Tag
InternScience 发布 Sage-A1.5，一个基于 Qwen3.6-35B-A3B 微调的 35B MoE 智能体模型，主打长程科学问题求解，具备智能体编排、自我验证与开放式科学发现能力，在多项基准上与 Kimi-K3、GLM-5.3、GPT-5.6-sol 等前沿模型竞争。
ECHOES 提出一种通过递归历史压缩与特权在线蒸馏来增强长时程 GUI 智能体状态保持能力的方法，无需在部署时保留历史截图即可取得有竞争力的性能，并发布了细粒度诊断基准 ECHOESBMK。
StarSkirmish Bench v0.2 在一小时时限内让各旗舰 LLM 用 C++ 编写 StarCraft: Brood War 神族机器人并与人类选手机器人对战，GPT-6 Astra 以明显优势居首，Gemini 4 Argon、Claude Opus 5.5 和 Claude Fable 5.1 紧随其后，且结果显示人类顶尖机器人仍远超所有模型。
Google DeepMind announces Gemini 4 Argon, a frontier model built for deep reasoning across long-horizon workflows in software engineering, enterprise knowledge work, and defensive cybersecurity, initially rolling out to trusted cyber defenders via the Fairwind Program before broader availability.
BAAI releases AREX-2, a 27B long-horizon self-improving agent model based on Qwen3.8 that learns to propose, measure, reflect, and revise solutions over multiple test-time rounds, achieving strong results on coding, ML engineering, and deep research benchmarks with only 27B parameters.
DeepRewind introduces an additive control layer that models a deep-research agent's epistemic state as a typed graph and uses a prompt-based world model plus consistency monitor to block and repair premature commitments, reducing them by 59.1% on DRBench and LiveDRBench.
SAGE is a framework that mitigates exploration and compounding biases in long-horizon reasoning for large language models through algebraic sparsification and hyperbolic structural guidance, achieving significant improvements across benchmarks.
This paper introduces QART, a quantum-classical hybrid architecture that enhances long-horizon reasoning in large language models by integrating quantum encoding and optimization, showing performance improvements on several benchmarks.
ContextPilot introduces a proactive context management framework for long-horizon agentic reasoning, using fine-grained reinforcement learning with branch sampling to improve performance and efficiency in maintaining compact working contexts.
Argus is a persistent, self-evolving agentic runtime designed for long-horizon reasoning, using Manager, Planner, Engineer, and Reviewer roles with verification-gated persistence and pivoting. It demonstrates strong results across seven benchmark arenas, including ~78% on SWE-Bench Pro, while reducing token usage after runtime self-evolution.
The paper introduces PRO-LONG, which uses a programmatic, lossless memory system for AI agents, storing all past actions in a structured text log. This approach significantly improves long-horizon reasoning and performance on ARC-AGI-3 games while using fewer tokens than stronger specialized systems.
LongMedBench is a new benchmark for evaluating LLM-based medical agents on long-horizon clinical decision-making using real EHR data from MIMIC-IV. It includes 335 patients with multiple visits and proposes evaluation suites for fact-based QA, temporal reasoning, and long-horizon decision-making.
This ICML paper introduces recursive models that recursively invoke themselves to solve subtasks in isolated contexts, proving they can surpass context-bounded autoregressive models for long-horizon reasoning. Experiments on SAT solving and Go game-tree search show improved accuracy with small active contexts.
Sakana AI launches its first commercial product, Sakana Marlin, an autonomous research assistant that completes strategy work in hours by generating structured slides and detailed reports.
This paper presents SPADER, a reinforcement learning framework for multi-answer QA that uses step-wise peer advantage for credit assignment and diversity-aware exploration rewards to improve recall of long-tail entities, achieving better performance on several benchmarks.
This paper proposes SAM, a state-adaptive memory framework that dynamically manages interaction histories for long-horizon agentic reasoning, enabling intent-driven recall without retraining the backbone model. It outperforms strong baselines across multiple benchmarks like BrowseComp and HLE.
The paper proposes the Map-then-Act Paradigm (MAP), a plug-and-play framework that shifts environmental understanding before execution in interactive LLM agents, achieving consistent gains across benchmarks and enabling frontier models to surpass near-zero baseline performance in 22 of 25 game environments.
HAGE introduces a weighted multi-relational memory framework that enables query-conditioned traversal over unified relational memory graphs, improving long-horizon reasoning accuracy through adaptive memory retrieval and reinforcement learning-based optimization.