Pith.

Artificial Intelligence · arXiv:2611.04772 v1 · submitted 3 Nov 2026 · 51 references

TactiNavi: Serving and Revising an LLM Game Coach with Explicit Strategy Conditions

Qi Liu, Xiaoyang Yuan, Zhuomeng Zhang, Yubin Ruan, Xinyi Mou, Wenjin Wang, Hanying Jiao, Di Wu

show 4 more authors

Jinke Li, Jingxuan Hong, Ke Feng, Zixun Sun

The paper says

“A deployed LLM coaching agent answers requests whose correct response depends on player state and game version, so advice that is right for one player can be wrong for the next.”

more of the abstract

We describe TactiNavi, the serving system behind the LLM coach of a commercial open-world game, which makes strategy applicability an explicit artifact shared by offline compilation, online execution, failure diagnosis, and release. An offline compiler turns versioned game evidence and expert guidance into strategy packages with explicit prerequisites, the serving plane evaluates those prerequisites as three-valued conditions over structured player state and records decision traces, and a control plane uses the resulting condition-rule dependencies to validate revisions on held-out traffic before release. In a retrospective evaluation of 1,122 logged coaching turns from three deployment periods, condition-bearing packages raise correctness from 57.9% to 82.0% over atomic facts and reduce strategy misapplication under changed player state from 18.1% to 10.4%, while trace-based claim verification reaches 86.4% precision at a median reply latency of 2.07 s. The deployment also showed that most apparent fixes are generation noise, because an unchanged-system rerun recovers 38.4% of failing turns and only one of 15 shipped revision units passes the release checks retrospectively. We distill these results into lessons on treating applicability as a first-class serving object, sizing traces for later diagnosis, and building the null model into the release path.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 4 Nov 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. The release-gate noise-floor result is the real contribution; the applicability claim is honest in the body but overstated in the abstract. This line is wrong

3 major objections · 4 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 51 references say.

The paper says, at §5.3, Table 3

The review did not quote this passage.

Pith says: Objection 1 of 3 · major

The 86.4% claim-verification precision is the least securely grounded headline number. The verification verdicts behind it are produced by the same self-hosted judge whose correctness labels agree with experts at κ=0.42 and whose error detection reaches only 55.6% precision and 52.6% recall (Appendix B.3), and the paper reports no expert adjudication of the verification verdicts themselves—only of whether the original doubts were founded.

more of objection 1

The reported length bias (15.9% false alarms for longer responses) is a plausible inflation mechanism for an elaborate trace-based method, and the precision figure rests on only 48 adjudicated doubts. Because Lesson 2 (traces must be sufficient for later diagnosis) and the abstract's 86.4% claim both depend on this number, the authors should either provide expert adjudication of the claim-verification verdicts or supply a quantitative sensitivity analysis showing how the judge's measured error rates and length bias would move the 86.4% figure.

The paper says, at §5.2, Table 2; abstract

The review did not quote this passage.

A classifier read this sentence of the paper as acknowledging the concern (probability 0.94):

“The guide comparison rests on 72 variant units from 36 base turns, and the exact McNemar test on stable pairs, with 5 discordant pairs favor- ing the package and none favoring the guide, gives p = 0 .062, so we read the violation reduction as consistent evidence of an effect rather than as a settled one.”

Pith says: Objection 2 of 3 · major

The abstract's 'reduce strategy misapplication under changed player state from 18.1% to 10.4%' is stated more strongly than the statistical support warrants. The package-versus-guide comparison rests on 72 variant units from 36 base turns with exactly 5 discordant pairs, giving McNemar p=0.062; the authors themselves read this as 'consistent evidence of an effect rather than a settled one' in §5.2 and again in the Section 6 limitations.

more of objection 2

Since the guide-matches-package-on-recorded-turns / diverges-under-state-change comparison is the load-bearing evidence for Lesson 1, the abstract and conclusion should carry the same 'not settled' qualification, or the evidence base (number of base turns) should be expanded before the reduction is presented as a headline result.

The paper says, at §5.2, §5.5, Appendix B.3

The review did not quote this passage.

A classifier read this sentence of the paper as acknowledging the concern (probability 0.68):

“The automated judge shares the Coach’s underlying model, which makes evaluation sensitive to shared judgment er- rors.”

Pith says: Objection 3 of 3 · major

The correctness and violation measurements all pass through the automated judge, which shares the Coach's underlying model (acknowledged in Section 6 as shared-judgment sensitivity). The paper quantifies the judge only partially: 89.0% correctness agreement (κ=0.48) on the held-out set, an 11.7% judge incorrect-response rate, and a 15.9% false-alarm rate for longer responses.

more of objection 3

What is missing is a bounded sensitivity analysis: under the judge's measured error rates, how much could the 24.1-point correctness gain and the 7.6-point violation reduction move? The direction of the bias is not obvious a priori—token-heavier package configurations could attract more length-based false alarms, which would move the two comparisons in opposite directions—so the magnitudes reported in the abstract should be accompanied by a quantitative bound derived from the judge's own calibration numbers.

4 minor objections

The paper says, at §5.3, Table 3

Pith says: minor objection 1 of 4

The paper notes that 36.9% of flagged cases are left unresolved when trace evidence is insufficient; it would help to state explicitly how unresolved cases enter the precision and recall denominators, since the treatment of that 36.9% materially affects the interpretation of the 52.8% recall figure.

The paper says, at Appendix C.1

Pith says: minor objection 2 of 4

The repair-ownership macro-F1 of 18.8% and repair-target Hit@1 of 36.7% are reported without discussion; given that §4.4 claims attribution names knowledge and execution repair targets, a sentence interpreting these low values (for example, their relation to the concentration of expert-attributed failures in reply execution) is needed.

show 2 more

The paper says, at §5.4, Table 4

Pith says: minor objection 3 of 4

The deployed unit that 'gains 25.0 points inside its declared scope and loses 20.8 points outside it' is not accompanied by the turn counts behind the in-scope and out-of-scope estimates; reporting those counts would allow the reader to judge the stability of the 20.8-point regression.

The paper says, at §2, §4.3, §4.4

Pith says: minor objection 4 of 4

The execution plane 'builds on' SGC (Liu et al., 2026) and the evaluator 'follows the DAAF formulation' (Yuan et al., 2026), both same-group arXiv preprints; a sentence making explicit what the present paper adds beyond those foundations (the compiled condition representation, the trace schema, and the release-gate protocol) would help readers situate the novelty.

What would settle it

Have game-domain experts adjudicate the verification outputs for all 130 player-doubt turns, or a stratified sample that includes all 48 already-expert-adjudicated doubts, labeling each TactiNavi outcome as correctly-resolved, incorrectly-resolved, or unresolved. Then recompute Table 3 precision, recall, and the 9/12 rejection count using expert verification labels instead of DeepSeek-V4-Flash verdicts. If expert-adjudicated precision drops materially below 86.4% or rejection of unfounded doubts falls below 9/12, the diagnosis claim weakens; if the numbers hold, the concern is answered.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 4 Nov 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic partly agrees 16 s · 22,653 tokens read, 1,989 written
  3. Editor wrote the verdict above 42 s · 23,790 tokens read, 5,218 written
  4. Referee major revision: 3 major objections, 4 minor 56 s · 22,897 tokens read, 9,142 written

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

Two things to know about this paper. The release-gate retrospective is the real contribution, and it is solid: rerunning the unchanged system recovers 38.4% of previously failing turns, the planted controls recover 39.1–40.7% with paired in-scope gains near zero, and only 1 of 15 shipped revision units would have passed the pre-fixed gates. One unit gains 25.0 points inside its declared scope while losing 20.8 outside it. That is a transferable result for anyone revising LLM agents, and it rests on executable checks and paired replays, not on the automated judge. The second thing is that the central positive claim is thinner than the abstract's framing: explicit three-valued conditions reduce misapplication from 18.1% to 10.4%, but that rests on 72 variants from 36 base turns, five discordant pairs, and McNemar p = 0.062. The prose guide ties the package on recorded turns, 82.5% vs 82.0%. The authors are honest about this in the body and limitations; the abstract is not.

What is genuinely new is the compiled condition lifecycle — prerequisites as objects that are evaluated online, recorded in traces, linked to evaluation rules, and used to scope release decisions. The integration is coherent, and the evaluation discipline is unusually careful: sealed T3, criteria fixed in advance, session-clustered bootstrap CIs, exact McNemar tests, planted entity-level vs mechanism-level controls, and plainly stated limitations. The paper also correctly separates synthesis from applicability, noting that atomic facts misapply rarely because they carry no strategy; that is a feature of the comparison, not a flaw, and the authors say so.

Soft spots, in proportion. The judge is a self-hosted DeepSeek-V4-Flash from the same family as the Coach, with kappa 0.42 against expert correctness labels and error-detection recall of 52.6%. Table 3's 86.4% claim-verification precision is the least secure headline number: the verification verdicts themselves are not expert-adjudicated, and the reported length bias (15.9% false alarms on longer responses) runs in the direction of inflating a trace-heavy method. The authors acknowledge shared-model sensitivity but do not quantify it. The 26.7% delayed-feedback localization survives this concern — it is concretely attributable to traces. Building on the same group's SGC and DAAF is not a problem here; the load-bearing measurements are anchored to external game evidence and executable checks.

Who this is for: people serving state-dependent LLM advice, and anyone measuring whether agent revisions actually help. It deserves a serious referee. The right outcome is revision: align the abstract with the p = 0.062 evidence, and either adjudicate a sample of verification verdicts or soften the precision claim.

The strongest objection

The paper's central system claim is that treating applicability as a first-class object improves coaching correctness and that trace-based diagnosis separates durable revisions from noise. The release-gate retrospective (Table 4: 38.4% unchanged-rerun recovery, 1/15 shipped units eligible) rests on paired replayed comparisons on a sealed T3 set, so it is well supported. The package-vs-atomic-facts correctness comparison (82.0% vs 57.9%) is also reasonably supported because it is large, the CI [16.4, 31.7] is tight, and the paper itself separates synthesis from applicability by noting the prose guide matches the package at 82.5%. The weakest load-bearing number is the diagnosis result in Table 3: 86.4% precision, 52.8% recall, and rejection of 9/12 unfounded doubts on only 48 expert-adjudicated doubts. Precision here is a property of the same self-hosted DeepSeek-V4-Flash judge that detects expert-labeled errors at only 55.6% precision and 52.6% recall (Appendix B.3, kappa = 0.42 with executable checks). Appendix B.3 reports judge-expert agreement on correctness labels, not on verification verdicts; the failure-diagnosis numbers are computed from the judge's own verification decisions, and the paper gives no expert adjudication of those decisions. The known length-bias result (15.9% false alarms for longer responses) is a plausible direction for inflating the precision of an elaborate trace-based method. The authors acknowledge shared-model sensitivity as a limitation in Section 6, but they do not quantify its size for Table 3. If the judge systematically flags longer, more elaborate verification outputs as correct, the 86.4% precision and the 9/12 rejection count could be largely instrument artifacts; the 26.7% delayed-feedback localization is smaller but concretely attributable to traces and would survive this concern.

The reader identifies the automated judge as the weakest assumption, which matches my concern, but I disagree on which number is most exposed. The reader emphasizes the knowledge-form comparisons; those are anchored by a large effect and by the paper's own admission that the package adds little over prose on recorded turns. The violation reduction is honestly reported with p = 0.062 and is already framed as consistent-evidence rather than settled. The release-gate retrospective is the most solid contribution and does not depend on the judge: 38.4% recovery is an unchanged-system paired replay, and 1/15 eligibility follows fixed rules on replayed correctness. The failure-diagnosis claim is the least secure because 86.4% precision, 52.8% recall, and 9/12 rejection are all direct outputs of a judge whose error-detection precision and recall (55.6% and 52.6%) are close to the range that would make those headline numbers unreliable, and no expert verification-verdict adjudication is reported. The 26.7% delayed-feedback localization is concrete and trace-dependent, so it survives the concern. Given the strong release-gate evidence and the authors' candid limitation statements, the conditional verdict is appropriate; the diagnosis claim should be re-read as provisional, not as a settled comparison.

The referee's report

The paper's methodological discipline is a genuine strength: a sealed T3 period, release criteria fixed in advance, six planted revision controls (entity-specific versus mechanism-based), session-clustered bootstrap CIs, stable-pair exact McNemar tests with Holm correction, ablations (removal of conditional rules, removal of turn localization), and unusually candid reporting of negative results (1/15 shipped revisions eligible, 38.4% rerun recovery, a knowledge-execution interaction of −9.6 points). If the headline measurements hold, the paper demonstrates a generic serving pattern and delivers a valuable caution to the agent-revision literature: apparent fixes are mostly generation noise, and recovery on failing turns is not evidence of repair. The central design argument—separating synthesis from applicability by showing the prose guide matches the package on recorded turns (82.5% vs 82.0%) yet diverges under state change—is informative.

Pith's reading of the paper (a machine's summary, not the authors')

TactiNavi claims that an LLM coaching agent whose advice depends on live player state and game version should treat a strategy's applicability as a first-class object rather than a sentence the model must interpret. The paper reports that compiling game knowledge into versioned packages with explicit, three-valued prerequisites lifted coaching correctness from 57.9% to 82.0% over atomic facts in logged turns from a commercial deployment, and cut strategy misapplication under changed player state from 18.1% to 10.4% compared with a prose guide carrying the same content. Decision traces written on the serving path made claim verification possible at 86.4% precision and localized 26.7% of delayed-feedback failures, which preceding-turn methods cannot do. The sharpest claim is about revision: rerunning the unchanged system recovered 38.4% of previously failing turns, and only one of 15 shipped revision units passed an independent release gate, so most apparent agent fixes are generation noise.

What carries the argument. The load-bearing object is the strategy package: a versioned unit of game knowledge carrying evidence, instructional text, and applicability conditions that evaluate to true, false, or unknown over the structured player-state snapshot, checked by deterministic conditional-knowledge rules before the reply model writes anything. Three supporting mechanisms carry the argument: decision traces that bind each response to the packages selected, condition outcomes, evidence, tool results, and rule versions; a condition–rule dependency index that gives any edit a known blast radius; and a release gate that replays the candidate against an unchanged-system baseline, decomposing validated gain into in-scope, out-of-scope, and unexposed groups and requiring a positive in-scope lower bound before release. The three-valued logic is what lets the system refuse to guess when state is missing, and the recorded condition outcomes are what let it name the first unsupported stage when a late complaint arrives.

Core claim. The paper's central claim is that applicability, not content, is the bottleneck for state-dependent advice, and that making prerequisites explicit artifacts changes what the system can measure and repair. A compiled strategy package carries conditions that evaluate to true, false, or unknown over a structured snapshot of player state and game version; only satisfied conditions permit a recommendation, violated ones trigger an alternative that names the unmet requirement, and unknown ones force a clarification instead of a guess. Because the same condition objects are linked to the execution rules that consume them and to the state fields they read, a decision trace can later attribute a complaint to knowledge, to a rule, or to state processing, and a revision knows its blast radius. On the evidence of one commercial deployment, the package form matches a prose guide on recorded turns (82.0% vs 82.5%) but beats it under changed state (10.4% vs 18.1% applicability violations), a comparison the authors note rests on 36 base turns and falls short of conventional significance. The release gate built from traces plus an unchanged-system rerun admits only 1 of 15 deployed revision units; an unchanged rerun recovers 38.4% of failing turns, the same order as planted revisions whose real gains were near zero.

The premise it rests on. The load-bearing premise is that the automated judge, a self-hosted model from the same family as the Coach that agrees with expert correctness labels at only kappa = 0.42 and detects expert-labeled errors at 55.6% precision and 52.6% recall, measures the reported differences faithfully, and the paper acknowledges this shared-model sensitivity as a limitation without quantifying it.

If the paper is right (Pith's inferences, not the authors')

Content quality and applicability control are separable engineering goals: a prose guide that matches packages on recorded turns still misapplies strategies at nearly double the rate once player state changes.

The release gate costs one reply generation per held-out turn because replay holds requests, histories, player states, planning decisions, and tool outputs fixed, and condition evaluation issues no model calls, making independent validation of every candidate affordable.

Recovery of previously failing turns is not evidence of repair: an unchanged system recovers 38.4% of failures, so any revision workflow that accepts a candidate because the reported case now passes will ship mostly inert changes, as 13 of the paper's 15 deployed units did.

Knowledge edits and execution-rule edits must cross the release boundary together: the measured knowledge–execution interaction of −9.6 points shows the combined change underperforms the sum of its separately measured parts.

Advice systems should treat missing state as unresolved rather than guessing: unknown-condition violations were the widest package-vs-guide gap (15.3% vs 26.4%), and the unknown outcome routes to retrieval or a clarifying question.

The 38.4% rerun-recovery figure implies a general caution for LLM agent improvement pipelines: any validation that compares a candidate against a failing prior run instead of against a replayed null will attribute generation noise to the fix, and the null-replay gate should transfer to any agent whose execution is deterministic given a recorded trace.

If the package-vs-guide violation gap reproduces under an independent judge, the same three-valued-condition pattern should transfer to other state-dependent advice domains such as finance, health, or configuration assistants, where guessing from missing data is costly and the same ask-don't-guess semantics apply.

The one revision unit that gained 25.0 points in scope while losing 20.8 outside it suggests that scope decomposition, not just an overall accuracy number, should be a standard report for any agent revision, because a single number can hide a redistribution of errors across traffic.

Since the violation gap concentrates in unknown state, the paper suggests but does not prove that the main benefit of explicit conditions is refusing to fabricate when data is missing rather than judging known situations better; a targeted experiment varying only the unknown-handling behavior could isolate that effect.

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

Every value above, in the review's public record (JSON)

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.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

A deployed LLM coaching agent answers requests whose correct response depends on player state and game version, so advice that is right for one player can be wrong for the next. We describe TactiNavi, the serving system behind the LLM coach of a commercial open-world game, which makes strategy applicability an explicit artifact shared by offline compilation, online execution, failure diagnosis, and release. An offline compiler turns versioned game evidence and expert guidance into strategy packages with explicit prerequisites, the serving plane evaluates those prerequisites as three-valued conditions over structured player state and records decision traces, and a control plane uses the resulting condition-rule dependencies to validate revisions on held-out traffic before release. In a retrospective evaluation of 1,122 logged coaching turns from three deployment periods, condition-bearing packages raise correctness from 57.9% to 82.0% over atomic facts and reduce strategy misapplication under changed player state from 18.1% to 10.4%, while trace-based claim verification reaches 86.4% precision at a median reply latency of 2.07 s. The deployment also showed that most apparent fixes are generation noise, because an unchanged-system rerun recovers 38.4% of failing turns and only one of 15 shipped revision units passes the release checks retrospectively. We distill these results into lessons on treating applicability as a first-class serving object, sizing traces for later diagnosis, and building the null model into the release path.

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 · 2026-11-04

The authors' main claim, as the classifier picked it out 0.49

“In a retrospective evaluation of 1,122 logged coaching turns from three deployment periods, condition-bearing packages raise correctness from 57.9% to 82.0% over atomic facts and reduce strategy misapplication under changed player state from 18.1% to 10.4%, while trace-based claim verification reaches 86.4% precision at a median reply latency of 2.07 s.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.04 · states limitation 0.89 · human participants 0.31 · controlled comparison 0.82 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.87 · mentions code or data availability 0.04

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.

SentenceRoleThe authors' own workStates the main conclusionStrength
A deployed LLM coaching agent answers requests whose correct response depends on player state and game version, so advice that is right for one player can be wrong for the next. background 0.98 0.24 0.04 asserted 1.00
We describe TactiNavi, the serving system behind the LLM coach of a commercial open-world game, which makes strategy applicability an explicit artifact shared by offline compilation, online execution, failure diagnosis, and release. contribution 0.99 0.87 0.13 asserted 1.00
An offline compiler turns versioned game evidence and expert guidance into strategy packages with explicit prerequisites, the serving plane evaluates those prerequisites as three-valued conditions over structured player state and records decision traces, and a control plane uses the resulting condition-rule dependencies to validate revisions on held-out traffic before release. method 0.99 0.91 0.11 asserted 0.99
In a retrospective evaluation of 1,122 logged coaching turns from three deployment periods, condition-bearing packages raise correctness from 57.9% to 82.0% over atomic facts and reduce strategy misapplication under changed player state from 18.1% to 10.4%, while trace-based claim verification reaches 86.4% precision at a median reply latency of 2.07 s. result 1.00 0.96 0.46 asserted 1.00
The deployment also showed that most apparent fixes are generation noise, because an unchanged-system rerun recovers 38.4% of failing turns and only one of 15 shipped revision units passes the release checks retrospectively. result 0.81 0.95 0.13 asserted 1.00
We distill these results into lessons on treating applicability as a first-class serving object, sizing traces for later diagnosis, and building the null model into the release path. interpretation 0.92 0.82 0.40 asserted 0.90

What kind of paper the classifier reports this is: method or tool 0.99 · evidence from analysis of existing data 0.54

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, 2026-11-04: 1 request, 5,264 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 (1)
Figure 1 of the paper
Figure 1. Overview of TactiNavi. The Strategic Knowledge Compiler supplies versioned strategy packages for

Figures are the authors', reproduced from the paper.

References (51)
  1. [1] React: Synergizing reasoning and acting in language models , author=. arXiv preprint arXiv:2210.03629 , year= on PithPith's reviewarXiv
  2. [2] Reflexion: Language agents with verbal reinforcement learning , author=. Advances in neural information processing systems , volume= on Pith
  3. [3] Retrieval-augmented generation for knowledge-intensive nlp tasks , author =. Advances in neural information processing systems , volume= on Pith
  4. [4] Agentbench: Evaluating llms as agents , author =. International Conference on Learning Representations , pages= on Pith
  5. [5] Toolformer: Language models can teach themselves to use tools , author=. Advances in neural information processing systems , volume= on Pith
  6. [6] Minedojo: Building open-ended embodied agents with internet-scale knowledge , author=. Advances in Neural Information Processing Systems , volume= on Pith
  7. [7] Pokellmon: A human-parity agent for pokemon battles with large language models , author=. arXiv preprint arXiv:2402.01118 , year= on PithPith's reviewarXiv
  8. [8] Raptor: Recursive abstractive processing for tree-organized retrieval , author=. International Conference on Learning Representations , pages= on Pith
  9. [9] Assisting in writing wikipedia-like articles from scratch with large language models , author=. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages= on Pith
  10. [10] Agentboard: An analytical evaluation board of multi-turn llm agents , author=. Advances in neural information processing systems , volume= on Pith
  11. [11] Self-refine: Iterative refinement with self-feedback , author =. Advances in neural information processing systems , volume= on Pith
  12. [12] Critic: Large language models can self-correct with tool-interactive critiquing , author =. International Conference on Learning Representations , pages= on Pith
  13. [13] Large language models cannot self-correct reasoning yet , author=. International conference on learning representations , pages= on Pith
  14. [14] Judging llm-as-a-judge with mt-bench and chatbot arena , author =. Advances in neural information processing systems , volume= on Pith
  15. [15] Yannakakis, Georgios N. and Spronck, Pieter and Loiacono, Daniele and Andr\'. Artificial and Computational Intelligence in Games , pages =. 2013 , volume =. doi:10.4230/DFU.Vol6.12191.45 , annote = on Pithdoi:10.4230/dfu.vol6.12191.45
  16. [16] Cognitive architectures for language agents , author=. arXiv preprint arXiv:2309.02427 , year= on PithPith's reviewarXiv
  17. [17] Generative agents: Interactive simulacra of human behavior , author=. Proceedings of the 36th annual acm symposium on user interface software and technology , pages= on Pith
  18. [18] Voyager: An open-ended embodied agent with large language models , author=. arXiv preprint arXiv:2305.16291 , year= on PithPith's reviewarXiv
  19. [19] Describe, Explain, Plan and Select: Interactive Planning with Large Language Models Enables Open-world Multitask Agents , author =. Proceedings of the 37th International Conference on Neural Information Processing Systems. Red Hook: Curran Associates Inc , pages= on Pith
  20. [20] Spring: Studying papers and reasoning to play games , author=. Advances in Neural Information Processing Systems , volume= on Pith
show 31 more
  1. [21] Retrieval augmented language model pre-training , author=. International conference on machine learning , pages=. 2020 , organization= on Pith
  2. [22] Self-rag: Learning to retrieve, generate, and critique through self-reflection , author=. International conference on learning representations , pages= on Pith
  3. [23] Active retrieval augmented generation , author=. Proceedings of the 2023 conference on empirical methods in natural language processing , pages= on Pith
  4. [24] Webgpt: Browser-assisted question-answering with human feedback , author=. arXiv preprint arXiv:2112.09332 , year= on PithPith's reviewarXiv
  5. [25] Toolllm: Facilitating large language models to master 16000+ real-world apis , author =. International Conference on Learning Representations , pages= on Pith
  6. [26] An llm compiler for parallel function calling , author=. arXiv preprint arXiv:2312.04511 , year= on PithPith's reviewarXiv
  7. [27] Webarena: A realistic web environment for building autonomous agents , author =. International Conference on Learning Representations , pages= on Pith
  8. [28] Gaia: a benchmark for general ai assistants , author=. International Conference on Learning Representations , pages= on Pith
  9. [29] Dspy: Compiling declarative language model calls into state-of-the-art pipelines , author =. International Conference on Learning Representations , pages= on Pith
  10. [30] Automatic prompt optimization with ``gradient descent'' and beam search , author=. Proceedings of the 2023 conference on empirical methods in natural language processing , pages= on Pith
  11. [31] G-eval: NLG evaluation using gpt-4 with better human alignment , author=. Proceedings of the 2023 conference on empirical methods in natural language processing , pages= on Pith
  12. [32] Hidden technical debt in machine learning systems , author=. Advances in neural information processing systems , volume= on Pith
  13. [33] The ML test score: A rubric for ML production readiness and technical debt reduction , author=. 2017 IEEE international conference on big data (big data) , pages=. 2017 , organization= on Pith
  14. [34] Next-generation agentic reinforcement learning systems enable self-evolving agents , author=. arXiv preprint arXiv:2607.01120 , year= on PithPith's reviewarXiv
  15. [35] Ouroboros: A self-developing frontier coding agent with reviewed core evolution , author=. arXiv preprint arXiv:2608.08311 , year= on PithPith's reviewarXiv
  16. [36] RRSI: Regularized recursive self-improvement of agent harnesses , author =. arXiv preprint arXiv:2609.24972 , year= on PithPith's reviewarXiv
  17. [37] ModularRSI: Modular and generalizable recursive harness self-improvement , author =. arXiv preprint arXiv:2609.14857 , year= on PithPith's reviewarXiv
  18. [38] GDPevo: Evaluating Agent Self-Evolution on Real Business Tasks , author =. arXiv preprint arXiv:2608.03764 , year= on PithPith's reviewarXiv
  19. [39] Which agent causes task failures and when? on automated failure attribution of llm multi-agent systems , author =. arXiv preprint arXiv:2505.00212 , year= on PithPith's reviewarXiv
  20. [40] Yuan, Xiaoyang and Liu, Qi and Ruan, Yubin and Mou, Xinyi and Zhang, Zhuomeng and Wang, Wenjin and Jiao, Hanying and Wu, Di and Xu, Mingye and Bin, Yi and others , journal= on Pith
  21. [41] Liu, Qi and Yuan, Xiaoyang and Ruan, Yubin and Zhang, Zhuomeng and Wang, Wenjin and Wu, Di and Xu, Mingye and Mou, Xinyi and Yin, Xingxi and Feng, Ke and others , journal=. State-Grounded Conditioning: Wrapping User-Facing on Pith
  22. [42] Game Coaching System Using 3D CNN and LLM , author=. 2026 IEEE 2nd International Conference on Consumer Technology (ICCT-Pacific) , pages=. 2026 , organization= on Pith
  23. [43] MOBA Coach: Exploring and analyzing multiplayer online battle arena data , author=. International Symposium on visual computing , pages=. 2021 , organization= on Pith
  24. [44] Crafting champions: an observation study of esports coaching processes , author=. Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems , pages= on Pith
  25. [45] Agents thinking fast and slow: A talker-reasoner architecture , author=. arXiv preprint arXiv:2410.08328 , year= on PithPith's reviewarXiv
  26. [46] ExInCOACH: Strategic exploration meets interactive tutoring for context-aware game onboarding , author=. Information Fusion , pages=. 2026 , publisher= on Pith
  27. [47] Self-evolving Agentic Customer Support System at LinkedIn , author=. arXiv preprint arXiv:2608.10224 , year= on PithPith's reviewarXiv
  28. [48] Audited skill-graph self-improvement for agentic llms via verifiable rewards, experience synthesis, and continual memory , author=. arXiv preprint arXiv:2512.23760 , year= on PitharXiv
  29. [49] RewardHackingAgents: Benchmarking Evaluation Integrity for LLM ML-Engineering Agents , author=. 2026 IEEE 42nd International Conference on Data Engineering Workshops (ICDEW) , pages=. 2026 , organization= on Pith
  30. [50] Reward hacking benchmark: measuring exploits in LLM agents with tool use , author=. arXiv preprint arXiv:2605.02964 , year= on PithPith's reviewarXiv
  31. [51] Deepseek-v4: Towards highly efficient million-token context intelligence , author =. arXiv preprint arXiv:2606.19348 , year= on PitharXiv

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.

No comments yet.

Sign in to comment