agentic-search

The authors announce Gear, an open-source, meta-learning-inspired framework described as "a scikit-learn for RSI," which separates search decisions from execution and evaluation so methods like GEPA, RRSI, and Harness² are easier to implement and compare. A working GEPA-style variant with shared failure diagnosis and staged evaluation showed pass-rate gains on AutomationBench and Terminal-Bench, plus a ~95% lower per-task API cost for GPT-5.6 Luna versus Astra using the same optimized harness.

Hyperresearch launches on Product Hunt, offering durable, verified agentic search for research workflows.

The paper introduces PatEx-Gym, the first framework for developing and evaluating agents for patent examination, including a 1K-task benchmark (PatEx-Bench), an annotation-free trajectory synthesis method (PASS), and a compact 4B PatEx-Agent trained via supervised fine-tuning that competes with much larger general-purpose agents.

ReGE 是一个将智能体历史执行经验合成为自指导信息的框架，使后续任务尝试无需更强教师模型或参数更新即可提升表现，在四个基准上平均提升约 9-10 个百分点，并大幅减少引导构建所需的 token 成本。

论文提出 SearchWeaver 框架，利用代理的执行轨迹和任务奖励自动联合优化语料表示、可执行工具与使用指令，从而改进 agentic search 的搜索接口设计。使用 GPT-5.4-nano 时，优化后的接口在 BrowseComp-Plus 上将答案准确率从 71.6% 提升至 78.4%，并在四个 BRIGHT 子集上将平均 nDCG@10 从 45.7 提升至 55.0。

Exa introduces ATLAS, a new benchmark for evaluating AI agents on search-intensive, real-world research tasks with verified golden answers and a refreshable automated pipeline. Key findings: no sub-$1 agent run achieved row F1 over 0.5, top agents still miss ~1/3 of golden results, and Exa's search backend defines the cost-performance Pareto frontier.

本文提出 MIRA（Musical Intent Refinement Agent），一种在测试时通过将用户意图细化为可验证 rubric 并进行轨迹感知树搜索来改进文本到音乐生成对齐的智能体，并发布了音乐专家标注的基准 MuRA-Bench，使开源生成器达到接近 Suno、Mureka 等商业系统的性能。

This paper introduces VHop, a data generation framework and benchmark for multi-step visual retrieval, and VHop-Router, an end-to-end trained autoregressive multi-step retriever that operates in visual latent space, boosting retrieval from under 5% to 76.3% and improving agentic search success by 52.7% while drastically cutting tokens and API payload.

This paper introduces decision-checkpoint auditing, a trajectory-level measurement method that classifies scientific-search agent failures into target exposure, inspection, and answer correctness stages, finding that keyword search and read-first policies achieve identical 24.6% accuracy despite differing failure distributions on AutoResearchBench Deep.

EvoDuet introduces a bilevel optimization framework that co-evolves solutions and web search queries with fixed LLM parameters, using a retrieval gate to decide when to fetch new documents versus reusing stored ones. Across 21 tasks it boosts OpenEvolve's normalized discovery gain notably (e.g., 61.3% to 82.3% with Gemini-3.8-Flash), surpassing prior best scores on eight tasks and generalizing across different evolutionary search scaffolds.

CorpusMap is an entity-based navigation layer for agentic search that organizes large document collections around recurring entities to improve evidence discovery and answer quality while reducing token usage.

AutoData introduces an AI agent that automates pre-training data selection for large language models by searching over executable algorithms, outperforming human-designed curation pipelines and improving downstream metrics.

ZGCM-1 is a fully open 7B LLM trained from scratch with FP8, MDP mid-training, and AI4AI, achieving performance comparable to Qwen3-235B on math and agentic search tasks.

ZGCM-1 is a 7B open foundation model trained from scratch with extreme efficiency, combining internal reasoning and external tool use for math and agentic search tasks, achieving competitive performance with much larger models like Qwen3-235B-A22B and GLM-5.1.

The EchoNet benchmark tests how open-weight AI models handle fake sources during agentic search, showing that models struggle most against fake majorities and revealing cost and performance differences.

This paper introduces ExTS, a tree-search policy for budget-constrained agentic search in LLM agents that improves over standard baselines in tasks like prompt optimization and code generation with an average +5.5% gain.

This paper introduces BioCheck Agent, an LLM-based agent that generates structured biomedical fact-checking reports using RL-enhanced agentic search, showing improved accuracy and reduced hallucinations compared to base models.

This paper introduces Conformalized Agentic Search (CAS), a framework that uses Conformal Prediction to enhance the reliability of search agents by adaptively retrieving documents and weighting policies during reinforcement learning, thereby improving accuracy and reducing redundant tool calls.

Mistral introduces Agentic Search, a multi-step retrieval loop that enhances AI system accuracy and efficiency in navigating complex documents, reducing latency and token use compared to traditional RAG methods.

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.