acceptodds
Withdrawn from ICLR 2027

Action-Equivalent Memory Sketches for Web Agents

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Web agents powered by large language models often fail not because the underlying model is too weak, but because the memory that conditions the policy is semantically faithful yet action-irrelevant. We formalize this gap as an action-equivalent memory sketch: a compact history representation that preserves exactly the information capable of changing the next action, operationalized as an eight-field structured sketch (goal, sub_goal, completed, failed, constraints, entity_bindings, pending, evidence) updated at every step by a single LLM call. Three results bound its content: a soundness theorem for policy equivalence, an information-theoretic lower bound on sketch size, and a minimality theorem showing the eight fields are the unique action-complete cover. We evaluate on Mind2Web test_website across three runs. On qwen-turbo (10 task, 12 conditions) ours reaches Step-SR 0.0964; the B3 diagnostic, which removes constraints and failed, drops Step-SR to 0.0723, a 25% relative decrease consistent with the minimality theorem. On GLM-5.3-Flash (10 task, 8 per-field ablations) the direction reverses: B3 reaches 0.1233 versus ours 0.1096, with sub_goal the largest up-move and constraints the only down-move, opposite to what the theorem predicts. Suspecting a sample-size artifact, we ran a 50-task replication restricted to B3 and ours (438 step-judgments per condition, horizons 2 to 16). The direction returns to ours above B3 at 0.1347 versus 0.1187, a 13.5% step-level gain. The two-backbone picture is that the sketch works on both backbones once the sample is large enough, and 10-task direction flips are sample-size artifacts.

Primary area:
reinforcement learning
Keywords:
web agents, memory, structured sketch

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_CM6GOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes Action-Equivalent Memory Sketches, an eight-field structured memory representation for LLM-based web agents that preserves action-relevant information. The authors provide theoretical analyses of action equivalence and memory minimality. Experiments on Mind2Web with two LLM backbones demonstrate improvements over selected baselines.

Critical strengths

The paper addresses a practical issue in long-horizon web agents, and the proposed memory structure is simple and easy to implement without additional training.

Critical weaknesses
  1. Concerns about manuscript quality. The paper raises concerns about potentially being produced through an automated research process (AutoResearch), given its unusually short main content of only six pages, complete absence of figures, and rough writing throughout. These issues raise concerns about the overall quality and rigor of the work.
  2. Insufficient experiments. The experimental evaluation is very limited, with only two result tables and no comprehensive comparisons with competitive existing methods. The results are insufficient to support the paper's claims.
Questions for authors

See above weaknesses.

Reviewer_MtmZOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper presents a new idea for web agent design. Web agents are LLM-powered systems capable of taking observations from a webpage and executing a multi-step goal on them. The authors claim that the web agents can fail even when high quality underlying LLMs because the "memory" that the agent uses to prompt the LLMs/policy, is faithful to the trajectory semantically but not relevant to determine the action that should be taken. They introduce a compact history representation within their web agent design consisting of 8 fields: goal, sub goal, completed, failed, constraints, entity bindings, pending, evidence. They use 10-50 task samples from the Mind2Web (https://huggingface.co/datasets/osunlp/Mind2Web) dataset's test split to show that adding this 8-field memory within their agent prompt achieves step success rate: Step-SR 0.0964 on model qwen-turbo, and 0.1096 on model GLM-5.3-Flash. On qwen-turbo, their setting achieves 18-44% higher results than other baseline and control settings that they test. However, their setting does not beat all other baselines with GLM-5.3-Flash.

Critical strengths

None

Critical weaknesses
  1. The contribution is not novel. Many existing works have explored maintaining various kinds of history and memory states for long-horizon web agents. A lot of such agent designs can be found on leaderboards like these (https://docs.google.com/spreadsheets/d/1M801lEpBbKSNwP-vDBkC_pF7LdyGU1f_ufZb_NWNBZQ/edit?gid=0#gid=0 for WebArena). Some examples are Jace AI by Zeta Labs, AgentLab by ServiceNow[1], AutoWebGLM[2] etc.
  2. The paper introduces 2 theories that are not sufficiently proven, for instance, Theorem 2 stating that this 8 field representation is the minimal set that is required to represent the trajectory state gets disproved by the result in Table 3 where removing some fields increases the success rates, violating minimality. Theorem 2 is quite strict in its formulation and needs further proofs before the community can use it.
  3. The numbers shown on Mind2Web are quite low, the sota agents on this benchmark attain much higher results. The original paper MindAct[3] itself had shown higher Step SRs. The community has also moved to newer variants of this benchmark, namely, Mind2Web 2 and Online Mind2Web.
  4. The paper will benefit from using the entire test split of Mind2Web instead of just using a smaller subset of the data. It would also be better to replicate these results on other web agentic benchmarks to prove the validity of the theorems and generalization of the proposed agent design.

[1]-The BrowserGym Ecosystem for Web Agent Research (https://github.com/ServiceNow/AgentLab, https://openreview.net/forum?id=5298fKGmv3) [2]-AutoWebGLM: A Large Language Model-based Web Navigating Agent (https://arxiv.org/pdf/2404.03648) [3]-MIND2WEB: Towards a Generalist Agent for the Web (https://arxiv.org/pdf/2306.06070)

Questions for authors
  1. L131: what’s A? Its not defined in the formula
  2. The references to Sections and Appendices are broken throughout the paper (eg, L51, L153)
  3. L46: Can we cite examples for each of these kind of agents, for example: agents using retrieval-based memory
  4. L21: The abstract forward-references sections, for example, B3 diagnostic. Similar forward references exist in the introduction as well without providing sufficient context of what these experiment names and numbers mean.
  5. L51: Can you clarify what12 conditions mean in "(10 task, 12 conditions)"?
  6. Can we clarify what was the setting for constructing the evaluation dataset in the paper? What was the sampling protocol, task difficulty etc?
Reviewer_2BheOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper is really unclear but from what I managed to understand it seems to focus on web agents based on frozen large language models. More specifically, given an action policy of the agent, π(a/s,H)\pi(a/s, H), where HH is the full history, the paper attempts to build a memory representation of HH using an eight-field structured sketch.

Critical strengths

None.

Critical weaknesses

This is a very poorly written paper (clearly below ICLR standards and heavily based on LLMs). There are significant presentation and writing issues that make the paper unclear: Unclear motivation/positioning w.r.t state-of-that-art; no details (both implementations and technical/mathematical details, etc.). Here are some examples that illlustrates how poorly this paper is written (among many others):

  • Many mathematical notations are introduced without proper definitions. For example, what is |M|_bytes in Theorem 1? What is variable A in the same Theorem. By the way, the theorem seems to be a standard information-theory result from the the well-known Thomas and Cover text book but it is not clear at all how this even relate to the proposed work. The information-theoretic bounds that the authors are providing seem to come out of the blue from information theory textbooks.

  • There are undefined procedures. For example, what is UpdtaeSketch in section 3.1. BTW, the authors used only 5 pages for the paper so there is room to detail the procedures.

Questions for authors

No further questions.

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

Related papers

Discussion (0)

Sign in to comment.