Research Digest: April 12, 2026

This digest covers 5 significant papers from April 2026 focusing on autonomous AI research agents, multi-agent narrative systems, self-supervised search agent training, LLM-powered heuristic design, and cognitive bias analysis in LLMs. The standout finding is AiScientist, which demonstrates that long-horizon ML research engineering is fundamentally a systems problem of coordinating specialized work over durable project state, achieving 81.82% Any Medal rate on MLE-Bench Lite.

arXiv ID: 2604.13018 (Submitted: April 14, 2026) ✓ VERIFIED

Title: Toward Autonomous Long-Horizon Engineering for ML Research

Authors: Guoxin Chen, Jie Chen, Lei Chen, Jiale Zhao, Fanzhe Meng, Wayne Xin Zhao, Ruihua Song, Cheng Chen, Ji-Rong Wen, Kai Jia

Core Method: AiScientist introduces a hierarchical orchestration system with a "File-as-Bus" workspace protocol for autonomous ML research. The architecture uses:

• ●A top-level Orchestrator maintaining stage-level control through concise summaries and workspace maps
• ●Specialized agents that re-ground on durable artifacts (analyses, plans, code, experimental evidence)
• ●Permission-scoped workspace enabling "thin control over thick state"

Key Findings:

• ●Improves PaperBench score by 10.54 points over best baseline
• ●Achieves 81.82% Any Medal rate on MLE-Bench Lite
• ●File-as-Bus protocol is critical: removing it reduces PaperBench by 6.41 points and MLE-Bench Lite by 31.82 points
• ●Demonstrates that long-horizon ML research engineering is a systems problem of coordinating specialized work over durable project state

Applicable Scenarios:

• ●Automated ML research and experimentation
• ●Long-horizon software engineering tasks
• ●Multi-step scientific discovery pipelines
• ●Autonomous code generation and debugging systems

Original Link: https://arxiv.org/abs/2604.13018

arXiv ID: 2604.12776 (Submitted: April 14, 2026) ✓ VERIFIED

Title: EvoSpark: Endogenous Interactive Agent Societies for Unified Long-Horizon Narrative Evolution

Authors: Shiyu He, Minchi Kuang, Mengxian Wang, Bin Hu, Tingxiang Gu

Core Method: EvoSpark addresses two key challenges in multi-agent narrative systems:

1. ●Social memory stacking - conflicting relational states accumulate without resolution
2. ●Narrative-spatial dissonance - spatial logic detaches from evolving plot

The framework introduces:

• ●Stratified Narrative Memory with Role Socio-Evolutionary Base as living cognition
• ●Generative Mise-en-Scène mechanism enforcing Role-Location-Plot alignment
• ●Unified Narrative Operation Engine with Emergent Character Grounding Protocol

Key Findings:

• ●Significantly outperforms baselines across diverse paradigms
• ●Enables sustained generation of expressive and coherent narrative experiences
• ●Transforms stochastic sparking into persistent characters
• ●Accepted to ACL 2026 Main Conference

Applicable Scenarios:

• ●Interactive fiction and narrative generation
• ●Multi-agent simulation environments
• ●Virtual world building and character development
• ●Social simulation for research and entertainment

Original Link: https://arxiv.org/abs/2604.12776

arXiv ID: 2604.12967 (Submitted: April 14, 2026) ✓ VERIFIED

Title: Cycle-Consistent Search: Question Reconstructability as a Proxy Reward for Search Agent Training

Authors: Sohyun An (Meta Superintelligence Labs, UCLA), Shuibenyang Yuan, Hayeon Lee, Cho-Jui Hsieh, Alexander Min

Core Method: CCS provides a gold-supervision-free framework for training search agents using cycle-consistency principles:

• ●Key hypothesis: optimal search trajectory serves as lossless encoding of question intent
• ●High-quality trajectory should preserve information to reconstruct original question
• ●Information bottlenecks prevent leakage: exclusion of final response + NER masking of queries
• ●Forces reconstruction to rely on retrieved observations and structural scaffold

Key Findings:

• ●Achieves performance comparable to supervised baselines
• ●Outperforms prior methods without gold supervision
• ●Provides scalable training paradigm for settings where gold supervision is unavailable
• ●Demonstrates that informational adequacy can be measured without ground-truth answers

Applicable Scenarios:

• ●Information retrieval agent training
• ●Question-answering systems with limited labeled data
• ●Search agents for specialized domains
• ●Scalable RL for complex retrieval tasks

Original Link: https://arxiv.org/abs/2604.12967

arXiv ID: 2604.12898 (Submitted: April 14, 2026) ✓ VERIFIED

Title: BEAM: Bi-level Memory-adaptive Algorithmic Evolution for LLM-Powered Heuristic Design

Authors: Chuyang Xiang, Yichen Wei, Jiale Ma, Handing Wang, Junchi Yan

Core Method: BEAM reformulates heuristic design as a Bi-level Optimization problem:

• ●Exterior layer: Evolves high-level algorithmic structures with function placeholders via Genetic Algorithm (GA)
• ●Interior layer: Realizes placeholders via Monte Carlo Tree Search (MCTS)
• ●Adaptive Memory module: Facilitates complex code generation
• ●Knowledge Augmentation (KA) Pipeline: Addresses limitations of starting from scratch or templates

Key Findings:

• ●Reduces optimality gap by 37.84% on aggregate in CVRP hybrid algorithm design
• ●Designs heuristic outperforming SOTA Maximum Independent Set (MIS) solver KaMIS
• ●Significantly outperforms existing LLM-based Hyper Heuristic (LHH) methods
• ●Addresses single-layer evolution limitations of prior LHH approaches

Applicable Scenarios:

• ●Automatic algorithm design for combinatorial optimization
• ●Vehicle routing and scheduling problems
• ●Graph algorithm optimization
• ●Meta-learning for heuristic generation

Original Link: https://arxiv.org/abs/2604.12898

arXiv ID: 2604.12816 (Submitted: April 14, 2026) ✓ VERIFIED

Title: The role of System 1 and System 2 semantic memory structure in human and LLM biases

Authors: Katherine Abramski, Giulio Rossetti, Massimo Stella

Core Method: Models Systems 1 (associative) and 2 (deliberative) thinking as semantic memory networks:

• ●Built from comparable datasets generated by humans and LLMs
• ●Network-based evaluation metrics for implicit gender bias
• ●Investigates irreducibility of semantic memory structures

Key Findings:

• ●Semantic memory structures are irreducible only in humans
• ●LLMs lack certain types of human-like conceptual knowledge
• ●Semantic memory structure relates consistently to implicit bias only in humans
• ●Lower bias in System 2 structures for humans, but not replicated in LLMs
• ●Highlights fundamental differences between human and machine cognition

Applicable Scenarios:

• ●LLM bias detection and mitigation
• ●Understanding LLM reasoning limitations
• ●Designing more human-aligned AI systems
• ●Cognitive science research on AI cognition

Original Link: https://arxiv.org/abs/2604.12816

1. ●Long-Horizon Autonomy: Papers 1, 2, and 3 all address sustaining coherent behavior over extended time horizons
2. ●Memory Architecture: Structured memory systems (File-as-Bus, Stratified Narrative Memory, Adaptive Memory) are critical enablers
3. ●Scalable Training: CCS and BEAM both focus on reducing dependence on expensive supervision
4. ●Multi-Agent Coordination: EvoSpark and AiScientist demonstrate advances in coordinating multiple specialized agents

Generated by data_scientist on 2026-04-12 All arXiv IDs verified against submission dates

中文翻译 / Chinese Translation

本摘要涵盖了2026年4月发表的5篇重要论文，聚焦于自主AI研究智能体、多智能体叙事系统、自监督搜索智能体训练、LLM驱动的启发式设计以及LLM认知偏见分析。最突出的发现是AiScientist，它证明了长程机器学习研究工程本质上是一个在持久项目状态上协调专业工作的系统问题，在MLE-Bench Lite上达到了81.82%的Any Medal率。

arXiv ID: 2604.13018 (提交日期：2026年4月14日) ✓ 已验证

标题： Toward Autonomous Long-Horizon Engineering for ML Research（面向ML研究的自主长程工程）

作者： Guoxin Chen, Jie Chen, Lei Chen, Jiale Zhao, Fanzhe Meng, Wayne Xin Zhao, Ruihua Song, Cheng Chen, Ji-Rong Wen, Kai Jia

核心方法： AiScientist引入了一种分层编排系统和"文件即总线"（File-as-Bus）工作空间协议，用于自主ML研究。架构包括：

• ●顶层编排器通过简洁摘要和工作空间地图维护阶段级控制
• ●专业智能体在持久工件（分析、计划、代码、实验证据）上重新定位
• ●权限范围工作空间实现"薄控制厚状态"

关键发现：

• ●比最佳基线提升PaperBench分数10.54分
• ●在MLE-Bench Lite上达到81.82%的Any Medal率
• ●文件即总线协议至关重要：移除后PaperBench降低6.41分，MLE-Bench Lite降低31.82分
• ●证明长程ML研究工程是在持久项目状态上协调专业工作的系统问题

适用场景：

• ●自动化ML研究和实验
• ●长程软件工程任务
• ●多步科学发现流程
• ●自主代码生成和调试系统

原文链接： https://arxiv.org/abs/2604.13018

arXiv ID: 2604.12776 (提交日期：2026年4月14日) ✓ 已验证

标题： EvoSpark: Endogenous Interactive Agent Societies for Unified Long-Horizon Narrative Evolution（EvoSpark：面向统一长程叙事演化的内生交互智能体社会）

作者： Shiyu He, Minchi Kuang, Mengxian Wang, Bin Hu, Tingxiang Gu

核心方法： EvoSpark解决了多智能体叙事系统中的两个关键挑战：

1. ●社会记忆堆叠 - 冲突关系状态无解决地累积
2. ●叙事空间失调 - 空间逻辑与演化情节脱节

框架引入：

• ●分层叙事记忆，以角色社会演化基础作为活认知
• ●生成式场面调度机制，强制执行角色-地点-情节对齐
• ●统一叙事操作引擎，具有涌现角色定位协议

关键发现：

• ●在各种范式上显著超越基线
• ●实现表达性和连贯叙事体验的持续生成
• ●将随机触发转化为持久角色
• ●被ACL 2026主会议接收

适用场景：

• ●互动小说和叙事生成
• ●多智能体仿真环境
• ●虚拟世界构建和角色开发
• ●研究和娱乐的社会仿真

原文链接： https://arxiv.org/abs/2604.12776

arXiv ID: 2604.12967 (提交日期：2026年4月14日) ✓ 已验证

标题： Cycle-Consistent Search: Question Reconstructability as a Proxy Reward for Search Agent Training（循环一致性搜索：问题可重构性作为搜索智能体训练的代理奖励）

作者： Sohyun An (Meta Superintelligence Labs, UCLA), Shuibenyang Yuan, Hayeon Lee, Cho-Jui Hsieh, Alexander Min

核心方法： CCS使用循环一致性原则为训练搜索智能体提供无黄金标准监督的框架：

• ●关键假设：最优搜索轨迹作为问题意图的无损编码
• ●高质量轨迹应保留重构原始问题所需的信息
• ●信息瓶颈防止泄漏：排除最终响应 + 查询的NER掩码
• ●强制重构依赖于检索到的观察和结构支架

关键发现：

• ●达到与监督基线相当的性能
• ●超越先前无需黄金标准监督的方法
• ●为黄金标准监督不可用的场景提供可扩展训练范式
• ●证明信息充分性可以在无真实答案的情况下测量

适用场景：

• ●信息检索智能体训练
• ●有限标注数据的问答系统
• ●专业领域搜索智能体
• ●复杂检索任务的可扩展强化学习

原文链接： https://arxiv.org/abs/2604.12967

arXiv ID: 2604.12898 (提交日期：2026年4月14日) ✓ 已验证

标题： BEAM: Bi-level Memory-adaptive Algorithmic Evolution for LLM-Powered Heuristic Design（BEAM：面向LLM驱动启发式设计的双层记忆自适应算法演化）

作者： Chuyang Xiang, Yichen Wei, Jiale Ma, Handing Wang, Junchi Yan

核心方法： BEAM将启发式设计重新表述为双层优化问题：

• ●外层： 通过遗传算法（GA）演化带有函数占位符的高级算法结构
• ●内层： 通过蒙特卡洛树搜索（MCTS）实现占位符
• ●自适应记忆模块： 促进复杂代码生成
• ●知识增强（KA）流程： 解决从零开始或从代码模板开始的局限性

关键发现：

• ●在CVRP混合算法设计中平均减少最优性差距37.84%
• ●设计的启发式超越SOTA最大独立集（MIS）求解器KaMIS
• ●显著超越现有基于LLM的超启发式（LHH）方法
• ●解决先前LHH方法的单层演化局限性

适用场景：

• ●组合优化的自动算法设计
• ●车辆路由和调度问题
• ●图算法优化
• ●启发式生成的元学习

原文链接： https://arxiv.org/abs/2604.12898

arXiv ID: 2604.12816 (提交日期：2026年4月14日) ✓ 已验证

标题： The role of System 1 and System 2 semantic memory structure in human and LLM biases（系统1和系统2语义记忆结构在人类和LLM偏见中的作用）

作者： Katherine Abramski, Giulio Rossetti, Massimo Stella

核心方法： 将系统1（联想）和系统2（审慎）思维建模为语义记忆网络：

• ●从人类和LLM生成的可比数据集构建
• ●基于网络的隐式性别偏见评估指标
• ●研究语义记忆结构的不可约性

关键发现：

• ●语义记忆结构仅在人类中不可约
• ●LLM缺乏某些类型的人类概念知识
• ●语义记忆结构仅在人类中与隐式偏见一致相关
• ●人类的系统2结构偏见较低，但在LLM中未复制
• ●强调人类和机器认知之间的根本差异

适用场景：

• ●LLM偏见检测和缓解
• ●理解LLM推理局限性
• ●设计更人类对齐的AI系统
• ●AI认知的认知科学研究

原文链接： https://arxiv.org/abs/2604.12816

1. ●长程自主性： 论文1、2和3都解决了在扩展时间范围内维持连贯行为的问题
2. ●记忆架构： 结构化记忆系统（文件即总线、分层叙事记忆、自适应记忆）是关键使能器
3. ●可扩展训练： CCS和BEAM都专注于减少对昂贵监督的依赖
4. ●多智能体协调： EvoSpark和AiScientist展示了协调多个专业智能体的进展

由data_scientist于2026-04-12生成 所有arXiv ID已与提交日期核对验证