The article "Hybrid AI Patterns: Thirteen Ways to Couple Language Models with Rules, and How to Choose Among Them" presents a catalogue of thirteen design patterns for integrating language models with rule-based systems. It details each pattern’s intent, motivation, applicability, participants, collaborations, consequences, implementation, known uses, and related patterns, and identifies five invariants and five failure patterns that apply to all combinations. A decision guide is provided to help select an appropriate pattern based on problem context, supported by measurements of performance and design costs.
By Siegfried Handschuh, G\"otz-Henrik Wiegand, Lorena Raichle
FilingLoop is an interactive benchmark and self‑evolving agent framework designed to evaluate and improve large language model agents in the context of corporate disclosure review and revision. It reconstructs 380 IPO review tasks from SEC correspondence, scoring agents against 6,099 requirements and auditing claims and revisions. The framework enables procedural learning across cases, converting feedback into role‑specific guidance without updating model weights, and demonstrates measurable gains in requirement fulfillment and comment resolution.
By Wentao Zhang, Yilei Zhao, Guangyan Gan, Shuo Sun, Lin William Cong
WeVisCut is a long‑form script‑to‑video system that unifies narration, editing, and tool execution into a single agent. It uses a four‑stage data pipeline that reconstructs editing intent from finished videos, generating over 130K training examples to train a 9B policy capable of coordinating production and recovering from failures. The system introduces a 100‑task benchmark with five metrics, outperforming its base model, matching Claude Opus 5 in performance while using far fewer tokens, and surpassing task‑matched native editors on the proposed metrics.
By Chenyu Mu, Wenming Weng, Qinrui Wang, Xingsong Ye, Hao Yu, Jun He, Chong Sun, Chen Li, Jing Lyu, Erkun Yang, Cheng Deng
AEM: Adaptive Entropy Modulation for Multi‑Turn Agentic Reinforcement Learning proposes a supervision‑free credit assignment technique that adjusts entropy dynamics at the response level during RL training. By aligning uncertainty estimation with the granularity of LLM agent actions and rescaling advantages based on a response‑level uncertainty proxy, AEM improves the exploration‑exploitation balance without relying on dense intermediate supervision. Experiments on ALFWorld, WebShop, and SWE‑bench‑Verified show consistent performance gains across models from 1.5B to 32B, including a 1.4% improvement in a state‑of‑the‑art software‑engineering RL framework.
By Haotian Zhao, Songlin Zhou, Yuxin Zhang, Stephen S. -T. Yau, Wenyu Zhang, Lun Tian, Tianshu Zhu, Yifeng Huang, Yucheng Zeng, Jingnan Gu, Daxiang Dong, Jianmin Wu
The paper introduces World Reward Models (WRMs), a new approach for machine‑learning agents that selects the best candidate program directly from execution results instead of predicting intermediate outcomes. WRMs, exemplified by the ForeCritic model, are trained on thousands of executed program pairs and achieve a 79.4% success rate in choosing the better program, outperforming general‑purpose LLM judges. When integrated into systems like AIDE, ML‑Master, and ForeAgent, ForeCritic reduces mean task time by 4.4–7.5× while maintaining or improving overall performance.
By Yuxiao Lee, Zherui Yang, Fan Liu, Hao Liu
The paper introduces a curriculum-driven proximal policy optimization framework for three‑dimensional autonomous rocket landing in stochastic atmospheric conditions. Learning is structured into far‑range approach, soft descent, and integrated landing stages, with techniques such as policy transfer, critic warm‑up, rehearsal, and progressive weather severity. In 100,000 evaluation episodes the policy achieved a 96.656% success rate, with performance decreasing to 88.07% under the highest weather‑severity band and small miss distances and low residual velocities at touchdown.
By Atin Chowdhury, Piyush Joshi
TreasuryQA Bench is a new benchmark that tests AI agents on real-world tasks involving searching, reading, and extracting exact figures from scanned U.S. Treasury PDFs stored on a shared drive. The benchmark presents 70 questions over 1,435 scanned documents, requiring agents to locate the relevant file via Dropbox’s Model Context Protocol, OCR the necessary pages, and return a precise numeric answer. Results from ten agents show a wide performance gap, with the best agent solving 64.8% of tasks and the worst only 16.7%, and scan quality is shown to significantly impact success rates.
By Ritiz Tambi, Sahar Shayegan, Joshua Meyer, Sun Kim, Ali Khan, Victor Shih, Mehdi Jamei, Andi Partovi
The paper introduces EasyProver, a Lean theorem prover that uses recursive agents, isolated contexts, asynchronous search, parallel tactic probing, and an elastic elaboration service. It achieves near‑saturation results on several benchmark suites, proving 669/672 PutnamBench problems, all six IMO 2026 formalizations, 597/600 NTP4VC verification conditions, and all 189 VERINA tasks. Additionally, the authors present Aegis, a pipeline that transforms competitive programming problems into harder verification tasks by generating safe Rust solutions, translating them to Lean, and producing 2,045 Rust‑Lean program pairs with 11,533 candidate specifications, enabling new benchmarks for provers and reinforcement learning.
By Debargha Ganguly, Divya Raghunathan, Dejan Jovanovi\'c, Vikash Singh, Soonho Kong, Aaron Tomb, Aman Goel, Ferhat Erata, Sam Bayless
The paper introduces Skeleton Coding, a method where developers jointly specify a program’s structure—its data types, modules, and functions—alongside behavioral tests. This approach allows the agent to first commit to a structure, narrowing the search space of possible programs, and then use tests to select the correct implementation. Empirical results across multiple models and benchmarks show that skeletons can dramatically reduce the number of required tests, improve accuracy, and lower implementation costs.
By Aditya Golatkar, Alessandro Achille, Renos Zabounidis, Vlad Sobal, Yangruibo Ding, Baishakhi Ray, Jonathan Buck, Guy Van den Broeck, Maryam Fazel, Pietro Perona, Wei Xia, Stefano Soatto
EPSBench is a benchmark that evaluates large language model agents on quarterly earnings forecasting by having them connect financial evidence to income‑statement line items for 149 firm‑quarters from 86 US companies. Agents read filings, earnings‑call transcripts, and news, estimate company‑specific drivers, and a shared accounting calculator converts these into GAAP diluted earnings per share (EPS). Six models were assessed, with three showing no significant difference from analyst consensus in point forecast error, while the others performed worse; nominal 80% prediction intervals contained reported EPS only 37–70% of the time, largely due to low operating‑income estimates.
By Yanbin Yin, Chuyu Xu, Liqin Ye, Jing Peng, Chao Zhang
The article describes a case study where a large language model (LLM) agent loop was used to optimize three Gemma‑4‑12B decode kernels for the FuriosaAI RNGD NPU, a sparsely documented Rust DSL accelerator. Over 12 days on a shared device, the loop built its own target knowledge, performed microbenchmarks, recorded compiler legality rules, and iteratively refined kernel designs through role‑based search and acceptance tests. The resulting kernels achieved a 9.293× geometric mean speedup over the official baseline, each within 1.25× of the theoretical cycle minimum for weight movement.
By Byonggon Chun
The paper reports a non‑systematic audit of 39 external sources up to August 16, 2026, and targeted updates through October 14, examining the incremental security and utility impact of consuming a file named /llms.txt while keeping ordinary retrieval constant. A 15‑tool study found no automatic requests in 45 trials, and a log study showed that 97% of domains with detected files received no recorded requests in May 2026. A pilot with 24 tasks across eight sites and 120 real model responses identified attacker‑target adoption in 21 of 24 raw adversarial responses, while the URLs‑only condition had no known adoptions; the pilot measures output changes conditioned on supplied evidence without live retrieval or external actions.
By Gangzhen Qian, Kainan Zhou, Chuhong Xu, Hang Xiao, Zhi Jing, Rong Bao, Lu Yi, Zhaoyi Li
Tracekit Box is an open‑source system that runs AI agents inside isolated Firecracker micro‑VMs with no network card, ensuring that every possible exit point—commands, outbound connections, secrets, model calls, and workspace patches—is controlled by a host‑side signer. The signer authorizes and signs each action, preventing unauthorized data leakage or code changes, and the VM’s final state is packaged into a verifiable .tkb bundle. The system has passed 13 end‑to‑end checks, 15 fuzz targets, and a demo run with 102 signed records, demonstrating robust isolation and auditability.
By Bravish Ghosh (Cygnux Labs)
The paper refines and empirically tests the Artificial Personality Model (APM), a hybrid framework that blends human‑centred and machine‑specific personality dimensions for conversational agents. Using three experiments with 395 participants, the authors operationalised the model with the Artificial Personality Questionnaire (APQ), finding that some human traits (e.g., agreeableness, conscientiousness) partially transfer while others (e.g., extraversion) do not, and identified new machine‑specific dimensions. The APQ showed promising psychometric properties, whereas the BFI‑2 performed poorly, and manipulations of artificial personality significantly affected user satisfaction, continuous use intention, and social presence.
By Alexander Dregger, Barbarella Petz, Erik Kuhn, Pia Volk, Andreas Oberweis, Stefan Hillmann, Jannik Str\"otgen
The study compares deterministic static application security testing (SAST) with AI-powered coding agents and security harnesses on the VulnBench 2.0 benchmark, which contains 108 curated vulnerabilities across 20 open‑source web applications. Nine productized configurations—including Snyk Code, Claude Code, Codex Security, and DeepSec—were evaluated under a blind protocol, with Snyk Code achieving 82.4% macro recall and Claude Code 79.3%. Together, the AI configurations identified 101 of the 108 vulnerabilities, surpassing any single tool, and also uncovered seven additional real vulnerabilities not present in the reference set, highlighting the complementary strengths of AI harnesses and deterministic SAST.
By Liran Tal, Chris Liu, Charlotte Fereday, Sebastian Roth, Brian Clark
KARMA is a training‑free framework that evolves six typed knowledge modules for industrial advertising agents while keeping model weights and harness code unchanged. It uses a deterministic Trace Verifier to map errors to candidate modules, a Critic to compare successful and failed trajectories, and an Evolver that applies atomic knowledge edits validated by a fixed gate. Evaluated on three LLM families across six e‑commerce domains, KARMA improves average hold‑out scores by 8.96 points (+29.7%) and enables transferable, inspectable knowledge adaptation.
By Xinhua Miao, Yuanzhen Rong, Weiru Zhang, Xiaoyi Zeng
The paper investigates how AI agents that generate smart contracts from plain‑English descriptions can lose intended behavior, a critical issue because smart contracts are immutable once deployed. The authors design an execution‑based evaluation that replays transactions on both the AI‑written contract and the original, measuring agreement and identifying discrepancies. They find that common grading tools miss many mistakes, and propose an elicitation‑first approach where the agent asks a short checklist of questions before coding, which significantly improves correctness across multiple AI models.
By Abhinav Goel, Agostino Capponi, Alfio Gliozzo, Chaitya Shah
The paper introduces LibStructBench, a document‑grounded benchmark designed to assess whether AI agents can accurately reconstruct sequencing‑library structures and molecular workflows from protocols. Experiments reveal that current state‑of‑the‑art agents often generate roughly correct reconstructions but struggle with exact molecular‑state representation and workflow connectivity. The study also finds that changing the underlying language model yields larger performance differences than changing agent harnesses, and that additional knowledge only modestly improves oligo‑family recovery without affecting overall workflow accuracy.
By Chi-Lam Poon, Michael Goneos, Zezheng Song, Xi Chen
VerifiedLab is a multi‑agent framework that transforms research records into active systems for laboratory experiments. It employs a team of AI scientists—principal investigator, librarian, specialists, gate, and critic—to reason over records field by field, verify claims against measured reliability, and decide whether to release or withhold information. Experiments on perovskite solar cells show that VerifiedLab can recover known physical trends, anticipate future device reports, and provide a trust metric for inherited literature.
By Yu Yao, Jiayi Dong, Yang Yang, Shuiwang Ji, Ju Li
The paper introduces an on‑policy training method for chain embedding retrievers used in multi‑hop retrieval‑augmented generation. By training the encoder on the states it actually visits—including those built on its own mistakes—the retriever learns to recover from errors and retrieve complete evidence more reliably than teacher‑forced training. Experiments show the approach outperforms graph‑based systems and a search agent, and it transfers to harder long‑chain variants of HotpotQA and 2WikiMultiHopQA without further training.
By Chenjie Qiu, Zixuan Zeng, Shaobin Shi, Bingfeng Chen, Dongsheng Luo, Boyan Xu