long-horizon-agents

该论文提出 RECAP 框架，通过匹配上下文干预、执行关系预测与独立策略认证，研究长程智能体何时应保留、压缩或删除历史失败记录，在 Hunyuan-A13B 上以匹配的续跑成功率将平均提示成本降低 2.13%。

LoopX addresses the hard engineering problem of whether a remote side has executed after a tool call timeout in long-horizon agents. It achieves idempotent recovery through stable operation/Job identities, reading back external facts before deciding how to recover, and compositional verification — and has open-sourced this local-first control plane for long-horizon agents.

本文提出一种逐步分解的评测方法，将语言模型在400步长程任务中的失败分为重复识别、局部处理与计数器追踪三类，发现重复识别错误随步数上升，而无关上下文注入会导致计数器追踪在所有种子中失败，外部精确提醒可恢复其对齐。

Interlock 将长时程任务合成为类型化完成图，通过 agent-as-judge 验证与自训练提升复杂环境下 LLM 智能体的任务完成率，在 AutomationBench、JobBench 和 Harvey LAB 上带来 3–7 个百分点的提升。

BizBench 推出一个多轮基准，基于一家领先支付公司的真实业务流程评估长时程智能体的中间执行有效性，结果显示前沿模型即使达成最终目标也频繁违反关键约束。

A paper examining a single-rollout asynchronous RL recipe combining the Muon optimizer, token-averaged loss, truncated importance sampling, and auxiliary penalties for training long-horizon repository-level issue resolution agents, showing improved training success over baselines across two model families.

论文提出 CarryOn，一种面向长时程智能体的对象导向上下文运行时，让智能体从可复用的任务进度继续工作而非反复压缩交互历史。在 OSWorld 2.0 和 BrowseComp-Plus 上分别提升 5.8 和 7.0 个百分点，同时 token 用量减少 51.9% 和 22.9%。

BELIEFWEAVER 是一个面向长时程智能体的失败恢复框架，通过维护关系型信念状态与隐式世界模型（IWM）评估修复假设，并将推理能力蒸馏至 9B 学生模型，在三个终端基准上显著提升最终得分并降低 54.8% 的教师 API 成本。

WorkRover is a framework for training long-horizon cowork agents through recursive environment evolution, introducing Factorial Compatibility Replay and Verifier-to-Turn Credit Assignment to improve training signals. Post-training Qwen3.5-35B-A3B into WorkRover-35B-A3B yields large gains on JobBench, OfficeQAPro, and GDPvalv2 benchmarks.

本文提出Tev-Gym——首个具有可验证、密集渐进奖励的长周期智能体训练数据集（7个领域1170个任务），并基于进度-势能奖励塑形（PPRS）训练出Tev-27B，在Agents' Last Exam计算与数学科学领域匹配GPT-6 Astra和Claude Opus 5的Pass@5表现，并可泛化至Terminal-Bench 3.0/4.0。

Meta Superintelligence Labs的研究者提出GitSwarm，一个通过共享Git仓库进行结构化持久记忆的异步多智能体系统，实现"compounding inference"，使中间成果可被后续计算继承、扩展与组合；在IMOProofBench-Advanced上一次运行解出全部30题，在ProgramBench上取得79.4%的成绩。

Remory 提出一种神经记忆网络，用有限的软记忆 token 补充文本摘要，使冻结的 LLM 在仅用 5.2% 输入位置的情况下逼近全上下文表现，并在 BrowseComp 和 Terminal-Bench 2.1 等长时程 agent 基准上带来一致提升、显著减少重复工具输出和错误。

APEX-Harm 是一个评估长时程 AI 代理在共享专业工作环境中遭受注入攻击脆弱性的基准，涵盖金融、咨询和法律任务，结果显示脚本类攻击成功率高达 73-84%，表明任务输出正确并不等于行为安全。

ProgramGym introduces a framework with Program-Env and Program-Agent for building scalable, verifiable training environments for long-horizon program reproduction agents, achieving notable performance gains when fine-tuning Qwen3.8-27B on the generated environments.

Opera is a verbal critic framework for long-horizon coding agents that treats each correction as a persistent note, tracking agent actions to distinguish mere compliance from actual resolution. It improves resolver rates by up to 15 percentage points across benchmarks and provides on-policy training data that boosts Qwen3.5-9B's SWE-Bench Pro performance without needing a critic at inference time.

REMORY introduces a neural memory network that supplements textual summaries with a bounded sequence of soft memory tokens for context compaction in long-horizon agents, enabling a frozen LLM to approach full-context performance using only 5.2% of input positions with consistent gains on Qwen3.8-27B and GLM-5.3-Flash.

The LoopX team shares research notes on running EdgeBench with astra as supervisor and 6.1-sol as executor, revealing evaluator-overfitting issues, benchmark design flaws, and structural requirements for long-range agent feedback and memory.

This paper introduces GHOST, a failure mode where long-horizon LLM agents violate safety constraints specified many turns earlier under benign interactions, finding an 11.5% occurrence rate on GPT-5.5, and proposes STAR-Guard, a two-layer defense combining historical safety-constraint restoration with a pre-execution audit that eliminates GHOST events in experiments.

The paper audits the shape of memory use in long-horizon language agents, finding heavy-tailed (core-tail) retrieval traces, and proposes Core-Tail World Model (CTWM), a rank-based memory controller that cuts prompt tokens by up to 24.48% on LongMemEval while preserving accuracy.

This NeurIPS 2026 workshop paper introduces Long-Transduction, a controlled diagnostic benchmark measuring how well models sustain stateful, context-dependent operations over long generations. Evaluating seven open-weight models, it finds severe degradation when scaling context length (up to 128K), varying input format, and increasing local task complexity.