Microsoft and Shanghai Jiao Tong University's Argus runtime enables AI agents to conduct autonomous research for days without dense human feedback, using an evidence-driven multi-agent loop across 27 campaigns and 1,548 hours, delivering verified results in AI systems, mathematics, chip design, and scientific discovery.
A joint survey from UIUC, Google, Stanford, UC Berkeley, and CMU introduces the 6A framework (Awareness, Anticipation, Agenda, Arbitration, Action, Adaptation) for proactive AI, analyzing how agents autonomously discover needs, decide when to act, and continuously adapt across 13 application domains while addressing evaluation and safety challenges.
Tencent's Terminal-RSI framework uses recursive self-rewrite and policy refinement phases to train a Qwen3.8-27B terminal agent that achieves 86.52% on Terminal-Bench 2.1, surpassing GPT-6 Sol and Kimi K3, while demonstrating improved long-horizon task completion through credit assignment and reward design innovations.
Renmin University introduces AEWM, an editing world model that classifies and revises noisy agent actions in real time, improving long-horizon task performance across search, terminal, and software engineering benchmarks without fabricating tool feedback.
The author evaluates Doubao-Seed-Evolving's latest upgrades across coding, agent, multimodal, and long-horizon skill execution using real-world tasks like PPT-to-website conversion, personal project management, SVG generation, 3D Rubik's cube animation, and a complex 30-minute web scraping skill, finding strong autonomous task decomposition, testing, and error recovery.
Recuris introduces a compact task‑state‑driven memory architecture and gated recursive self‑improvement, enabling agents to use and evolve memory more reliably and delivering large, consistent gains from 3B open‑source models up to frontier models such as Claude Opus 5 across multiple long‑horizon benchmarks.
The article introduces DeepSeek Harness and Alibaba’s LongHorizon‑Harness, explains their Manage‑Execute‑Audit (MEA) loop for explicit task‑state management, and shows benchmark improvements—WeaveBench up to 80.7%, OSWorld 3×, Terminal‑Bench 77.2%—while analyzing token costs, compute allocation, and case studies of failure recovery.
Researchers introduce Skill‑Entropy, a metric quantifying the difficulty of switching between reasoning skills in long‑horizon tasks, build the 558‑skill Skill²‑Bench, and show that Skill‑Entropy‑RL training dramatically improves cross‑skill performance of LLMs such as Qwen3, closing the gap observed in standard benchmarks.
GLM‑5.2 replaces the Group Relative Policy Optimization (GRPO) algorithm with a critic‑based PPO approach for long‑horizon tasks, arguing that GRPO’s group comparison breaks down on variable‑length trajectories, a shift that has sparked vigorous debate across the reinforcement‑learning community.
The Deli AutoResearch SKILL, now open‑sourced, presents a three‑layer framework that tackles cognitive loops, stalling, and runtime fragility in long‑horizon tasks by persisting state, detecting stalls, and using a heartbeat watchdog, and it includes a paper‑writing skill with self‑play experiments that achieve self‑rated scores up to 8.6.
SaaS‑Bench evaluates AI agents on 23 real SaaS applications and 106 cross‑app, long‑horizon tasks, revealing that even the strongest model, Claude Opus 4.7, passes fewer than four percent of tasks and exposing four structural failure modes that separate benchmark scores from true office productivity.
HiF-VLA introduces a motion‑centric bidirectional spatiotemporal reasoning framework with a joint‑expert module that simultaneously predicts future visual motion and generates high‑precision action sequences, eliminating visual redundancy, cutting inference latency and memory usage, and achieving superior success rates on long‑horizon benchmarks such as CALVIN and LIBERO‑LONG.
Stanford, Berkeley and Nvidia introduce LLM‑as‑a‑Verifier, a verification framework that scales verification compute, uses fine‑grained score tokens, repeated checks and criteria decomposition to boost agent performance, eliminate scoring ties and achieve SOTA results on Terminal‑Bench, surpassing Claude Mythos and GPT‑5.5 while improving safety in long‑horizon tasks.
The MANSION framework creates a highly realistic, multi‑scene simulation that lets robots train for long‑duration, cross‑environment tasks, dramatically cutting real‑world trial costs and narrowing the sim‑to‑real gap for embodied intelligence.
LongHorizonUI tackles the steep success‑rate drop of GUI agents on tasks longer than 10‑15 steps by introducing three tightly coupled modules—enhanced perception, deep reflective decision, and compensatory execution—and validates the approach on the new LongGUIBench benchmark with consistent performance gains across both app and game scenarios.
