over both static datasets and dynamic environments. Empirical results demonstrate that MagicAgent-32B and MagicAgent-30B-A3B deliver superior performance, achieving accuracies of $75.1\%$ on Worfbench, $55.9\%$ on NaturalPlan, $57.5\%$ on $τ^2$-Bench, $86.9\%$ on BFCL-v3, and $81.2\%$ on ACEBench, as well as strong results on our in-house MagicEval benchmarks. These results substantially outperform existing sub-100B models and even surpass leading closed-source models.

Mnemis: Dual-Route Retrieval on Hierarchical Graphs for Long-Term LLM Memory

AI Memory, specifically how models organizes and retrieves historical messages, becomes increasingly valuable to Large Language Models (LLMs), yet existing methods (RAG and Graph-RAG) primarily retrieve memory through similarity-based mechanisms. While efficient, such System-1-style retrieval struggles with scenarios that require global reasoning or comprehensive coverage of all relevant information. In this work, We propose Mnemis, a novel memory framework that integrates System-1 similarity search with a complementary System-2 mechanism, termed Global Selection. Mnemis organizes memory into a base graph for similarity retrieval and a hierarchical graph that enables top-down, deliberate traversal over semantic hierarchies. By combining the complementary strength from both retrieval routes, Mnemis retrieves memory items that are both semantically and structurally relevant. Mnemis achieves state-of-the-art performance across all compared methods on long-term memory benchmarks, scoring 93.9 on LoCoMo and 91.6 on LongMemEval-S using GPT-4.1-mini.

X-MAP: eXplainable Misclassification Analysis and Profiling for Spam and Phishing Detection

Misclassifications in spam and phishing detection are very harmful, as false negatives expose users to attacks while false positives degrade trust. Existing uncertainty-based detectors can flag potential errors, but possibly be deceived and offer limited interpretability. This paper presents X-MAP, an eXplainable Misclassification Analysis and Profilling framework that reveals topic-level semantic patterns behind model failures. X-MAP combines SHAP-based feature attributions with non-negative matrix factorization to build interpretable topic profiles for reliably classified spam/phishing and legitimate messages, and measures each message's deviation from these profiles using Jensen-Shannon divergence. Experiments on SMS and phishing datasets show that misclassified messages exhibit at least two times larger divergence than correctly classified ones. As a detector, X-MAP achieves up to 0.98 AUROC and lowers the false-rejection rate at 95% TRR to 0.089 on positive predictions. When used as a repair layer on base detectors, it recovers up to 97% of falsely rejected correct predictions with moderate leakage. These results demonstrate X-MAP's effectiveness and interpretability for improving spam and phishing detection.

A consequence of failed sequential learning: A computational account of developmental amnesia

Developmental amnesia, featured with severely impaired episodic memory and almost normal semantic memory, has been discovered to occur in children with hippocampal atrophy. This unique combination of characteristics seems to challenge the understanding that early loss of episodic memory may impede cognitive development and result in severe mental retardation. Although a few underlying mechanisms have been suggested, no computational model has been reported that is able to mimic the unique combination of characteristics. In this study, a cognitive system is presented, and developmental amnesia is demonstrated computationally in terms of impaired episodic recall, spared recognition and spared semantic learning. Impaired sequential/spatial learning ability of the hippocampus is suggested to be the cause of such amnesia. Simulation shows that impaired sequential leaning may only result in severe impairment of episodic recall, but affect neither recognition ability nor semantic learning. The spared semantic learning is inline with the view that semantic learning is largely associated with the consolidation of episodic memory, a process in which episodic memory may be mostly activated randomly, instead of sequentially. Furthermore, retrograded amnesia is also simulated, and the result and its mechanism are in agreement with most computational models of amnesia reported previously.

SciAgentGym: Benchmarking Multi-Step Scientific Tool-use in LLM Agents

Scientific reasoning inherently demands integrating sophisticated toolkits to navigate domain-specific knowledge. Yet, current benchmarks largely overlook agents' ability to orchestrate tools for such rigorous workflows. To bridge this gap, we introduce SciAgentGym, a scalable interactive environment featuring 1,780 domain-specific tools across four natural science disciplines, supported by a robust execution infrastructure. Complementing this, we present SciAgentBench, a tiered evaluation suite designed to stress-test agentic capabilities from elementary actions to long-horizon workflows. Our evaluation identifies a critical bottleneck: state-of-the-art models struggle with complex scientific tool-use. Even for a leading model like GPT-5, success rates drop sharply from 60.6% to 30.9% as interaction horizons extend, primarily due to failures in multi-step workflow execution. To address this, we propose SciForge, a data synthesis method that models the tool action space as a dependency graph to generate logic-aware training trajectories. By fine-tuning on these trajectories, our SciAgent-8B outperforms the significantly larger Qwen3-VL-235B-Instruct while exhibiting positive cross-domain transfer of scientific tool-use capabilities. These results underscore the promising potential of next-generation autonomous scientific agents.