CogniWeb: Dual-Process Web-Agent Architecture

• CogniWeb is a modular web-agent architecture that applies dual-process cognition to alternate between fast reactive control and slow deliberative planning.
• The paper demonstrates that integrating offline imitation learning with online reinforcement improves performance, achieving a 43.96% success rate with 75% token reduction compared to slow-only systems.
• A dynamic switch mechanism orchestrates system transitions based on task complexity, ensuring adaptive error recovery and efficient navigation in high-entropy web environments.

Searching arXiv for CogniWeb and closely related web-agent literature. arxiv_search(query="^^^^^^^^3CogniWeb web agent dual-process cognition3^^^^^^^^", max_results=^^^^3submittedDate3CogniWeb web agent dual-process cognition3^^^^, sort_by="^^^^^^^^3submittedDate3 Searching by exact arXiv id and title for confirmation. arxiv_search(query="^^^^^^^^3(Liu et al., 7 Aug 2025) OR \3^^^^^^^^"Cognitive Duality for Adaptive Web Agents\"", max_results=5, sort_by="relevance") CogniWeb is a modular web-agent architecture that operationalizes dual-process cognition for autonomous web navigation by adaptively toggling between a fast reactive subsystem and a slow deliberative subsystem. Introduced in "Cognitive Duality for Adaptive Web Agents" (&&&^^^^3CogniWeb web agent dual-process cognition3^^^^&&&), it is positioned as a unifying framework for integrating offline imitation learning with online exploration in high-entropy, dynamic web environments characterized by combinatorially explosive action spaces. In evaluation on WebArena, CogniWeb is reported to achieve a 43.96% success rate while reducing token usage by 75%, emphasizing an explicit capability–efficiency trade-off rather than pure maximization of task success (&&&^^^^3CogniWeb web agent dual-process cognition3^^^^&&&).

^^^^3submittedDate3^^^^. Conceptual basis

CogniWeb is directly motivated by the dual-process theory of human cognition. Its decomposition distinguishes System ^^^^3submittedDate3^^^^ ("Fast"), described as intuitive, heuristic, rapid, and reactive, from System ^^^^3(Liu et al., 7 Aug 2025) OR \3^^^^ ("Slow"), described as deliberative, logical, slower, and capable of multi-step reasoning, reflection, and planning. Within web navigation, this division is intended to separate routine interface manipulation from recovery, long-horizon planning, and adaptation to novelty or execution failure (&&&^^^^3CogniWeb web agent dual-process cognition3^^^^&&&).

The framework formalizes action selection as a mixture of two sub-policies:

PRESERVED_PLACEHOLDER_^^^^3CogniWeb web agent dual-process cognition3^^^^

where PRESERVED_PLACEHOLDER_^^^^3submittedDate3^^^^ is the System ^^^^3submittedDate3^^^^ policy, PRESERVED_PLACEHOLDER_^^^^3(Liu et al., 7 Aug 2025) OR \3^^^^ is the System ^^^^3(Liu et al., 7 Aug 2025) OR \3^^^^ policy, and $\lambda_t \in [0,1]$ controls switching at each step. The switch can be determined by learned predictors or hand-coded heuristics. The paper further frames web navigation as a complexity-weighted optimization problem, arguing that tasks cluster into separable difficulty regimes and therefore naturally favor dynamic partitioning between heuristic and reflective control (&&&^^^^3CogniWeb web agent dual-process cognition3^^^^&&&).

A common misconception is that CogniWeb merely ensembles two models. The formulation is narrower and more operational: the architecture is built around stepwise arbitration conditioned on task complexity and execution status. This suggests a control-theoretic view in which switching is central, not ancillary.

^^^^3(Liu et al., 7 Aug 2025) OR \3^^^^. Architectural organization

CogniWeb comprises three principal modules: System ^^^^3submittedDate3^^^^, System ^^^^3(Liu et al., 7 Aug 2025) OR \3^^^^, and a switch mechanism. System ^^^^3submittedDate3^^^^ executes web actions through learned heuristics or pattern recognition and is optimized for minimal reasoning and direct output. In the reported implementation, this fast mode can be realized with fine-tuned Phi-3-mini-^^^^3submittedDate3(Liu et al., 7 Aug 2025) OR \3^^^^8k-instruct or gemma-3-^^^^3submittedDate3^^^^b-it, reflecting an explicit emphasis on fast inference (&&&^^^^3CogniWeb web agent dual-process cognition3^^^^&&&).

System ^^^^3(Liu et al., 7 Aug 2025) OR \3^^^^ uses a larger instruction-following LLM, exemplified by GPT-4o, and is prompted for chain-of-thought reasoning and reflection. It handles multi-step planning, mistake analysis, and strategy adaptation. Two memory structures are associated with this mode. Working memory is defined as the last $k=10$ actions, providing expanded short-horizon context, and episodic memory stores self-generated experience summaries accumulated in a replay pool. These memories support reflection over prior trajectories rather than purely myopic action selection (&&&^^^^3CogniWeb web agent dual-process cognition3^^^^&&&).

The switch mechanism dynamically chooses which subsystem to invoke at each decision point. Its implementation is hybrid: heuristic rules are combined with prompt-based exemplars for switching scenarios. The paper gives a concrete trigger: three failed or redundant attempts cause a transition from System ^^^^3submittedDate3^^^^ to System ^^^^3(Liu et al., 7 Aug 2025) OR \3^^^^ for deeper analysis. This makes the switch neither purely learned nor purely hard-coded; it is an adaptive policy over cognitive modes.

3. Learning regime and optimization logic

CogniWeb’s design explicitly unifies two learning paradigms that are often treated separately in web-agent work. System ^^^^3submittedDate3^^^^ is specialized through offline imitation learning, using supervised fine-tuning and reranking over large collections of expert demonstrations. The reported training sources include MiniWoB++ and Mind^^^^3(Liu et al., 7 Aug 2025) OR \3^^^^Web. This subsystem is therefore aligned with behavior cloning and low-latency action production (&&&^^^^3CogniWeb web agent dual-process cognition3^^^^&&&).

System ^^^^3(Liu et al., 7 Aug 2025) OR \3^^^^ is optimized through online reinforcement learning or planning. It is intended for shifting dynamics, error recovery, and tasks requiring multi-step strategy construction. Its prompts instruct the model to analyze progress, reflect on errors, and adapt strategies, while leveraging working memory and episodic memory. In the paper’s terms, the switch manages the adaptive hybridization of these offline and online learning regimes (&&&^^^^3CogniWeb web agent dual-process cognition3^^^^&&&).

This division of labor matters methodologically. System ^^^^3submittedDate3^^^^ absorbs high-frequency action priors from demonstrations, whereas System ^^^^3(Liu et al., 7 Aug 2025) OR \3^^^^ is reserved for low-frequency or failure-contingent reasoning. A plausible implication is that CogniWeb treats deliberation as a scarce computational resource rather than a default mode of inference.

4. Execution flow and memory updates

The episode loop described in the appendix begins with an early-stop check over the trajectory. If early stopping is triggered and the agent is not in timeout, the slow system is invoked for failure analysis; otherwise a stop action is emitted. If no early stop is triggered, the switch selects either fast or slow processing for the next action (&&&^^^^3CogniWeb web agent dual-process cognition3^^^^&&&).

The control flow can be summarized as follows. At each step, the agent consults the switch; straightforward states are routed to System ^^^^3submittedDate3^^^^, while stuck or complex states are routed to System ^^^^3(Liu et al., 7 Aug 2025) OR \3^^^^. The action is appended to the trajectory, the environment is stepped, and termination is checked. After each transition, the trajectory is evaluated, summarized by System ^^^^3(Liu et al., 7 Aug 2025) OR \3^^^^, and the resulting experience is appended to episodic memory. The paper therefore treats memory maintenance as part of the online control loop rather than as a separate offline consolidation stage (&&&^^^^3CogniWeb web agent dual-process cognition3^^^^&&&).

This execution model clarifies another potential misunderstanding: System ^^^^3(Liu et al., 7 Aug 2025) OR \3^^^^ is not invoked only at the beginning of hard tasks. It can enter mid-trajectory for diagnosis, reflection, and repair, which makes CogniWeb an adaptive interleaving architecture rather than a static task-level router.

5. Empirical performance on WebArena

The reported evaluation uses WebArena, comprising 8^^^^3submittedDate3(Liu et al., 7 Aug 2025) OR \3^^^^ diverse web tasks, with ^^^^3(Liu et al., 7 Aug 2025) OR \3^^^^ evaluation epochs. Two metrics are emphasized: Success Rate, defined as the fraction of tasks successfully completed, and Tokens per Trajectory, used as a proxy for compute and efficiency (&&&^^^^3CogniWeb web agent dual-process cognition3^^^^&&&).

System ^^^^3submittedDate3^^^^	System ^^^^3(Liu et al., 7 Aug 2025) OR \3^^^^	Success Rate (%)	Tokens per Traj.
Phi-3-mini + SFT	gpt-4o + reason. + refl.	43.96	393.89
gemma-3-^^^^3submittedDate3^^^^b-it + SFT	gpt-4o + reason. + refl.	4^^^^3CogniWeb web agent dual-process cognition3^^^^.^^^^3submittedDate3^^^^5	4^^^^3CogniWeb web agent dual-process cognition3(Liu et al., 7 Aug 2025) OR \3^^^^.9^^^^3(Liu et al., 7 Aug 2025) OR \3^^^^
gpt-4o	gpt-4o + reason. + refl.	4^^^^3submittedDate3^^^^.99	387.^^^^3submittedDate3CogniWeb web agent dual-process cognition3^^^^
Not Used	gpt-4o + reason. + refl.	46.^^^^3CogniWeb web agent dual-process cognition3^^^^6	^^^^3submittedDate3Cognitive Duality for Adaptive Web Agents3CogniWeb web agent dual-process cognition33^^^^.83
Not Used	gpt-4o + reason.	^^^^3(Liu et al., 7 Aug 2025) OR \3(Liu et al., 7 Aug 2025) OR \3^^^^.4^^^^3submittedDate3^^^^	^^^^3submittedDate3relevance3(Liu et al., 7 Aug 2025) OR \3submittedDate3^^^^.94
gpt-4o	Not Used	^^^^3submittedDate3^^^^5.64	^^^^3submittedDate3^^^^67.99

The central reported result is that Phi-3-mini + SFT for System ^^^^3submittedDate3^^^^ combined with gpt-4o + reasoning + reflection for System ^^^^3(Liu et al., 7 Aug 2025) OR \3^^^^ attains 43.96% success with 393.89 tokens per trajectory. A slow-only configuration reaches 46.^^^^3CogniWeb web agent dual-process cognition3^^^^6% success but requires ^^^^3submittedDate3Cognitive Duality for Adaptive Web Agents3CogniWeb web agent dual-process cognition33^^^^.83 tokens per trajectory, described as nearly 4× more tokens. Conversely, a fast-only configuration is highly efficient at ^^^^3submittedDate3^^^^67.99 tokens per trajectory but attains only ^^^^3submittedDate3^^^^5.64% success. The article’s empirical argument is therefore not that dual-process control dominates every baseline on raw accuracy, but that it produces an efficient operating point close to the top-performing slow-only regime (&&&^^^^3CogniWeb web agent dual-process cognition3^^^^&&&).

The ablation discussion reinforces this interpretation. Removing System ^^^^3submittedDate3^^^^ or System ^^^^3(Liu et al., 7 Aug 2025) OR \3^^^^ degrades either efficiency or accuracy, and removing episodic memory or self-reflection in System ^^^^3(Liu et al., 7 Aug 2025) OR \3^^^^ causes performance to collapse from approximately 4^^^^3CogniWeb web agent dual-process cognition3^^^^% to approximately ^^^^3(Liu et al., 7 Aug 2025) OR \3CogniWeb web agent dual-process cognition3^^^^%. Qualitative examples further show that fast-to-slow switching enabled recovery from errors and successful multi-stage solutions such as identifying discounts or performing post analysis on forums (&&&^^^^3CogniWeb web agent dual-process cognition3^^^^&&&).

6. Significance, limitations, and relation to adjacent work

CogniWeb’s stated significance lies in the principled unification of offline imitation learning for speed and online exploration/reasoning for adaptability. Its modularity also means that improvements in offline datasets, base models, or switching logic can be incorporated without redesigning the full control stack. The reported 75% reduction in token usage at near-top accuracy is presented as a practical argument for lower-cost deployment of LLM-based web agents (&&&^^^^3CogniWeb web agent dual-process cognition3^^^^&&&).

At the same time, the reported numbers indicate a clear limitation: a pure slow-mode system still yields the highest listed success rate, albeit at much higher token cost. CogniWeb should therefore be understood as an explicit optimization over efficiency and capability, not as proof that dual-process arbitration universally improves absolute task completion.

In adjacent work, "Web-CogReasoner: Towards Knowledge-Induced Cognitive Reasoning for Web Agents" (&&&^^^^3submittedDate3^^^^5&&&) decomposes web-agent ability into knowledge content learning and cognitive processes, structured through Factual, Conceptual, and Procedural knowledge. That framework emphasizes curriculum-style knowledge acquisition and knowledge-driven chain-of-thought, whereas CogniWeb emphasizes adaptive switching between fast and slow control modes. Taken together, these works suggest two complementary trajectories in web-agent research: one centered on control arbitration between cognitive modes, and another centered on structured knowledge acquisition as a substrate for reasoning (&&&^^^^3submittedDate3^^^^5&&&).

A broader implication is that CogniWeb belongs to a line of work treating web navigation not as a monolithic sequence-modeling problem but as a heterogeneous control problem with distinct operational regimes. Within that view, heuristic reactivity, reflective planning, memory, and switching policy are first-class architectural objects rather than implementation details.