Tool use, function calling, orchestration and the protocols that let models act rather than only answer.
Vision-language-model-based embodied agents can complete instructed tasks but often violate safety constraints in the process, a problem recently framed as interactive safety. Training such agents to act safely is difficult, since safety and task success are distinct objectives, and safety arises only at a small number of safety-critical steps within a trajectory.
Long-term multi-agent systems continuously accumulate the memories produced by different agents. Existing memory methods typically treat retrieved memories as independent evidence and combine them through voting or weighting.
The paper introduces the Scientific Data Skill (SciDSK), an agent‑ready representation that packages dataset‑specific knowledge and operational guidance as a reusable skill. SciDSK integrates dataset descriptions, scientific context, file organization, usage procedures, quality checks, and provenance information while keeping the data in its original repository. The authors define a structured specification, build a construction pipeline, and launch the Scientific Data Skill Bank to publish SciDSK resources across six scientific disciplines, demonstrating improved agent‑driven dataset discovery and interpretation through evaluation benchmarks.
The review examines how Large Language Models (LLMs) enhance nanophotonics design by providing semantic interfaces, code generation, and tool orchestration. It traces the evolution from classical neural networks to transformer-based models and categorizes LLM applications into surrogate models that map structure to spectrum and agentic systems that generate code and orchestrate simulations for closed-loop optimization. The article also highlights potential cross-disciplinary uses of LLMs in materials science and wireless communications, and envisions future multimodal foundation models that actively collaborate in autonomous scientific discovery.
FM‑Bench is a new benchmark that tests large language model agents on long‑horizon decision making by having them run a football club for 20 in‑game years. The agent must manage a squad, trade players, negotiate contracts, invest in facilities and youth, set lineups, and respond to a board that can fire it, all using 26 tools and roughly 340–400 decision stops, with a deterministic engine producing a final score without human or LLM judges. The benchmark includes a solo track where each of 15 frontier models competes against a frozen scripted world, and an Arena track where the same models plus a scripted anchor share one 20‑year world, allowing the first head‑to‑head evaluation at this scale. whyItMatters":"FM‑Bench provides a rigorous, large‑scale test of sustained, cumulative decision‑making in language‑model agents, revealing that managerial strategy—not computational scale or vendor—drives performance over long horizons."
The paper studies optimal placement of millimeter-wave base stations in a realistic, non-convex campus layout using deep reinforcement learning. It compares four DRL methods—single-agent DQN, multi-agent partitioned DQN, single-agent DDPG, and multi-agent partitioned DDPG—and finds that the multi-agent DDPG approach achieves full coverage, a Jain's fairness index of 0.94, and superior performance in dense scenarios with 400 users. The multi-agent DDPG also converges more efficiently than single-agent methods.
The paper proposes a systematic framework for creating a "Map of Datasets in Engineering Design and Systems Engineering" (EDSE) to address the fragmented and inaccessible nature of existing datasets. It introduces a multi‑dimensional taxonomy that classifies datasets by domain, lifecycle stage, data type, and format, and presents an interactive discovery tool built on a knowledge graph data model. The authors analyze the current data landscape, identify underrepresented areas such as early‑stage design and system architecture, and suggest strategies for curation and sustainability to build a dynamic, community‑driven resource.
The paper argues that many failures in large language model–based multi‑agent systems stem from concurrency control issues rather than coordination or communication problems. It highlights how concurrent reads and writes to shared state, coupled with long inference windows, lead to stale data, lost updates, and inconsistent outcomes. The authors propose that MAS frameworks incorporate explicit concurrency controls—such as conflict detection, isolation guarantees, and structured resource access—to make concurrency a primary design concern.
The paper introduces rEDMRec, a method that compresses a large language model’s reasoning about user preferences and item comparisons into a compact, editable memory. This memory, organized into four channels—long‑term preference, short‑term context, item perception, and counterfactual hard‑negative comparisons—can be updated by an LLM controller and queried by a lightweight student LLM for ranking, eliminating the need to re‑run the expensive teacher model for each request. Experiments on ML‑1M, Amazon Beauty, and Steam datasets show that rEDMRec consistently outperforms zero‑shot, few‑shot, RAG, and GraphRAG baselines, achieving up to a 13.3% improvement in HR@1 on ML‑1M.
The paper introduces the Environment State-Text Injection (ESTI) attack, a novel method that manipulates the textual representation of environment states in large language model‑driven embodied agents without altering user instructions, model parameters, or executors. ESTI re‑frames adversarial goals as false state evidence that aligns with the current environment, thereby influencing both planning and execution through object properties, spatial relations, affordances, task‑stage rules, and execution feedback. The authors also present ESTI‑Bench, a benchmark that evaluates attack propagation across the planning‑to‑execution closed loop, and demonstrate that ESTI outperforms existing baselines on multiple embodied task datasets, achieving up to 89.32% higher planning‑level and 43.69% higher execution‑level attack success rates.
Brain2Qwerty v2 is a model that decodes natural sentences from real‑time magnetoencephalography (MEG) recordings, achieving an average word error rate of 39% across 22,000 sentences typed by nine subjects. The model uses character, word, and sentence‑level representations and shows that decoding accuracy improves log‑linearly with more data, narrowing the gap to intracranial brain‑computer interfaces. AI contributes by replacing hand‑crafted event detection with deep learning, fine‑tuning large language models for semantic extraction, and employing AI agents to refine the decoding pipeline through automated code development.
The paper introduces Verifiable Latent Alignments (VLA), a framework that monitors and steers hidden communication channels between language‑model agents. VLA links private latent states to public actions via event identifiers, enabling causal analysis. Experiments on a multi‑agent auction benchmark show high detection accuracy and effective mitigation of collusion, even without training on attack examples.
The paper "Jailbreaking in the Haystack" introduces NINJA, a jailbreak technique that exploits long-context language models by appending benign, model-generated content to harmful user goals. It demonstrates that the position of harmful goals within the context is crucial for safety, and shows that NINJA significantly boosts attack success rates on models such as LLaMA, Qwen, Mistral, and Gemini. Unlike previous methods, NINJA is low-resource, transferable, less detectable, and compute‑optimal, revealing that carefully crafted benign long contexts can expose fundamental vulnerabilities in modern LMs.
Redakto is a new tool designed to anonymize text before it is processed by large language models (LLMs). It offers state‑of‑the‑art redaction of personally identifiable information (PII) and pseudonymization, accessible via a web interface, REST APIs, and model context protocol hooks. The authors evaluate its performance on legal and medical datasets, showing that anonymized texts retain utility comparable to the originals, enabling LLM tasks without significant loss of effectiveness.
The Perspective reviews the rapid growth of agentic AI systems in computational chemistry, noting an increase from a handful in 2024 to about fifty by August 2026. These systems are evolving from assisting with specific tasks to autonomously designing, executing, and analyzing in‑silico experiments, even drafting manuscripts. While fully autonomous AI scientists are not yet realized and human oversight remains, the trend toward commoditized generalist agents suggests a future where specialized systems may become obsolete, prompting reflection on the field’s direction and priorities.
The paper introduces Interval POMDP Shielding for agents with imperfect perception, aiming to prevent unsafe actions when sensor readings may be misclassified. By estimating perception uncertainty from finite labeled data, the authors construct confidence intervals and model the system as a finite Interval Partially Observable Markov Decision Process. They propose an algorithm that computes a conservative belief set, enabling a runtime shield that guarantees, with high probability, that any action allowed by the shield meets a specified safety lower bound. Experiments on four case studies demonstrate that this shielding approach outperforms state‑of‑the‑art baselines in safety.
TrojanGYM is an LLM‑driven framework that automatically generates diverse hardware Trojan (HT) insertions to expose blind spots in learning‑based detectors. It uses multiple large language models to propose and refine RTL modifications, while an agentic loop with syntactic checks, functional verification, and GNN‑based detectors iteratively improves the HT designs. The authors also present Robust‑GNN4TJ, a more robust detector that improves detection rates on TrojanGYM benchmarks, and demonstrate that TrojanGYM can achieve up to 68.75% evasion against modern GNN detectors on SRAM, AES‑128, UART, and RISC‑V RTL designs.
Eureka is a task‑conditioned Meta‑Agent architecture that transforms long‑horizon scientific tasks into dynamic obligation graphs with explicit acceptance semantics. During execution it constructs Macro‑Agents equipped with specialized state, memory, operators, tools, verifiers, and local topology, using receding‑horizon planning, architecture promotion, and minimal‑sufficient compilation. The system demonstrates strong empirical performance, completing all 170 recursive tasks, generating 3,948 certificates without false acceptances, and achieving significant reductions in input size, recomputation, and consistent serialization across 16,000 concurrent executions.
The paper investigates the limitations of post-training AI agents that can autonomously train large language models. It distinguishes between execution-level capability—making adjustments within a chosen training strategy—and strategy-level capability—revising the overall approach based on new evidence. Analysis of many public post-training runs shows that agents lock into a strategy early and then only perform local tweaks, regardless of task. Experiments with experience scaffolds, human guidance, and extra compute improve execution but do not enable strategy reevaluation, indicating that agents lack a mechanism to spontaneously reassess their strategy during training.
ChainWorld is a new benchmark that composes atomic OSWorld tasks into long‑horizon desktop workloads, creating 347 chains of length two to four that compare two renderings of the same task sequence. In single‑turn evaluation all tasks are presented together in one prompt, while in multi‑turn evaluation tasks are revealed one at a time. Across four current computer‑use agents, maximum chain completion is 31 %, with multi‑turn evaluation improving completion for three models but both protocols remaining challenging and exposing different failure profiles.