WebCoach: Self-Evolving Web Agents with Cross-Session Memory Guidance

Genglin Liu; Hejie Cui; Sarah Zhang; Sha Li; Shijie Geng; Tianyi Liu; Xin Liu

REVIEW 5 major objections 5 minor 2 cited by

WebCoach: Self-Evolving Web Agents with Cross-Session Memory Guidance

T0 review · 5 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read WebCoach, a memory-and-coach layer wrapped around any web agent, lifts task success for a 38B model from 47% to 61% on live web tasks without retraining, while keeping step counts flat.

desk verdict Plausible but not established: the 14-point gain rests on self-reported success and no error bars, so the paper merits refereeing, not citation yet. read the letter →

arxiv 2511.12997 v2 pith:AKIMIPVL submitted 2025-11-17 cs.AI cs.CL

Genglin Liu , Shijie Geng , Sha Li , Hejie Cui , Sarah Zhang , Xin Liu , Tianyi Liu This is my paper · ORCID

classification cs.AIcs.CL

keywords webagentscross-sessionmemoryself-evolvingepisodicretrieval-augmentedcoachingcontinuallearningnavigationexternalstore

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that web agents do not need to be retrained to get better at browsing: they can improve by remembering and being advised by their own past episodes. It introduces WebCoach, a plug-in layer of three modules — a condenser that summarizes navigation logs, an external store that indexes those summaries, and a coach that retrieves relevant memories and injects short advice mid-task. On the live WebVoyager benchmark, pairing WebCoach with a 38B open-source model raises success from 47% to 61% while holding or reducing the number of steps, and lifts a 32B model to near-GPT-4o level. The central claim is that experience-driven memory, not scale or retraining, is what closes much of the gap between open-source and proprietary web agents.

What carries the argument

The central mechanism is a three-part memory loop operating outside the agent's policy: (1) WebCondenser, a small LLM that turns raw (observation, action, reward) traces into a fixed schema of summary text, an embedding, and success/failure metadata; (2) External Memory Store, a vector database with approximate nearest-neighbor search that stores only completed episodes and retrieves top-K experiences by cosine similarity and recency; (3) Coach, an 8B LLM that, given the current partial-trace summary and K=5 retrieved memories, returns a JSON decision to either stay silent or inject one-to-two sentences of advice as a system message before the agent's next action.

What would settle it

Run WebCoach in its dynamic self-experience mode on a held-out set of websites and task templates that never appear in the External Memory Store, and compare success against the no-memory baseline; if the gain over baseline shrinks to near zero, the reported improvement is largely in-benchmark leakage rather than transferable cross-session learning.

Watch

Extended reading notes

Core claim

The paper's central claim is that a lightweight, model-agnostic wrapper can give web navigation agents persistent episodic memory and selectively inject advice at runtime, producing continual self-improvement without any parameter update. WebCoach standardizes raw navigation traces into concise summaries (WebCondenser), stores completed episodes in a vector memory indexed by embedding similarity and recency (External Memory Store), and calls a small 8B LLM (Coach) that decides whether to interrupt the agent with task-specific guidance such as 'avoid this link, previous agents got stuck in a loop here.' The authors report consistent gains across three open-source backbones, with the largest j

Load-bearing premise

The dynamic self-experience evaluation assumes that memories pulled from the same benchmark's other tasks—often on the same websites and task templates, since only same-task-ID examples are excluded—are a fair measure of self-evolution; if the memory is instead exploiting the benchmark's task distribution, the 14-point gain would overstate what a deployed agent would gain from its own past interactions.

Editorial extensions

If this is right

Any web agent, regardless of backbone, could improve over time by accumulating and consulting its own completed trajectories, meaning continual learning can be achieved without retraining.
Gains concentrate in semantically complex, multi-step sites (Apple, arXiv, BBC) while simpler click-heavy domains show little change, implying memory helps most where planning and element disambiguation matter.
Self-generated memory beats externally seeded memory, so agents learn most from trajectories produced by their own policy and representation space, not from demonstrations of a different model.
Provided the underlying agent is above a cognitive threshold (the 7B model fails, 32B/38B benefit), the coaching signal amplifies with agent scale; memory guidance pays off at the boundary of partial competence.
The 8B coach performs comparably to a much larger proprietary LLM as a coach, and a DPO fine-tune was deemed unnecessary, so the entire memory layer can run on small open-weight models.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural extension is to apply the same condenser-store-coach loop to other interactive agents—mobile UIs, desktop automation, or robotics—wherever raw action traces can be standardized; the paper's abstractions are domain-agnostic.
The reported 14-point gain may overstate real-world self-evolution because the memory is populated from the same benchmark's tasks, and only same-task-ID memories are excluded; trajectories from the very websites the agent is about to visit can be retrieved, so part of the gain could come from in-benchmark leakage rather than transferable learning.
Because retrieval uses summary embeddings, the framework inherits the condenser LLM's summarization quality and the embedding model's semantic coverage; failures in either would degrade advice even if the store and coach are sound.
Average completion time rises substantially (e.g., roughly 150 seconds added) due to Condenser and Coach inference, though step counts drop; a latency-aware analysis could determine whether the throughput trade-off is acceptable in production.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The 14-point success gain for Skywork-38B is the thing to know about this paper. It is a plausible result, but you should not read it as established: the success signal comes from the agent itself, not from WebVoyager's human rubric, and there are no uncertainty estimates anywhere. That combination means the headline number could be inflated.

What the paper does well is straightforward. WebCoach is a clean, model-agnostic wrapper: a condenser normalizes logs, a vector store holds completed episode summaries, and a coach retrieves top-k memories and decides when to inject advice into the base agent's system prompt. That is a reasonable extension of reflexion-style memory work. The evaluation has real scope—all 643 WebVoyager tasks, live browser, three open-source backbones and a GPT-4o ceiling—and the per-subtask tables in the appendix let you see where the gains come from. The result that self-generated memory beats externally seeded memory is the most interesting finding. Steps per task stay flat or drop, which suggests the gains are not from extra exploration. Code is available.

The soft spots are proportionate to the claims. The self-evaluation issue is the biggest. Section 3.1 says the browser-use agent checks the final state against the query to decide success, and the appendix trace shows the agent simply outputting 'success: true.' There is no spot-check against human labels, so the coached condition and the baseline could have different false-positive rates. The 14-point gain could be partly a measurement artifact. The paper needs a human-judged subset or an independent judge applied to both conditions.

Second, the 'consistently improves' claim in the abstract is contradicted by the 7B results in Table 1, where success drops from 0.328 to 0.288–0.311 in every WebCoach condition. The paper does acknowledge this in Section 4, so the authors are not hiding it, but the abstract still overstates. With no confidence intervals or significance tests, the close differences between conditions could be run-to-run noise.

Third, leakage control excludes only the same task ID. Retrieval can still surface memories from the same domain with nearly identical task templates, so the dynamic memory condition is partly memorizing the benchmark. A held-out memory split or cross-domain test would be needed to show the gains generalize.

Fourth, there is no comparison against existing memory baselines like Reflexion, Mem0, or A-Mem, so the contribution beyond the specific combination is hard to calibrate.

Who gets value from this: people building low-cost memory layers on top of existing web agents. The architecture is easy to understand and reuse. This deserves a serious referee, but the evaluation needs substantial revision before the central claim is credible.

Referee Report

5 major / 5 minor

Summary. The paper introduces WebCoach, a model-agnostic memory-and-coaching layer for web-navigation agents. WebCoach uses a WebCondenser to compress trajectories, an external memory store (EMS) for retrieval, and a Coach LLM that injects advice into the actor's prompt. The authors evaluate on all 643 WebVoyager tasks with three open-weight base models (Qwen2.5-VL-7B, Qwen2.5-VL-32B, Skywork-r1v3-38B) and report that WebCoach improves success rates, e.g., 47.3%→61.4% for Skywork-38B, while roughly maintaining step counts. They also claim that a self-expanding memory store outperforms a frozen memory seeded with GPT-4o trajectories, and that smaller models with WebCoach approach GPT-4o-level performance.

Significance. If the results are robust, the paper makes a practically useful contribution: a lightweight, model-agnostic way to give web agents persistent cross-session memory without retraining. The full-benchmark evaluation on live websites, the per-subtask breakdowns, and the release of code are strengths. However, the headline claims rest on evaluation choices that need substantially more support: success labels come from the agent itself, the dynamic-memory evaluation draws on the same benchmark distribution it is tested on, and no uncertainty quantification is provided. These issues are load-bearing for the paper's central claims, so the manuscript currently overstates its evidence.

major comments (5)

[§3.1, §4, Appendix A] The success signal used for the headline metric is the browser-use agent's own completion check ('we utilize browser-use agent's evaluation capability by checking the resulting state of their last action against the initial user query'). No human validation or external judge is described. Since WebCoach changes the agent's behavior and injects task-specific advice, the same model that performs the task also certifies success; the Appendix trace shows the main agent returning 'success: true' directly. If coached agents are more likely to declare success, the reported gain could be partly a measurement artifact. The authors should validate a random subset against human/external labels and, ideally, report false-positive rates per condition.
[§3.2, 'Memory Retrieval and Leakage Control', Tables 1 and 5] The dynamic self-experience evaluation populates the EMS from WebVoyager tasks on the same 15 websites and similar task templates, and retrieval excludes only the same task ID. This means the agent is evaluated on a task distribution that is represented in its memory store; the later tasks in the sequence can retrieve memories of near-identical tasks from the same domain. The reported 14-point Skywork gain therefore conflates genuine cross-session learning with in-benchmark distribution leakage. A cleaner evaluation would hold out entire domains or task templates, or seed the memory exclusively from outside the evaluation benchmark, and would report how the gain changes as memory is withheld.
[Abstract, §4, Table 1] The abstract says WebCoach 'consistently improves the performance of browser-use agents across three different LLM backbones,' but the results show Qwen-VL-7B at 0.328 baseline vs. 0.311 with dynamic self-experience WebCoach (Table 1). The text in §4 acknowledges this ('the 7B backbone does not benefit'), but the abstract and introductory framing are not corrected. This is a direct contradiction of a core claim and must be fixed. The framing should be scale-dependent, not 'consistent.'
[Table 1, Tables 3–5] No error bars, confidence intervals, or significance tests are reported anywhere, despite 643-task runs. For example, Qwen-VL-32B goes from 319/643 to 367/643, a difference of 48 successes; the 7B regression is 21 successes. Without at least binomial confidence intervals or a paired test, it is impossible to tell which of these differences are beyond run-to-run noise. The paper's central quantitative claims need this reporting. Multiple seeds, or a bootstrap, would strengthen the conclusions considerably.
[Table 2 vs. Table 1 / Figure 4] The baseline success rate for Qwen-VL-7B is inconsistent: Table 2 shows 0.344 (221/643), while Table 1 and Figure 4 show 0.328. Since the 7B result is already a key negative case, this discrepancy must be resolved. The corrected value affects the interpretation of the 'no benefit' conclusion.

minor comments (5)

[§2.1, §2.3] The WebCondenser and Coach are said to use 'a small LLM (≤8B)' and 'an 8B LLM,' but no specific model, prompt template, or temperature is given. For reproducibility, these details should be provided.
[Figure 4] The legend labels are confusing: 'WebCoach (GPT-4o)', 'WebCoach (Qwen3-8b)', 'WebCoach with Self-exp', and 'GPT-4o baseline' do not map clearly to the table columns. Clarify which color/hatch corresponds to each condition.
[§3.1] The sentence 'WebVoyager complements this by enabling large-scale crawling and evaluation' seems to misstate the benchmark's role; WebVoyager is the benchmark, not a crawling tool. Rephrase.
[§4, 'Effect of memory source'] The claim that self-generated embeddings result in 'smoother reasoning continuity' is not supported by quantitative evidence; if this is a hypothesis, label it as such, or provide an analysis (e.g., embedding distances or failure-mode comparisons).
[§2.2, Figure 2] The retrieval latency is measured for only 600 trajectories. The claim that HNSW supports 'millions of episodes' is extrapolation; this is fine as a design note, but should not be stated as an empirical result.

Circularity Check

2 steps flagged · score 6.0 of 10

The 14-point self-evolution gain is supported by in-benchmark memory and self-certified success labels.

fitted input called prediction [Section 3.2 'Memory Retrieval and Leakage Control'; Table 1 'Self (Dynamic)' row; Section 4 'Effect of memory source and self-experience']
"‘During evaluation, retrieval explicitly excludes any episode whose WebVoyager task ID matches that of the current task ... The EMS returns only experiences from distinct tasks or from prior runs of different subtasks.’ (Sec 3.2); ‘Dynamic EMS (Qwen3-8B coach): each main agent iteratively updates its own trajectories into the memory store’ (Table 1)."

WebCoach's self-evolution condition is evaluated on WebVoyager while the EMS is filled with WebVoyager trajectories from the same evaluation run. The exclusion only removes same-task-ID episodes, not same-domain or similar-template tasks, so later test tasks are coached with memories of earlier test tasks from the same sites. The headline 47%→61% dynamic self-experience gain is therefore partly a measure of in-benchmark memorization; the agent's memory is literally constructed from the test distribution it is then scored on.
self definitional [Section 3.1 'Data'; Appendix A 'Main agent returned results']
"‘For evaluation, we utilize browser-use agent’s evaluation capability by checking the resulting state of their last action against the initial user query, and determine the success of the task.’ (Sec 3.1); ‘[Main agent returned results:] is done: true, success: true’ (Appendix A)."

The outcome variable supporting the headline is not an external ground-truth label; it is the browser-use agent’s own determination. WebCoach is a wrapper that injects advice into that same agent, so the model that performs the task also certifies completion. The Appendix shows the coached agent simply returns “success: true”. Thus the 14-point gain can be produced by a change in self-evaluation rather than verified task completion; without human-rubric validation or a false-positive-rate comparison across conditions, the measured success is endogenous to the system being evaluated.

full rationale

WebCoach is an empirical systems paper, not a derivation, so most classical circularity patterns (self-citations, imported uniqueness, ansatz-by-citation) are absent. The load-bearing problem is in the evaluation design. First, the dynamic EMS condition populates memory from the same WebVoyager test tasks it is scored on, excluding only identical task IDs; this is training-on-the-test-set, so the claimed self-evolution advantage is partly memorization of the benchmark distribution. Second, success is determined by the browser-use agent’s own evaluation capability, with the Appendix showing the coached agent returning “success: true”; the headline improvement is therefore measured with the same model that is being modified, not an independent judge. These two issues make the central quantitative claim partially circular by construction, even though the system itself is a real engineering contribution and the frozen-memory results may retain some independent value. No self-citation is load-bearing. Score 6 reflects partial circularity of the central empirical claim.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim depends on retrieval quality (embedding assumption), coach reliability, self-evaluation validity, and the experimental protocol for memory. No new scientific entities are invented. The largest hidden burden is the evaluation protocol: memory is populated from other tasks in the same WebVoyager benchmark, which may leak task distribution.

free parameters (3)

top-k retrieved experiences = 5
The Coach receives the top-K=5 similar past experiences. The value is chosen by hand ('we pick k=5 because presenting five closest past experiences gives the coach enough varied examples') and could affect the quality of advice.
embedding model and dimension = OpenAI 1536-d embedding
The WebCondenser uses a 1536-d OpenAI embedding vector for summaries. The paper says this block could be replaced, but the choice affects retrieval quality and is not sweeped.
per-action timeout and step cap = 30s timeout, 50-step cap
These evaluation constraints bound runtime and prevent degenerate loops; they are experimental choices that can affect success rates for tasks requiring more steps.

assumptions (4)

domain assumption Embedding similarity in summary space corresponds to task-relevant similarity for web navigation.
The entire retrieval mechanism rests on cosine similarity over OpenAI embeddings of text summaries. If similar tasks do not produce similar summaries, retrieval will be irrelevant. No retrieval-quality analysis is provided.
domain assumption The Coach (an 8B LLM) can reliably generate correct and actionable advice from the retrieved summaries.
The Coach is prompted to decide when to intervene and what to say. The paper does not verify that the advice is factually correct or that it does not mislead the actor; only the final success rate is measured.
domain assumption The browser-use agent's self-evaluation of task success is a valid proxy for true task completion.
The paper states 'we utilize browser-use agent's evaluation capability by checking the resulting state of their last action against the initial user query.' This is not human-verified and could be biased toward the agent's own behavior.
standard math FAISS/HNSW approximate nearest-neighbor search returns the true top-k neighbors often enough for the system to work.
The paper relies on the standard FAISS HNSW library; this is uncontroversial but is an implementation assumption about retrieval quality.

how reviews work

0 comments

Cite this review

Pith. "Pith review of WebCoach: Self-Evolving Web Agents with Cross-Session Memory Guidance." pith.science (2026). https://pith.science/paper/AKIMIPVL

@misc{pith2026251112997,
  author       = {Pith},
  title        = {Pith review of: WebCoach: Self-Evolving Web Agents with Cross-Session Memory Guidance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AKIMIPVL}},
  note         = {Machine review of arXiv:2511.12997}
}

read the original abstract

Multimodal LLM-powered agents have recently demonstrated impressive capabilities in web navigation, enabling agents to complete complex browsing tasks across diverse domains. However, current agents struggle with repetitive errors and lack the ability to learn from past experiences across sessions, limiting their long-term robustness and sample efficiency. We introduce WebCoach, a model-agnostic self-evolving framework that equips web browsing agents with persistent cross-session memory, enabling improved long-term planning, reflection, and continual learning without retraining. WebCoach consists of three key components: (1) a WebCondenser, which standardizes raw navigation logs into concise summaries; (2) an External Memory Store, which organizes complete trajectories as episodic experiences; and (3) a Coach, which retrieves relevant experiences based on similarity and recency, and decides whether to inject task-specific advice into the agent via runtime hooks. This design empowers web agents to access long-term memory beyond their native context window, improving robustness in complex browsing tasks. Moreover, WebCoach achieves self-evolution by continuously curating episodic memory from new navigation trajectories, enabling agents to improve over time without retraining. Evaluations on the WebVoyager benchmark demonstrate that WebCoach consistently improves the performance of browser-use agents across three different LLM backbones. With a 38B model, it increases task success rates from 47% to 61% while reducing or maintaining the average number of steps. Notably, smaller base models with WebCoach achieve performance comparable to the same web agent using GPT-4o.

Figures

Figures reproduced from arXiv: 2511.12997 by the authors.

**Figure 1.** Overview of the WebCoach framework. WebCoach augments web-browsing agents with persistent, cross-session memory through an External Memory Store (EMS) and a retrievalaugmented coaching mechanism. The Condenser converts raw navigation histories into standardized summaries stored in EMS, from which the Coach retrieves relevant prior experiences to provide taskspecific guidance to the main web agent. This design enab… view at source ↗

**Figure 2.** Retrieval speed at k for the EMS with 600 trajectories. Repeat each 200 times to measure the consistency. Most runs end up averaging between 9.0 and 9.5 ms for k ranging from 1 to 10. Given the Condenser embedding et for the current partial trace, EMS returns the top-K past experiences using score(et, ei) = e ⊤ t ei ∥et∥2 ∥ei∥2 . Cold start. A key feature of the EMS is its generality: it is agnostic to the model, … view at source ↗

**Figure 3.** Asynchronous Evaluation of WebVoyager. WebCoach’s asynchronous evaluation pipeline distributes the 15 subdomains in WebVoyager (e.g., Amazon, Apple, ArXiv) across parallel evaluation queues to maximize throughput and GPU utilization. Yellow boxes indicates in-progress tasks, green indicates completed tasks, and blue indicates tasks that are waiting in the queue. Our limited compute supports running 5 tasks in parall… view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Performance comparison across base models. WebCoach consistently improves browser-use agents’ reasoning and robustness across different backbones. The framework achieves higher success rates with equal or fewer average steps, while maintaining efficient completion time…

**Figure 5.** Figure 5: Step 1 Screenshot 15 [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]

**Figure 6.** Figure 6: Step 2 Screenshot [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]

**Figure 7.** Figure 7: Step 3 Screenshot 16 [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]

**Figure 8.** Figure 8: Step 4 Screenshot B APPENDIX B: FULL PER-SUBTASK RESULTS Subtask GPT-4o Qwen-VL-32B Qwen-VL-7B Skywork-r1v3 SR S/T Time Steps SR S/T Time Steps SR S/T Time Steps SR S/T Time Steps Allrecipes 0.889 40/45 67 7.4 0.644 29/45 254 12.3 0.356 16/45 142 16.1 0.556 25/45 204 1…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

The Horizon Gap: Planning, Memory, Execution, Training, and Evaluation for Long-Horizon LLM Agents
cs.CL 2026-08 conditional novelty 5.0 of 10

A survey of 1,547 papers defines the 'horizon gap' and documents that long-horizon agent research is converging on trajectory-level process signals instead of outcome-only scores.
Automatic Hard Example Synthesis with Multi-Level Agentic Data Curation
cs.AI 2026-07 conditional novelty 5.0 of 10

Test-time retrieval of committee-disagreement-mined synthetic images cuts a safety classifier's false-negative rate on a hard HoliSafe subset from 41.2% to 24.5%.

Reference graph

Works this paper leans on

6 extracted references · cited by 2 Pith papers

[1]

Repeated Navigation to MacBook Air Page: The user repeatedly navigates to the MacBook Air product page, likely due to a misstep or a need to verify the correct page
[2]

Persistent Scrolling on MacBook Air Page: The user continues to scroll down the MacBook Air product page in an effort to find the color options section, indicating that they haven’t found the information yet
[3]

Repeated Search Bar Interactions: The user frequently interacts with the search bar, possibly due to confusion about the page structure or uncertainty about the location of the color options. Similarity score: 0.8895191550254822 [2 more relevant retrieved experiences omitted] [Main agent returned results:] is done: true, success: true, long-term memory: T...
[4]

name: Navigation to Product Page, description: The user navigated from the Apple homepage to the TV and Home section and then to the HomePod mini product page, which is a key step in accessing the product information
[5]

name: Color Option Identification, description: The user successfully identified the available colors (Blue, Yellow, Midnight, White, and Orange) on the product page, which is the main goal of the task
[6]

name: Scrolling and Searching for Color Options, description: The user attempted to locate the color options by scrolling through the product page, which eventually led them to the color selection section. embedding: (...omitted...) Figure 5: Step 1 Screenshot 15 Preprint Figure 6: Step 2 Screenshot Figure 7: Step 3 Screenshot 16 Preprint Figure 8: Step 4...

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [1]

Repeated Navigation to MacBook Air Page: The user repeatedly navigates to the MacBook Air product page, likely due to a misstep or a need to verify the correct page

[2] [2]

Persistent Scrolling on MacBook Air Page: The user continues to scroll down the MacBook Air product page in an effort to find the color options section, indicating that they haven’t found the information yet

[3] [3]

Repeated Search Bar Interactions: The user frequently interacts with the search bar, possibly due to confusion about the page structure or uncertainty about the location of the color options. Similarity score: 0.8895191550254822 [2 more relevant retrieved experiences omitted] [Main agent returned results:] is done: true, success: true, long-term memory: T...

[4] [4]

name: Navigation to Product Page, description: The user navigated from the Apple homepage to the TV and Home section and then to the HomePod mini product page, which is a key step in accessing the product information

[5] [5]

name: Color Option Identification, description: The user successfully identified the available colors (Blue, Yellow, Midnight, White, and Orange) on the product page, which is the main goal of the task

[6] [6]

name: Scrolling and Searching for Color Options, description: The user attempted to locate the color options by scrolling through the product page, which eventually led them to the color selection section. embedding: (...omitted...) Figure 5: Step 1 Screenshot 15 Preprint Figure 6: Step 2 Screenshot Figure 7: Step 3 Screenshot 16 Preprint Figure 8: Step 4...