Writing study: Triple-BERT: Do We Really Need MARL for Order Dispatch on Ride-Sharing Platforms?

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Zijian Zhao, Sen Li · Official ICLR 2026 oral record

Observed writing choices, not acceptance rules. This study examines the inspected manuscript version. It does not establish why the paper was accepted or that its presentation choices should be copied unchanged.

Figure, table and section ledger

Editorial contribution type: Architecture & representation design (primary) · Learning & optimization algorithms.

Layout extraction, not a required paper template. Counts are distinct caption labels; a multi-panel figure counts once. Section spans are estimated from heading positions, including floats and whitespace. Method and limitations.

Canonical PDF

29 physical pages

Before references / appendix

≈ 10.52 page-equivalents

Main figures / tables

3 / 2

Attached appendix figures / tables

3 / 8

Attached appendix span

≈ 12.89 page-equivalents

Before references / appendix: ≈ 36.28%References / intervening material: ≈ 19.28%Attached appendix: ≈ 44.45%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Section-by-section lengths and source pages

The title/abstract/front matter precede the first numbered heading. A span ends at the next detected top-level heading, not at the last sentence; end-page whitespace is included. References and statements are listed separately. These are not measurements of text-only or artwork area.

Top-level heading	Starts on PDF page	Estimated span	Share of entire PDF
1 INTRODUCTION	1	≈ 1.12 pp.	3.86%
2 PROBLEM SETUP	2	≈ 1.48 pp.	5.1%
3 METHODOLOGY	4	≈ 5.04 pp.	17.38%
4 EXPERIMENT	9	≈ 1.39 pp.	4.79%
5 CONCLUSION	10	≈ 0.44 pp.	1.52%
ETHICS STATEMENT	11	≈ 0.32 pp.	1.1%
REPRODUCIBILITY STATEMENT	11	≈ 0.1 pp.	0.34%
REFERENCES	11	≈ 5.58 pp.	19.24%
A ACTION SPACE SIZE	17	≈ 0.32 pp.	1.1%
B BIPARITE GRAPH CONSTRUCTION	17	≈ 0.79 pp.	2.72%
C POLICY GRADIENT PROOF	18	≈ 1.3 pp.	4.48%
D TRAINING PROCESS	19	≈ 1.96 pp.	6.76%
E EXPERIMENT DETAILS	21	≈ 4.62 pp.	15.93%
F RELATED WORK	26	≈ 3.89 pp.	13.41%

All detected figure and table locations

Main text

• Figure 1 · p. 3
• Figure 2 · p. 4
• Table 1 · p. 9
• Figure 3 · p. 9
• Table 2 · p. 10

Attached appendix

• Figure 4 · p. 19
• Figure 5 · p. 19
• Table 3 · p. 21
• Table 4 · p. 23
• Table 5 · p. 24
• Table 6 · p. 24
• Table 7 · p. 24
• Table 8 · p. 25
• Table 9 · p. 25
• Table 10 · p. 25
• Figure 6 · p. 28

Argument and organization

1. Dispatch MDP

   Makes the application constraints explicit before introducing the network.

2. Architecture and two-stage training

   Separates scalable representation from optimization and initialization.

3. Simulation, transfer and ablations

   Tests the system across demand regimes and discloses metric trade-offs.

Problem, claim and contribution

The provocative title is usefully qualified in its first-page footnote: the final controller is single-agent, but MARL pretraining remains essential. This prevents the framing from erasing a dependency that the ablation later confirms.

Title and footnote1, p.1; §3.3, pp.6–8 · Main paper

Argument and section order

A complete MDP specification explains changing order counts, unavailable drivers and reward terms before the method. This application-first order helps readers understand why ordinary fixed-action TD3 is unsuitable.

§2, pp.2–4 · Main paper

Figures, captions and visual explanation

Figure 2 uses magnified callouts for worker encoding, QK attention and critic pooling while preserving the central actor-to-assignment-to-critic path. Driver and order colors make the fused sequence interpretable.

Figure2, p.4 · Main paper

Tables, comparisons and evidence

The result tables separate global reward from service, pickup and detour metrics. Supplementary bars show gains growing with demand while delivery time can worsen, an important counterweight to a single aggregate improvement.

Table2, p.10; Table4, p.23; Figure6, pp.28–29 · Main paper

Appendix and reproducibility

Simulator assumptions, adapted baselines, time splits, hardware and decision latency are all recoverable. However, July17 training spans only one 30-minute interval, and the 60km/h default is a simulation assumption rather than observed traffic realism.

Appendices E.1–E.5, pp.21–24 · Main paper

Clarity and writing conventions

The loss and policy-gradient notation needs an implementation check: main-text TD errors omit the square used in algorithms, and the appendix's elasticity for ax^b should be b, not ab. A heuristic noisy matching policy should not be presented as a fully established policy-gradient guarantee.

Eq.8, p.8; AppendixC, p.18; Algorithms1–2, pp.20–21 · Main paper

What to reuse—and what not to infer

Transferable lesson: For an application paper, define the operating constraints first, show the complete decision path, disclose simulator assumptions, and report operational trade-offs alongside the headline reward.

Boundary: This is simulation on real trip records, not an online deployment. Parameter sharing does not make the whole attention/matching pipeline linear-time. BERT denotes the bidirectional architecture here, not language pretraining, and LoRA is only an analogy rather than an implemented KD component.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-09-19

Reading status

Full text + selected visuals inspected

Main text

pp.1–10, §§1–5; ethics, reproducibility and references pp.11–15

Appendix / supplement

pp.16–29, contents and appendices A–F, algorithms, simulator and all supplementary result panels

Visually inspected PDF pages

4, 28

• official ICLR 2026 proceedings PDF · 29 pages