Vision-language-action (VLA) policies can learn to follow execution preferences, such as where to grasp an object or how to move around an obstacle, from demonstrations annotated with preference labels. Providing these annotations for every new task, however, is costly. We study whether learned preference control can transfer to new tasks whose demonstrations contain the desired behaviors but lack preference annotations. We find that a policy trained with preference annotations can follow preferences on a new task before adaptation but achieves limited task success. Fine-tuning on target demonstrations without these annotations improves task success while weakening preference control. Motivated by this finding, we introduce Self-Labeled Preference Transfer (SPT). SPT jointly trains a shared vision-language backbone for preference-conditioned action prediction and preference recognition on annotated source demonstrations. It then freezes the recognizer to preserve its cross-task preference recognition capability and uses it to self-label target demonstrations for policy fine-tuning. Across five types of execution preferences in simulation, the recognizer achieves 96.4% accuracy, and SPT reaches 84.8% preference following and 89.2% task success, compared with 59.4% and 88.0% for naive finetuning. We further extend to multiple preferences, achieving 93.8% preference following compared with 42.5% of baselines. In real-world experiments, SPT increases preference following from 52.5% to 92.5% relative to naive finetuning across two task transfers. These results show that SPT can support preference adaptation across tasks and objects while maintaining success, without additional manual preference annotations.
Agent skills, reusable instruction packages centred on a SKILL.md file, extend AI agents on demand, and a single line in a widely copied skill reaches every agent that loads it. Some lines are oversight-loosening directives: they tell the agent to proceed without approval, to bypass checks with --no-verify, or to keep working without checking in. Although such directives can remove a checkpoint from every agent that loads them, no prior work has examined them. We present the first large-scale study of oversight-loosening directives, analyzing 1.53 million English skills from GitSkills. We split skills into units (sentences, bullets, table cells, and code lines), develop a codebook that labels oversight directives by polarity, mechanism, and scope, label 2,097 units by hand, and train a detector to scale the analysis. Our results show that about one in seventeen skills (about 88,000 skills) contains a loosening directive. These directives most often loosen oversight through permission flags, and most of them (70.1%) apply unconditionally rather than only in specific contexts such as CI. Compared with skills that have the same amount of oversight language, skills that run shell code, install packages, or use sudo have 1.7-1.8 times the odds of containing a loosening directive, an association that holds across all our robustness checks, and they do not compensate with more tightening. Loosening spreads through a few upstream skills rather than through edits by their adopters: 2.4% of the skill families that carry loosening account for half of all copies of loosening-carrying skills, and 1.1% of distinct loosening units for half of all loosening occurrences. Open-source skill scanners catch at most 46% of loosening directives, with a precision below 0.4, while the LLM behind them, prompted with our codebook instead, catches 61% at a precision of 0.71.
Kazi Amit Hasan, Enrong Pan, Yuan Tian, Mariam Guizani, Steven H. H. Ding
Open-source vulnerability mitigation spans upstream libraries and the downstream projects that depend on them. Upstream, a vulnerability progresses through a public report when one exists, a fix, a fixed release, and NVD disclosure. Downstream, a project mitigates it by adopting the fixed release, and the timeliness of this adoption is measured relative to these events. Prior work studies the two sides separately, so it is unclear whether downstream exposure stems from delayed upstream release, delayed downstream adoption, or both. We address this gap by following 832 CVEs in the Maven ecosystem from the upstream fix to its adoption by each exposed downstream project. We classify upstream fixes as \textit{silent} or \textit{transparent}, recover the order of report, fix, release, and NVD disclosure, and link vulnerable artifacts to their downstream dependents. We then compare this order with downstream adoption and dependabot usage. We find that delay arises on both sides. Upstream, the fixed release comes before NVD disclosure for about 82\% of CVEs, for both silent and transparent fixes. For the other 17.9\%, downstream projects have no fix to adopt for days to more than four months after disclosure. Downstream, exposure is concentrated, with the top 10\% of CVEs accounting for 72.1\% of it. Of 1,677 observed adoptions, 85.3\% happen after NVD disclosure, rising to 88.3\% for silent fixes released before disclosure. Most adoptions come through routine dependency updates. Adoptions with dependabot version updates have a median release-to-adoption delay of 10.4 days, compared with 110.6 days without dependabot. Yet version updates were configured at adoption time for only 8.5\% of total adoptions. We translate these findings into implications for upstream maintainers, downstream developers, and dependency-automation tools.
Baihui Liu National University of Defense Technology, Kaiyuan Tian National University of Defense Technology, Wei Wang National University of Defense Technology, Zhaoning Zhang National University of Defense Technology, Linbo Qiao, Dongsheng Li National University of Defense Technology
Mixture-of-Experts (MoE) has become a dominant architecture for scaling large language models due to their sparse activation mechanism. However, the substantial number of expert activations creates a critical latency bottleneck during inference, especially in resource-constrained deployment scenarios. Existing approaches that reduce expert activations potentially lead to severe model performance degradation. In this work, we introduce the concept of activation budget as a constraint on the number of expert activations and propose Alloc-MoE, a unified framework that optimizes budget allocation coordinately at both the layer and token levels to minimize performance degradation. At the layer level, we introduce Alloc-L, which leverages sensitivity profiling and dynamic programming to determine the optimal allocation of expert activations across layers. At the token level, we propose Alloc-T, which dynamically redistributes activations based on routing scores, optimizing budget allocation without increasing latency. Extensive experiments across multiple MoE models demonstrate that Alloc-MoE maintains model performance under a constrained activation budget. Especially, Alloc-MoE achieves 1.15\times prefill and 1.34\times decode speedups on DeepSeek-V2-Lite at half of the original budget.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Zixuan Weng Pennsylvania State University, Jinghuai Zhang University of California, Los Angeles, Kunlin Cai University of California, Los Angeles, Ying Li, Peiran Wang, Yuan Tian University of California, Los Angeles
Large language models (LLMs) often exhibit undesirable behaviors, such as safety violations and hallucinations. Although inference-time steering offers a cost-effective way to adjust model behavior without updating its parameters, existing methods often fail to be simultaneously effective, utility-preserving, and training-efficient due to their rigid, one-size-fits-all designs and limited adaptability. In this work, we present FineSteer, a novel steering framework that decomposes inference-time steering into two complementary stages—conditional steering and fine-grained vector synthesis—allowing fine-grained control over when and how to steer internal representations. In the first stage, we introduce a Subspace-guided Conditional Steering (SCS) mechanism that preserves model utility by avoiding unnecessary steering. In the second stage, we propose a Mixture-of-Steering-Experts (MoSE) mechanism that captures the multimodal nature of desired steering behaviors and generates query-specific steering vectors for improved effectiveness. Through tailored designs in both SCS and MoSE, FineSteer maintains robust performance on general queries while adaptively optimizing steering vectors for targeted inputs in a training-efficient manner. Extensive experiments on safety and truthfulness benchmarks show that FineSteer outperforms the state-of-the-art methods in overall performance (e.g., a 7.6% improvement on TruthfulQA over Llama-3), achieving stronger steering performance with minimal utility loss. The code is available at https://github.com/YukinoAsuna/FineSteer
尚无 AI 解读，展示原文摘要
cs.CL · cs.CV
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表