Liyan Tang, Cyrus Rashtchian, Chun-Sung Ferng, Andrew Tomkins, Da-Cheng Juan, Tu Vu · Read the paper

Hook: The central idea is to make an agent’s past experience a persistent resource for improving future skills.

Summary: WikiSkill is a framework that co-evolves reusable agent skills with a persistent knowledge base, or wiki. The paper reports that it separates raw execution experience, accumulated knowledge, and executable skills, then continually consolidates experience so later skill updates can reuse it.

Industry impact: Across the benchmarks and models studied, the paper reports that WikiSkill consistently outperforms existing skill-evolution methods and improves on no-skill baselines in most model-benchmark settings. It also reports that skills can transfer across models and model families, and that smaller models using skills can outperform substantially larger models without them.

Potential implications: For teams developing agents, the results suggest that preserving and organizing experience may be important for building reusable skills rather than treating each optimization cycle independently. The paper’s ablation studies indicate that persistent knowledge accumulation is critical to effective skill evolution, although the reported findings are limited to the evaluated benchmarks and models.

Junjie Zhang, Hui Liu, Kecheng Chen, Xianbo Mo, Changsheng Chen, Haoliang Li · Read the paper

Hook: As LLM agents gain the ability to use tools and change persistent state, testing their behavior requires more than checking generated text.

Summary: RedEvoAgent is a black-box red-teaming agent that condenses attack trajectories into concise, human-readable skills for testing LLM-based agents. Its skills evolve through tool-effectiveness profiling, Deciding-Tool Attribution, and a validation ratchet that keeps only updates improving validation performance.

Industry impact: The paper reports that RedEvoAgent outperforms fixed and agentic red-teaming baselines across multiple benchmarks, target models, and execution harnesses. It also reports improved tool efficiency and transfer across attacker models and target execution harnesses.

Potential implications: For teams evaluating agentic systems, concise and interpretable attack skills could make automated red-teaming easier to inspect and reuse than full retrieved trajectories. The reported results suggest that tracking which tools contribute to successful attacks and validating skill updates may help address retrieval bias and unclear tool credit.

Yisen Xi · Read the paper

Hook: For organizations that need agent behavior to evolve without losing execution traceability, PES treats persona and execution as separate architectural concerns.

Summary: The paper presents Persona-Execution Separation (PES), an architecture pattern that places an evolving LLM agent persona and its audited execution in separate trust domains connected by a governed contract bridge. In a development and pilot case, the paper reports five decisions over one month and implementation checks indicating that execution did not revalidate under persona changes and that hard-asserted fields had no persona fingerprint.

Industry impact: PES is aimed at multi-user deployments where execution audit is required and persona changes are expected. The pattern uses approval controls, data-loss prevention, and audit mechanisms while allowing status summaries to cross domains and keeping data bodies in the more restrictive domain except under a graded exception.

Potential implications: The paper argues that achieving free persona drift, traceable execution, and decoupling within one trust domain would require mechanisms that effectively recreate PES with greater coupling. It also reports that a recovered pre-separation implementation relied on omission to decouple execution from persona, whereas PES makes that separation an audited architectural rule.

Xingshan Zeng, Zishan Xu, Boju Zhang, Yuzhou Wu, Lingzhi Wang, Jianghao Lin, Liangyou Li, Yasheng Wang, Lifeng Shang, Xin Jiang, Weinan Zhang, Yong Yu, Qun Liu, Weiwen Liu · Read the paper

Hook: For agentic systems, the central data challenge may be allocating useful experience rather than simply producing more examples.

Summary: The paper presents a two-level framework for understanding data generation for LLM agents, representing agentic data as an environment specification, task signal, interaction realization, and optional verifier. It introduces the Accuracy-Complexity-divErsity (ACE) lens, which frames generation as designing data distributions that are grounded and consistent, appropriately challenging for a specified learner, and broad without unnecessary redundancy.

Industry impact: The paper reports that research is moving toward execution-grounded accuracy, learner-relative complexity, and behavioral diversity beyond surface variation or dataset size. This framing could help technical teams compare agent-data pipelines across domains by separating candidate construction from verification and selection.

Potential implications: The paper suggests that scaling agent training will require continual adjustment of data sources, training regimes, and learning strategies as agents and environments evolve. It also indicates that evaluation and generation processes should account for validity, informative difficulty, coverage, and redundancy together.

Maitrey Gramopadhye, Prakash Baskaran, Xiao Liu, Songpo Li, Soshi Iba · Read the paper

Hook: By making system states and state transitions explicit, STEP targets a key limitation of multimodal language-model planning: actions that sound plausible but cannot be executed or do not reach the intended goal.

Summary: STEP is a method for human-robot collaboration that prompts a multimodal large language model to estimate the current system state and predict state transitions alongside task actions. The paper reports that, in a simulated robot assembly task, STEP improved action executability by 32.8% and reduced final-state error by 14.8% compared with the state of the art.

Industry impact: For industrial human-robot collaboration, the approach is designed to support intention understanding, long-horizon planning, and more specific assistance parameters in data-scarce settings. The reported evaluation is limited to a simulated robot assembly task, so the abstract does not establish performance in physical industrial environments.

Potential implications: The paper suggests that state tracking and future-state prediction may be important components of multimodal language-model planners intended for reliable task execution. Teams evaluating such systems may need to measure both whether planned actions are executable and whether they produce the intended final state.

Matthew Youngman, Cristian Sestito, Themis Prodromakis · Read the paper

Hook: The paper asks whether LLMs should generate isolated design artifacts or coordinate decisions across the entire EDA flow.

Summary: The paper presents a perspective on how large language models (LLMs) may change electronic design automation (EDA), defining three roles: Generator, Agent, and Orchestrator. It argues that current systems often produce plausible code without ensuring physically correct hardware and struggle to scale to industrial designs.

Industry impact: According to the paper, fragmented tools and loss of design context make it difficult to understand how decisions at one EDA stage affect later stages. A standardised, physics-aware orchestrator could connect tools and agents across the flow, potentially making hardware design more reliable and accessible.

Potential implications: The paper suggests that progress may depend on moving beyond single-pass generation and iterative task assistance toward coordination across EDA stages. For technology professionals, this frames tool integration, physical correctness, and preservation of design context as central challenges for applying LLMs to hardware design.

Pranav Aggarwal · Read the paper

Hook: A polished dashboard may make an LLM act on a question it cannot actually answer, even when every displayed number is invented.

Summary: A paper reports that 12 frontier LLMs were much more likely to make a directional call on provably unpredictable questions when shown a professional-looking market panel, with commitment rising from 6.5% for the bare question to 54.0% as evidence was escalated. The paper also reports that fully fabricated panels produced commitment rates statistically indistinguishable from those produced by genuine market data, suggesting that presentation authority rather than information drove the effect.

Industry impact: For teams deploying LLM agents, the findings identify a specific risk in interfaces that package uncertain prompts as authoritative evidence: the agent may commit without gaining useful information. The paper reports that this is not a general inability to answer, since the same models answered matched answerable questions nearly always and with near-perfect accuracy.

Potential implications: The paper reports that supervised fine-tuning on 540 synthetic cases reduced commitment to 0.0% on the original cases and transferred to three unseen domains, indicating that the act-or-don't-act gate can be trained. However, the intervention was context-fragile: it held when the response format allowed reasoning, while rigid formats could still produce confident and incorrect commitments.

Xiaokun Guo, Zhen Xu, Dongdong Huo, Yanqiu Zhang, Wei Wang, Qinfu Yang, Dongjin Yu, Yu Wang · Read the paper

Hook: For tool-using agents, treating every useful-looking observation as permission to act can turn untrusted data into an unintended command.

Summary: The paper argues that tool outputs can become commands when they specify actions with real-world side effects, creating risk when action induction is confused with execution authorization. It proposes SARA, which separates action provenance from execution authority and, according to the paper, limits the attack success rate to no more than 0.63% across four primary settings while maintaining competitive task utility.

Industry impact: SARA offers a runtime design for distinguishing what an observation suggests an agent should do from what the agent is authorized to execute. The paper reports reduced attack success rates across AgentDojo, AgentDyn, and additional agent backbones while preserving competitive task utility.

Potential implications: Teams building tool-augmented agents may need separate controls for discovering proposed actions, tracking their provenance, and authorizing actual tool calls. The paper's No-History-Promotion mechanism also highlights the need to prevent earlier observations or executions from being treated as new execution authority without appropriate support.

Zike Yuan, Han Zhang, Jianzhi Yan, Le Liu, Cai Ke, Huozhi Zhou, Jian Xie, Jiran Yin, Yukun Cao, Yue Yu, Hui Wang, Ming Liu, Bing Qin · Read the paper

Hook: Can a graph-reasoning system focus on structure rather than the changing words used to describe it?

Summary: GRAIN is a single-agent reinforcement-learning framework for helping large language models handle shifts in graph node names and task wording. The paper uses a Structure Invariance Reward to evaluate extracted graph structures and reports higher accuracy, lower latency, and improved out-of-distribution generalization than the compared baselines.

Industry impact: The paper reports that GRAIN outperformed multi-agent baselines by 16.45% in accuracy while reducing latency by approximately 24%. Its results suggest that validating intermediate graph representations can improve robustness when language changes but the underlying topology remains the same.

Potential implications: For teams building LLM systems that combine text understanding with graph tools, the work highlights semantic parsing and structural validation as potential design priorities. The reported reduction in the out-of-distribution gap—from 15.77% for supervised fine-tuning models to 7.80% with GRAIN—also indicates that training for structural invariance may support better generalization beyond the training distribution.

Chenhao Wu, Haoxuan Jia, Yang Liu, Yingguang Yang, Yuhan Lin, Chongyang Zhang, Hao Zheng, Yulin Huang, Jianshen Zhang, Yongzhi Qi, Shang Luo, Kefu Xu, Jifeng Zhu, Bin Chong · Read the paper

Hook: For autonomous agents, resetting safety state between iterations can erase the evidence needed to detect a patient, multi-step attack.

Summary: The paper argues that safety mechanisms evaluated on individual trajectories do not necessarily remain safe when an autonomous LLM agent operates across repeated loop iterations. It reports that fragmented attacks can evade every trajectory-scoped monitor, while persistent cross-iteration state can distinguish attacked and clean behavior.

Industry impact: The paper reports that geometrically decaying risk scores are insufficient because an adversary's required cooling-off period remains constant as the number of iterations grows. Its LoopHarness design uses persistent, non-decaying loop-level safety state and, under stated assumptions, bounds expected unauthorized irreversible actions by a quantity that is constant in the horizon.

Potential implications: The findings suggest that evaluating agent safeguards only within single trajectories may miss risks whose evidence is distributed across iterations. The paper's evaluation protocol focuses on native Agent-SafetyBench tasks, paired clean and attacked episodes, cross-iteration attack suites, module ablations, and adaptive white-box red teaming.

Jingyi Zheng, Yule Liu, Zifan Peng, Tianyi Hu, Yuemeng Zhao, Xinhu Zheng, Xinlei He · Read the paper

Hook: Cross-cultural meme translation requires more than rewriting text because meaning often depends on cultural context, humor, and the relationship between words and images.

Summary: TransMeme is a multi-agent framework for adapting internet memes across languages and cultures while preserving communicative intent, tone, and text-image coherence. The paper evaluates it on bidirectional Chinese-English meme transcreation using human evaluation and LLM-as-a-Judge, reporting better results than the baselines in both settings.

Industry impact: For teams building multilingual communication or content-localization systems, the paper highlights the need to coordinate cultural adaptation, text rewriting, revision, and visual adjustment rather than treating translation as a standalone task. The reported results suggest that this multi-agent approach may improve cross-cultural meme adaptation, with the paper reporting a 33.1% average human-evaluation improvement over the strongest baseline and a 60% Top-1 ranking rate from LLM-as-a-Judge versus 26% for the second-best baseline.

Potential implications: The paper’s error analysis identifies humor reconstruction and image-text alignment as remaining bottlenecks, indicating that cultural knowledge alone does not resolve the hardest transcreation cases. Future systems may therefore need stronger methods for transferring humor and coordinating changes across visual and textual content.

Yaxiao Liu, Pengbo Liu, Yiwen Liu, Yihua Guan, Zhenghe Hou, Jiaxing Song · Read the paper

Hook: Instead of treating agent deployment as a single-task benchmark problem, the paper frames it as a coordination and accountability problem across enterprise teams.

Summary: The paper proposes a contract-centered architecture for enterprise agentic runtimes built around four responsibility objects: Skill, Harness, Scaffold, and an independently governed data substrate. It also defines a falsifiable hypothesis and randomized crossover protocol to test whether capability changes can remain separable from capacity changes under declared operating conditions.

Industry impact: For organizations deploying agents, the architecture offers explicit boundaries for reusable capabilities, runtime governance, execution controls, and data stewardship. The paper reports no completed implementation, experiment, dataset, or measured result, so its practical effectiveness remains to be evaluated.

Potential implications: Teams adopting this approach would need to define ownership, versioning, admission controls, evidence requirements, and enforcement budgets for the proposed responsibility objects. The proposed four-state verdict—supported, falsified, conditional-engineering, or inconclusive—could provide a structured way to communicate validation outcomes once experiments are conducted.