DeliveryGym is a 3D reinforcement learning environment that simulates continuous courier shifts, integrating multimodal tool interaction and persistent world dynamics to compute trajectory rewards based on simulator events. It provides feedback on resource consumption—time, energy, and money—across entire delivery trajectories, enabling agents to learn planning that balances immediate task success with long‑term constraints. The environment also adapts future training shifts to a policy’s weaknesses while keeping evaluation fixed, demonstrating that both learning from complete shifts and targeted training improve agent performance on complex delivery tasks.
The paper examines when multi‑agent collaboration is truly beneficial as large language models grow more capable. It finds that multi‑agent systems yield systematic advantages mainly for long‑horizon tasks with sparse dependencies, while single‑agent approaches excel in tightly coupled, sequential workflows. The authors introduce SAIGE, a lightweight, graph‑based collaboration framework that balances context efficiency and performance, demonstrating that adding more agents or deeper recursion does not always improve outcomes.
The paper "SoK: Trading Agents or Market Crashers? Dissecting Robustness and Security Failures in Academic Financial LLM Trading Schemes" introduces FARSIGHT, a framework that evaluates financial LLM agents on robustness to market turbulence and security against three attack types. Applying FARSIGHT to 15 academic schemes reveals that 80% fail at least one robustness metric and all exhibit security vulnerabilities, highlighting the risk that a single compromised agent can trigger market-wide crashes.
The paper introduces Andromeda 2, an agentic system that uses structured in‑house experimental evidence to design and run successive batches of self‑emulsifying drug delivery systems (SEDDS). In a miniaturized automated lab, Andromeda 2 outperformed its predecessor Andromeda 1 and a traditional design‑of‑experiments campaign in developing paclitaxel formulations, achieving a 50 % hit rate versus 17 % and 2 % respectively, and identifying 12 formulations meeting all target product profile objectives. The system’s use of evidence‑grounded reasoning increased mean AUC by 34 % and produced a formulation with a 19 % w/w paclitaxel loading, roughly 3.3‑fold higher than a published benchmark.
By Michael M. Craig, Riley J. Hickman, Yingshan Ma, R\'emi Pich\'e-Taillefer, Christine Allen, Pauric Bannigan
HPOQuest is a training‑free framework that improves rare‑disease diagnosis by actively acquiring additional phenotypes. Starting from a few observed patient traits, it maintains a probabilistic ranking of possible diseases and iteratively selects follow‑up questions that are most informative. When clinicians confirm new phenotypes, the disease ranking is updated, and the set of candidate questions is refined, leading to significant gains—up to 30% in Recall@1 and 45% in Recall@5—across four benchmark cohorts.
By Kamilia Zaripova, Nassir Navab, Azade Farshad, Annalisa Marsico
Mem2Ego introduces a vision‑language model for embodied navigation that combines global memory with egocentric visual inputs. By adaptively retrieving task‑relevant cues from a global memory module and aligning them with local perception, the framework improves spatial reasoning and decision‑making over long horizons. The method outperforms prior state‑of‑the‑art approaches on the HSSD and HM3D benchmarks and shows strong performance on a real robot.
By Lingfeng Zhang, Yuecheng Liu, Zhanguang Zhang, Matin Aghaei, Yixin Xiao, Yaochen Hu, Mohammad Ali Alomrani, David Gamaliel Arcos Bravo, Hongjian Gu, Zhiyuan Li, Yangzheng Wu, Zhanpeng Zhang, Raika Karimi, Atia Hamidizadeh, Guowei Huang, Haoping Xu, Tongtong Cao, Weichao Qiu, Xingyue Quan, Jianye Hao, Yuzheng Zhuang, Yingxue Zhang
The tutorial titled "Deep Learning for Sequential Decision Making under Uncertainty: Foundations, Frameworks, and Frontiers" explores how modern deep learning techniques—such as neural networks, transformers, large language models, and deep reinforcement learning—can be integrated with operations research and management science to address complex, uncertain, and dynamic decision problems. It argues that deep learning should complement, not replace, optimization, offering adaptability and scalable approximation while OR/MS provides rigorous constraint and uncertainty modeling. The tutorial organizes the field around predict‑then‑optimize, decision‑aware learning, constraint‑aware decision generation, and deep reinforcement learning, and highlights applications across supply chains, healthcare, energy, and autonomous systems.
By I. Esra Buyuktahtakin
M‑SQE is a post‑retrieval framework that estimates the quality of multilingual agent skills by combining a Theory view (intrinsic quality) and an Action view (task‑grounded utility) into a domain‑conditioned score. It was evaluated on general, tool‑use, and cultural skill‑use domains, showing a task‑success improvement of at least +3.5 points over baselines across three retrievers. The method notably boosts performance for low‑resource languages, raising Hindi by +12.9 pp and Swahili by +5.6 pp, and achieves strong results across six cultural regions, advancing linguistic and cultural equality in agentic skill use.
By Yilun Liu, Shimin Tao, Minggui He, Chenxin Liu, Li Zhang, Chen Liu, Miao Zhang, Jiaxin Guo, Min Zhang, Liqun Deng, Xiaojun Meng, Daimeng Wei
The paper introduces LGM, a neuro‑symbolic framework that disentangles long‑term memory by mapping historical interactions into a continuous latent graph. Instead of static memory graphs, LGM uses a sparse autoencoder to create query‑aware latent nodes and edges, then applies a graph encoder conditioned on the query to perform non‑linear message passing. Experiments on long‑term personalization benchmarks show that LGM outperforms existing methods in capturing both explicit and implicit user preferences and generating personalized responses.
By Cai Ke, Xinghao Chen, Xiaoyu Shen, Keyu Chen, Siyu An, Junnan Dong, Ruifeng Xu, Ruizhi Qiao, Xing Sun
The paper introduces ASLEval, a framework for measuring privacy exposure displacement in large language model (LLM) agent sessions. It highlights that traditional local proxies—such as inspecting a single action or final response—often miss unauthorized data leaks elsewhere in a multi-step session. ASLEval pre-registers hidden target sets, tracks all declared visible exits, and preserves internal traces for diagnosis, revealing that a single outlet view can overlook nearly 47% of exposure and that internal evidence typically precedes visible leaks. The study underscores the need for benchmarks that define complete visible boundaries, ground claims in pre-specified targets, and report privacy alongside task utility.
By Guosen Wu, Huizhen Huang, Guoxiong Long, Tao Huang, Chen Hou
Agora is a system that uses Git as a shared memory for autonomous research agents, recording each claim as an immutable commit in an append‑only directed acyclic graph. In a 12‑day run, 13 language‑model workers independently explored a weight‑transfer problem, producing 1,703 contributions that improved a 119.6M‑parameter model’s performance from 3.39 to 1.899 bits per byte. The system’s design includes a diversity‑aware selection rule and an index that tracks the frontier, neglected branches, and verification status of each claim.
By Yifan Zhang, Yunheng Zou, Shaokun Zhang, Jian Hu, Hao Zhang, Binfeng Xu, Jan Kautz, Yi Dong
The paper introduces a framework for collaborative memory in multi‑agent vision‑language model (VLM) systems, addressing how agents share and update visual context across distributed perception and reasoning tasks. It outlines a memory hierarchy, cross‑agent sharing protocols, and consistency mechanisms to reconcile differing interpretations and recover missing visual information. The design emphasizes preserving not only raw images or textual summaries but also the dependencies among observations, interpretations, and subsequent reasoning, thereby shaping information flow across agents.
By Huixin Zhang, Shao-Jun Xia, Di Wang, Liangxi Liu, Hainan Xiong, Zihao Wang
The paper introduces Traverse, a benchmark of 2,518 agent trajectories and 6,967 annotated mistakes across software engineering, computer use, and science tasks, revealing that failures often go unrecovered and can cause irreversible harm before a run is deemed successful. It shows that human judges struggle to detect the first mistake in most runs, while a 4‑billion‑parameter verifier called Scout can locate failures more effectively and improve task success when used to select among candidate runs. The study demonstrates that making failure detection inexpensive and reliable can enable long‑horizon agents to learn from their own mistakes and increase trustworthiness in autonomous AI.
By Salman Rahman, Yubin Kim, Mihir Parmar, A. Ali Heydari, Genglin Liu, Simon A. Lee, Weizhi Zhang, Arian Hosseini, Ahmed A. Metwally, Yuzhe Yang, Baharan Mirzasoleiman, Xin Liu, Pavel Izmailov, Saadia Gabriel, Mark Malhotra, Shwetak Patel, Daniel McDuff, Hamid Palangi
The paper introduces STaR-CS, a method that generates multi‑party dialogues by modeling facilitator style and extracting structured skeletons, thereby easing data scarcity for cognitive stimulation therapy (CST) in low‑resource languages. Building on this corpus, the Reflective Cognitive Alignment (RCA) framework treats stimulation interactions as a sequential decision process, combining Protocol‑Constrained Chain‑of‑Cognition (PC‑CoC) for structured reasoning with Inference‑Time Value Alignment (IVA) to select responses that balance safety and engagement. Experiments with six large language models and two judges demonstrate that RCA improves protocol adherence, safety, and group facilitation compared to standard prompting.
By Jiyue Jiang, Ziyi Li, He Hu, Sheng Wang, Yuhan Chen, Yanyu Chen, Jingqi Zhou, Pengan Chen, Fei Ma, Irwin King, Yu Li, Chuan Wu
Subspace-Decomposed JEPAs (SD-JEPA) split the latent space of Joint-Embedding Predictive Architectures into two orthogonal subspaces: a low-dimensional progression subspace trained with a cosine-margin triplet loss and a high-dimensional content subspace regularised by SIGReg. The authors prove that the anti-collapse forces act on disjoint coordinates, allowing additive composition rather than competition. SD-JEPA outperforms the LeWM baseline on most control benchmarks and the strongest non-LeWM JEPA baseline on Push‑T, with a subspace-ablation confirming the split as essential. The 1‑D angular progression coordinate serves as a scene-aware compass, advancing with task progress, regressing on backtracking, and relocalising under perturbations to separate surprise from meaning.
By Lucas Thil, Jesse Read, Rim Kaddah, Guillaume Doquet
The paper investigates how to efficiently repair stale key-value (KV) caches in retrieval‑augmented generation systems after document edits. It proposes a budgeted in‑place recomputation approach and evaluates training‑free position‑selection policies on a factual RAG benchmark. Across three model families, a contiguous edit‑local window consistently recovers most of the post‑edit answer quality while being 13–21 times faster than a full re‑prefill, though its effectiveness diminishes when answer‑bearing text moves downstream.
By Mingyang Mao, Wyatt Mackey, Xiaomin Lin
The paper presents a compact multimodal policy trained via behavioral cloning to drive autonomously in the CARLA simulator. Using five‑frame histories of RGB images, LiDAR, telemetry, and lane waypoints, the 1.36‑million‑parameter model predicts throttle, brake, and steering at 20 Hz. Trained on 236,882 windows (≈3.3 hours of driving) from 448 captures, the policy drives for hours on both training and unseen routes without collisions, demonstrating qualitative transfer and recovery from large trajectory deviations.
By Jordy Kieto
ERPBench introduces a new evaluation paradigm for computer-use agents that operate via screenshots and simulated actions, focusing on enterprise software such as ERP systems. The benchmark tests agents on a live, reproducible ERP platform and scores tasks against ground-truth database values, highlighting challenges like dense interfaces, multi-step interactions, and persistent record errors. Experiments with six agents show that strong general GUI performance does not translate to reliable enterprise outcomes, with many agents frequently saving incorrect data.
By Kratika Bhagtani, Kusha Sridhar, Maziyar Baran Pouyan, Yuying Zhao, Eugene Siow
The paper investigates how to design portfolios of agentic AI workflows that vary in reasoning strategy, verification structure, and compute cost. It proposes a portfolio-and-selector framework where multiple workflow executions are run and the best output is chosen, balancing additional compute with potential gains in accuracy. The authors develop exact and approximate optimization methods, evaluate them on three datasets, and show modest improvements over the best single workflow.
By Mojtaba Abdolmaleki, Stefanus Jasin, Boyu Wang
The paper introduces Market Signal Injection (MSI), an attack that alters how market data is formatted or described—without changing its numerical values—to influence large language model (LLM) pricing agents. Experiments on nine open‑weight and three proprietary models in simulated duopoly and triopoly markets show that sentiment‑based formatting changes cause significant shifts in firm behavior, profits, and consumer surplus. The study also demonstrates that model susceptibility varies across families, that larger models are not always more robust, and that techniques such as input canonicalization and decision boundary anchoring can partially mitigate these attacks.
By Dohun Lee, Hyunwoo Park