acceptodds
Under review as a conference paper at ICLR 2027

Answer2Path: Outcome-Guided Acquisition of Reusable Web-Agent Experience

Abstract

Successful web-agent trajectories are costly to acquire, and terminal success alone does not show whether an agent gathered evidence or learned a reusable procedure. Yet many tasks already have a known answer or desired final state, typically used only for evaluation. Can a small pool of such outcomes guide trajectory acquisition and seed further supervision? We propose Answer2Path (A2P), a closed-loop framework that uses outcomes during exploration and expands supervision from verified webpage evidence. For queries, the agent seeks a path supporting the known answer; for modifications, the desired state guides action and post-state verification. We compare when the same outcome is revealed and whether it explicitly guides evidence seeking. A condition-blind companion records evidence, and independently validated page facts become provenance-preserving task–outcome pairs. Fixed workflow induction then tests whether acquired trajectories help on held-out tasks without outcome access. On 61 WebArena Shopping queries over three seeds, A2P raises pooled official success from 40.4% to 84.7%; at fixed K, induced workflows raise held-out success from 88.3% to 97.2%. Across six modification families, eligible source success rises from 22.9% to 54.2%, while source tokens per success fall by 75.0%. Late guidance has the highest aggregate yield in a two-family placement study, and a two-round single-family pilot validates backend-free, evidence-grounded task expansion. Together, these results show how limited outcome supervision can improve procedural experience acquisition and seed further verified tasks.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Web Agents, Outcome Supervision, Trajectory Acquisition, Workflow Transfer, Evidence-Grounded Data Flywheel

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_oPreOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Answer2Path uses known answers or desired states to guide web-agent trajectory acquisition. It audits webpage evidence, induces reusable workflows, and evaluates them on held-out tasks without outcomes. Experiments on WebArena Shopping also examine guidance timing, modification tasks, and a small loop for generating new task–outcome pairs.

Critical strengths
  1. Complementing reverse task synthesis such as OS-Genesis, the paper tests whether known outcomes yield experience that remains useful after those outcomes are removed. Fixed-K query transfer improves from 88.3% to 97.2% with identical workflow induction, and improves across all three source seeds.
  2. The symmetric evidence check also improves supported yield from 29.0% to 45.4%, beyond official-success gains alone. Leakage audits, token accounting, modification tasks, and independent scaffold/backbone replications give the acquisition finding substantial practical support.
Critical weaknesses
  1. The main fixed-K study does not include Full GT-only acquisition. Smaller placement controls are informative, but the added transfer value of explicit path-seeking guidance over simply revealing outcomes remains less certain than the overall outcome-conditioning benefit.
  2. The semantic evidence audit is incomplete: 124/155 combines literal passes with semantic review of failures. The symmetric mechanical audit supports improvement, but this provisional figure cannot establish the full rate of genuinely evidence-supported trajectories.
  3. Modification transfer includes A2P-only compact compilation. As disclosed, it evaluates the combined pipeline; it cannot independently confirm that better acquired trajectories caused the transfer gain. The clean query result carries the stronger evidence here.
  4. Pooled tests reuse tasks across seeds and nested K values, with shared workflows. The positive per-seed results are reassuring, but the clustering unit and effective sample size need clarification before interpreting the reported significance.
  5. Minor corrections: Table 11's A2P + same-family gain is 14.2 percentage points, rather than 14.1. Figure 3 should distinguish flywheel admission from the main experiments' eligibility rule. Keep the seven-pair expansion result explicitly framed as a feasibility pilot.
Questions for authors
  1. Can the existing Full GT-only trajectories be evaluated with the same fixed-K protocol to clarify the incremental value of path-seeking guidance?
  2. What does a condition-blind semantic audit of a matched sample from both methods show, including literal-check passes and failures?
  3. How do the confidence intervals and paired tests account for repeated tasks, nested K values, and shared workflows?
  4. How much of modification transfer remains when both conditions use identical workflow compilation?
Reviewer_j213OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

Answer2Path reveals a known answer or desired final state during web-agent exploration to guide the agent toward an evidence-supported path. Successful trajectories are then turned into reusable workflows. Across 61 Shopping query tasks and three seeds, official success rises from 40.4% to 84.7%. In a fixed-K experiment holding the number of successful source trajectories constant, success on held-out tasks without answer access rises from 88.3% to 97.2%

Critical strengths
  • The paper separates two questions: whether guidance yields more successful trajectories, and whether the same number of acquired trajectories is more useful. The fixed-K comparison is particularly valuable.
  • It examines query and modification tasks, as well as when and how outcome information is supplied. It records webpage evidence and provenance, then tests induced workflows on held-out tasks without revealing the outcome.
Critical weaknesses
  • The reported 124/155 evidence-supported query successes come from a provisional two-stage audit. The 83 trajectories that passed a literal check were not reassessed using the same semantic-review standard.
  • In the modification transfer experiment, only Answer2Path workflows receive additional compact compilation. The increase from 63.3% to 80.0% therefore cannot be attributed solely to better source trajectories.
  • The experiments focus on WebArena Shopping. The task-expansion demonstration covers one family and two rounds, while the query fixed-K baseline is already strong at 88.3%.
Questions for authors
  • What is the fixed-K modification transfer result when both conditions use exactly the same workflow compilation?
  • Could both conditions’ successful trajectories be sampled and blindly audited under one semantic-evidence standard?
Reviewer_XheSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a simple idea for constructing reusable agentic workflows based on an existing web-agent scaffold and the Agent Workflow Memory (AWM) method. It directly provides a known answer or desired outcome to guide the agent in collecting successful trajectories, which are then summarized into reusable workflows using AWM. The paper also compares different ways of introducing the answer, including early versus late and passive versus guided conditions, to study how answer placement affects trajectory acquisition and downstream transfer. Experiments are conducted on WebArena Shopping query and modification tasks across multiple seeds and functional families. The results show improvements over the no-answer baseline in acquisition success, token efficiency, evidence-supported yield, and fixed-K transfer to held-out tasks where the answer is no longer provided. Finally, the paper presents a small data-flywheel pilot, which extracts new task–outcome pairs from verified webpage evidence and uses them for subsequent trajectory acquisition and workflow memory updates.

Critical strengths

This paper has two strengths:

  1. The proposed approach is simple and easy to integrate with existing web-agent systems. The paper also clearly distinguishes acquisition efficiency from the transfer quality of the acquired experience through its fixed-K evaluation.

  2. The empirical evaluation is relatively careful and comprehensive, covering query and modification tasks, different seeds, task families, model backbones, and agent scaffolds. It reports not only success rates but also transfer performance, token efficiency, and several mixed or negative results.

Critical weaknesses

I have two main concerns that prevent me from supporting acceptance in its current form:

  1. It is unclear whether the gains come from evidence-guided acquisition or simply from providing the answer. In the main experiment, A2P differs from the no-answer baseline in both outcome exposure and explicit evidence-seeking guidance. Therefore, the large improvement in source-task success may mainly come from reducing the search space by directly providing the answer. The full-GT and answer-placement ablations are helpful, but they are conducted on a smaller subset or under a different protocol. In the matched supervision ladder, Full GT achieves comparable immediate success to A2P, while the additional gain in mechanically evidence-supported trajectories is relatively modest. Since the outcome-only condition is not evaluated under the same main-scale fixed-K transfer setting, it remains unclear whether evidence-guided acquisition produces substantially more reusable experience beyond making the source task easier.
  2. The current evidence for generality and scalability is limited. The fixed-K query baseline already achieves 88.3% success, leaving relatively limited room for improvement, and most transfer experiments are conducted within the same functional family. The weaker cross-family result suggests that the induced workflows may mainly capture family-specific procedures. The modification transfer comparison is also affected by applying compact workflow compilation only to A2P. Finally, the data flywheel is demonstrated for only two rounds in one task family, producing seven unique task–outcome pairs and five downstream evidence-supported acquisitions. These results demonstrate that the pipeline can operate, but do not yet provide strong evidence for broad transfer or scalable supervision expansion.
Questions for authors
  1. Can you evaluate a Full-GT outcome-only baseline under the same 61-task and fixed-K transfer protocol as A2P? This would help separate the contribution of evidence-seeking guidance from the effect of directly providing the answer.

  2. Can you apply the same semantic evidence audit to the no-answer, Full-GT, and A2P conditions? Can you also report fixed-K transfer using only evidence-qualified trajectories?

  3. Can you provide more cross-family or cross-website transfer results? Is the induced workflow still useful when the page structure or navigation procedure changes?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.