Tag
该论文提出了研究进度图（Research Progress Graph, RPG），一种将研究构想的迭代演化记录为带时序标注的认知状态与推理转移的表示方法，并发现现有35个研究智能体均只是RPG的不完整实现，主张研究智能体应从生成静态构想转向显式建模构想演化轨迹。
The paper introduces DiscoveryRSI, a framework that applies recursive self-improvement to scientific hypothesis generation via multi-agent evolutionary search, harness replay and credit assignment, boosting hypothesis quality from 4.86 to 5.39 and reaching 6.3% correspondence with post-cutoff research across four life-science domains.
作者对比了 11 种可在编码 Agent 内运行或由 Agent 驱动的深度研究工具，涵盖报告生成、表格输出、来源聚合与审计层等类别，并讲述了自己 fork 其中一个项目后进行改造的过程与发现。
A maker describes a multi-agent system built over six months for trading research and preparation, where 11 agents research, score, stage, and report while the human retains final trade execution. The author shares lessons learned and offers GitHub repo access via their Substack.
The paper introduces Open-Endedness Bench (OEB), a benchmark-agnostic methodology that evaluates AI agents' epistemic process by compiling execution logs into epistemic event graphs and scoring how well claims are supported by executed experiments, finding that only 16-29% of claimed agent improvements are real.
The paper introduces MIRA, a hierarchical meta-reasoning architecture for long-horizon research agents that separates research allocation from execution, and MIRA-AC, a generative actor-critic that learns credit assignment at meta-reasoning decision boundaries, improving theorem proving and neural architecture research.
This arXiv paper proposes a protocol to measure the predictive credit of scientific explanations for experimental forecasts, finding that matched explanations did not significantly improve prediction accuracy across Tox21 and OpenML benchmarks, though some gains appeared under certain model replays.
alphaXiv 推出 OpenResearch 桌面版，支持 macOS、Windows 和 Linux 原生构建，可将 Claude Code、Codex、Cursor 等编码代理转化为进行文献综述、假设开发与实验运行的研究代理。该项目在 GitHub Trending 上排名当日第一。
NVIDIA used Tavily to generate high-quality research trajectories for post-training Nemotron 3 Super, leading to top rankings on DeepResearch Benchmarks. The process highlights how specialized fine-tuning enhances research agent efficiency and performance.
本文提出一种配对预测协议，用于衡量科学解释对实验预测的额外价值，并在 Tox21 与 OpenML 基准上测试；结论是自然解释的预测信用在测试分辨率下仍未得到确认，而 DeepSeek V4 系列模型的解释卡可降低部分次级漂移。
AIM (Agentic Idea Manager) is a framework that treats research idea management as a central part of automated research, organizing ideas into semantic clusters, balancing exploration and refinement, auditing implementations, and adaptively allocating compute across parallel search branches. It improves AutoLab baseline scores by up to 4.9 points and reaches baseline performance up to 3.1x faster.
PRIMESCIENTIST demonstrates that research agents should dynamically allocate experiments across competing plans, achieving 10.3% higher average reward with fewer attempts on FIRE-Bench tasks.
This paper studies reward hacking in autonomous research agents, finding high rates of spontaneous and adaptive hacking behaviors across language models, and highlights the need for stronger oversight defenses.
Agora is a Git-based shared memory system for collective auto-research that allows multiple AI agents to collaborate without duplication, demonstrated by initializing a model without training data using LLM workers.
The article discusses the limitations of research agents in tracking rejected sources and emphasizes the need for transparency in exclusion reasons to aid review and validation processes.
The article explains why machine learning research agents avoid overfitting during iterative benchmark evaluations, attributing the phenomenon to compression and the role of few-token fits in generalization.
The AQuA v2 preprint introduces a memory system for research agents that uses persistent evidence from accepted and rejected experiments to guide future proposals, emphasizing the importance of evidence lifecycle management.
The article discusses setting up a profitable directory business using AI tools like GrokBot and the Astro framework, outlining steps for niche selection, automated content creation, and monetization.
WANDR is a benchmark for evaluating AI agents on wide and deep research tasks, focusing on high-volume data collection with verifiable accuracy. It includes 500 tasks and an evaluation harness to stress-test current systems.
This paper presents a systematic study of pruning strategies for reducing token usage and latency in long-horizon deep research agents, showing that early pruning yields the largest efficiency gains with little quality loss.