The paper introduces Patterns of Past Rewards (PPR), a lightweight, algorithm‑agnostic online change‑point detector for cooperative multi‑agent reinforcement learning. PPR smooths agents’ return streams, highlights recent changes, and applies a statistical drift detector to flag significant shifts. Experiments in a custom Speaker‑Listener environment show that PPR balances detection speed and alarm stability, outperforming both a smoothed‑return baseline and a raw‑return detector.
By Fatemeh Saberi Khomami, Julita Vassileva
MedProb is a lightweight probing framework that predicts multiple-choice medical visual question answering (Med‑VQA) answers directly from frozen vision‑language model (VLM) representations, avoiding free‑text generation. On datasets such as PATH‑VQA, SLAKE, and VQA‑RAD, MedProb extracts more answer‑relevant signal than prompting and outperforms both medical VLMs and agentic systems. The approach also narrows the performance gap between small and large models, shows that medical adaptation does not consistently improve linear decodability, and reveals positional biases in both prompting and generation.
By Erfan Nourbakhsh, Ke Yang, Anthony Rios
Harbor Adapters is a unified evaluation infrastructure that ports over 80 agentic benchmarks, enabling arbitrary agents to be tested across complex environments. The authors performed a large‑scale evaluation of 8 models on 54 benchmarks, using Terminus‑2 and three native harnesses, revealing detailed agent capabilities and failure modes. They also created Harbor‑Index, a curated set of 82 challenging tasks from 29 benchmarks, designed to be affordable yet comprehensive, with the best model achieving a 28.0% pass rate.
By Lin Shi (Audrey), Haowei Lin (Audrey), Zixuan Zhu (Audrey), Xiaoyue Zhou (Audrey), Xiang Li (Audrey), Xiangning Lin (Audrey), Yaxuan Deng (Audrey), Han Xu (Audrey), Yuangang Li (Audrey), Shanda Li (Audrey), Zizhao Chen (Audrey), Hanwen Xing (Audrey), Harsh Raj (Audrey), Bo Chen (Audrey), Quan Shi (Audrey), Steven Dillmann (Audrey), Yipeng Gao (Audrey), Puneesh Khanna (Audrey), Ruofan Lu (Audrey), Chao Beyond Zhou (Audrey), Michael Yang (Audrey), Robert Zhang (Audrey), Siyuan Chai (Audrey), Jiayu Chang (Audrey), Yizhao Chen (Audrey), Xiaokun Chen (Audrey), Yiwei Dai (Audrey), Wenting Yang (Audrey), Hange Liu (Audrey), Minghao Liu (Audrey), Zihan Wang (Audrey), Adnan El Assadi (Audrey), Benedikt Stroebl (Audrey), E. Kelly Buchanan (Audrey), Han Meng (Audrey), Junwei He (Audrey), Longxuan Yu (Audrey), Radin Shayanfar (Audrey), Yukyung Lee (Audrey), Zhikang Dong (Audrey), Allen G Hart (Audrey), Anjiang Wei (Audrey), Anurag Kashyap (Audrey), Arpandeep Khatua (Audrey), Audrey Jixin Zheng (Audrey), Chengrui Ma (Audrey), David Heineman (Audrey), Dubing Chen (Audrey), Hai-Anh Trinh (Audrey), Haishuo Fang (Audrey), Hefan Zhang (Audrey), Hui Shen (Audrey), Issa Sugiura (Audrey), Jiankai Sun (Audrey), Jiechao Gao (Audrey), Junhong Lin (Audrey), Junnan Li (Audrey), Kai Yang (Audrey), Lei Hsiung (Audrey), Maoyu Wang (Audrey), Mengze Tang (Audrey), Nabil Omi (Audrey), Negin Raoof (Audrey), Nicholas Edwards (Audrey), Octavia Guo (Audrey), Orfeas Menis Mastromichalakis (Audrey), Pengliang Ji (Audrey), Przemys{\l}aw Hejman (Audrey), Qi Qi (Audrey), Qunshu Lin (Audrey), Richard Zhuang (Audrey), Rui Yang (Audrey), Ruichen Zheng (Audrey), Ryan Marten (Audrey), Shaghayegh Fazliani (Audrey), Shizheng Hou (Audrey), Sicong Jiang (Audrey), Sijie Li (Audrey), Song Bian (Audrey), Terry Yue Zhuo (Audrey), Tianqing Wu (Audrey), Tom Tang (Audrey), Wanjia Zhao (Audrey), Weihao Xuan (Audrey), Wenhua Liang (Audrey), Xian Liu (Audrey), Xin Lan (Audrey), Xuan Zhang (Audrey), Xuandong Zhao (Audrey), Yanchuan Tang (Audrey), Yifan Jiang (Audrey), Yijiang Li (Audrey), Yitong Guan (Audrey), Yizhi Li (Audrey), Yonghui Liu (Audrey), Yuheng Tang (Audrey), Yujun (Audrey), Mao, Yunfei Zhao, Yuxin Wang, Yuxuan Tang, Zhenheng Tang, Zhifei Li, Ziruo Wang, Ziyu She, Kaiyuan Liu, Iheb Chaabane, Yuxin Tang, Xiangyi Li, Andy Konwinski, Boxuan Li, Leon Liangyu Chen, Alex Dimakis, Nicholas Carlini, Soroush Vosoughi, Di He, Etash Guha, Benjamin Feuer, Mike Merrill, Ludwig Schmidt, Alex Shaw
The paper reviews the evolution of AI recruitment systems from simple profile matching to complex, multi‑stage workflows that retrieve evidence, compare candidates, and execute actions. It analyzes 40 representative works, highlighting transitions from similarity to reciprocal suitability, from single models to compound workflows, and from offline predictions to evidence‑aligned evaluation. The authors identify persistent gaps—such as confounded behavioral labels, limited data validity, and lack of privacy assessment—and propose a staged mapping for defensible evaluation and an agenda for auditable, evidence‑grounded systems.
By Ziyi Zhao, Guanzheng Wei
Large language models (LLMs) are increasingly used in high‑stakes real‑world systems such as financial markets. This study demonstrates that enhancing individual LLM capability can actually worsen system‑level outcomes by making models behave more similarly, leading to correlated actions that increase risk. Using an agent‑based simulation of LLM traders, the authors show that while higher capability can reduce market risk when reasoning is accurate, it can amplify risk when agents share misinformation, revealing a capability paradox.
By Jillian Ross, Eric So, Zoe De Simone, Charles Pozniak, Andrew W. Lo
MaxKernel is a multi‑agent system designed to generate high‑performance custom kernels for TPUs. It offers three paradigms: a Human‑in‑the‑Loop agent for collaborative design, an Autonomous agent that runs a fully automated optimization loop, and a Graph‑Based Autonomous Search for global exploration. All paradigms share specialized sub‑agents for planning, implementation, debugging, testing, and profiling, and the system consistently matches expert hand‑tuned baselines on the JaxBench suite and real‑world workloads.
By Shangkun Wang, Nina Cai, Charles Hoong, Julian Walker, Gerson Kroiz, George Vanica, Deepak Patil, Andi Gavrilescu, Hassan Sipra, Sethu Sankaran
SiLR introduces a structure‑preserving admission and process reward mechanism for large language model (LLM) tool agents. Unlike traditional scalar‑score gates that can trap agents in plateau trajectories, SiLR shadow‑executes each proposal and admits it based on a product order over branch‑level violation states, ensuring safe and recoverable actions. Experiments on Gym‑ANM and CityLearn benchmarks show SiLR consistently recovers all multi‑action episodes and outperforms scalar gates, while also providing a robust reward signal for policy learning.
By Chenyu Zhou, Qiliang Jiang, Shuning Wu, Xu Zhou
The paper introduces $ au^ au$-Bench, a benchmark that turns the construction of AI agents into a measurable task. In this environment a developer agent receives real business records, client requirements, a production API, an existing codebase, and constraints on cost and models, and must deliver a complete customer‑service agent. The benchmark evaluates performance by deploying the agent against simulated users, revealing that current state‑of‑the‑art models achieve only 23.9% success while an expert‑written reference scores 82.2%.
By Quan Shi, Keshav Dhandhania, Karthik Narasimhan, Victor Barres
ERPBench is a benchmark that evaluates large language model agents in enterprise decision-making through a six‑round ERP simulation covering pricing, production, procurement, inventory, finance, and market competition. It tests the same 100 problems in two market ecologies—Solo, where agents compete against rule‑based opponents, and Arena, where six agents compete together—producing 1,200 model trajectories across 7,200 decision rounds. Results show that model performance varies by ecology, with DeepSeek best in Solo and Gemini best in Arena, and only 21 of 100 problems yield the same top performer across both settings.
By Xinran Zhang, Pengrui Lu, Lyumanshan Ye, Pengfei Liu
The study examined biomedical research articles from 2022 to March 2026 that employed large language models (LLMs). It found that 42% of the most frequently used models were already retired or scheduled to retire within two years of publication, with a median retirement interval of 538 days. This high rate of model deprecation threatens the reproducibility of biomedical AI research.
By Nathan Wolfrath, Meghan Conroy, Thomas Kosten, Dave Bell, Bhabishya Neupane, Jonah Kindel, Anjishnu Banerjee, Priya Deshpande, Bradley Taylor, Anai N. Kothari
The paper introduces the DevRev NL2SQL benchmark, featuring 900 execution‑verified queries that test natural‑language‑to‑SQL systems on nested, graph‑like enterprise schemas, and proposes the Semantic Depth Score (SDS) as a rubric for analytical reasoning depth. It also presents a cost‑aware, single‑generation agentic architecture that includes schema selection, metadata retrieval, and error‑repair components tailored to these complex schemas. On the DevRev benchmark, the system achieves 91.7% answer correctness, outperforming the next‑best baseline by 54.6 percentage points, and remains competitive on the Spider 2.0 Snowflake dataset.
By Yoga Sri Varshan Varadharajan, Ajay Yadav, Ritesh Goru, Prateek Chaudhury, Constantine Caramanis, Prateek Jain, Divyateja Pasupuleti, Sunil Kumar Pandey
SQL-Zero is a self‑playing framework that trains a Text‑to‑SQL agent without any human‑annotated natural‑language/SQL pairs. It uses a challenger and a solver, both based on the same large language model, to generate SQL queries that are only validated by execution against the database. On the BIRD dataset, self‑play improves performance over a zero‑shot baseline by 6.6 to 7.3 points and shows promising transfer to unseen Spider databases, especially at 3B parameter scale.
By Daniel Machado Pedrozo, Julia Soares Dollis, Bryan Lincoln Marques de Oliveira, Vinicius Alboneti Aguiar, S\'avio Salvarino Teles de Oliveira, Telma Woerle de Lima Soares
ElderBench is a new benchmark designed to evaluate autonomous mobile agents in scenarios that reflect the natural language patterns of older adults. It comprises 249 smartphone tasks collected from older users across 20 applications, highlighting linguistic differences such as indirect speech, referential ambiguity, and under-specified requests compared to existing GUI benchmarks. The study shows that mainstream GUI agents and Vision‑Language Models perform significantly worse on these elderly‑oriented instructions, and it offers insights into how specific language patterns lead to failures.
By Weide Zhan, Qumu Shaqu, Yuanqing Liu, Peng Zhang, Jiahao Liu, Kam Him Lam, Ning Gu, Zhan Hu, Tun Lu
The paper introduces a hybrid framework that uses large language models (LLMs) to assist in designing behavioural and scenario specifications for an agent‑based model of solar photovoltaic adoption by Irish dairy farms. It integrates bounded behavioural rubrics—conservative, balanced, and optimistic—with structured scenario specifications into a calibrated ABM, preserving the original techno‑economic adoption mechanism while adding controlled behavioural modulation and scenario‑driven uncertainty analysis. Experiments across various policy settings and Monte Carlo simulations show stable, economically plausible outcomes, with up to a 13% increase in behavioural adoption compared to a logistic baseline, without causing unrealistic saturation dynamics.
By Iias Faiud, Hossein Khaleghy, Michael Schukat, Karl Mason
The paper introduces RSM‑full, an online clustered‑memory pipeline that uses a cosine‑gated max‑member merge rule and an atom‑aware grouped context packer to improve compact‑memory performance for large language model deployments. On the AMA‑Bench benchmark, RSM‑full achieves 83 % of full‑context quality while using only 32 % of the token cost at a 4 k budget, outperforming streaming‑clustered baselines by 3.5–6.0 pp. Similar gains are observed on the RealMem persona‑memory benchmark, where RSM‑full surpasses Budget‑RAG, matches BM25‑RAG, and outperforms Streaming‑Proto and A‑MEM.
whyItMatters":"RSM‑full demonstrates that careful design of memory merging and context packing can substantially reduce token costs while maintaining high quality, addressing practical constraints of latency, cost, and context limits in long‑horizon LLM deployments."
By Jiahe Geng, Jinpeng Wang, Kun Yuan
Artificial Intelligence now underpins investment workflows from data and prediction to execution and tool use, yet its technical prowess does not automatically translate into profitability. A comprehensive review of public research up to 31 August 2026 across equities, ETFs, crypto spot, perpetual futures, and on‑chain markets shows real progress in prediction, text processing, portfolio design, and workflow integration, but evidence for durable net performance remains thin. The study highlights that factors such as temporal contamination, survivorship bias, weak benchmarks, implementation costs, and venue mechanics can erode alpha, and no single AI architecture has proven to deliver persistent, cross‑regime, capacity‑aware net alpha.
"whyItMatters":"The findings underscore that while AI advances are evident, investors must rigorously test and govern AI systems to avoid overestimating their profitability potential."
By Linsen Zhu, Mengqing Cai
arXiv:2609. 05019v1 Announce Type: new Abstract: Agents tend to optimize, select, or constrain execution structures before decisive runtime outcomes are observed.
By Tianxing Wang, Mingming Zhao, Shuai Huang, Huiyang Xu, Chaoyue Niu, Shengzhong Liu, Fan Wu
RISE (Recursive Improvement via Self-Extrapolating Policy Distillation) is a new method that builds a synthetic teacher from a language model’s own RLVR training trajectory. By extrapolating the displacement between the current checkpoint and a trailing anchor in parameter or logit space, RISE transforms sparse outcome-based updates into dense token-level targets without external models or privileged conditioning. The approach recursively refines the student model, combining RLVR and on‑policy distillation, and demonstrates superior performance across mathematical reasoning, STEM, code generation, and multi‑turn agentic tasks.
By Yang Li, Semih Yavuz, Shafiq Joty
The study examines whether agents in multi‑agent systems can be swapped without affecting performance, assuming interchangeability. Eight independently formed teams of agents were created from a single base model, and agents were traded between teams while tracking task scores and communication efficiency. Swapping agents had minimal impact on task scores but increased communication costs by 16–63 % and revealed that agents are more interchangeable in outcomes than in coordination, with larger effects after longer formation histories.
By Jianxin Gao, Tianyi Yu, Linna Deng, Runze Li, Zining Wang
arXiv:2609. 05339v1 Announce Type: new Abstract: Model upgrades are routine; memory migrations are not.
By Ankit Goyal, Jaideep Ray