❌

Normal view

Received — 17 March 2026 ⏭

Efficient and Interpretable Multi-Agent LLM Routing via Ant Colony Optimization

✇cs.AI, q-bio.NC updates on arXiv.org
By: Xudong Wang · Chaoning Zhang · Jiaquan Zhang · Chenghao Li · Qigan Sun · Sung-Ho Bae · Peng Wang · Ning Xie · Jie Zou · Yang Yang · Hengtao Shen
16 March 2026 at 12:00
arXiv:2603.12933v1 Announce Type: new Abstract: Large Language Model (LLM)-driven Multi-Agent Systems (MAS) have demonstrated strong capability in complex reasoning and tool use, and heterogeneous agent pools further broaden the quality--cost trade-off space. Despite these advances, real-world deployment is often constrained by high inference cost, latency, and limited transparency, which hinders scalable and efficient routing. Existing routing strategies typically rely on expensive LLM-based selectors or static policies, and offer limited controllability for semantic-aware routing under dynamic loads and mixed intents, often resulting in unstable performance and inefficient resource utilization. To address these limitations, we propose AMRO-S, an efficient and interpretable routing framework for Multi-Agent Systems (MAS). AMRO-S models MAS routing as a semantic-conditioned path selection problem, enhancing routing performance through three key mechanisms: First, it leverages a supervised fine-tuned (SFT) small language model for intent inference, providing a low-overhead semantic interface for each query; second, it decomposes routing memory into task-specific pheromone specialists, reducing cross-task interference and optimizing path selection under mixed workloads; finally, it employs a quality-gated asynchronous update mechanism to decouple inference from learning, optimizing routing without increasing latency. Extensive experiments on five public benchmarks and high-concurrency stress tests demonstrate that AMRO-S consistently improves the quality--cost trade-off over strong routing baselines, while providing traceable routing evidence through structured pheromone patterns.

Test-Time Strategies for More Efficient and Accurate Agentic RAG

✇cs.AI, q-bio.NC updates on arXiv.org
By: Brian Zhang · Deepti Guntur · Zhiyang Zuo · Abhinav Sharma · Shreyas Chaudhari · Wenlong Zhao · Franck Dernoncourt · Puneet Mathur · Ryan Rossi · Nedim Lipka
16 March 2026 at 12:00
arXiv:2603.12396v1 Announce Type: cross Abstract: Retrieval-Augmented Generation (RAG) systems face challenges with complex, multihop questions, and agentic frameworks such as Search-R1 (Jin et al., 2025), which operates iteratively, have been proposed to address these complexities. However, such approaches can introduce inefficiencies, including repetitive retrieval of previously processed information and challenges in contextualizing retrieved results effectively within the current generation prompt. Such issues can lead to unnecessary retrieval turns, suboptimal reasoning, inaccurate answers, and increased token consumption. In this paper, we investigate test-time modifications to the Search-R1 pipeline to mitigate these identified shortcomings. Specifically, we explore the integration of two components and their combination: a contextualization module to better integrate relevant information from retrieved documents into reasoning, and a de-duplication module that replaces previously retrieved documents with the next most relevant ones. We evaluate our approaches using the HotpotQA (Yang et al., 2018) and the Natural Questions (Kwiatkowski et al., 2019) datasets, reporting the exact match (EM) score, an LLM-as-a-Judge assessment of answer correctness, and the average number of turns. Our best-performing variant, utilizing GPT-4.1-mini for contextualization, achieves a 5.6% increase in EM score and reduces the number of turns by 10.5% compared to the Search-R1 baseline, demonstrating improved answer accuracy and retrieval efficiency.

CognitionCapturerPro: Towards High-Fidelity Visual Decoding from EEG/MEG via Multi-modal Information and Asymmetric Alignment

✇cs.AI, q-bio.NC updates on arXiv.org
By: Kaifan Zhang · Lihuo He · Junjie Ke · Yuqi Ji · Lukun Wu · Lizi Wang · Xinbo Gao
16 March 2026 at 12:00
arXiv:2603.12722v1 Announce Type: cross Abstract: Visual stimuli reconstruction from EEG remains challenging due to fidelity loss and representation shift. We propose CognitionCapturerPro, an enhanced framework that integrates EEG with multi-modal priors (images, text, depth, and edges) via collaborative training. Our core contributions include an uncertainty-weighted similarity scoring mechanism to quantify modality-specific fidelity and a fusion encoder for integrating shared representations. By employing a simplified alignment module and a pre-trained diffusion model, our method significantly outperforms the original CognitionCapturer on the THINGS-EEG dataset, improving Top-1 and Top-5 retrieval accuracy by 25.9% and 10.6%, respectively. Code is available at: https://github.com/XiaoZhangYES/CognitionCapturerPro.

Cheers: Decoupling Patch Details from Semantic Representations Enables Unified Multimodal Comprehension and Generation

✇cs.AI, q-bio.NC updates on arXiv.org
By: Yichen Zhang · Da Peng · Zonghao Guo · Zijian Zhang · Xuesong Yang · Tong Sun · Shichu Sun · Yidan Zhang · Yanghao Li · Haiyan Zhao · Wang Xu · Qi Shi · Yangang Sun · Chi Chen · Shuo Wang · Yukun Yan · Xu Han · Qiang Ma · Wei Ke · Liang Wang · Zhiyuan Liu · Maosong Sun
16 March 2026 at 12:00
arXiv:2603.12793v1 Announce Type: cross Abstract: A recent cutting-edge topic in multimodal modeling is to unify visual comprehension and generation within a single model. However, the two tasks demand mismatched decoding regimes and visual representations, making it non-trivial to jointly optimize within a shared feature space. In this work, we present Cheers, a unified multimodal model that decouples patch-level details from semantic representations, thereby stabilizing semantics for multimodal understanding and improving fidelity for image generation via gated detail residuals. Cheers includes three key components: (i) a unified vision tokenizer that encodes and compresses image latent states into semantic tokens for efficient LLM conditioning, (ii) an LLM-based Transformer that unifies autoregressive decoding for text generation and diffusion decoding for image generation, and (iii) a cascaded flow matching head that decodes visual semantics first and then injects semantically gated detail residuals from the vision tokenizer to refine high-frequency content. Experiments on popular benchmarks demonstrate that Cheers matches or surpasses advanced UMMs in both visual understanding and generation. Cheers also achieves 4x token compression, enabling more efficient high-resolution image encoding and generation. Notably, Cheers outperforms the Tar-1.5B on the popular benchmarks GenEval and MMBench, while requiring only 20% of the training cost, indicating effective and efficient (i.e., 4x token compression) unified multimodal modeling. We will release all code and data for future research.

GeoChemAD: Benchmarking Unsupervised Geochemical Anomaly Detection for Mineral Exploration

✇cs.AI, q-bio.NC updates on arXiv.org
By: Yihao Ding · Yiran Zhang · Chris Gonzalez · Eun-Jung Holden · Wei Liu
16 March 2026 at 12:00
arXiv:2603.13068v1 Announce Type: cross Abstract: Geochemical anomaly detection plays a critical role in mineral exploration as deviations from regional geochemical baselines may indicate mineralization. Existing studies suffer from two key limitations: (1) single region scenarios which limit model generalizability; (2) proprietary datasets, which makes result reproduction unattainable. In this work, we introduce \textbf{GeoChemAD}, an open-source benchmark dataset compiled from government-led geological surveys, covering multiple regions, sampling sources, and target elements. The dataset comprises eight subsets representing diverse spatial scales and sampling conditions. To establish strong baselines, we reproduce and benchmark a range of unsupervised anomaly detection methods, including statistical models, generative and transformer-based approaches. Furthermore, we propose \textbf{GeoChemFormer}, a transformer-based framework that leverages self-supervised pretraining to learn target-element-aware geochemical representations for spatial samples. Extensive experiments demonstrate that GeoChemFormer consistently achieves superior and robust performance across all eight subsets, outperforming existing unsupervised methods in both anomaly detection accuracy and generalization capability. The proposed dataset and framework provide a foundation for reproducible research and future development in this direction.

Scaling Generalist Data-Analytic Agents

✇cs.AI, q-bio.NC updates on arXiv.org
By: Shuofei Qiao · Yanqiu Zhao · Zhisong Qiu · Xiaobin Wang · Jintian Zhang · Zhao Bin · Ningyu Zhang · Yong Jiang · Pengjun Xie · Fei Huang · Huajun Chen
16 March 2026 at 12:00
arXiv:2509.25084v3 Announce Type: replace-cross Abstract: Data-analytic agents are emerging as a key catalyst for automated scientific discovery and for the vision of Innovating AI. Current approaches, however, rely heavily on prompt engineering over proprietary models, while open-source models struggle to face diverse-format, large-scale data files and long-horizon, multi-step reasoning that real-world analytics demands. This paper introduces DataMind, a scalable data synthesis and agent training recipe designed to build generalist data-analytic agents. DataMind tackles three key challenges in building open-source data-analytic agents, including insufficient data resources, improper training strategy, and unstable code-based multi-turn rollout. Concretely, DataMind applies 1) a fine-grained task taxonomy and a recursive easy-to-hard task composition mechanism to increase the diversity and difficulty of synthesized queries; 2) a knowledge-augmented trajectory sampling strategy followed by model-based and rule-based filtering; 3) a dynamically adjustable training objective combining both SFT and RL losses; 4) a memory-frugal and stable code-based multi-turn rollout framework. Built on DataMind, we curate DataMind-12K, a high-quality trajectory set spanning diverse domains, task categories, and data file formats for data-analytic tasks. Trained on DataMind-12K, our DataMind-14B achieves state-of-the-art with an average score of 71.16% on multiple data analysis benchmarks, outperforming the strongest proprietary baselines DeepSeek-V3.1 and GPT-5. Our DataMind-7B also performs best among all open-source models with a score of 68.10%. We also incorporate some empirical insights gained from our exploratory trials into the analysis experiments, aiming to provide actionable insights about agentic training for the community. We will release DataMind-12K and DataMind-7B,14B for the community's future research.

Think with 3D: Geometric Imagination Grounded Spatial Reasoning from Limited Views

✇cs.AI, q-bio.NC updates on arXiv.org
By: Zhangquan Chen · Manyuan Zhang · Xinlei Yu · Xufang Luo · Mingze Sun · Zihao Pan · Xiang An · Yan Feng · Peng Pei · Xunliang Cai · Ruqi Huang
16 March 2026 at 12:00
arXiv:2510.18632v4 Announce Type: replace-cross Abstract: Though recent advances in vision-language models (VLMs) have achieved remarkable progress across a wide range of multimodal tasks, understanding 3D spatial relationships from limited views remains a significant challenge. Previous reasoning methods typically rely on pure text (e.g., topological cognitive maps) or on 2D visual cues. However, their limited representational capacity hinders performance in specific tasks that require 3D spatial imagination. To address this limitation, we propose 3DThinker, a framework that can effectively exploits the rich geometric information embedded within images while reasoning, like humans do. Our framework is the first to enable 3D mentaling during reasoning without any 3D prior input, and it does not rely on explicitly labeled 3D data for training. Specifically, our training consists of two stages. First, we perform supervised training to align the 3D latent generated by VLM while reasoning with that of a 3D foundation model (e.g., VGGT). Then, we optimize the entire reasoning trajectory solely based on outcome signals, thereby refining the underlying 3D mentaling. Extensive experiments across multiple benchmarks show that 3DThinker consistently outperforms strong baselines and offers a new perspective toward unifying 3D representations into multimodal reasoning. Our code is available at https://github.com/zhangquanchen/3DThinker.

Multimodal Continual Learning with MLLMs from Multi-scenario Perspectives

✇cs.AI, q-bio.NC updates on arXiv.org
By: Kai Jiang · Siqi Huang · Xiangyu Chen · Jiawei Shao · Hongyuan Zhang · Ping Luo · Xuelong Li
16 March 2026 at 12:00
arXiv:2511.18507v3 Announce Type: replace-cross Abstract: Multimodal large language models (MLLMs) deployed on devices must adapt to continuously changing visual scenarios such as variations in background and perspective, to effectively perform complex visual tasks. To investigate catastrophic forgetting under real-world scenario shifts, we construct a multimodal visual understanding dataset (MSVQA), covering four distinct scenarios and perspectives: high-altitude, underwater, low-altitude, and indoor environments. Furthermore, we propose UNIFIER (mUltimodal coNtInual learning with MLLMs From multi-scenarIo pERspectives), a continual learning (CL) framework designed to address visual discrepancies while learning different scenarios. Compared to existing CL methods, UNIFIER enables knowledge accumulation within the same scenario and mutual enhancement across different scenarios via Vision Representation Expansion (VRE) and Vision Consistency Constraint (VCC). Experimental results show that UNIFIER improves the last-step VQA scores by 2.70%~10.62% and the last-step F1 scores by 3.40%~7.69% compared to the state-of-the-art method, QUAD, in 20-step cross-scenario continual learning tasks. MSVQA dataset is available at https://huggingface.co/datasets/Kaij00/MSVQA.

OpenVision 3: A Family of Unified Visual Encoder for Both Understanding and Generation

✇cs.AI, q-bio.NC updates on arXiv.org
By: Letian Zhang · Sucheng Ren · Yanqing Liu · Xianhang Li · Zeyu Wang · Yuyin Zhou · Huaxiu Yao · Zeyu Zheng · Weili Nie · Guilin Liu · Zhiding Yu · Cihang Xie
16 March 2026 at 12:00
arXiv:2601.15369v2 Announce Type: replace-cross Abstract: This paper presents a family of advanced vision encoder, named OpenVision 3, that learns a single, unified visual representation that can serve both image understanding and image generation. Our core architecture is simple: we feed VAE-compressed image latents to a ViT encoder and train its output to support two complementary roles. First, the encoder output is passed to the ViT-VAE decoder to reconstruct the original image, encouraging the representation to capture generative structure. Second, the same representation is optimized with contrastive learning and image-captioning objectives, strengthening semantic features. By jointly optimizing reconstruction- and semantics-driven signals in a shared latent space, the encoder learns representations that synergize and generalize well across both regimes. We validate this unified design through extensive downstream evaluations with the encoder frozen. For generation, we test it under the RAE framework: ours substantially surpasses the standard CLIP-based encoder (e.g., gFID: 1.87 vs. 2.54 on ImageNet). For multimodal understanding, we plug the encoder into the LLaVA-1.5 and LLaVA-NeXT framework: it performs comparably with a standard CLIP vision encoder (e.g., 63.3 vs. 61.2 on SeedBench, and 59.2 vs. 58.1 on GQA). We provide empirical evidence that generation and understanding are mutually beneficial in our architecture, while further underscoring the critical role of the VAE latent space. We hope this work can spur future research on unified modeling.
Received — 14 March 2026 ⏭

Integrated Network Toxicology and Metabolomics Elucidate Mechanisms of Carbosulfan-Induced Respiratory Toxicity in Rats

✇(Multiomics OR Omics) AND (Lung OR gastric OR Hepatocellular)
By: Xian Ju · Di Liang · Hongyu Su · Yachun Zhang · Zhenyu Liang · Yiheng Liu · Wenqi Zhao · Dan Zhang · Zhe Chen · Keming Yun
14 March 2026 at 18:00

Int J Mol Sci. 2026 Feb 25;27(5):2170. doi: 10.3390/ijms27052170.

ABSTRACT

Carbosulfan is a widely used carbamate insecticide, yet its mechanisms of respiratory toxicity remain poorly understood. This study integrated network toxicology, untargeted metabolomics, and molecular docking to systematically investigate the potential mechanisms of carbosulfan-induced respiratory toxicity in male Sprague Dawley rats. Rats were administered a single oral dose of carbosulfan (125 or 250 mg/kg) and assessed after 12 h. Exposure resulted in significant pathological lung damage, characterized by disrupted alveolar architecture, inflammatory cell infiltration, and increased serum levels of the pro-inflammatory cytokines IL-6, IL-1β, and TNF-α. Network toxicology analysis identified 51 potential targets associated with respiratory toxicity, with core targets including SRC, EGFR, PTGS2, CXCL8, CYP3A4, and NR3C1. Enriched pathways were primarily related to neuroactive ligand-receptor interaction, VEGF signaling, and arachidonic acid metabolism. Untargeted metabolomics revealed significant metabolic perturbations in pathways central to antioxidant defense and energy homeostasis, including glutathione metabolism, the tricarboxylic acid cycle, and arginine biosynthesis. Molecular docking confirmed stable in silico binding affinities between carbosulfan and the predicted core targets. Integrative analysis suggests that carbosulfan exposure is associated with respiratory damage, potentially through interconnected mechanisms involving oxidative stress, inflammation, and disruption of cell signaling and metabolic enzyme systems. However, given the acute high-dose nature of the model and the interpretative integration of multi-omics data, these findings should be considered hypothesis-generating. This study provides a novel system-level perspective on carbosulfan-induced respiratory toxicity and highlights key pathways and targets for future validation in chronic exposure models.

PMID:41828400 | PMC:PMC12984169 | DOI:10.3390/ijms27052170

Integrated Network Toxicology and Metabolomics Elucidate Mechanisms of Carbosulfan-Induced Respiratory Toxicity in Rats

✇Omics In Lung
By: Xian Ju · Di Liang · Hongyu Su · Yachun Zhang · Zhenyu Liang · Yiheng Liu · Wenqi Zhao · Dan Zhang · Zhe Chen · Keming Yun
14 March 2026 at 18:00

Int J Mol Sci. 2026 Feb 25;27(5):2170. doi: 10.3390/ijms27052170.

ABSTRACT

Carbosulfan is a widely used carbamate insecticide, yet its mechanisms of respiratory toxicity remain poorly understood. This study integrated network toxicology, untargeted metabolomics, and molecular docking to systematically investigate the potential mechanisms of carbosulfan-induced respiratory toxicity in male Sprague Dawley rats. Rats were administered a single oral dose of carbosulfan (125 or 250 mg/kg) and assessed after 12 h. Exposure resulted in significant pathological lung damage, characterized by disrupted alveolar architecture, inflammatory cell infiltration, and increased serum levels of the pro-inflammatory cytokines IL-6, IL-1β, and TNF-α. Network toxicology analysis identified 51 potential targets associated with respiratory toxicity, with core targets including SRC, EGFR, PTGS2, CXCL8, CYP3A4, and NR3C1. Enriched pathways were primarily related to neuroactive ligand-receptor interaction, VEGF signaling, and arachidonic acid metabolism. Untargeted metabolomics revealed significant metabolic perturbations in pathways central to antioxidant defense and energy homeostasis, including glutathione metabolism, the tricarboxylic acid cycle, and arginine biosynthesis. Molecular docking confirmed stable in silico binding affinities between carbosulfan and the predicted core targets. Integrative analysis suggests that carbosulfan exposure is associated with respiratory damage, potentially through interconnected mechanisms involving oxidative stress, inflammation, and disruption of cell signaling and metabolic enzyme systems. However, given the acute high-dose nature of the model and the interpretative integration of multi-omics data, these findings should be considered hypothesis-generating. This study provides a novel system-level perspective on carbosulfan-induced respiratory toxicity and highlights key pathways and targets for future validation in chronic exposure models.

PMID:41828400 | PMC:PMC12984169 | DOI:10.3390/ijms27052170

A clinical environment simulator for dynamic AI evaluation

✇Nature Medicine
By: Luyang Luo · Sung Eun Kim · Xiaoman Zhang · Julius M. Kernbach · Roshan Kenia · Julian N. Acosta · Larry A. Nathanson · Adrian D. Haimovich · Adam Rodman · Ethan Goh · Jonathan H. Chen · Nigam H. Shah · David A. Kim · James Zou · Faisal Mahmood · Jakob Nikolas Kather · Matthew Lungren · Vivek Natarajan · Eric J. Topol · Pranav Rajpurkar
12 March 2026 at 08:00

Nature Medicine, Published online: 12 March 2026; doi:10.1038/s41591-026-04252-6

The authors propose a framework for clinical AI evaluation within simulated digital hospital environments that capture the evolving constraints, and cascading effects, of clinical decisions.

Risk-adaptive therapy guided by dynamic ctDNA in nasopharyngeal carcinoma

✇Nature - Issue - nature.com science feeds
By: Jiawei Lv · Dan-Xue Zheng · Jin-Hui Liang · Ning Zhang · Zu-Lu Ye · Xu-Dong Xu · Melvin. L. K. Chua · Lu-Lu Zhang · Zi-Ming Du · Zi-Chen Zhang · Wen-Fei Li · Ling-Long Tang · Lei Chen · Yan-Ping Mao · Rui Guo · Yu-Pei Chen · Li Lin · Yuan Zhang · Xu Liu · Cheng Xu · Zhi-Xuan Li · Ling-Xin Xu · Pan-Yang Yang · Kun Chen · Deng Bin · Tian-Sheng Gao · Jian-Ye Yan · Lu-Si Chen · Shao Hui Huang · Hong-Yun Zhao · Shu-Bin Hong · Yu-Sheng Jie · Hui-Ling Huang · Xu-Hua Tang · Jing-Ping Yun · Li-Zhi Liu · Li Tian · Hao-Jiang Li · Ji-Bin Li · Guan-Qun Zhou · Jun Ma · Ying Sun
11 March 2026 at 08:00

Nature, Published online: 11 March 2026; doi:10.1038/s41586-026-10244-w

A clinical trial testing whether monitoring ctDNA clearance during treatment for nasopharyngeal cancer could be used to inform decisions about an individual’s subsequent therapeutic programme shows promising results.

A mechanism to initiate emergency type 2 myelopoiesis

✇Nature - Issue - nature.com science feeds
By: Alexandre Fagnan · Cristina Di Genua · Yiran Meng · Roy Drissen · Zishan Zhang · Bowen Zhang · Padraic G. Fallon · Vassilis Pachnis · Erika J. Mancini · Fränze Progatzky · Claus Nerlov
11 March 2026 at 08:00

Nature, Published online: 11 March 2026; doi:10.1038/s41586-026-10256-6

Myelopoiesis in response to a parasitic worm infection and the mechanism selective to this form of parasite are revealed.
Received — 11 March 2026 ⏭

GameVerse: Can Vision-Language Models Learn from Video-based Reflection?

✇cs.AI, q-bio.NC updates on arXiv.org
By: Kuan Zhang · Dongchen Liu · Qiyue Zhao · Jinkun Hou · Xinran Zhang · Qinlei Xie · Miao Liu · Yiming Li
10 March 2026 at 12:00
arXiv:2603.06656v1 Announce Type: cross Abstract: Human gameplay is a visually grounded interaction loop in which players act, reflect on failures, and watch tutorials to refine strategies. Can Vision-Language Models (VLMs) also learn from video-based reflection? We present GameVerse, a comprehensive video game benchmark that enables a reflective visual interaction loop. Moving beyond traditional fire-and-forget evaluations, it uses a novel reflect-and-retry paradigm to assess how VLMs internalize visual experience and improve policies. To facilitate systematic and scalable evaluation, we also introduce a cognitive hierarchical taxonomy spanning 15 globally popular games, dual action space for both semantic and GUI control, and milestone evaluation using advanced VLMs to quantify progress. Our experiments show that VLMs benefit from video-based reflection in varied settings, and perform best by combining failure trajectories and expert tutorials-a training-free analogue to reinforcement learning (RL) plus supervised fine-tuning (SFT).

Revealing Behavioral Plasticity in Large Language Models: A Token-Conditional Perspective

✇cs.AI, q-bio.NC updates on arXiv.org
By: Liyuan Mao · Le Yu · Jing Zhou · Chujie Zheng · Bowen Yu · Chang Gao · Shixuan Liu · An Yang · Weinan Zhang · JunYang Lin
10 March 2026 at 12:00
arXiv:2603.08398v1 Announce Type: cross Abstract: In this work, we reveal that Large Language Models (LLMs) possess intrinsic behavioral plasticity-akin to chameleons adapting their coloration to environmental cues-that can be exposed through token-conditional generation and stabilized via reinforcement learning. Specifically, by conditioning generation on carefully selected token prefixes sampled from responses exhibiting desired behaviors, LLMs seamlessly adapt their behavioral modes at inference time (e.g., switching from step-by-step reasoning to direct answering) without retraining. Based on this insight, we propose Token-Conditioned Reinforcement Learning (ToCoRL), a principled framework that leverages RL to internalize this chameleon-like plasticity, transforming transient inference-time adaptations into stable and learnable behavioral patterns. ToCoRL guides exploration with token-conditional generation and keep enhancing exploitation, enabling emergence of appropriate behaviors. Extensive experiments show that ToCoRL enables precise behavioral control without capability degradation. Notably, we show that large reasoning models, while performing strongly on complex mathematics, can be effectively adapted to excel at factual question answering, which was a capability previously hindered by their step-by-step reasoning patterns.

Your Agent May Misevolve: Emergent Risks in Self-evolving LLM Agents

✇cs.AI, q-bio.NC updates on arXiv.org
By: Shuai Shao · Qihan Ren · Chen Qian · Boyi Wei · Dadi Guo · Jingyi Yang · Xinhao Song · Linfeng Zhang · Weinan Zhang · Dongrui Liu · Jing Shao
10 March 2026 at 12:00
arXiv:2509.26354v2 Announce Type: replace Abstract: Advances in Large Language Models (LLMs) have enabled a new class of self-evolving agents that autonomously improve through interaction with the environment, demonstrating strong capabilities. However, self-evolution also introduces novel risks overlooked by current safety research. In this work, we study the case where an agent's self-evolution deviates in unintended ways, leading to undesirable or even harmful outcomes. We refer to this as Misevolution. To provide a systematic investigation, we evaluate misevolution along four key evolutionary pathways: model, memory, tool, and workflow. Our empirical findings reveal that misevolution is a widespread risk, affecting agents built even on top-tier LLMs (e.g., Gemini-2.5-Pro). Different emergent risks are observed in the self-evolutionary process, such as the degradation of safety alignment after memory accumulation, or the unintended introduction of vulnerabilities in tool creation and reuse. To our knowledge, this is the first study to systematically conceptualize misevolution and provide empirical evidence of its occurrence, highlighting an urgent need for new safety paradigms for self-evolving agents. Finally, we discuss potential mitigation strategies to inspire further research on building safer and more trustworthy self-evolving agents. Our code and data are available at https://github.com/ShaoShuai0605/Misevolution . Warning: this paper includes examples that may be offensive or harmful in nature.

Adaptation of Agentic AI: A Survey of Post-Training, Memory, and Skills

✇cs.AI, q-bio.NC updates on arXiv.org
By: Pengcheng Jiang · Jiacheng Lin · Zhiyi Shi · Zifeng Wang · Luxi He · Yichen Wu · Ming Zhong · Peiyang Song · Qizheng Zhang · Heng Wang · Xueqiang Xu · Hanwen Xu · Pengrui Han · Dylan Zhang · Jiashuo Sun · Chaoqi Yang · Kun Qian · Tian Wang · Changran Hu · Manling Li · Quanzheng Li · Hao Peng · Sheng Wang · Jingbo Shang · Chao Zhang · Jiaxuan You · Liyuan Liu · Pan Lu · Yu Zhang · Heng Ji · Yejin Choi · Dawn Song · Jimeng Sun · Jiawei Han
10 March 2026 at 12:00
arXiv:2512.16301v3 Announce Type: replace Abstract: Large language model (LLM) agents are moving beyond prompting alone. ChatGPT marked the rise of general-purpose LLM assistants, DeepSeek showed that on-policy reinforcement learning with verifiable rewards can improve reasoning and tool use, and OpenClaw highlights a newer direction in which agents accumulate persistent memory and reusable skills. Yet the research landscape remains fragmented across post-training, retrieval, memory, and skill systems. This survey studies these developments under a single notion of \emph{adaptation}: improving an agent, its tools, or their interaction after pretraining. We organize the field with a four-paradigm framework spanning agent adaptation and tool adaptation. On the agent side, A1 (tool-execution-signaled) and A2 (agent-output-signaled) improve the agent itself through supervised fine-tuning, preference optimization, and reinforcement learning with verifiable rewards. On the tool side, T1 (agent-agnostic) provides reusable pre-trained modules any agent can call, while T2 (agent-supervised) uses the agent's outputs to train memory systems, skill libraries, or lightweight subagents. Using this framework, we review post-training methods, adaptive memory architectures, and agent skills; compare their trade-offs in cost, flexibility, and generalization; and summarize evaluation practices across deep research, software development, computer use, and drug discovery. We conclude by outlining open problems in agent-tool co-adaptation, continual learning, safety, and efficient deployment.

ARLArena: A Unified Framework for Stable Agentic Reinforcement Learning

✇cs.AI, q-bio.NC updates on arXiv.org
By: Xiaoxuan Wang · Han Zhang · Haixin Wang · Yidan Shi · Ruoyan Li · Kaiqiao Han · Chenyi Tong · Haoran Deng · Renliang Sun · Alexander Taylor · Yanqiao Zhu · Jason Cong · Yizhou Sun · Wei Wang
10 March 2026 at 12:00
arXiv:2602.21534v2 Announce Type: replace Abstract: Agentic reinforcement learning (ARL) has rapidly gained attention as a promising paradigm for training agents to solve complex, multi-step interactive tasks. Despite encouraging early results, ARL remains highly unstable, often leading to training collapse. This instability limits scalability to larger environments and longer interaction horizons, and constrains systematic exploration of algorithmic design choices. In this paper, we first propose ARLArena, a stable training recipe and systematic analysis framework that examines training stability in a controlled and reproducible setting. ARLArena first constructs a clean and standardized testbed. Then, we decompose policy gradient into four core design dimensions and assess the performance and stability of each dimension. Through this fine-grained analysis, we distill a unified perspective on ARL and propose SAMPO, a stable agentic policy optimization method designed to mitigate the dominant sources of instability in ARL. Empirically, SAMPO achieves consistently stable training and strong performance across diverse agentic tasks. Overall, this study provides a unifying policy gradient perspective for ARL and offers practical guidance for building stable and reproducible LLM-based agent training pipelines.
❌