❌

Reading view

↗️

PANDO: Efficient Multimodal AI Agents via Online Skill Distillation

✇cs.AI, q-bio.NC updates on arXiv.org
By: Yubo Li · Yidi Miao · Yuntian Shen · Yuxin Liu
arXiv:2605.24785v2 Announce Type: new Abstract: Recent advances in multimodal web agents often rely on increased inference-time computation, including rollout search, verifier passes, offline skill discovery, and specialist model stacks. This raises a central question: can a web agent become more efficient as it accumulates experience, rather than more expensive? We first analyze trajectories from VisualWebArena and identify three recurring sources of inefficiency: repeat-action loops, hidden discovery costs, and low prompt-cache reuse. We then introduce PANDO, a single-rollout online skill-distillation framework that maintains a structured Skill Library and combines progress reflection, confidence-based skill demotion, hierarchical routing, visual compression, and cache-aware prompting. On the full set of 910 VisualWebArena tasks, PANDO achieves a 58.3% success rate, outperforming SGV (54.0%) and our WALT reproduction (45.2%), while using 58% fewer tokens than SGV and 61% fewer tokens than WALT, without any pre-evaluation discovery budget. A 300-task ablation further shows that rules and routines provide most of the success gains, while routing, compression, and cache-aware prompting convert the larger skill library into lower marginal token cost. Finally, we introduce three trajectory-level efficiency metrics -- Action Repetition Rate, Step Overhead Ratio, and Prompt Cache Utilization -- to make efficiency visible beyond terminal success.
  •  
  • ↗️
↗️

L2IR: Revealing Latent Intent in Graph Fraud Detection

✇cs.AI, q-bio.NC updates on arXiv.org
By: Jinsheng Guo · Zhenhao Weng · Yibo Liu · Yan Qiao · Meng Li
arXiv:2605.26040v1 Announce Type: new Abstract: Graph fraud detection has long depended on Graph Neural Networks (GNNs) to propagate and aggregate information across relational data. A critical obstacle in practice, however, is that fraudsters frequently disguise themselves by forging numerous connections with benign users, causing fraud signals to be progressively diluted during neighborhood aggregation and undermining detection reliability. While recent efforts have used Large Language Models (LLMs) to provide rich semantic cues for fraud detection, the underlying intent behind suspicious connections remains insufficiently explored. Compounding this issue, the scarcity of annotated fraud samples makes it difficult to train detectors that remain robust under heavy camouflage. To address these gaps, we propose L2IR, an LLM-driven Latent Intent Revealing framework for graph fraud detection. By uncovering latent intent from both user behaviors and suspicious connections, L2IR extracts intent-aware representations from raw behavioral traces and reasons about the true purpose behind individual connections, effectively distinguishing supportive links from misleading ones. It further incorporates adaptive self-training to enhance robustness under limited supervision. Evaluations on two real-world datasets characterized by pervasive camouflage demonstrate that L2IR surpasses strong baselines and can function as a plug-in enhancement for a range of GNN-based detectors, improving AUPRC by up to 8.27%.
  •  
  • ↗️
↗️

VaaWIT: Visual-Aware Adaptation of Large Language Models for Multilingual Web Image Translation

✇cs.AI, q-bio.NC updates on arXiv.org
By: Bo Li · Ronghao Chen · Ningyuan Deng · Huacan Wang · Shaolin Zhu · Lijie Wen
arXiv:2605.24675v1 Announce Type: cross Abstract: Translating text embedded in Web images is crucial for improving content accessibility and cross-lingual information retrieval, particularly within social media and e-commerce domains. Although Large Vision-Language Models (LVLMs) have advanced multimodal understanding, applying them to Web image translation remains challenging due to the visual representation gap: standard encoders often prioritize high-level semantics over the fine-grained visual details required for recognizing diverse character morphologies. To address this challenge, we propose VaaWIT, an end-to-end framework that adapts Large Language Models for multilingual Web image translation. The framework introduces two key technical contributions: (1) a Dual-Stream Attention Module (DSAM), which facilitates bidirectional interaction between multilingual semantic features and detailed visual representations, thereby synthesizing unified features robust to textual variations; and (2) a Visual-Aware Adapter (VAA), a parameter-efficient fine-tuning strategy that dynamically injects these fused visual cues into the frozen LLM backbone. This design enables the model to align the visual context with linguistic reasoning effectively while minimizing computational costs. Extensive experiments on eight tasks on three public benchmarks demonstrate that VaaWIT significantly outperforms state-of-the-art (SOTA) open-source baselines and achieves competitive performance against proprietary models. These results validate the efficacy of integrating fine-grained visual perception into LLMs for complex Web content analysis.
  •  
  • ↗️
↗️

Mix-MoE: Improving Multilingual Machine Translation of Large Language Models through Mixed MoEs

✇cs.AI, q-bio.NC updates on arXiv.org
By: Bo Li · Tianyu Dong · Shaolin Zhu · Deyi Xiong
arXiv:2605.24681v1 Announce Type: cross Abstract: Large Language Models (LLMs) have shown great promise in multilingual machine translation (MT), even with limited bilingual supervision. However, fine-tuning LLMs with parallel corpora presents major challenges, namely parameter interference. To address these issues, we propose Mix-MoE, a mixed Mixture-of-Experts framework designed to train LLMs for multilingual MT. Our framework operates in two distinct stages: (1) post-pretraining with MoE on monolingual corpora, and (2) post-pretraining with MoE on parallel corpora. Crucially, we divide the MoE layers into two specialized groups: Language Model Experts (LM Experts) and Machine Translation Experts (MT Experts). LM Experts are designed to capture and retain the monolingual knowledge learned by the pre-trained LLM. MT Experts, on the other hand, are specifically trained to acquire and store bilingual translation knowledge. Furthermore, to facilitate effective interaction between these specialized experts and leverage potential underlying structural patterns in text, we introduce a routing mechanism enhanced by Fourier Transform features derived from model representations. The experimental results demonstrate that Mix-MoE excels in multilingual MT, significantly outperforming existing baselines and showing notable progress in mitigating parameter interference.
  •  
  • ↗️
↗️

CONF-KV: Confidence-Aware KV Cache Eviction with Mixed-Precision Storage for Long-Horizon LLM

✇cs.AI, q-bio.NC updates on arXiv.org
By: Yubo Li · Yidi Miao
arXiv:2605.24786v1 Announce Type: cross Abstract: Long-horizon LLM inference turns the key--value (KV) cache into the dominant GPU memory consumer and makes per-token attention increasingly expensive. Many common eviction policies use static recency windows or historical attention, leaving unused a signal computed on every decoding step: the model's current uncertainty. We introduce CONF-KV, a KV-cache manager that converts the next-token distribution into a scalar confidence score and uses it to choose the per-step cache budget, retaining more context when the model is uncertain and pruning aggressively when it is confident. Within each budget, tokens are ranked by a composite of accumulated attention mass and recency, while a protected recent window preserves local coherence. We combine the policy with blockwise online-softmax attention, mixed FP16/INT8 storage, and a pyramidal per-layer budget variant. Across four model families and generated lengths up to 4K, CONF-KV stays near the footprint of a fixed 512-token sliding window while remaining within 1.5--2.1 perplexity points of full KV. On Needle-in-a-Haystack up to 32K tokens, CONF-KV reaches 91.4% retrieval accuracy versus 53.8% for sliding windows and 80.6% for H2O; on 75 VisualWebArena tasks it retains 95.3% of full-KV success at 2.8 times lower peak memory.
  •  
  • ↗️
↗️

Agent Learning via Early Experience

✇cs.AI, q-bio.NC updates on arXiv.org
By: Kai Zhang · Xiangchao Chen · Bo Liu · Tianci Xue · Zeyi Liao · Zhihan Liu · Xiyao Wang · Yuting Ning · Zhaorun Chen · Xiaohan Fu · Jian Xie · Yuxuan Sun · Boyu Gou · Qi Qi · Zihang Meng · Jianwei Yang · Ning Zhang · Xian Li · Ashish Shah · Dat Huynh · Hengduo Li · Zi Yang · Sara Cao · Lawrence Jang · Shuyan Zhou · Jiacheng Zhu · Huan Sun · Jason Weston · Yu Su · Yifan Wu
arXiv:2510.08558v3 Announce Type: replace Abstract: A long-term goal of language agents is to learn and improve through their own experience, ultimately outperforming humans in complex, real-world tasks. However, training agents from experience data with reinforcement learning remains difficult in many environments, which either lack verifiable rewards (e.g., websites) or require inefficient long-horizon rollouts (e.g., multi-turn tool use). As a result, most current agents rely on supervised fine-tuning on expert data, which is challenging to scale and generalizes poorly. This limitation stems from the nature of expert demonstrations: they capture only a narrow range of scenarios, and expose the agent to limited environment diversity. We address this limitation with a middle-ground paradigm we call early experience: interaction data generated by the agent's own actions, where the resulting future states serve as supervision without reward signals. Within this paradigm, we study two strategies of using such data: (1) implicit world modeling, which uses collected states to ground the policy in environment dynamics; and (2) self-reflection, where the agent learns from its suboptimal actions to improve reasoning and decision-making. Evaluation across eight diverse environments and multiple model families shows that our approaches consistently improve effectiveness and out-of-domain generalization, highlighting the value of early experience. Moreover, in environments with verifiable rewards, our results provide promising signals that early experience offers a strong foundation for subsequent reinforcement learning, making it a practical bridge between imitation learning and fully experience-driven agents.
  •  
  • ↗️
↗️

Safety in Embodied AI: A Survey of Risks, Attacks, and Defenses

✇cs.AI, q-bio.NC updates on arXiv.org
By: Xiao Li · Xiang Zheng · Yifeng Gao · Xinyu Xia · Yixu Wang · Xin Wang · Ye Sun · Yunhan Zhao · Ming Wen · Jiayu Li · Zixing Chen · Xun Gong · Yi Liu · Yige Li · Yutao Wu · Cong Wang · Jun Sun · Yixin Cao · Zhineng Chen · Jingjing Chen · Tao Gui · Qi Zhang · Zuxuan Wu · Xipeng Qiu · Xuanjing Huang · Tiehua Zhang · Zhipeng Wei · Kun Wang · Xinfeng Li · Hanxun Huang · Sarah Erfani · James Bailey · Jianping Wang · Chaowei Xiao · Ran He · Bo Li · Xingjun Ma · Yu-Gang Jiang
arXiv:2605.02900v2 Announce Type: replace-cross Abstract: Embodied Artificial Intelligence (Embodied AI) integrates perception, cognition, planning, and interaction into agents that operate in open-world, safety-critical environments. As these systems gain autonomy and enter domains such as transportation, healthcare, and industrial or assistive robotics, ensuring their safety becomes both technically challenging and socially indispensable. Unlike digital AI systems, embodied agents must act under uncertain sensing, incomplete knowledge, and dynamic human-robot interactions, where failures can directly lead to physical harm. This survey provides a comprehensive and structured review of safety research in embodied AI, examining attacks and defenses across the full embodied pipeline, from perception and cognition to planning, action and interaction, and agentic system. We introduce a multi-level taxonomy that unifies fragmented lines of work and connects embodied-specific safety findings with broader advances in vision, language, and multimodal foundation models. Our review synthesizes insights from over 500 papers spanning adversarial, backdoor, jailbreak, and hardware-level attacks; attack detection, safe training and robust inference; and risk-aware human-agent interaction. This analysis reveals several overlooked challenges, including the fragility of multimodal perception fusion, the instability of planning under jailbreak attacks, and the trustworthiness of human-agent interaction in open-ended scenarios. By organizing the field into a coherent framework and identifying critical research gaps, this survey provides a roadmap for building embodied agents that are not only capable and autonomous but also safe, robust, and reliable in real-world deployment.
  •  
  • ↗️
↗️

ProcCtrlBench: Evaluating Process-Level Defects and Control Preservation in LLM Coding Agents

✇cs.AI, q-bio.NC updates on arXiv.org
By: Jiawei He · Jie Jia · Chenbo Liu · Chaoyi Xue · Yapeng Song · Xikai Yang · Dong Sun
arXiv:2605.20251v4 Announce Type: replace-cross Abstract: Existing benchmarks for LLM coding agents primarily evaluate final outcomes. While useful for measuring overall capability, these metrics provide limited visibility and often miss defects that arise during execution. We present ProcCtrlBench, a benchmark for execution-process evaluation in LLM coding agents. ProcCtrlBench organizes recurrent execution defects into a reusable ontology covering 11 defect types in 4 categories, and evaluates agent trajectories through standardized process evidence rather than final outcomes alone. To support comparison across heterogeneous agents, ProcCtrlBench standardizes raw logs into a unified trajectory representation and reports calibrated scorecards over process-level findings. In addition, ProcCtrlBench uses control preservation as a way to quantify execution-process quality, capturing whether execution remains interpretable, interruptible, correctable, reversible, and able to hand back authority when needed. We evaluate ProcCtrlBench on 200 cases sampled from three benchmarks: AndroidBench, TerminalBench, and SWE-bench-Verified. Results show that ProcCtrlBench can be instantiated with useful reliability, provides more stable semantics than direct thresholding, and reveals meaningful differences in execution quality that are often overlooked by conventional outcome-based evaluation.
  •  
  • ↗️
↗️

A framework for building a synthetic cell from the SynCell Asia Initiative

✇Nature Biotechnology - Issue - nature.com science feeds
By: Zhuojun Dai · Wataru Aoki · Pimchai Chaiyen · Matthew Wook Chang · Xixian Chen · Zibo Chen · Byung-Kwan Cho · Nan-Nan Deng · Shishen Du · Jee Loon Foo · Julius Fredens · Meifang Fu · Xiaojun Han · Tomohisa Hasunuma · Norikazu Ichihashi · Nurulfiza Mat Isa · Daisuke Kiga · Dong-Myung Kim · Jongmin Kim · Akihiko Kondo · Yuishin Kosaka · Kobkul Laoteng · Joongoo Lee · Jian Li · Junbai Li · Ming Li · Sierin Lim · Jianbo Liu · Xingguo Liu · Yuan Lu · Nan Luo · Tomoaki Matsuura · Wataru Mizunashi · Yan Qiao · Ahmad Bazli Ramzi · Sang Woo Seo · Yoshihiro Shimizu · Kwanwoo Shin · Liangfei Tian · Wanwipa Vongsangnak · Baojun Wang · Cheng-Cai Zhang · Xian-En Zhang · Guoping Zhao · Bo Zheng · Yaoqi Zhou · Chenli Liu

Nature Biotechnology, Published online: 26 May 2026; doi:10.1038/s41587-026-03153-w

Building a living cell from scratch requires overcoming a bottleneck that has remained unresolved despite decades of progress: orchestrating the spatiotemporal integration of core functional modules. To tackle this barrier, the SynCell Asia Initiative outlines a strategy for developing core functional modules followed by their systems-level integration through the establishment of a centralized, artificial intelligence (AI)-driven biofoundry.
  •  
  • ↗️
↗️

Integrative multi-omics analysis identifies stromal-immune crosstalk as a determinant of immunotherapy efficacy and establishes a prognostic signature in gastric cancer

✇Omics in Gastric
By: Fan Zhang · Qingqing Zhang · Shuai Shao · Xiaobo Li · Yi Cheng · Xu Cao · Xiaotang Yu · Zhengming Gao

Comput Biol Chem. 2026 Apr 23;124(Pt 1):109095. doi: 10.1016/j.compbiolchem.2026.109095. Online ahead of print.

ABSTRACT

Immune checkpoint inhibitors like pembrolizumab exhibit variable efficacy in metastatic gastric cancer (GC). This study aimed to identify molecular drivers of pembrolizumab response, explore mechanisms of immune checkpoint inhibitors (ICIs) efficacy, and develop a prognostic signature. Transcriptomic analysis of pembrolizumab-treated GC (TIGER database) identified 165 response-associated differentially expressed genes (DEGs). Functional annotation and single-cell RNA sequencing (scRNA-seq) data from the Gene Expression Omnibus (GEO) revealed that responder-upregulated genes (R-DEGs) were enriched in immune activation pathways and mainly localized to CD8 + T/NK cells. In contrast, non-responder-upregulated genes (D-DEGs) were linked to extracellular matrix (ECM) remodeling and mainly expressed in fibroblasts/endothelial cells. CellChat analysis demonstrated that key DEGs mediate immune-stromal crosstalk via MHC-I and collagen/laminin signaling. A prognostic signature (Lasso-StepCox[forward] Riskscore; LSR: APOD, APOH, BATF2, GJA1, MAGED1, SLC5A1, SLCO2A1, VWF, VCAN) was derived and validated in four independent GC cohorts from the GEO and Cancer Genome Atlas (TCGA) database. Multi-omics analyses showed that LSR-high tumors exhibited aggressive clinicopathological features, increased stromal components, reduced cytotoxic immune infiltration, diminished tumor mutational burden (TMB), and poorer prognosis. Immunohistochemistry (IHC) and spatial transcriptomics in GC showed that stromal VWF/VCAN expression correlates with reduced CD8⁺ T cell granzyme B expression, suggesting T cell dysfunction. High VWF expression in GC predicted poor survival, and a combined VWF/VCAN score showed enhanced prognostic stratification. This study highlights stromal-immune crosstalk as a driver of pembrolizumab resistance and provides a signature as a clinical tool for prognosis and personalized therapy in metastatic GC.

PMID:42068630 | DOI:10.1016/j.compbiolchem.2026.109095

  •  
  • ↗️
↗️

Mummified early Permian reptile reveals ancient amniote breathing apparatus

✇Nature - Issue - nature.com science feeds
By: Robert R. Reisz · Ethan D. Mooney · Tea Maho · David Mazierski · Xu Chu · Joseph J. Bevitt · Yao-Chang Lee · Pei-Yu Huang · Xiaobo Li · Jun Chen

Nature, Published online: 08 April 2026; doi:10.1038/s41586-026-10307-y

A mummified fossil of the early Permian reptile Captorhinus reveals the potential ancestral amniote breathing mechanism and its impact on terrestrial vertebrate evolution.
  •  
  • ↗️
↗️

ShieldNet: Network-Level Guardrails against Emerging Supply-Chain Injections in Agentic Systems

✇cs.AI, q-bio.NC updates on arXiv.org
By: Zhuowen Yuan · Zhaorun Chen · Zhen Xiang · Nathaniel D. Bastian · Seyyed Hadi Hashemi · Chaowei Xiao · Wenbo Guo · Bo Li
arXiv:2604.04426v1 Announce Type: new Abstract: Existing research on LLM agent security mainly focuses on prompt injection and unsafe input/output behaviors. However, as agents increasingly rely on third-party tools and MCP servers, a new class of supply-chain threats has emerged, where malicious behaviors are embedded in seemingly benign tools, silently hijacking agent execution, leaking sensitive data, or triggering unauthorized actions. Despite their growing impact, there is currently no comprehensive benchmark for evaluating such threats. To bridge this gap, we introduce SC-Inject-Bench, a large-scale benchmark comprising over 10,000 malicious MCP tools grounded in a taxonomy of 25+ attack types derived from MITRE ATT&CK targeting supply-chain threats. We observe that existing MCP scanners and semantic guardrails perform poorly on this benchmark. Motivated by this finding, we propose ShieldNet, a network-level guardrail framework that detects supply-chain poisoning by observing real network interactions rather than surface-level tool traces. ShieldNet integrates a man-in-the-middle (MITM) proxy and an event extractor to identify critical network behaviors, which are then processed by a lightweight classifier for attack detection. Extensive experiments show that ShieldNet achieves strong detection performance (up to 0.995 F-1 with only 0.8% false positives) while introducing little runtime overhead, substantially outperforming existing MCP scanners and LLM-based guardrails.
  •  
  • ↗️
↗️

FileGram: Grounding Agent Personalization in File-System Behavioral Traces

✇cs.AI, q-bio.NC updates on arXiv.org
By: Shuai Liu · Shulin Tian · Kairui Hu · Yuhao Dong · Zhe Yang · Bo Li · Jingkang Yang · Chen Change Loy · Ziwei Liu
arXiv:2604.04901v1 Announce Type: cross Abstract: Coworking AI agents operating within local file systems are rapidly emerging as a paradigm in human-AI interaction; however, effective personalization remains limited by severe data constraints, as strict privacy barriers and the difficulty of jointly collecting multimodal real-world traces prevent scalable training and evaluation, and existing methods remain interaction-centric while overlooking dense behavioral traces in file-system operations; to address this gap, we propose FileGram, a comprehensive framework that grounds agent memory and personalization in file-system behavioral traces, comprising three core components: (1) FileGramEngine, a scalable persona-driven data engine that simulates realistic workflows and generates fine-grained multimodal action sequences at scale; (2) FileGramBench, a diagnostic benchmark grounded in file-system behavioral traces for evaluating memory systems on profile reconstruction, trace disentanglement, persona drift detection, and multimodal grounding; and (3) FileGramOS, a bottom-up memory architecture that builds user profiles directly from atomic actions and content deltas rather than dialogue summaries, encoding these traces into procedural, semantic, and episodic channels with query-time abstraction; extensive experiments show that FileGramBench remains challenging for state-of-the-art memory systems and that FileGramEngine and FileGramOS are effective, and by open-sourcing the framework, we hope to support future research on personalized memory-centric file-system agents.
  •  
  • ↗️
↗️

SoSBench: Benchmarking Safety Alignment on Six Scientific Domains

✇cs.AI, q-bio.NC updates on arXiv.org
By: Fengqing Jiang · Fengbo Ma · Zhangchen Xu · Yuetai Li · Zixin Rao · Bhaskar Ramasubramanian · Luyao Niu · Bo Li · Xianyan Chen · Zhen Xiang · Radha Poovendran
arXiv:2505.21605v3 Announce Type: replace-cross Abstract: Large language models (LLMs) exhibit advancing capabilities in complex tasks, such as reasoning and graduate-level question answering, yet their resilience against misuse, particularly involving scientifically sophisticated risks, remains underexplored. Existing safety benchmarks typically focus either on instructions requiring minimal knowledge comprehension (e.g., ``tell me how to build a bomb") or utilize prompts that are relatively low-risk (e.g., multiple-choice or classification tasks about hazardous content). Consequently, they fail to adequately assess model safety when handling knowledge-intensive, hazardous scenarios. To address this critical gap, we introduce SoSBench, a regulation-grounded, hazard-focused benchmark encompassing six high-risk scientific domains: chemistry, biology, medicine, pharmacology, physics, and psychology. The benchmark comprises 3,000 prompts derived from real-world regulations and laws, systematically expanded via an LLM-assisted evolutionary pipeline that introduces diverse, realistic misuse scenarios (e.g., detailed explosive synthesis instructions involving advanced chemical formulas). We evaluate frontier models within a unified evaluation framework using our SoSBench. Despite their alignment claims, advanced models consistently disclose policy-violating content across all domains, demonstrating alarmingly high rates of harmful responses (e.g., 84.9% for Deepseek-R1 and 50.3% for GPT-4.1). These results highlight significant safety alignment deficiencies and underscore urgent concerns regarding the responsible deployment of powerful LLMs.
  •  
  • ↗️
↗️

InfoTok: Information-Theoretic Regularization for Capacity-Constrained Shared Visual Tokenization in Unified MLLMs

✇cs.AI, q-bio.NC updates on arXiv.org
By: Lv Tang · Tianyi Zheng · Bo Li · Xingyu Li
arXiv:2602.01554v2 Announce Type: replace-cross Abstract: Unified multimodal large language models (MLLMs) aim to unify image understanding and image generation within a single framework, where a shared visual tokenizer serves as the sole interface that maps high-dimensional images into a limited token budget for downstream multimodal reasoning and synthesis. However, existing shared-token designs are largely architecture-driven and lack an explicit criterion for what information should be preserved to simultaneously support semantic abstraction and visual detail. In this paper, we adopt a capacity-constrained perspective, viewing the shared tokenizer as a compute-bounded learner whose finite representational budget should prioritize reusable structure over hard-to-exploit high-entropy variations and redundancy. Motivated by this view, we propose \textbf{\textit{InfoTok}}, an information-regularized tokenization mechanism grounded in the Information Bottleneck (IB) principle. InfoTok explicitly controls information flow from images to shared tokens to multimodal outputs by imposing mutual-information (MI) constraints that enforce a principled trade-off between compression and task relevance, while also encouraging cross-modal consistency. Because MI is intractable for high-dimensional visual representations, we instantiate InfoTok with practical, differentiable dependence estimators, including a variational IB formulation and a Hilbert Schmidt Independence Criterion (HSIC) based alternative. Integrated into three representative unified MLLMs without introducing any additional training data, InfoTok consistently improves both image understanding and generation performance. These results support information-regularized visual tokenization as a sound basis for token learning in unified MLLMs.
  •  
  • ↗️
↗️

DR-LoRA: Dynamic Rank LoRA for Fine-Tuning Mixture-of-Experts Models

✇cs.AI, q-bio.NC updates on arXiv.org
By: Guanzhi Deng · Bo Li · Ronghao Chen · Xiujin Liu · Zhuo Han · Huacan Wang · Lijie Wen · Linqi Song
arXiv:2601.04823v5 Announce Type: replace Abstract: Mixture-of-Experts (MoE) has become a prominent paradigm for scaling Large Language Models (LLMs). Parameter-efficient fine-tuning methods, such as LoRA, are widely adopted to adapt pretrained MoE LLMs to downstream tasks. However, existing approaches typically assign identical LoRA ranks to all expert modules, ignoring the heterogeneous specialization of pretrained experts. This uniform allocation leads to a resource mismatch: task-relevant experts are under-provisioned, while less relevant ones receive redundant parameters. To address this, we propose DR-LoRA, a Dynamic Rank LoRA framework for fine-tuning pretrained MoE models. Specifically, DR-LoRA initializes all expert LoRA modules with a small active rank and uses an expert saliency score, which combines routing frequency and gradient-based rank importance, to identify which experts would benefit most from additional capacity. It then periodically expands the active ranks of the task-critical expert LoRA, progressively constructing a heterogeneous rank distribution tailored to the target task. Experiments on three MoE models across six tasks show that DR-LoRA consistently outperforms LoRA and other strong baselines, demonstrating that task-adaptive heterogeneous rank allocation is an effective strategy to improve active capacity utilization in MoE fine-tuning.
  •  
  • ↗️
↗️

Learning to Learn-at-Test-Time: Language Agents with Learnable Adaptation Policies

✇cs.AI, q-bio.NC updates on arXiv.org
By: Zhanzhi Lou · Hui Chen · Yibo Li · Qian Wang · Bryan Hooi
arXiv:2604.00830v2 Announce Type: replace-cross Abstract: Test-Time Learning (TTL) enables language agents to iteratively refine their performance through repeated interactions with the environment at inference time. At the core of TTL is an adaptation policy that updates the actor policy based on experience from previous episodes, thereby improving future behavior. Existing methods rely on fixed, hand-crafted adaptation policies rather than optimizing them for downstream improvement. We argue that optimal adaptation policies should be learned from task environments, not hand-engineered based on human intuition. To achieve this, we introduce Meta-TTL, a framework that formulates the discovery of effective adaptation policies as a bi-level optimization problem. Within this framework, the inner loop executes the standard TTL process, measuring how effectively a candidate adaptation policy helps an agent correct errors across sequential episodes. Guided by the agent's performance, the outer loop employs evolutionary search over a diverse distribution of training tasks to iteratively refine the adaptation policy. We evaluate Meta-TTL on Jericho and WebArena-Lite across both in-distribution (ID) and out-of-distribution (OOD) settings, using multiple meta-agent backbones. Results on both benchmarks show that Meta-TTL consistently outperforms hand-crafted baselines, suggesting that the optimized adaptation policy encodes transferable strategies that generalize beyond the training task distribution.
  •  
  • ↗️
↗️

Xuanwu: Evolving General Multimodal Models into an Industrial-Grade Foundation for Content Ecosystems

✇cs.AI, q-bio.NC updates on arXiv.org
By: Zhiqian Zhang · Xu Zhao · Xiaoqing Xu · Guangdong Liang · Weijia Wang · Xiaolei Lv · Bo Li · Jun Gao
arXiv:2603.29211v1 Announce Type: new Abstract: In recent years, multimodal large models have continued to improve on general benchmarks. However, in real-world content moderation and adversarial settings, mainstream models still suffer from degraded generalization and catastrophic forgetting because of limited fine-grained visual perception and insufficient modeling of long-tail noise. In this paper, we present Xuanwu VL-2B as a case study of how general multimodal models can be developed into an industrial-grade foundation model for content ecosystems. The model adopts a compact InternViT-300M + MLP + Qwen3 1.7B architecture, balancing fine-grained visual perception, language-semantic alignment, and deployment cost within an approximately 2B-parameter budget. To balance business specialization with the retention of general capabilities, we developed a data iteration and curation mechanism and trained the model through a progressive three-stage pipeline: pre-training, mid-training, and post-training. Ablation studies and offline business evaluations show that Xuanwu VL-2B achieves an average score of 67.90 across seven OpenCompass multimodal metrics (vs. 64.27 for InternVL 3.5 2B), an average recall of 94.38% over seven independent business moderation tasks, and a weighted overall recall of 82.82% on policy-violating text in challenging adversarial OCR scenarios, outperforming Gemini-2.5-Pro (76.72%). These results show that, under a limited parameter budget, Xuanwu VL-2B achieves a practical balance among business alignment, visual perception, general capability retention, and deployment cost.
  •  
  • ↗️
↗️

The Model Says Walk: How Surface Heuristics Override Implicit Constraints in LLM Reasoning

✇cs.AI, q-bio.NC updates on arXiv.org
By: Yubo Li · Lu Zhang · Tianchong Jiang · Ramayya Krishnan · Rema Padman
arXiv:2603.29025v1 Announce Type: cross Abstract: Large language models systematically fail when a salient surface cue conflicts with an unstated feasibility constraint. We study this through a diagnose-measure-bridge-treat framework. Causal-behavioral analysis of the ``car wash problem'' across six models reveals approximately context-independent sigmoid heuristics: the distance cue exerts 8.7 to 38 times more influence than the goal, and token-level attribution shows patterns more consistent with keyword associations than compositional inference. The Heuristic Override Benchmark (HOB) -- 500 instances spanning 4 heuristic by 5 constraint families with minimal pairs and explicitness gradients -- demonstrates generality across 14 models: under strict evaluation (10/10 correct), no model exceeds 75%, and presence constraints are hardest (44%). A minimal hint (e.g., emphasizing the key object) recovers +15 pp on average, suggesting the failure lies in constraint inference rather than missing knowledge; 12/14 models perform worse when the constraint is removed (up to -39 pp), revealing conservative bias. Parametric probes confirm that the sigmoid pattern generalizes to cost, efficiency, and semantic-similarity heuristics; goal-decomposition prompting recovers +6 to 9 pp by forcing models to enumerate preconditions before answering. Together, these results characterize heuristic override as a systematic reasoning vulnerability and provide a benchmark for measuring progress toward resolving it.
  •  
  • ↗️
❌