LongCat-DeepResearch Technical Report

arXiv 2026 53.5 method

TLDR

LongCat-DeepResearch combines an enhanced LongCat model with a multi-agent workflow to produce evidence-grounded research reports, evaluated on multiple benchmarks.

Reasoning

The paper presents a clear system design with planning, parallel investigation, and section-level revision, supported by benchmark evaluations. Weaknesses include reliance on an in-house benchmark for ranking and mixed effects from planning refinement, limiting broader validation.

Read-first score

Read-first score 53.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 58.

Recency 8%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 25%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=benchmark,evaluation

Topical relevance 42%

48.3

Uses existing LLM keyword relevance scores normalized to 0-100. AI scientist,automated scientific discovery,autonomous research agent,automated research,literature review agent,survey generation,automated experimentation,experiment design agent,AI for scientific research,paper writing agent,research automation,scientific discovery agent

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 44.

Keyword Scores

automated research

research automation

autonomous research agent

AI for scientific research

literature review agent

paper writing agent

survey generation

automated scientific discovery

scientific discovery agent

AI scientist

experiment design agent

automated experimentation

Deep Analysis

Innovations

Combines an enhanced LongCat model with a multi-agent workflow for evidence-grounded report generation.
Separates global planning from detailed investigation and coordinates revision at the section level.
Uses multiple planning agents to explore external sources and refine an actionable ResearchSpec plan.
Research agents investigate and draft assigned sections in parallel, gathering additional evidence in separate contexts as analyses develop.
Global review guides targeted local revisions, reducing reliance on repeated full-report rewriting.
The workflow supports constructing research tasks and trajectories for mid-training and post-training of LongCat general-purpose models.

Methodology

LongCat-DeepResearch uses multiple planning agents to explore external sources and refine a ResearchSpec plan, then research agents draft assigned sections in parallel while gathering additional evidence in separate contexts. After assembly, global review directs targeted local revisions rather than repeated full-report rewriting. Evaluation is conducted on DeepResearchBench, DeepResearchBench II, ResearchRubrics, and an in-house benchmark.

Key Results

LongCat-DeepResearch achieves 55.25 on DeepResearchBench, 51.35 on DeepResearchBench II, and 79.83 on ResearchRubrics; on an in-house benchmark it scores 76.04, ranking second among four compared systems. Development-set analyses show benefits from combining planning perspectives, mixed effects from further planning refinement, and improved average automatic readability preference from additional editing across two benchmarks, with different trends on each.

Limitations

Further planning refinement has mixed effects.
Additional editing improves average automatic readability preference across two benchmarks, but with different trends on each.

Resources

Paper