PaperlayerView exact revision on arXiv

Source-linked AI summary

JEV-Star: Fast, Low-Cost StarCraft II Control with Language-Model Planning

Weiyu Ma, Liangbing Zhao, Yongcheng Zeng, Jian Zhao

arXiv:2609.27331v1cs.GT

TL;DR

StarCraft II controllers can choose legal affordable actions yet fail to sustain longer-term commitments such as saving to expand. JEV-Star combines frequent JEV action selection with persistent GPT-6 planning, winning four full games including two against Lv7 while retaining a 0.422-second median JEV response time.

  • Problem

    A controller can choose legal, affordable actions yet fail to save enough to expand or carry an attack or retreat decision through, motivating separation of immediate choices from strategy maintenance.

    Source: §1 · p. 1

  • Method

    JEV selects frequent structured actions while GPT-6 supplies persistent goals, spending priorities, and army-posture conditions.

    Source: §1 · p. 1, §3.2 · p. 2

  • Results

    Four full-game wins include two victories against Lv7; mean enemy elimination rises from 16.50% to 37.69% across 35 battle maps.

    Source: §1 · p. 1, §6 · p. 15

  • Takeaways & Limitations

    JEV provides inexpensive action selection while GPT-6 supplies persistent strategic guidance and accounts for most model expenditure.

    Source: §6 · p. 15

    The comparison evaluates complete systems with differing candidate descriptions and execution handling, so it does not isolate planning's contribution.

    Source: §5 · p. 14

Abstract

from arXiv · show

We present JEV-Star, a StarCraft II controller that defeats the strongest non-cheating built-in AI, Lv7, by combining fast JEV action selection with persistent GPT-6 planning. The combined system wins four full games at Lv5--Lv7, including two Lv7 victories with different seeds, while retaining a median JEV response time of 0.422 seconds. Mean estimated model cost across these games is USD~3.71 per game: USD~0.15 for JEV and USD~3.56 for GPT-6. We compare this system with an initial JEV-only controller in full-game macro control and multi-unit micromanagement. The standalone controller reaches a 20-minute limit against Lv2 without expanding. Across 35 battle maps with three episodes per map and controller, the combined system raises mean enemy elimination from 16.50\% to 37.69\% and wins from 3 to 7 out of 105. Replay frames and decision logs document resource reservation, persistent economic goals, and stable army objectives in successful full games. The results demonstrate a practical division between inexpensive, subsecond decisions and longer-horizon planning. The comparison evaluates complete systems; concurrent interface improvements mean that planning's contribution is not isolated by a controlled ablation. Code is available at https://github.com/sc2musa/Jev_Star

1 Introduction

StarCraft II control requires immediate legal actions as well as commitments that preserve long-term goals such as saving to expand. JEV-Star separates frequent JEV action selection from persistent GPT-6 strategic guidance, and the combined controller records two victories against Lv7.

Introduction

  • Immediate legal actions can still fail to save enough for expansion or sustain an attack or retreat decision, motivating separate action selection and strategy maintenance.The section frames these as failures of carrying long-term commitments through repeated decisions.Source: §1 · p. 1
  • JEV selects among structured alternatives, while GPT-6 supplies goals, spending priorities, and conditions for changing army posture.JEV handles frequent action selection; the intended division preserves fast, low-cost interaction while applying longer-horizon planning where one-step decisions are insufficient.Source: §1 · p. 1
  • Two victories against Lv7, the strongest non-cheating built-in difficulty, are reported for the combined controller.The comparison is between JEV-only and JEV + GPT-6 and also measures outcomes, latency, and complete-game model cost.Source: §1 · p. 1

2 Related work

The related work spans StarCraft II control, cooperative combat, language-model interfaces and persistent context, interactive evaluation, and cost- and execution-oriented methods. JEV-Star's stated distinction is inference-time use of hosted models with native action choices and an explicit plan retained across decisions.

Related work

  • StarCraft II research includes interfaces, full-game agents, offline learning, and cooperative combat methods such as QMIX and MAPPO [5] [6] [7] [12] [15] [16] [18] [20] [23] [26].The passage names SC2LE, TStarBots, AlphaStar, TStarBot-X, StarCraft Commander, and AlphaStar Unplugged across those areas.Source: §2 · p. 2
  • LLM Play SC2 and LLM-PySC2 establish language-model interfaces for game control [9] [11], while ReAct, Reflexion, Voyager, and SayCan connect reasoning or persistent context to executable actions [1] [17] [22] [25].AgentBench and WebArena emphasize evaluating interactive outcomes [10] [27]; JEV-Star uses native choices for concrete actions and retains an explicit plan.Source: §2 · p. 2
  • FrugalGPT and RouteLLM reduce cost through model selection, while ReWOO and LLMCompiler organize reasoning and tool execution [4] [8] [13] [24].JEV-Star assigns frequent actions to JEV and occasional plans to GPT-6, evaluated through performance, latency, and per-game expenditure.Source: §2 · p. 2

3 Method

JEV selects executable actions from structured observations, while GPT-6 adds persistent goals that guide subsequent choices. Macro and micro control use separate adapters, and their interaction timing differs.

3.1 Structured action selection with JEV

  • For macro control, JEV selects one feasible high-level command from structured state and recent outcomes, with game code handling placement and low-level execution.Observations cover economy, technology, army, enemies, and recent actions; candidates include production, construction, research, expansion, combat, retreat, and waiting.Source: §3.1 · p. 2
  • For micromanagement, JEV chooses legal actions for each living unit from a shared team observation, with concurrent batches and centralized shared vision.Alternatives include movement, attacks, healing where supported, and stopping; this is not decentralized MARL. The JEV-only method has no separate persistent budget, so an expansion may remain hidden until resources accumulate.Source: §3.1 · p. 2

3.2 Persistent planning with GPT-6

  • GPT-6 proposes validated goals that remain active across action decisions, while JEV continues selecting actions using the plan as context.A failed or delayed planning request leaves the prior valid plan available. Macro plans set economic and army objectives, and resource reservations can suppress lower-priority spending while saving for expansion; emergency supply or defense can override restrictions. Micro plans supply conditional tactics while preserving legal unit choices, but unit actions lack joint reservations for damage or healing targets.Source: §3.2 · pp. 2–3, Figure 1 · p. 3

3.3 Control timing

  • Macro control runs asynchronously, whereas micromanagement advances eight game loops only after a complete concurrent team decision returns.Full-game JEV requests are rate-limited to one per wall-clock second, and planning is checked on a nominal 60-game-second interval and at relevant events. In micro, the slowest required response determines the next step, so response latency and episode wall time measure different parts of the system. Neither method fine-tunes JEV or GPT-6.Source: §3.3 · pp. 3–4

4 Experiments

Across full-game control and micromanagement, JEV + GPT-6 improves on the initial JEV-only system, while the experiments also expose limitations in sample size, runtime, and tactical execution. The results pair persistent strategic guidance with frequent JEV decisions, but compare complete systems rather than isolating planning alone.

4.1 Setup and measurements

  • The evaluation compares one JEV-only full game with four combined full games and 105 micromanagement episodes per method across 35 shared maps.Candidate descriptions and execution handling also changed between controllers, and the combined micromanagement episodes share one plan per map; the comparison is not a controlled planner ablation.Source: §4.1 · p. 4

4.2 Full-game performance: defeating the strongest non-cheating AI

  • JEV + GPT-6 wins all four completed non-cheating full games, including two victories against Lv7; JEV-only reaches its 20-minute limit without winning.The two Lv7 seeds demonstrate successful play at the highest non-cheating difficulty, but the sample is too small to estimate a reliable win rate. The separate Lv8 stress test ends in defeat.Source: §4.2 · pp. 4–5

4.3 Full-game decisions: build order, technology, and unit allocation

  • JEV-only waits in 78.4% of decisions versus 68.1% for the combined controller, while the combined runs submit 156 Zealot and 149 Stalker production orders.The combined runs also submit Immortal orders in all four wins; candidate availability differs, so the observed unit allocation does not by itself establish model preference.Source: §4.3 · pp. 5–7
  • All four combined games follow Gateway, Cybernetics Core, early expansion, and Robotics development, while JEV-only expands neither during its 20-minute run.The combined runs start Warp Gate research at 2:42–2:52 versus 3:20 for JEV-only; in the Lv7 seed-1 game, the expansion order comes at 22 workers despite a plan targeting about 19.Source: §4.3 · p. 7

4.4 Micromanagement: greater enemy elimination, limited wins

  • Across 105 episodes, mean enemy elimination rises from 16.50% to 37.69% and wins from 3 to 7, but the combined system still loses 98 episodes.Improved elimination does not reliably translate into survival or victory; the comparison evaluates complete systems, and the three combined episodes per map share a plan.Source: §4.1 · p. 4, §4.4 · pp. 7–8, Table 3 · p. 8

4.5 Why does planning help?

  • JEV-only reached 68 workers but stayed on one base for 20 minutes, peaking at 335 minerals—below the 400 needed for a Nexus; across four combined-controller wins, a second Nexus appeared at 2:50–3:05.The expansion priority is implemented through a spending reservation.Source: §4.5 · p. 10
  • None of the four combined full-game wins contains an attack-to-retreat reversal within two game seconds, versus five in the JEV-only game; micromanagement traces still show missed movement and repeated healing targets.The plan uses separate attack and retreat thresholds and a commitment window, but different opponents and other controller changes preclude a causal estimate; tactical counts identify recurring behavior, not optimal-action labels.Source: §4.5 · pp. 11–12

4.6 Response time and cost per game

  • Median JEV latency is 0.406 seconds for standalone macro control and 0.422 seconds with planning, while the four combined full-game wins average USD 3.7103 in model cost.GPT-6 accounts for USD 3.5604 of that average. Mean micro wall time rises from 25.78 to 46.23 seconds per episode, and available timing records do not isolate the contributions of planning, setup, and longer battles.Source: §4.6 · pp. 12–13

5 Discussion and limitations

The traces support persistent commitments as an explanation for observed behavior, but the comparison does not isolate planning’s contribution. Evidence is bounded by differing systems and opponents, narrow settings, small samples, and planning cost.

5 Discussion and limitations

  • Persistent goals and adapter rules make resource reservations and army-posture commitments operational, and traces support this proposed mechanism.Saving for an expansion can suppress lower-priority spending; attack thresholds and minimum commitment intervals reduce reversals.Source: §5 · p. 14
  • The comparison is observational: candidate descriptions, execution handling, and macro opponents differ, so planning’s contribution is not isolated.The logged progression from accepted goals to submitted actions supports the proposed mechanism, but does not establish how much of the gain comes from planning.Source: §5 · p. 14
  • The scope is one macro race, matchup, and map, plus centralized micro observations and separate macro and micro implementations; low battle win counts remain a weakness.The paper also notes a custom starting-worker count.Source: §5 · p. 14
  • GPT-6 dominates observed full-game model cost; decision-wait medians exclude planning, whereas episode wall time includes planning and other runtime stages.Episode wall time also includes initialization, game interaction, and cleanup.Source: §5 · p. 14, Figure 11 · p. 15

6 Conclusion

JEV-Star combines rapid structured action selection with persistent planning, defeating Lv7 and improving battle-map outcomes over the initial JEV-only controller.

6 Conclusion

  • JEV-Star defeats Lv7, StarCraft II’s strongest non-cheating built-in AI, through rapid structured decisions and persistent planning.The four full-game wins retain a 0.422-second median JEV response and cost an estimated USD 3.71 per game.Source: §6 · p. 15
  • Mean enemy elimination rises from 16.50% to 37.69% across 35 battle maps compared with the initial JEV-only controller.Resource reservation and stable army objectives help explain the observed behavior.Source: §6 · p. 15
  • JEV provides inexpensive action selection, while GPT-6 supplies persistent strategic guidance and accounts for most model expenditure.Estimated cost per game is USD 0.15 for JEV and USD 3.56 for GPT-6.Source: §6 · p. 15

A.1 A paired micromanagement example

The paired mmmt replays contrast the controllers’ battle outcomes, while the accompanying traces show persistent economic and tactical guidance in the combined system.

A.1 A paired micromanagement example

  • In the first mmmt episode, JEV-only loses with two allied units remaining, while JEV + GPT-6 wins with nine; the pair is illustrative, not the full map result.Both episodes start with 22 allied and 22 enemy units; the combined controller’s guidance calls for an infantry screen, supported tanks, and healing wounded units.Source: Figure 12 · p. 17, §A.1 · p. 17

A.2 Battle progression in the paired replay

  • The paired replay curves end at their own episode boundaries: JEV-only has two allied and 20 enemy units left, while JEV + GPT-6 has nine allied and two enemies.Both battles start with 22 units on each side; these trajectories describe one pair, while the aggregate evaluation covers 35 maps.Source: §A.2 · p. 18, Figure 13 · p. 18

A.3 Full-game decision trace

  • The Lv7 seed-1 decision trace records a plan targeting 30 workers and a second base, followed by a Nexus choice at 400 minerals with 391-ms latency.The recorded choice submits a BUILD NEXUS order.Source: §A.3 · p. 19

A.4 Micromanagement plan and outcome

  • For mmmt, the accepted plan directs infantry to screen tanks and Medivacs, withdraws badly wounded units toward healing, and is reused across three episodes.Its first episode ends in a win with nine allies and two enemies alive; the remaining enemies are non-attacking support units.Source: §A.4 · p. 19

B All recorded micromanagement victories

The galleries document all ten recorded micromanagement wins across six maps, including wins by both controllers on several maps.

B All recorded micromanagement victories

  • The galleries cover all ten recorded wins across six maps, pairing combat frames with terminal states.Terminal counts match episode summaries and include non-attacking support units.Source: B All recorded micromanagement victories · p. 19
  • JEV + GPT-6 wins three displayed episodes on 2c_vs_64zg, 8m, and MMM, with final allied/enemy counts of 1/0, 3/0, and 6/1.The remaining enemy in MMM is a non-attacking support unit.Source: Figure 14 · p. 20
  • On 2s3z, JEV-only wins twice and JEV + GPT-6 once; the three victories end with 3/0, 1/0, and 2/0 allied/enemy counts.The gallery retains this unfavorable comparison for the combined controller.Source: Figure 15 · p. 21
  • Both controllers record a bane_vs_bane victory, ending with three and one allied units, respectively, and no enemies remaining.The JEV-only and combined-system episodes end at 7.14 and 6.07 seconds.Source: Figure 16 · p. 22
  • JEV + GPT-6 records two mmmt victories, each ending with 11 allies and two surviving non-attacking support enemies in the second episode or nine allies and two in the first.The second episode lasts 25.36 seconds.Source: Figure 17 · p. 23

C Opening construction and conditional unit choices

The construction and unit-choice tables describe submitted actions and conditional opportunities in the standalone and combined full-game runs.

C Opening construction and conditional unit choices

  • Table 7 lists construction commands submitted during the first four game minutes, not building completion times.It includes repeated Pylons and Gateways; Probe production and research are excluded, and both games start with eight workers.Source: Table 7 · p. 24, C Opening construction and conditional… · p. 24
  • Table 8 conditions unit selections on the candidate sets actually offered, so row denominators differ and are not disjoint opportunities.Resource reservation, production priorities, prerequisites, and idle producers affect those candidate sets; the table describes complete controllers, not preferences under identical alternatives.Source: C Opening construction and conditional… · p. 24, Table 8 · p. 24
  • The action profile contains 1,020 standalone and 2,871 combined-system action events; response records without a corresponding action event are excluded.Three standalone and 34 combined-system response records are excluded, and the independent Lv8 game remains outside the main action and technology comparison.Source: C Opening construction and conditional… · p. 24

D Per-map micromanagement results

The per-map comparison covers 35 maps and three episodes per method, with wins and mean enemy elimination reported across all 105 episodes.

D Per-map micromanagement results

E Cost accounting and reproducibility

Cost estimates use recorded model usage and separate the four-game Lv5–Lv7 average from the completed Lv8 stress test. The section also specifies accounting rates, data provenance, and reproduction requirements.

E Cost accounting and reproducibility

  • The main combined macro cost average covers four Lv5–Lv7 games, while the completed Lv8 stress test is reported separately.Table 10 lists individual full-game costs estimated from known token usage.Source: E Cost accounting and reproducibility · p. 26, Table 10 · p. 26
  • JEV input is priced at USD 0.042 per million tokens, while GPT-6 uncached input, cached input, and output cost USD 10, 1, and 50 per million tokens, respectively.Requests above 272,000 input tokens use multipliers of two for input and 1.5 for output; rates are applied per response before aggregation.Source: E Cost accounting and reproducibility · p. 26
  • Unknown costs for attempts without recorded usage are omitted from known-usage estimates, and reused micro plans are costed across three episodes per map.The Lv8 stress test and a canceled planning attempt lacked usage; neither is treated as free, and micro costs would be higher per game without reuse.Source: Table 11 · p. 26, E Cost accounting and reproducibility · p. 26
  • The analysis files include 210 battle-episode records, response-level latency and usage, full-game trajectories, and accepted plans linked to frozen summaries and hashed log prefixes.Request latency uses responses with known usage; wall time comes from complete run summaries and includes setup and cleanup.Source: E Cost accounting and reproducibility · p. 26
  • Reproduction requires the specified StarCraft II build, maps, vision and ability settings, and action adapters; the initial and combined catalogs contain 72 and 73 commands.Candidate and execution improvements are part of the system comparison, and no model weights are trained.Source: E Cost accounting and reproducibility · p. 26

Related papers

Loading 2609.27331v1…