PaperlayerView exact revision on arXiv

Source-linked AI summary

AgentOccam: A Simple Yet Strong Baseline for LLM-Based Web Agents

Ke Yang, Yao Liu, Sapana Chaudhary, Rasool Fakoor, Pratik Chaudhari, George Karypis, Huzefa Rangwala

arXiv:2410.13825v2cs.AIcs.CL

TL;DR

Web agents face a mismatch between task-specific observations and actions and the knowledge encoded during LLM training. AgentOccam aligns those spaces with a simple workflow, reporting improvements over prior WebArena approaches and a plain web agent. The authors present this as evidence of strong zero-shot web-task performance without extra examples or agent roles.

  • Problem

    LLM web agents face a mismatch between task-specific observations and actions and the knowledge encoded during their training, while web tasks require varied-page information extraction and embodied actions.

    Source: §1 · pp. 1–2

  • Method

    AgentOccam aligns web observations and actions with LLM capabilities by simplifying actions, refining observations, and adding branch/prune planning actions.

    Source: §1 · p. 2

  • Results

    AgentOccam surpasses SteP by 9.8 absolute points and WebPilot by 5.9 absolute points on WebArena.

    Source: §6 · p. 10

  • Takeaways & Limitations

    The authors argue that a simple agent architecture can support better generalizability, with observation/action alignment orthogonal to future agent-strategy advances.

    Source: §1 · p. 3, §6 · p. 10

    The formulation considers only the text modality and models web state transitions as deterministic.

    Source: §3 · p. 4

Abstract

from arXiv · show

Autonomy via agents using large language models (LLMs) for personalized, standardized tasks boosts human efficiency. Automating web tasks (like booking hotels within a budget) is increasingly sought after. Fulfilling practical needs, the web agent also serves as an important proof-of-concept example for various agent grounding scenarios, with its success promising advancements in many future applications. Prior research often handcrafts web agent strategies (e.g., prompting templates, multi-agent systems, search methods, etc.) and the corresponding in-context examples, which may not generalize well across all real-world scenarios. On the other hand, there has been limited study on the misalignment between a web agent's observation/action representation and the pre-training data of the LLM it's based on. This discrepancy is especially notable when LLMs are primarily trained for language completion rather than tasks involving embodied navigation actions and symbolic web elements. Our study enhances an LLM-based web agent by simply refining its observation and action space to better align with the LLM's capabilities. This approach enables our base agent to significantly outperform previous methods on a wide variety of web tasks. Specifically, on WebArena, a benchmark featuring general-purpose web interaction tasks, our agent AgentOccam surpasses the previous state-of-the-art and concurrent work by 9.8 (+29.4%) and 5.9 (+15.8%) absolute points respectively, and boosts the success rate by 26.6 points (+161%) over similar plain web agents with its observation and action space alignment. We achieve this without using in-context examples, new agent roles, online feedback or search strategies. AgentOccam's simple design highlights LLMs' impressive zero-shot performance on web tasks, and underlines the critical role of carefully tuning observation and action spaces for LLM-based agents.

1 INTRODUCTION

The introduction argues that web agents struggle when their observations and actions diverge from LLM training, and proposes aligning those representations rather than adding complex policies. AgentOccam combines three changes and reports gains over prior and plain agents without examples, search, or extra roles.

1 INTRODUCTION

  • Web-task agents can underperform when task-specific observation and action representations diverge from knowledge encoded during LLM training.Web tasks require extracting information from varied pages and issuing embodied actions such as scrolling, clicking, and hovering, which are less common in LLM training data.Source: §1 · pp. 1–2
  • AgentOccam reduces non-essential actions, refines observations by removing irrelevant elements and restructuring content, and adds branch/prune planning actions for navigation and history replay.The components use generic rules for markup-language pages without test-benchmark task information.Source: §1 · p. 2
  • 9.8 absolute points (+29.4%) over the previous state of the art and 5.9 absolute points (+15.8%) over concurrent work on WebArena are reported for AgentOccam.The introduction also reports a 26.6-point (+161%) success-rate increase over a similar plain web agent and a 2.4-point (+4.6%) lead over the previous best on WebVoyager’s definite-answer subset.Source: §1 · p. 2
  • AgentOccam avoids in-context examples, online search or sampling, specialized prompting templates, and additional agent roles.The authors describe observation/action alignment as orthogonal to agent strategies and potentially combinable with future advances in those strategies.Source: §1 · pp. 2–3

2 RELATED WORK

Prior web-agent work has improved performance through policy structures, examples, search, or trained models, but these approaches can require task-specific information, repeated inference, or training. The paper situates its evaluation in WebArena and real-world WebVoyager tasks.

2 RELATED WORK

  • Prior web agents use auxiliary modules, specialized roles, examples, and search or reflection strategies, which can depend on predefined control or increase inference cost through repeated trials.The related-work discussion notes that repeated sampling or resetting can limit use when failed tasks cannot be revoked.Source: §2 · p. 3
  • Fine-tuned web models offer adaptivity, but task-specific model sizes may not match powerful closed-source models, and earlier trained agents may lag LLM-based agents in performance or generalizability.The paper places fine-tuned approaches beyond its scope.Source: §2 · p. 3
  • The evaluation uses WebArena for text-based tasks across varied websites and intents, with further generalizability assessment on real-world WebVoyager tasks with golden answers.WebArena is described as a suite of simulated sites, while WebVoyager supplies real-world tasks.Source: §2 · p. 4

3 PROBLEM FORMULATION

The paper models web interaction as a POMDP and frames its central problem as improving an LLM-based policy through observation and action mappings. These mappings process inputs and outputs while leaving the base LLM policy in place.

3 PROBLEM FORMULATION

  • Web interaction is formalized as a POMDP with observations, hidden states, actions, deterministic state transitions, rewards, an initial task distribution, and a discount factor.The paper considers only text observations; the discount factor is typically 1 when reward is assigned only at episode end.Source: §3 · p. 4
  • The observation and action space alignment problem is to optimize rule-based mappings f and g around the base LLM policy, rather than adding a control program or helper modules.The policy is expressed as π(g(at)|ht) = πLLM(at|f(ht)); f processes observations and g processes actions.Source: §3 · p. 4

4 METHOD

AgentOccam uses a simple workflow in which a general-purpose LLM receives web observations and directly produces actions, aligning web tasks with LLMs’ text-centric training. It refines both actions and observations, while using a planning tree to organize task execution and memory.

4 METHOD

  • The method avoids adding modules or hierarchical structures: a general-purpose LLM API receives web observations and its outputs serve directly as actions.Source: §4 · p. 4

4.1 ACTION SPACE ALIGNMENT

  • The action space removes distracting or embodiment-heavy commands and adds branch and prune actions for LLM-generated plans across multi-path tasks.Branch decomposes objectives into subgoals; prune lets the agent abandon a sub-plan and seek alternatives.Source: §4.1 · pp. 5–6

4.2 OBSERVATION SPACE ALIGNMENT

  • Previous interactions and current page content dominate observation tokens, motivating shorter page descriptions and selective history to reduce context length.Longer context increases inference cost and makes relevant information harder to extract accurately.Source: §4.2 · pp. 6–7
  • The agent condenses web observations by merging duplicate descriptions of interactive elements and converting tables and lists to Markdown while retaining their information.Source: §4.2 · p. 7
  • History replay retains pivotal page elements with their tree context, then uses the planning tree to exclude observations from earlier plans.Pivotal nodes’ ancestors, siblings, and descendants preserve hierarchy, nearby context, and detail.Source: §4.2 · p. 7

5 EXPERIMENTAL RESULTS AND ANALYSIS

On WebArena, AgentOccam reaches a new state of the art while ablations show how action and observation refinements affect performance, token use, and task completion. Combining it with other strategies has mixed results.

Overall performance

  • 43.1% overall success rate raises AgentOccam from the previous and concurrent best of 37.2%, a 15.8% relative improvement on WebArena.The benchmark includes 812 tasks across four website domains.Source: §5 · pp. 7–8, Table 2 · p. 8

Action-space refinement

  • Removing rarely used actions significantly improves performance across tested websites and reduces wandering, shifting action use toward click and type.The action statistics report frequencies across tasks in each experiment setting.Source: §5 · p. 8, Table 3 · p. 9

Observation-space refinement

  • Disabling scrolling and passing the entire page proves advantageous, especially for GitLab and Reddit, but increases observation tokens.Source: §5 · p. 9
  • Simplifying page elements reduces context tokens and boosts performance across task types, except when GitLab tasks lead the agent to overlook simpler solutions.The authors give task 394 as an example of the GitLab exception.Source: §5 · p. 9

History replay

  • Selective replay reduces steps needed for task completion but lowers Shopping and Reddit success rates by 3.2 and 6.0 points, respectively.The cited step comparison is summarized as average steps per task across WebArena sites.Source: §5 · p. 9, Table 5 · p. 10

Planning

  • Branch and prune let the agent generate plans and exclude history outside the current sub-plan, improving performance across nearly all websites while reducing observation tokens.The actions are primarily used to correct a failed strategy and try an alternative path.Source: §5 · p. 9

Combining strategies

  • AGENTOCCAM + SteP outperforms standalone SteP but trails the base agent, while AGENTOCCAM + Judge raises overall success rate by 2.6%.SteP can help when its strategy fits, but mismatched guidance lowers AGENTOCCAM + SteP's overall success rate by 2 points; the judge selects among generated action candidates.Source: §5 · p. 10, Table 6 · p. 10

6 CONCLUSION

AgentOccam refines action and observation spaces for LLMs trained primarily on static text, using a simple policy workflow. It outperforms previous and contemporary approaches on WebArena, and the authors argue that simplicity supports better generalizability.

6 CONCLUSION

  • AgentOccam exceeds SteP by 9.8 absolute points and WebPilot by 5.9 absolute points on WebArena.The conclusion presents the result as evidence that the simple workflow does not compromise performance.Source: §6 · p. 10
  • AgentOccam refines its action and observation spaces to be more comprehensible for LLMs primarily trained on static text.The conclusion describes its policy workflow as requiring no extra modules, additional LLM calls, or in-context examples.Source: §6 · p. 10
  • The authors argue that a simple agent architecture supports better generalizability and aim to provide a foundation for future web-agent research.They frame this conclusion as echoing Occam’s razor.Source: §6 · p. 10

A COMPARISON OF THE VANILLA AND THE ALIGNED ACTION SPACE

The comparison identifies WebArena’s action space and AgentOccam’s aligned action space; the text describes removing some actions and adding workflow-management actions.

A COMPARISON OF THE VANILLA AND THE ALIGNED ACTION SPACE

B PIVOTAL NODES AND THE PLANNING TREE CLARIFICATIONS

The section distinguishes selective replay of web elements from selective replay of entire pages, and notes a multisite-task qualification.

B PIVOTAL NODES AND THE PLANNING TREE CLARIFICATIONS

B.1.1 INTUITION AND BACKGROUND

Pivotal nodes focus page history on web elements potentially useful for completing a task. The prompt asks the agent to identify and rank relevant current-page elements for its observation highlight.

B.1.1 INTUITION AND BACKGROUND

  • Pivotal nodes are page elements potentially useful for task completion, including informative elements and elements that enable navigation to crucial states.The method retains associated ancestor, sibling, and descendant nodes to create a more succinct, less noisy observation for focused working history.Source: §B.1.1 · p. 15

B.2 PROMPT, PIVOTAL NODE SECTION

  • The prompt requests an observation highlight in a specified response format.Source: §B.2 · p. 15
  • The prompt asks for relevant current-page element IDs, including elements useful if later failure requires restoring to the current step, ranked by relevance and potential value.It instructs the agent to omit elements from previous pages and select a higher-level element when most of its child nodes are crucial.Source: §B.2 · p. 15

B.4 AGENTOCCAM ’S WORKFLOW REGARDING THE PIVOTAL NODES

The workflow identifies pivotal web elements to focus interaction history on task-relevant content, then uses a planning tree to organize plans and selectively replay prior pages.

B.4 AGENTOCCAM ’S WORKFLOW REGARDING THE PIVOTAL NODES

  • Pivotal nodes retain crucial task information while omitting irrelevant noise from the interaction history.The agent selects pivotal nodes and associated nodes to construct the updated history used with the current observation.Source: §B.4 · pp. 15–16

B.5 SELECTIVELY REPLAYING ENTIRE PAGES WITH THE PLANNING TREE (FIGURE 4 STEP 3)

  • The planning tree supports selective replay of past pages rather than retaining every page in the task history.The paper introduces the tree as a way to selectively replay past pages.Source: §B.5 · p. 16

B.5.1 INTUITION AND BACKGROUND

  • The active plan scopes navigation like a function in an execution stack, while pruning returns the task to a previous plan state.The tree is continually available in the prompt; branch creates subplans and prune restores a prior state.Source: §B.5.1 · p. 16, §B.5.2 · p. 16

B.5.2 PROMPT, PLANNING TREE SECTION

  • The prompt lets the agent branch from a parent plan to create a subplan, or prune to resume a chosen earlier plan state.When neither planning command is needed, the prompt supplies navigation specifications.Source: §B.5.2 · p. 16

B.6 EXAMPLE TASK OBJECTIVE

The example starts with the user’s issue-search request as the root plan, then decomposes it into scoped subplans whose navigation history can be selectively replayed.

B.7 AGENTOCCAM ’S WORKFLOW REGARDING THE PLANNING TREE

  • The task objective becomes the root plan, and the agent branches into navigating to the Issues page before searching for the target issue.The example objective is to open the latest issue whose title includes “feature” and check whether it is closed.Source: §B.7 · pp. 16–17
  • The agent creates separate subplans to find and check the latest matching issue and to open it, then performs navigation for each active plan.The example searches for “feature,” sorts issues, and then proceeds to the plan for opening the latest matching issue.Source: §B.7 · p. 17
  • Scoping navigation to the active plan lets the agent replay only the corresponding actions and web observations as task history.In the example, two new subplans share a parent plan, shaping the planning tree.Source: §B.7 · p. 17

C AGENTOCCAM’S GENERALIZABILITY TO REAL-WORLD WEBSITES

On the WebVoyager definite-answer subset, AgentOccam performs better than Agent-E overall, though their strengths differ by website.

C AGENTOCCAM’S GENERALIZABILITY TO REAL-WORLD WEBSITES

  • AgentOccam performs better than Agent-E on WebVoyager’s definite-answer subset, with Agent-E stronger on Wolfram Alpha and AgentOccam stronger on Google Search and Hugging Face.The comparison uses WebVoyager tasks on real-world web environments.Source: Table 9 · p. 17, C AGENTOCCAM’S GENERALIZABILITY TO… · p. 18
  • The evaluated WebVoyager subset contains 129 definite-answer tasks across 13 real-world web environments.The authors use golden-answer questions to avoid subjective human evaluation and omit other tasks whose answers may change over time.Source: C AGENTOCCAM’S GENERALIZABILITY TO… · p. 18
  • The authors replicate Agent-E on that subset with the same definite-answer evaluators and run each task once for both agents.Agent-E is the baseline because it had reported the previous state of the art on the full WebVoyager benchmark using human assessments.Source: C AGENTOCCAM’S GENERALIZABILITY TO… · p. 18

D AGENTOCCAM’S GENERALIZABILITY TO BASE MODEL FAMILIES

A WebArena development-set study reports that AgentOccam’s alignment components benefit performance with Gemini as well as GPT-4-turbo; the appendix also details evaluator rectifications.

D AGENTOCCAM’S GENERALIZABILITY TO BASE MODEL FAMILIES

  • The Gemini ablations use a 190-task WebArena development set sampled from task clusters, with GPT-4-turbo results on the same set for comparison.Each task is run once, with restarts when the simulator fails.Source: D AGENTOCCAM’S GENERALIZABILITY TO… · pp. 18–19
  • Each AgentOccam alignment component contributes to overall performance with Gemini, supporting generalization across the tested base models.The study uses Gemini-1.5-flash, trained with different data from the GPT model family, for its ablations.Source: D AGENTOCCAM’S GENERALIZABILITY TO… · pp. 18–19
  • Removing non-essential actions encourages more active use of click and type, while disabling scrolling helps when key information is not initially visible.These component effects are reported for the Gemini model.Source: D AGENTOCCAM’S GENERALIZABILITY TO… · p. 19
  • Observation simplification reduces token counts, while selective replay within a page reduces the steps needed to complete tasks.Planning and selective replay of past pages also lets the agent organize workflow and return to an earlier stage after failed attempts.Source: D AGENTOCCAM’S GENERALIZABILITY TO… · p. 19

E.1 RECTIFICATION CATEGORIZATION

  • Evaluator changes are limited to cases judged erroneous; the authors retain original evaluators when task definitions or metrics may be misleading.They identify label errors and improper evaluation-function selection as two categories of modifications.Source: §E.1 · p. 20
  • Rectifications address erroneous evaluator definitions or answer typos, as well as matching functions that reject equivalent or more complete answers.Examples include a URL evaluator that accepts only one URL for identical page content and exact matching that rejects “December 15th” when “Dec 15th” is specified.Source: §E.1 · pp. 20–21

E.2 DETAILS

This section details task-level reference answers and evaluation checks across varied web tasks, including exact, fuzzy, must-include, and programmatic checks for Reddit posting, product reviews, and administrative workflows, alongside captions for action, step-count, and observation-token statistics tables.

E.2 DETAILS

  • Shown travel-time/distance tasks include Task 16 (fuzzy match: driving 2min, walking 16min), Task 97 (must include 914km), and Tasks 363–365 and 367 (748m, 1.7km, 2.2km, and 1.4km exact matches).Task 363 also has a must-include answer of 778m.Source: §E.2 · pp. 21, 24
  • Other evaluation entries check specific answer strings or required content, including names, repository URLs, product details, and page-status answers.Checks include exact-match or must-include criteria, with fuzzy-match alternatives in some tasks.Source: §E.2 · pp. 22–23
  • Product tasks specify category-specific listing URLs and price ordering, while order tasks instantiate a shared prompt template with different statuses.The listed categories include video games, health and household, competitive swimwear, and home and kitchen; order statuses include completed, cancelled, pending, processing, and out of delivery.Source: §E.2 · pp. 23–24
  • Programmatic checks target selected page elements or text for web-admin, forum, and Reddit tasks, including page titles, descriptions, sidebar fields, and post content.For instance, tasks 415–418 inspect merge-request or project-page content, while tasks 583–584 check forum fields.Source: §E.2 · pp. 24–25
  • Reddit task checks specify required wording for a city-living question, a Pittsburgh opinion post, and virtual meetup announcements about books, Harry Potter, Big Little Lies, and racing cars.The meetup checks include dates and require the post to identify the event as virtual.Source: §E.2 · p. 26
  • Other Reddit checks cover a Tears of Kingdom meetup and posts asking how machine learning, Midjourney, open-source LLMs, or large language models could help a field.These checks specify the subreddit or forum and required topic wording.Source: §E.2 · pp. 26–27
  • Product-review checks require refund language, a statement that an item broke after three days of use, and matching item identifiers and prices.The specified identifier-price pairs include B087QJN9W1 with 000000180, B09P7BFL4H with 161, B0041MSF2S with 180, and B003FVW3VA with 148.Source: §E.2 · pp. 27–28
  • Administrative-task checks require a completed-status filter and sales-report date ranges covering 2022, 2023, and May 1, 2021 through March 31, 2022.The date fields are checked for exact start and end dates.Source: §E.2 · p. 28
  • Task 786's reference answers include 412 and 414.Source: §E.2 · p. 28
  • The appendix includes tables captioned as action statistics, average steps per task across WebArena sites, and average observation tokens per step across those sites.The supplied passages give table scopes but no numerical table values.Source: Table 14 · p. 28, Table 15 · p. 28, Table 16 · p. 28

F ADDITIONAL EXPERIMENT DETAILS

The appendix reports how compound policies change AgentOccam’s WebArena success patterns and supplies additional trial, ablation, and implementation details. Its prompt and action specifications describe how AgentOccam plans and interacts with web pages.

F ADDITIONAL EXPERIMENT DETAILS

  • Table 17 reports AgentOccam’s WebArena success rates for an ablation study; the appendix says these figures are also visualized in Figure 5.The supplied caption identifies the metric and benchmark but gives no ablation values.Source: Table 17 · p. 29, F ADDITIONAL EXPERIMENT DETAILS · p. 29
  • AgentOccam finishes tasks with the fewest steps, often 9, and the three agents have comparable orders of magnitude in token consumption.The appendix also notes that the agents actively issue environment-changing actions such as click and type.Source: F ADDITIONAL EXPERIMENT DETAILS · pp. 28–29
  • AgentOccam + SteP gains success on shopping tasks with detailed instructions but falls short on shopping-admin and Reddit tasks, while AgentOccam + Judge has similar patterns with some denser success blocks.The Judge pattern is attributed to behavior rectification by its action-generation and selection pipeline.Source: F ADDITIONAL EXPERIMENT DETAILS · p. 29, Figure 6 · p. 29

G.1 AGENTOCCAM

  • AgentOccam’s prompt supplies the task, current step, page observations, prior plans, and interaction history, and asks for one action in a specified format.The prompt also asks the agent to highlight relevant current-page element IDs and rank them by relevance and potential value.Source: §G.1 · p. 31
  • The planning commands branch and prune create a subplan linked to a parent plan or return to a previous plan state; navigation commands cover actions such as click, type, go back, note, stop, and go home.The specifications direct branch and prune to change the plan rather than interact with the web environment.Source: §G.1 · p. 31
  • The observation example represents a Magento dashboard with numbered web elements, including links, text, tabs, and tables.The example includes dashboard data such as product, customer, and search-term tables.Source: §G.1 · pp. 32–33

G.2 JUDGE USED IN AGENTOCCAM + JUDGE EXPERIMENTS

The Judge prompt assesses plan progress and candidate web actions before selecting one action. It distinguishes planning commands, which modify the prior plan, from commands that interact with the web environment.

G.2 JUDGE USED IN AGENTOCCAM + JUDGE EXPERIMENTS

  • The Judge evaluates candidate actions by value and risk, considering best- and worst-case outcomes, then selects one action.Its prompt frames selection as maximizing future objective-fulfillment reward while minimizing risk.Source: §G.2 · p. 33
  • Branch and prune modify the previous plan section without interacting with the web environment.The prompt separately directs the Judge to select only one action.Source: §G.2 · p. 33
  • The prompt asks the Judge to review why plans have not been fulfilled or the objective remains unmet, supporting that assessment with evidence from objectives, observations, and actions.The prescribed format itemizes milestones and why each was unsuccessful.Source: §G.2 · p. 33

H FUTURE DIRECTION: TRAINING EVOLVING WEB AGENTS

The authors propose moving beyond heuristic action and observation alignment built for non-trainable base models. They suggest adaptable agents that refine their web mappings and skills as web layouts and human needs evolve.

H FUTURE DIRECTION: TRAINING EVOLVING WEB AGENTS

  • The current action and observation alignment is based on human heuristics and assumes that base models are not trainable.This assumption excludes some potentially useful web actions or element formats that could be learned with minimal examples.Source: H FUTURE DIRECTION: TRAINING EVOLVING… · p. 33
  • Adding scrolling could reduce observation volume and better match real websites, where shopping sites may load more products as users scroll.The authors suggest interaction traces could help an LLM learn scrolling without extensive training.Source: H FUTURE DIRECTION: TRAINING EVOLVING… · p. 33
  • The authors envision agents autonomously refining observation and action mappings and skills to adapt to changing web layouts and human needs.They suggest lifting the non-trainability assumption might improve performance and reduce operational costs.Source: H FUTURE DIRECTION: TRAINING EVOLVING… · p. 33
Loading 2410.13825v2…