Computer Science and Game Theory · arXiv:2609.27331 v1 · submitted 23 Sep 2026 · 27 references
JEV-Star: Fast, Low-Cost StarCraft II Control with Language-Model Planning
The paper says
“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.”
more of the abstract
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
from the abstract
Pith says
Machine review · deepseek-v4-flash · 24 Sep 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. An honest, well-documented engineering result that beats Lv7 at low cost, but the paper's central mechanism claim about persistent planning is explicitly not isolated by a controlled comparison. This line is wrong
3 major objections · 5 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 27 references say.
The paper says, at Abstract and Conclusion (b283); §5 (b8, b272, b275)
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.95):
“The comparison evaluates complete systems; concurrent interface improvements mean that plan- ning’s contribution is not isolated by a controlled ablation.”
Pith says: Objection 1 of 3 · major
The headline claim asserts that the Lv7 victories are achieved 'through a combination of rapid structured decisions and persistent planning,' but the comparison is between complete systems that differ simultaneously on planner presence, candidate descriptions, execution handling, and opponent difficulty.
more of objection 1
The authors acknowledge this limitation (abstract; [b8], [b272], [b275]), and the trace evidence in Table 4 shows the planner behaved as designed—it does not show that planning caused the wins. The paper's own Table C data ([b123], [b124]) make the confound concrete: Sentry appears in 58 standalone candidate sets and never in any combined set, while Immortal appears in 3 standalone versus 34 combined decisions; a pure candidate-set change, or reserve and posture rules implemented in the adapter without GPT-6, could plausibly produce the observed production shift. To support the causal phrasing, the authors should either add a matched planner-on/off ablation (same executor, candidate descriptions, and opponent difficulty, differing only in the planner) or reword the conclusion and abstract to present a system-level demonstration with a trace-supported hypothesis about the mechanism.
The paper says, at Table 1; §4.2
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.83):
“The standalone and combined controllers also differ in candidate descriptions and execution handling, and the macro opponents are not matched.”
Pith says: Objection 2 of 3 · major
The full-game evidence for the headline result is two Lv7 wins (plus one each at Lv5 and Lv6) with no win-rate estimate, and the JEV-only baseline is a single Lv2 game that times out. The authors acknowledge the small sample ([b273]) and the unmatched opponents ([b272]), but the contrast as presented—'JEV-only reaches the time limit against Lv2' versus 'JEV+GPT-6 beats Lv5–Lv7'—conflates method with difficulty.
more of objection 2
As it stands, one cannot tell whether JEV-only at Lv7, or the combined system at Lv2, would behave differently. Because the paper is framed as a difficulty milestone, the resolution is either a few matched-difficulty runs of each controller, even if non-winning, or an explicit sentence stating that no matched-difficulty comparison was performed and that the milestone claim applies only to the combined system.
The paper says, at §4.4; Table 9
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.69):
“Four successful macro games and three micro episodes per map cannot establish broad reliability.”
Pith says: Objection 3 of 3 · major
The micromanagement summary (mean elimination 16.50% to 37.69%, wins 3 to 7 of 105) is reported as an aggregated mean over episodes that are not independent samples of the planning method: the three episodes per map share a single plan, as the authors note ([b273]).
more of objection 3
No significance test accompanies the aggregate, and Table 9 shows the combined controller underperforming on at least two maps (2s3z: wins drop from 2 to 1; bane_vs_bane: elimination drops from 100% to 77.78%). The per-map layout of Table 9 is the appropriate level of inference, and the text should present the result as a per-map paired comparison (25 maps improved, seven tied, three declined) with an explicit statement that the aggregated means are descriptive and not significance-tested, rather than leading with the pooled percentages.
5 minor objections
The paper says, at Abstract and §4.2
Pith says: minor objection 1 of 5
The label 'strongest non-cheating built-in AI' depends on the difficulty enumeration of reference [3]; consider adding a note that the Lv8 stress test (privileged vision) is outside this label, which the paper already separates in its presentation.
The paper says, at Title page; References
Pith says: minor objection 2 of 5
The author list contains an errant space in 'Y ongcheng Zeng', and reference [22] has the same issue in 'V oyager'; both should be corrected.
show 3 more
The paper says, at §3.2, Eq. (2)
Pith says: minor objection 3 of 5
The notation for the plan context and the previous plan is clear, but the superscript 'eAt' in the equation is unexplained at first use; it should be defined at the equation site as the candidate set with plan-derived constraints.
The paper says, at §4.5
Pith says: minor objection 4 of 5
The healing-rejection and cooldown statistics (794 continued attacks out of 930 movement opportunities; 243 of 378 repeated healing targets) are the only mechanism-level micro evidence; consider promoting them from the shoulder of §4.5 into a short dedicated paragraph, since they are the paper's strongest evidence that planning alone does not resolve coordination.
The paper says, at Figure 9 and Table 5
Pith says: minor objection 5 of 5
With a logarithmic x-axis and three pooled curves, the percentile values are only available in Table 5; labeling the P50 and P95 markers directly on the figure would improve readability.
What would settle it
Run a matched full-game ablation on AltitudeLE at Lv7 with at least 10 seeds per arm, holding fixed the combined system's exact executor, candidate descriptions, execution handling, and game settings in every arm, and toggling only the planner: (i) planner off, with plan context, resource-reservation, and posture rules removed; (ii) planner on; (iii) planner on but resource reservation disabled. If arm (i) wins Lv7 at a comparable rate to arm (ii), or arm (iii) matches arm (ii), then the planner is not the causal mechanism behind the Lv7 victories, and the mechanism claim must be restated as a system-level, not a planning-level, result.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 24 Sep 2026
- Reader holds up, with conditions
- Skeptic agrees
- Editor wrote the verdict above
- Referee major revision: 3 major objections, 5 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
You should read this as an engineering demonstration, not as a controlled study of why it works. The two Lv7 wins are on the table, with replay frames, decision logs, and a plan-decision-execution trace that matches what the system claims to do. The cost and latency numbers are also concrete: subsecond JEV actions, GPT-6 planning at roughly thirty seconds, around USD 3.71 per full game by standard-API-equivalent accounting. The action-versus-planner division of labor is a genuinely useful pattern for LLM-agent work, and the absence of fine-tuning is a real practical point.
The paper is also unusually honest. The abstract and Section 5 say plainly that the two compared controllers differ in candidate descriptions, execution handling, and opponent difficulty, so planning's contribution is not isolated by a controlled ablation [b8, b272]. On my read of the candidate tables, that is not a minor caveat. The standalone run never has Immortals offered in practice, while the combined system selects them in 26 of 34 opportunities; Sentries disappear entirely from the combined candidate sets [b123, b124]. Those candidate-set changes alone could shift the production mix substantially. And the macro baseline is a single Lv2 game that hits the time limit against an unmatched opponent, so the full-game comparison is suggestive, not probative. The authors say as much: the data "motivate the planning hypothesis without isolating its entire contribution" [b275]. The stress-test reader's concern lands, but the paper already concedes it.
The soft spots beyond the ablation: two Lv7 seeds give no win-rate estimate, the micro episodes share one plan per map so the three combined-method episodes are not independent, and the cost figures are estimates from recorded token usage rather than billing statements. Those are real weaknesses, though each is disclosed. Lv7 is also not grandmaster-level play; the benchmark is the built-in AI, and the Lv8 privileged-opponent loss is reported rather than hidden.
The paper is worth engaging with because it ships a testable system, documents its conditions, and distinguishes model intent, submitted commands, and observed game state better than most LLM-agent papers. The right next step is exactly what the authors propose: a matched planner-on/off ablation with the same executor, same candidate sets, and more seeds. That experiment would determine whether the persistent-planning claim survives. I would not block peer review over the missing ablation, but I would make it a condition for acceptance, and I would ask the authors to report a win-rate estimate over a larger seed set. If they can run that, this becomes a solid reference point for cheap action selection plus expensive planning.
The strongest objection
The strongest claim has two components: the empirical demonstration (JEV+GPT-6 wins two Lv7 games at low cost) and the mechanism attribution (persistent planning's resource-reservation and posture rules cause the wins). The empirical component is credible on the trace evidence: decoded replays, the plan-decision-execution log in Table 4, and honest divergence notes such as the expansion ordered at 22 workers against the plan's target of 19. The mechanism component is the load-bearing premise, and it is untested. The only full-game comparison in the paper differs on four dimensions at once: planner present/absent, candidate descriptions, execution handling, and opponent difficulty (Lv2 for baseline vs Lv5-Lv7 for the combined system), as the authors state directly ([b8], [b272]). The paper's own data show the candidate sets changed materially: Sentry is offered in 58 standalone decisions and never appears in any combined candidate set ([b124]); Immortal is offered at only 3 standalone decisions versus 34 combined decisions ([b123]). A shift in what is even offered to the action selector, with no planner at all, could produce the observed production-mix change; the combined system's reserve and commitment rules could themselves be implemented in the adapter without GPT-6. The reservation trace (plan accepted at 1:34, 400 minerals at 2:56, Nexus ordered at 2:57, observed at 3:05) shows planning behaved as designed, but it cannot rule out that the concurrent candidate and execution changes, or luck at n=2, would have produced the same outcome. The authors acknowledge this directly ([b275]: 'The current data motivate the planning hypothesis without isolating its entire contribution'). Because the paper's stated contribution is the cheap-action/expensive-planner division, the mechanism attribution, not the raw Lv7 win, is what must hold for the novelty claim.
The reader's weakest_assumption and my read converge on the same point: the strongest claim decomposes into an empirical component (JEV-Star can defeat Lv7) and a mechanism component (persistent planning causes those defeats), and the mechanism component is not isolated from concurrent changes to candidate descriptions, execution handling, and opponent matching. I agree with that diagnosis. The empirical component is supported by good-faith evidence worth crediting: decoded replay frames, per-decision demand logs, the plan-decision-execution trace in Table 4, frank reporting of the Lv8 stress-test loss, explicit statements that a time limit is not a victory, and acknowledgment that micro win counts remain low. These are falsifiable, reproducible traces rather than vague self-reports. The gap is specifically causal attribution: the full-game comparison is uncontrolled, and the paper's own candidate-offer counts (Sentries and Immortals) show the offered action sets diverged materially between the two systems, so the production shift alone does not demonstrate planner efficacy. In the micro setting opponents are matched, yet the same candidate and execution changes still confound planner attribution there, and the three combined-method episodes share one map plan. I did not find an additional concern that would force the verdict toward REJECT: the demonstration itself is genuine as far as the record goes, and even if planning's marginal contribution proved small, the system-level result — a subsecond action layer plus occasional planner defeating Lv7 at roughly USD 3.71 per game — would still stand, though it would lose the stated division-of-labor interpretation. The reader's CONDITIONAL verdict already encodes exactly this controlling condition, so I recommend no change. The concrete test I propose isolates the planner while holding every other system component and the opponent fixed, and would settle whether the mechanism claim lands.
The referee's report
The empirical result, if it holds, is a useful difficulty milestone: defeating Lv7 at an estimated cost of a few dollars per game using an inference-time division of labor, with no weight training or fine-tuning. The verification apparatus is a genuine strength: outcomes are checked against decoded replays, cost figures are reconstructed from recorded token usage at published standard rates, per-map micro results are given in full, and the authors include honest divergence notes (e.g., the expansion executed at 22 workers against a plan target of 19). Equally important, the key limitations—unmatched macro opponents, small sample, shared per-map plans, unpriced usage, and the absence of a controlled ablation—are stated explicitly in the abstract and Section 5. The unresolved issue is framing: the conclusion asserts the causal mechanism ('through... persistent planning'), whereas the evidence supports a system-level demonstration with a plausible, trace-supported but untested mechanism. That distinction, rather than the empirical claims themselves, is what requires correction.
Pith's reading of the paper (a machine's summary, not the authors')
The paper claims that a StarCraft II controller can beat Lv7 — the strongest non-cheating built-in opponent — by dividing control between a cheap model (JEV) that picks the next legal action on every decision and an occasional planner (GPT-6) that sets persistent economic and army goals. The combined system won four full games at Lv5–Lv7, including two Lv7 victories on different seeds, at a median action-response time of 0.422 seconds and an estimated mean of USD 3.71 in model charges per game, with no weight training. Against the initial action-only controller, the combined version more than doubled mean enemy elimination in micromanagement, from 16.50% to 37.69%, and raised wins from 3 to 7 out of 105. Why care: it offers a low-cost path to strong real-time strategic play by separating rapid, subsecond decisions from slower, rarer deliberation.
What carries the argument. The central mechanism is a division of labor written as $a_t = \mathrm{JEV}(x_t, z_k, \mathcal{A}_t)$, where JEV picks an identifier from a legal candidate set on nearly every decision while a planner returns guidance $z_k$ that stays active across many decisions and enters both the model's context and the adapter's candidate rules. Two operating mechanisms make the plan operational inside the fast loop: resource reservation, which suppresses lower-priority spending until a target such as a 400-mineral Nexus is affordable, and army-posture persistence, which uses separate attack and retreat thresholds plus a commitment window to eliminate rapid reversals, cutting the standalone run's five within-two-second reversals to zero in the wins.
Core claim. On its own terms, the paper establishes that JEV-Star, pairing the fast structured action selector JEV with persistent GPT-6 planning, defeats Lv7, the highest non-cheating difficulty of StarCraft II's built-in AI, in two full games across two seeds, and in additional games at Lv5 and Lv6, while keeping median action latency at 0.422 seconds and estimated model spend near USD 3.71 per game. It attributes the wins to commitments the planner supplies that local action selection cannot form on its own: a resource reservation that suppresses lower-priority spending until 400 minerals accumulate for a Nexus, and separated attack and retreat thresholds with a minimum commitment interval that remove the attack-to-retreat reversals the standalone controller showed. The standalone controller is the diagnosed contrast: it reaches 68 workers yet stays on one base for 20 minutes, because affordable purchases keep draining its minerals before any expansion ever becomes affordable.
The premise it rests on. The load-bearing premise is that the Lv7 wins come from the planner's resource-reservation and posture mechanisms rather than from the concurrent changes the paper itself says it could not ablate apart — improved candidate descriptions, different execution handling, and harder opponents for the combined system.
If the paper is right (Pith's inferences, not the authors')
If correct, a full-game real-time strategy controller can run on subsecond action selection plus occasional planner calls (median 27.6 seconds in macro control), at about USD 3.71 per game with no weight training, a cost and latency profile an ordinary research group can afford.
The planner dominates cost at roughly 96% of the per-game total, so lowering replanning frequency or reusing plans more widely reduces expenditure more than speeding up JEV.
Persistent planning fixes commitment failures — saving for an expansion and holding an attack — but does not fix micromanagement coordination, since enemy elimination rises sharply while wins stay at 7 of 105, leaving tactical execution as the bottleneck.
The separate Lv8 stress test against a privileged-vision opponent ends in defeat with no attack command ever issued, bounding the demonstrated result to non-cheating opponents until economic and military balance improves.
A two-speed template of a cheap reactive layer plus a rarer deliberative layer sharing one interface could generalize to other real-time decision settings such as robot control, network management, or trading; the paper does not test that transfer.
Counting how often the reservation actually blocks a purchase that would have delayed the Nexus would quantify the planner's causal leverage per game; the paper's logs could support such a count but do not report it.
If the planner's value concentrates in a handful of commitment moments, those decisions could be precomputed or cached, potentially shrinking the planner's USD 3.56 share of per-game cost well below the reported mean.
Because one shared micro plan serves all three episodes of a map, the plan sample is not independent across those episodes, so true planner variability within a map remains untested.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
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
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-09-24
The authors' main claim, as the classifier picked it out 0.86
“The results demonstrate a practical division between inexpensive, subsecond decisions and longer-horizon planning.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.05 · states limitation 0.97 · human participants 0.03 · controlled comparison 0.94 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.96 · mentions code or data availability 0.98
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| 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. | contribution 0.96 | 0.93 | 0.72 | asserted 1.00 |
| 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. | result 1.00 | 0.96 | 0.27 | asserted 1.00 |
| 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. | result 1.00 | 0.95 | 0.06 | asserted 1.00 |
| We compare this system with an initial JEV-only controller in full-game macro control and multi-unit micromanagement. | method 0.77 | 0.90 | 0.03 | asserted 0.80 |
| The standalone controller reaches a 20-minute limit against Lv2 without expanding. | result 0.99 | 0.89 | 0.03 | asserted 1.00 |
| 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. | result 1.00 | 0.95 | 0.13 | asserted 1.00 |
| Replay frames and decision logs document resource reservation, persistent economic goals, and stable army objectives in successful full games. | result 0.59 | 0.92 | 0.07 | asserted 0.99 |
| The results demonstrate a practical division between inexpensive, subsecond decisions and longer-horizon planning. | interpretation 1.00 | 0.76 | 0.58 | asserted 1.00 |
| The comparison evaluates complete systems; concurrent interface improvements mean that planning's contribution is not isolated by a controlled ablation. | limitation 1.00 | 0.74 | 0.05 | asserted 0.99 |
| Code is available at https://github.com/sc2musa/Jev_Star | contribution 0.33 | 0.56 | 0.02 | asserted 0.80 |
What kind of paper the classifier reports this is: method or tool 0.98 · evidence from observation or experiment 0.87
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-24: 1 request, 7,552 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (17)
Figures are the authors', reproduced from the paper.
References (27)
- [1] Michael Ahn, Anthony Brohan, Noah Brown, et al. Do as i can, not as i say: Grounding language in robotic affordances.arXiv preprint arXiv:2204.01691, 2022. URLhttps://arxiv.org/abs/2204.01691 on PithPith's reviewarXiv
- [2] Diogo Almeida. Introducing system one models & jev. TypeSafe AI, 2026. URL https://typesafe. ai/blog/introducing-system-one-models-and-jev . Published September 15, 2026; accessed September 22, 2026 on Pith
- [3] Blizzard Entertainment. StarCraft II api protocol: built-in opponent difficulty levels, 2026. URL https:// github.com/Blizzard/s2client-proto/blob/master/s2clientprotocol/sc2api.proto. Dif- ficulty enumeration; accessed September 22, 2026 on Pith
- [4] Lingjiao Chen, Matei Zaharia, and James Zou. FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance.arXiv preprint arXiv:2305.05176, 2023. URL https: //arxiv.org/abs/2305.05176 on PithPith's reviewarXiv
- [5] Yue Deng, Yan Yu, Weiyu Ma, Zirui Wang, Wenhui Zhu, Jian Zhao, and Yin Zhang. SMAC-Hard: Enabling mixed opponent strategy script and self-play on SMAC.arXiv preprint arXiv:2412.17707, 2024. URL https://arxiv.org/abs/2412.17707 on PithPith's reviewarXiv
- [6] Benjamin Ellis, Jonathan Cook, Skander Moalla, Mikayel Samvelyan, Mingfei Sun, Anuj Mahajan, Jakob N. Foerster, and Shimon Whiteson. SMACv2: An Improved Benchmark for Cooperative Multi- Agent Reinforcement Learning.arXiv preprint arXiv:2212.07489, 2022. URL https://arxiv.org/ ... on PithPith's reviewarXiv
- [7] Lei Han, Jiechao Xiong, Peng Sun, Xinghai Sun, Meng Fang, Qingwei Guo, Qiaobo Chen, Tengfei Shi, Hongsheng Yu, Xipeng Wu, and Zhengyou Zhang. TStarBot-X: An Open-Sourced and Comprehensive Study for Efficient League Training in StarCraft II Full Game.arXiv preprint arXiv:2011.1... on PithPith's reviewarXiv
- [8] Sehoon Kim, Suhong Moon, Ryan Tabrizi, Nicholas Lee, Michael W. Mahoney, Kurt Keutzer, and Amir Gholami. An LLM Compiler for Parallel Function Calling.arXiv preprint arXiv:2312.04511, 2023. URL https://arxiv.org/abs/2312.04511 on PithPith's reviewarXiv
- [9] Zongyuan Li, Yanan Ni, Runnan Qi, Lumin Jiang, et al. LLM-PySC2: Starcraft II learning environment for Large Language Models.arXiv preprint arXiv:2411.05348, 2024. URL https://arxiv.org/abs/ 2411.05348 on PithPith's reviewarXiv
- [10] Xiao Liu, Hao Yu, Hanchen Zhang, Yifan Xu, et al. AgentBench: Evaluating LLMs as Agents.arXiv preprint arXiv:2308.03688, 2023. URLhttps://arxiv.org/abs/2308.03688 on PithPith's reviewarXiv
- [11] Weiyu Ma, Qirui Mi, Yongcheng Zeng, Xue Yan, Yuqiao Wu, Runji Lin, Haifeng Zhang, and Jun Wang. Large language models play StarCraft II: Benchmarks and a chain of summarization approach.arXiv preprint arXiv:2312.11865, 2023. URL https://arxiv.org/abs/2312.11865. Version 3, June 2024 on PithPith's reviewarXiv
- [12] Michaël Mathieu, Sherjil Ozair, Srivatsan Srinivasan, Caglar Gulcehre, et al. AlphaStar Unplugged: Large-Scale Offline Reinforcement Learning.arXiv preprint arXiv:2308.03526, 2023. URL https: //arxiv.org/abs/2308.03526 on PithPith's reviewarXiv
- [13] Isaac Ong, Amjad Almahairi, Vincent Wu, Wei-Lin Chiang, Tianhao Wu, Joseph E. Gonzalez, M Waleed Kadous, and Ion Stoica. RouteLLM: Learning to Route LLMs with Preference Data.arXiv preprint arXiv:2406.18665, 2024. URLhttps://arxiv.org/abs/2406.18665 on PithPith's reviewarXiv
- [14] OpenAI. gpt-6-astra model documentation and pricing, 2026. URL https://developers.openai. com/api/docs/models/gpt-6-astra. Standard API rates; accessed September 22, 2026 on Pith
- [15] Tabish Rashid, Mikayel Samvelyan, Christian Schroeder de Witt, Gregory Farquhar, Jakob Foerster, and Shimon Whiteson. QMIX: Monotonic Value Function Factorisation for Deep Multi-Agent Reinforcement Learning.arXiv preprint arXiv:1803.11485, 2018. URLhttps://arxiv.org/abs/1803.11485 on PithPith's reviewarXiv
- [16] Mikayel Samvelyan, Tabish Rashid, Christian Schröder de Witt, Gregory Farquhar, Nantas Nardelli, Tim G. J. Rudner, Chia-Man Hung, Philip H. S. Torr, Jakob Foerster, and Shimon Whiteson. The StarCraft multi- agent challenge.arXiv preprint arXiv:1902.04043, 2019. URL https://arx... on PithPith's reviewarXiv
- [17] Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Language Agents with Verbal Reinforcement Learning.arXiv preprint arXiv:2303.11366, 2023. URLhttps://arxiv.org/abs/2303.11366 on PithPith's reviewarXiv
- [18] Peng Sun, Xinghai Sun, Lei Han, Jiechao Xiong, Qing Wang, Bo Li, Yang Zheng, Ji Liu, Yongsheng Liu, Han Liu, and Tong Zhang. TStarBots: Defeating the Cheating Level Builtin AI in StarCraft II in the Full Game.arXiv preprint arXiv:1809.07193, 2018. URLhttps://arxiv.org/abs/1809.07193 on PithPith's reviewarXiv
- [19] TypeSafe AI. Models and pricing, 2026. URL https://docs.typesafe.ai/models. Accessed September 22, 2026 on Pith
- [20] Oriol Vinyals, Timo Ewalds, Sergey Bartunov, et al. StarCraft II: A new challenge for reinforcement learning.arXiv preprint arXiv:1708.04782, 2017. URLhttps://arxiv.org/abs/1708.04782 on PithPith's reviewarXiv
show 7 more
- [21] Oriol Vinyals, Igor Babuschkin, Wojciech M. Czarnecki, et al. Grandmaster level in StarCraft II using multi-agent reinforcement learning.Nature, 575:350–354, 2019. doi: 10.1038/s41586-019-1724-z. URL https://www.nature.com/articles/s41586-019-1724-z on Pithdoi:10.1038/s41586-019-1724-z
- [22] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. V oyager: An Open-Ended Embodied Agent with Large Language Models.arXiv preprint arXiv:2305.16291, 2023. URLhttps://arxiv.org/abs/2305.16291 on PithPith's reviewarXiv
- [23] Xiangjun Wang, Junxiao Song, Penghui Qi, Peng Peng, Zhenkun Tang, Wei Zhang, Weimin Li, Xiongjun Pi, Jujie He, Chao Gao, Haitao Long, and Quan Yuan. SCC: an efficient deep reinforcement learning agent mastering the game of StarCraft II.arXiv preprint arXiv:2012.13169, 2020. UR... on PithPith's reviewarXiv
- [24] Binfeng Xu, Zhiyuan Peng, Bowen Lei, Subhabrata Mukherjee, Yuchen Liu, and Dongkuan Xu. ReWOO: Decoupling Reasoning from Observations for Efficient Augmented Language Models.arXiv preprint arXiv:2305.18323, 2023. URLhttps://arxiv.org/abs/2305.18323 on PithPith's reviewarXiv
- [25] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. ReAct: Synergizing Reasoning and Acting in Language Models.arXiv preprint arXiv:2210.03629, 2022. URL https://arxiv.org/abs/2210.03629 on PithPith's reviewarXiv
- [27] Shuyan Zhou, Frank F. Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Tianyue Ou, Yonatan Bisk, Daniel Fried, Uri Alon, and Graham Neubig. WebArena: A Realistic Web Environment for Building Autonomous Agents.arXiv preprint arXiv:2307.13854, 2023. URL https:/... on PithPith's reviewarXiv
- [2021] URLhttps://arxiv.org/abs/2103.01955 on PithPith's reviewarXiv
Cited by 4 Pith papers
Reviewed papers on Pith that cite this one. The full citation record
- NIOCIS: Typed-Decision Next-Hop Routing with Node-Conditioned Option Spaces for Tool-Use Agents cs.CR · Oct 2026
- Can Jev be Your Q or Policy in Reinforcement Learning? cs.LG · Oct 2026
- OmniMed-Jev: Calibrating LVLM Confidence for Trustworthy Medical Multimodal Decisions via System One cs.LG · Sep 2026
- Typed Decision Models: An Early Evidence Audit and Evaluation Checklist cs.CL · Sep 2026
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment