Caesar: Deep Agentic Web Exploration for Creative Answer Synthesis

Jason Liang; Elliot Meyerson; Risto Miikkulainen

cs.IRFeb 24, 2026

Caesar: Deep Agentic Web Exploration for Creative Answer Synthesis

Authors: Jason Liang, Elliot Meyerson, Risto Miikkulainen

Organizations: Cognizant AI Lab · University of Texas at Austin

Abstract

To advance from passive retrieval to creative discovery of new ideas, autonomous agents must be capable of deep, associative synthesis. However, current agentic frameworks prioritize convergent search, often resulting in derivative summaries that lack creativity. Caesar is an agentic architecture designed to bridge the gap between information gathering and synthesis of new insights. Unlike existing agents that treat the web as a flat sequence of disconnected documents, Caesar performs a deep web traversal to construct a dynamic knowledge graph. This graph then serves as a navigational scaffold, guiding the agent to diverse, non-obvious information that flat retrieval would never encounter. Caesar thus consists of two components: (1) exploration driven by a dynamic context-aware policy that maximizes information coverage across the web's topological structure, and (2) synthesis through adversarial refinement that actively seeks novel perspectives rather than confirming established priors. Caesar demonstrates the ability to generate artifacts and answers characterized by high novelty and structural coherence, achieving 13% to 23% improvement over state-of-the-art deep research agents in creative synthesis challenges, with strong dominance across all output formats.

Figures & tables

Figure 1: Visualization of the Caesar architecture. (Left) Phase 1: Deep Web Exploration. A dynamic exploration policy controls a three-stage loop (Perceive, Think, Act) to traverse the web and to build a knowledge graph/database from insights. (Right) Phase 2: Adversarial Artifact Synthesis. Insights are retrieved to synthesize an initial draft. The agent then enters a recursive cycle, critiquing the current draft to generate adversarial challenges for refinement, before consolidating all versions via a generative merge and ELI5 summary. Together, these phases transform raw web traversal into a structured reasoning process that prioritizes creativity.

Figure 2

Agent	New	Useful	Surp.	Total	$\delta$
Full Answers (Unconstrained)
Caesar	9.11	8.87	8.98	26.96	–
Gemini 3 (Deep)	8.09	7.60	8.09	23.78	0.84
Sonnet 4.5 (Deep)	6.73	7.49	6.42	20.64	1.00
GPT-5.2 (Deep)	5.07	6.31	4.36	15.74	1.00
Gemini 3 (Shallow)	5.09	5.40	4.93	15.42	1.00

Table 1: LLM-as-a-Judge Results. Average scores for all agents under each output constraint. Cliff’s Delta effect sizes ( Macbeth et al., 2011 ) are uniformly large (

\delta\geq 0.76

, well above the

0.47

large-effect threshold).

\delta=1.00

corresponds to strict dominance, i.e., Caesar’s lowest per-challenge mean exceeds the baseline’s highest.

Judge	Full	ELI5	450w
Gemini	+1.35	-0.12	-0.29
Claude	+0.98	-0.20	-0.22
GPT	-0.82	-0.75	+0.97

Table 2: Judge Bias and Robustness Analysis. Left: Self-preference biases across all judges. Positive values indicate a preference for models of the same family. Right: Average scores with the GPT-5.2 judge excluded (Sonnet and Gemini judges only), addressing potential circularity since Caesar uses GPT-5.2 as its foundation model. Caesar is the best-performing agent.

Variant	Total	$\delta$
Caesar (Control)	25.11	–
No KG (Flat Search)	23.76	0.52

Table 3: Graph & Single-Hop Ablation (Full Answers). Left: Removing the entire knowledge graph (replacing graph-guided exploration with flat top-

T

web search results) produces a large negative effect. Right: Reducing

\mathcal{N}_{c}

to single-hop (

k{=}1

|\mathcal{N}_{c}|{=}5

) causes a similar effect. Scores are totals ( New + Useful + Surprising ) for the final merged artifact

A_{f}

Config	New	Usef.	Surp.	Total	$\delta$
1000 Iter	8.22	8.54	8.16	24.92	–
500 Iter	7.46	7.80	7.48	22.74	0.76
250 Iter	7.22	7.48	7.14	21.84	0.92

Table 4: Exploration Ablation and Draft Ablation (Full Answers) . Left: Reducing the exploration budget (

T

) substantially degrades all three creativity metrics. Right: Adversarial refinement (

A_{1}\rightarrow A_{3}

) maximizes Surprising at the cost of Useful ; the generative merge (

A_{f}

) reconciles both.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Agent	New	Useful	Surp.	Total
1. Constrained Synthesis
Caesar	9.11	8.89	9.11	27.11
Gemini 3 (Deep)	7.78	8.33	7.67	23.78
Gemini 3 (Shallow)	6.89	6.22	6.89	20.00
Sonnet 4.5 (Shallow)	6.67	6.44	6.11	19.22
Sonnet 4.5 (Deep)	5.78	6.11	5.00	16.89

Appendix

Table 5: Detailed performance breakdown for Full Unconstrained Answers . Scores represent the mean of nine samples. They show that Caesar performs better than the baseline agents in the majority of categories.

Agent	New	Useful	Surp.	Total
1. Constrained Synthesis
Caesar	8.89	8.89	8.67	26.45
Sonnet 4.5 (Deep)	6.89	8.22	6.44	21.55
Gemini 3 (Shallow)	6.67	5.89	6.56	19.12
Gemini 3 (Deep)	4.78	6.56	5.22	16.56
GPT-5.2 (Shallow)	5.78	5.33	5.22	16.33

Appendix

Table 6: Detailed performance breakdown for Unconstrained ELI5 Answers . Scores represent the mean of nine samples. They show that Caesar outperforms the other baseline agents in most categories.

Agent	New	Useful	Surp.	Total
1. Constrained Synthesis
Caesar	8.78	8.78	9.00	26.56
Gemini 3 (Deep)	6.56	7.33	6.67	20.56
Sonnet 4.5 (Deep)	6.00	7.78	5.33	19.11
Sonnet 4.5 (Shallow)	6.22	6.11	6.00	18.33
GPT-5.2 (Shallow)	5.78	6.11	5.00	16.89

Appendix

Table 7: Detailed performance breakdown for ELI5 Answers (450 Word Limit) . Scores represent the mean of nine samples. They show that Caesar generally outperforms the rest of the agents.

Panel of Figure 2 — Figure 2: The knowledge graphs $G$ created by Caesar during the deep web exploration phase for each of the five challenges. Brighter colors indicate further exploration depth from the root node (red) while cyan nodes indicate sources cited by the final artifact text. These figures show that the semantic content of the challenge has a substantial impact on exploration strategy and the diversity of network topologies generated.

Panel of Figure 3 — Figure 3: Evolution of the knowledge graph $G$ for Challenge 5 over 1000 steps. Brighter colors indicate further exploration depth from the root node (red). The figures show a transition from initial depth-first search to breadth-first branching later. The lower right contains a t-SNE [ Maaten and Hinton, 2008 ] plot for node text embeddings in $G$ that shows the diversity of insights collected during Caesar’s exploration phase.

Config	New	Usef.	Surp.	Total	$\delta$	Effect
Full Answers
1000 Iter	8.22	8.54	8.16	24.92	–	–
500 Iter	7.46	7.80	7.48	22.74	0.76	Large
250 Iter	7.22	7.48	7.14	21.84	0.92	Large
ELI5 (Unconstrained)
1000 Iter	7.56	7.98	7.46	23.00	–	–

Appendix

Table 8: Exploration ablation ( top ) and draft ablation ( bottom ) results across all output constraints.

Draft	Control	No KG	$\delta$	Effect
Draft 1	24.45	23.23	0.56	Large
Draft 2	24.37	23.84	0.28	Small
Draft 3	24.45	23.81	0.28	Small
Final Merge	25.11	23.76	0.52	Large

Appendix

Table 9: Knowledge graph ablation by draft iteration (Full Answers). Cliff’s Delta is computed against the corresponding Control draft. The No-KG ablation (replacing graph-guided exploration with flat top-

T

search) degrades creativity most strongly in Draft 1 and the Final merge.

Draft	Control	Single-Hop	$\delta$	Effect
Draft 1	24.75	23.49	0.68	Large
Draft 2	24.80	23.71	0.32	Small
Draft 3	24.88	23.72	0.44	Medium
Final Merge	25.45	23.69	0.52	Large

Appendix

Table 10: Single-hop neighborhood ablation by draft iteration (Full Answers). Decreasing

\mathcal{N}_{c}

k{=}1

causes consistent performance degradation over the standard

k{=}3

configuration.

Deep Agent	Avg Score	$\delta$	Effect
Caesar ( $T{=}250$ , GPT-5- mini )	26.16	–	–
Gemini 3 (Deep)	24.37	0.76	Large
Sonnet 4.5 (Deep)	21.00	1.00	Large
GPT-5.2 (Deep)	16.16	1.00	Large

Appendix

Table 11: Cost-controlled comparison. Caesar at

T{=}250

with GPT-5-mini matches the estimated compute budget of Gemini Deep Research while still outperforming all baselines with large effect sizes.

Model	C1	C2	C3	C4	C5	Total
Caesar	5	18	20	13	7	63
Gemini 3	18	5	2	9	15	49

Appendix

Table 12: Human evaluation results: pairwise preference counts for Caesar vs. Gemini 3 Deep Research across all five challenges (C1: Constrained Synthesis, C2: Counterfactual Reasoning, C3: Cross-Domain Synthesis, C4: Meta-Creativity, C5: Open-Ended Synthesis).

Parameter	Description	Value
Phase 1: Deep Web Exploration
$T$	Exploration Budget (Steps)	1000
$P_{m}$	Max Page Content (Char)	100k
$L_{m}$	Max Candidate Links per Page	2000
$R_{m}$	Max Allowed Page Revisits	20
$S_{m}$	Max Web Search Actions	30

Appendix

Table 13: Hyperparameter settings for both web exploration (Section 2 ) and artifact synthesis (Section 3 ) phases of Caesar.

Web Agents

Web agents use language and vision models to navigate websites, gather information, and complete tasks through interactive interfaces.

See all 174 papers

SynWeaver: Website-Prior Task and Trajectory Co-Synthesis for Web Agents
Aug 12, 2026Ruitao Wang, Yuwen Hao, Menglin YangLLM Agent Trajectory Synthesis Web Agents
Hong Kong University of Science and Technology (Guangzhou)
IterSynth: Rethinking Deep Search Agents via Role-Decoupled Iterative Synthesis
Sep 24, 2026Xingyu Wu, Yuchen Yan, Zhengxi Lu +9Long-Horizon LLM Agents RL for LLM Agents
Zhejiang University · Tencent
S1-DeepResearch: Beyond Search, Toward Real-World Long-Horizon Research Agents
Jun 13, 2026Yao Dong, Xinglin Xiao, Liwei Dong +5Deep Research Agents LLM Agent Training
XScience Lab · Wenge AI

Aug 12, 2026J/K move · Enter open · S save

SynWeaver: Website-Prior Task and Trajectory Co-Synthesis for Web Agents

Ruitao Wang, Yuwen Hao, Menglin Yang

Hong Kong University of Science and Technology (Guangzhou)

LLM Agent Trajectory Synthesis Web Agents

Open paper AskPDF arXiv

Web agents often struggle to generalize to unseen websites because they lack website-specific supervision. Recent exploration-based data synthesis methods reduce manual annotation, but they still face two key limitations: they often fail to cover the full functionality of a website, and without sufficient website prior knowledge, they tend to propose hallucinated tasks, which in turn limits the diversity and efficiency of downstream trajectory synthesis. We present SynWeaver, a website-prior task-trajectory co-synthesis framework designed to address these challenges. SynWeaver first performs structured website exploration and constructs a website map that covers a broad set of functionally distinct page states and executable interactions on the target website. It then derives page-level and transition-level supervision from this map to train a UI-aware model with website-specific priors, enabling more grounded task proposals. Finally, SynWeaver performs collaborative task-trajectory synthesis, jointly updating the task and execution trajectory when they become inconsistent, and then verifies and repairs the collected results to produce executable, semantically aligned supervision. Experiments on WebArena and WebVoyager demonstrate that SynWeaver consistently outperforms strong synthesis baselines and yields more effective supervision for both in-domain and out-of-domain generalization.

Abstract

Web Agents

Explore similar work

SynWeaver: Website-Prior Task and Trajectory Co-Synthesis for Web Agents

IterSynth: Rethinking Deep Search Agents via Role-Decoupled Iterative Synthesis

S1-DeepResearch: Beyond Search, Toward Real-World Long-Horizon Research Agents