Deep Research Agents (DRAs) aim to answer complex questions by searching the web, checking evidence, and synthesizing conclusions across heterogeneous sources. We introduce a category-theoretic framework for evaluating and improving such agents. The framework treats deep research as a structured mapping from user intent to evidence-grounded conclusions, making retrieval traces, cross-source alignment, and final synthesis explicit. Guided by this view, we derive a mechanism-aware benchmark of 296 bilingual questions. The benchmark targets four structural skills central to real research: following multi-hop evidence chains, verifying claims across sources, re-ordering fragmented information, and rejecting unsupported assumptions. We evaluate 16 frontier systems with human verification and find that these structural tasks remain highly challenging: the best system reaches only 19.9% average accuracy. The results show that strong agents can sometimes reorganize evidence and detect false premises, but still struggle with long-horizon synthesis and intersection-heavy verification. Beyond evaluation, the same theory also leads to practical system improvements. We instantiate theory-guided interventions such as tracked search, which preserves retrieval traces, and category tools, which add explicit verification and synthesis steps. These interventions yield measurable gains in API-based deep research systems. Our work therefore provides both a challenging benchmark and concrete design guidance for building more reliable research agents.
Figures & tables
Figure 1: An illustrative picture of 4 categorical evaluation, definitions can be found in Section 2 .
Figure 2: An illustrative picture of categorical view of the deep research agent workflow.
Benchmark
Sequential Tracing (Chains)
Multi-Source Synthesis (Pullbacks)
Substructure Disentanglement (Re-ordering)
Ontological Probing (Yoneda)
Theory-Based
BrowseComp
✓
✗
✗
✗
✗
WebShaper
✓
✓
✗
✗
✓
DeepResearch Bench
✓
✓
✗
✗
✗
Finance Agent Benchmark
✓
✓
✗
✗
✗
FinSearchComp
✓
✓
✗
✗
✗
Ours
✓
✓
✓
✓
✓
Table 1: Comparison of our proposed benchmark against existing evaluation frameworks.
Figure 3: Mechanism-grouped performance heatmap across four structural task types. Rows are grouped by model mechanism, and columns report average success rates (%) for each task type and the total average. Darker cells indicate stronger performance; gold outlines mark the best score within each column.
Figure 4: An illustrative overview of the scoring pipeline.
Figure 5: Search–reasoning diagnostic profiles across evaluated systems. Models are grouped by mechanism and sorted within each group by diagnostic strength. The left panel reports Search Score ( S ), and the right panel reports Reasoning Score ( R ); values denote average diagnostic scores (%).
Figure 6: Performance comparison between DeepResearchBench and our benchmark. Grok Deep Research and OpenAI Deep Research are closed-source end-to-end products; the remaining systems are our API-based implementations. Brackets indicate the gains from adding category tools on top of tracked search.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Category
Task ID
S Score
R Score
Primary Bottleneck & Rationale
Type I:
Task 1
90%
10%
Adversarial Retrieval: Overcoming lexical collision to find off-domain fragments; logic is trivial counting once the node is reached.
Linear Chain
Task 2
80%
20%
Schema Awareness: High reliance on querying specific vertical databases; reasoning is limited to simple temporal matching.
Task 3
85%
15%
Long-Tail Search: Tracing obscure numerical ranks to break index shortcuts; reasoning is basic historical traceback.
Type II:
Task 4
50%
50%
Balanced Synthesis: Equal challenge in retrieving disjoint, long-tail reports ( S ) and aligning dynamic schemas to solve algebraic deltas ( R ).
Pullback (V-Struct)
Task 5
30%
70%
Parametric Prior Override: Retrieval is standard; extreme R pressure to reject pre-training biases and resolve identities strictly via rigid IDs.
Task 6
30%
70%
Meta-Cognitive Triggering: Retrieving baseline data is trivial; the bottleneck is the unprompted realization and mathematical execution of latent regulatory constraints.
Appendix
Table 2: Detailed breakdown of the four-fold categorical assessment tasks, highlighting the distinct performance bottlenecks between Search ( S ) and Reasoning ( R ).
Model
Type I
Type II
Task 1
Task 2
Task 3
Task 4
Task 5
Task 6
Task 7
Task 8
Reasoning Models
Qwen3 w/Reasoning
0.0
0.0
5.0
0.0
0.0
10.0
10.0
0.0
GPT-5 w/Reasoning
0.0
0.0
0.0
2.2
0.0
0.0
20.0
0.0
Claude-4.5-Sonnet w/ Reasoning
4.0
0.0
0.0
2.2
0.0
0.0
0.0
0.0
Doubao w/Reasoning
4.0
0.0
0.0
0.0
8.3
0.0
0.0
0.0
Appendix
Table 3: Detailed performance comparison on specific tasks (Type I and Type II). Scores are presented as percentages (%).
Model
Type III
Type IV
Task 9
Task 10
Task 11
Task 12
Task 13
Task 14
Task 15
Reasoning Models
Qwen3 w/Reasoning
0.0
0.0
0.0
0.0
0.0
95.7
10.0
GPT-5 w/Reasoning
0.0
0.0
0.0
0.0
0.0
60.9
0.0
Claude-4.5-Sonnet w/ Reasoning
5.0
0.0
0.0
0.0
0.0
47.8
0.0
Doubao w/Reasoning
15.0
0.0
0.0
0.0
0.0
34.8
0.0
Appendix
Table 4: Detailed performance comparison on specific tasks (Type III and Type IV). Scores are presented as percentages (%).