acceptodds
Under review as a conference paper at ICLR 2027

SAGE: Structured Strategic Reasoning for Efficient LLM Game Playing

Abstract

A strong LLM strategic agent should reason prospectively over uncertain futures, adapt its strategy to opponents’ behavioral tendencies, and continuously recalibrate its decision process from interaction experience. However, incorporating these sources in free-form reasoning could lead to unsupported strategic assumptions, inconsistent opponent estimates, and harmful interference from irrelevant historical interactions. To address these issues, we propose SAGE, a training-free inference-time framework that structures LLM strategic reasoning around three coordinated operations: anchor, adapt, and recalibrate. SAGE first anchors reasoning to an equilibrium policy that provides a strategically valid prior. It then conditions deviations from this anchor on a soft belief over opponent behavioral tendencies, enabling opponent-specific exploitation. Finally, SAGE distills strategically related interactions into counterfactual hypotheses about previously missing considerations, allowing past experience to recalibrate the model’s reasoning. We evaluate SAGE on three repeated imperfect-information games: Leduc Hold’em, Liar’s Dice, and Goofspiel, against various opponent types in each game. Compared with reasoning-intensive LLM agents, including Suspicion-Agent, ReTA, Agent-Pro, EMO, and Hypothetical Minds, SAGE achieves up to a 127.6% payoff improvement in Liar’s Dice while reducing input and output token usage by up to 80% and 90%, respectively. In direct match-up play, it attains non-negative mean payoff against 5/10, 8/10, and 8/10 evaluated opponents in Leduc Hold’em, Liar’s Dice, and Goofspiel, respectively, while using relatively fewer tokens. Code is available at https://anonymous.4open.science/r/SAGE-23C2.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
LLM game playing agent

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_XatNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1(−1)
Read review
Paper summary

This paper proposes SAGE, a training-free inference-time framework for strategic LLM agents in repeated imperfect-information games. SAGE structures decision making into three components: (1) anchoring the LLM to a solver-derived equilibrium policy, (2) adapting its decisions using an online belief over opponent behavioral tendencies, and (3) recalibrating reasoning using counterfactual hypotheses extracted from relevant past interactions. The method is evaluated on Leduc Hold'em, Liar's Dice, and Goofspiel against a range of LLM and non-LLM baselines. The experiments suggest that SAGE can improve payoff while using substantially fewer inference tokens than several reasoning-intensive baselines.

Critical strengths
  1. The paper is generally clear and well organized. The motivation and the anchor-adapt-recalibrate framework are easy to follow, and the roles of the three components are explained clearly.

  2. The empirical evaluation is fairly extensive. The paper evaluates the method across three different games, a broad set of baselines and opponent types, and additionally includes ablations, efficiency analysis, model-scaling experiments, and opponent-shift experiments.

  3. The experimental results are presented effectively overall. Apart from the readability issue with Figure 1, the remaining figures and tables are generally informative and make it relatively easy to understand the performance, efficiency, and behavior of the proposed method.

Critical weaknesses

W1. The comparison does not fully separate the benefit of structured LLM reasoning from access to a strong external strategic prior.

SAGE is presented as a structured inference-time reasoning framework, but its main configuration is additionally provided with an approximate Nash-equilibrium policy computed by a well-trained MCCFR solver. This gives SAGE a strong game-specific strategic reference that many reasoning-intensive baselines must instead approximate through search or free-form reasoning.

The paper does include useful controls, including LLM(Eq) and experiments with weaker strategic references, so I do not view the use of an external solver itself as problematic. However, these results also indicate that the quality of the external reference contributes substantially to performance. It would therefore be useful to more clearly separate the improvement attributable to the proposed reasoning framework from that provided by the solver-derived strategic prior.

W2. The evaluation lacks a matched non-LLM adaptive baseline that combines the same equilibrium anchor with online opponent-specific optimization.

SAGE's main advantage over the equilibrium baseline comes from deviating from a strong solver-derived policy based on an online belief over opponent tendencies. However, the current baselines separate these ingredients: MCCFR and LLM(Eq) provide equilibrium guidance without comparable opponent-specific optimization, while planning and opponent-modeling baselines do not receive the same strategic anchor.

This makes it difficult to determine whether the gains specifically require structured LLM reasoning, or whether they could be obtained by a more direct algorithmic procedure that uses the same equilibrium policy and opponent belief to compute an approximate best response. A particularly informative baseline would therefore hold the equilibrium reference and opponent belief fixed, but replace the LLM adaptation step with dynamic search or expected-payoff maximization against the estimated opponent distribution.

W3. The evaluation does not examine reasoning-enabled / thinking-mode models.

All reported LLM experiments appear to use inference with thinking mode disabled, including the model-scaling study over Qwen3 backbones. This is a reasonable choice for controlling inference cost, but it leaves open how SAGE interacts with models that already perform substantially more explicit reasoning. This seems particularly relevant here because SAGE is itself designed to structure the reasoning process by providing an equilibrium anchor, opponent belief, and historical recalibration signals.

I do not view this as a major weakness, but an additional comparison with thinking mode enabled would make the evaluation more complete. It would be useful to know whether stronger native reasoning complements SAGE, reduces its marginal benefit, or changes which of its components are most useful.

W4. Figure 1 is difficult to read due to its dense layout.

Figure 1 contains several small panels with tightly packed labels, annotations, and text, making parts of the figure difficult to parse at normal viewing size. Since the figure is important for motivating the three failure modes that SAGE is designed to address, improving its readability would strengthen the presentation. Enlarging the text, reducing visual clutter, or splitting the figure into separate panels could help.

This paper is quite borderline. Now I will slightly lean forward rejection but if rebuttal resloves my problems, I think it has good chance for acceptance.

Questions for authors

Questions for Authors

  1. How much of SAGE's improvement should be attributed to the external equilibrium prior versus the proposed structured reasoning framework itself?
    The strong MCCFR reference appears to contribute substantially to performance in some games. Could the authors more explicitly quantify the incremental gain of SAGE over an agent with access to the same equilibrium reference, and clarify whether the computational cost of obtaining this reference is included in the reported efficiency comparison?

  2. How would SAGE compare with a non-LLM adaptive baseline that has access to the same equilibrium policy and opponent belief?
    In particular, could the authors evaluate an algorithmic baseline that uses the estimated opponent distribution to perform dynamic search or approximate expected-payoff maximization / best-response computation? This would help determine whether the gains specifically require LLM-based structured reasoning.

  3. Does SAGE remain beneficial when the backbone uses a reasoning/thinking mode?
    The current experiments appear to use thinking mode disabled, including the model-scaling study. Since SAGE explicitly structures the model's reasoning process, it would be informative to know whether stronger native reasoning complements SAGE, reduces its marginal benefit, or changes which components are most useful.

  4. How does the opponent-belief mechanism behave when the opponent is not well represented by the predefined behavioral prototypes?
    Could the authors clarify how SAGE handles genuinely novel or out-of-distribution opponent strategies, and whether performance degrades when none of the predefined prototypes provides a good match?

  5. Is the counterfactual-hypothesis formulation itself necessary beyond simply providing relevant historical information?
    The ablation shows that removing historical recalibration hurts performance, but it does not isolate the proposed counterfactual representation. Could the authors compare against matched alternatives that use the same retrieved interactions and similar context budget, such as concise lessons, summaries, or direct trajectory replay?

4 comments

We thank the reviewer for the constructive assessment. We have added controls for unstructured reasoning and historical lessons, and comparisons with Bayes' Bluff and Restricted Nash Response (RNR).

W1 / Q1. Contribution of the strategic prior versus structured reasoning. The equilibrium reference provides a strategic anchor, while SAGE structures how the LLM combines it with opponent beliefs and historical information. Against the common opponent styles, SAGE's mean payoff is 1.139, 0.494 and 2.744 across Leduc, Liar's Dice and Goofspiel, compared with 0.953, 0.394 and 1.411 for LLM(Eq). In the new unstructured control, the payoff difference between full SAGE and the control is +0.2047, +0.1402 and +0.1839, respectively (Table 3). Together, these results show the additional benefit of SAGE's structured reasoning. With weaker MCCFR and code-based search, Leduc payoffs are 0.825 and 0.842 versus 1.139 with the original reference.

W2 / Q2. Non-LLM adaptive baselines. We have added Bayes' Bluff and RNR to Tables 1 and 2. Against common Leduc opponent styles, Bayes' Bluff and RNR obtain mean payoffs of 1.05 and 1.20, compared with 1.139 for SAGE; in direct matches against them, SAGE obtains +0.25 and +0.10, respectively. These results provide algorithmic opponent-adaptation comparisons as reference. Nevertheless, our main goal is to mitigate instability in LLM reasoning by structuring the use of strategic and historical information.

W3 / Q3. Thinking mode. Appendix E reports an additional comparison with thinking mode enabled and disabled against common opponent types. With thinking enabled, SAGE achieves mean payoffs of 1.411, 0.500 and 2.558 in Leduc, Liar's Dice and Goofspiel, respectively, compared with 0.778, 0.223 and 1.408 for Vanilla. With thinking disabled, SAGE achieves 1.139, 0.494 and 2.744 versus 0.778, 0.217 and 1.444 for Vanilla. Thus SAGE remains ahead of Vanilla with thinking enabled, and the change from enabling thinking is quite small.

Q4. Unrepresented opponents. All opponents in the direct-match evaluation are previously unseen agents, including Bayes' Bluff and RNR (Table 2). The predefined behavioral prototypes serve as anchors for interpreting observed behavior; they are not a list of opponents that SAGE has previously encountered. SAGE updates its belief from the interaction history and adapts to each opponent using the same set of anchors. These direct matches therefore evaluate SAGE against previously unseen agents, complementing the common-style and style-shift evaluations.

Q5. Historical hypotheses versus simpler representations. We added a history-lessons control. Full SAGE exceeds this control by +0.1283, +0.0846 and +0.1071 mean payoff per game in Leduc, Liar's Dice and Goofspiel, respectively (Table 3). This comparison shows the benefit of counterfactual recalibration over the tested history-lessons alternative in all three games.

W4 / Figure 1. We revised the preliminary-study presentation and clarified its accompanying discussion. We will ensure the final PDF's labels remain readable at normal viewing size.

Reviewer_XatNOpenReview ↗

Response

W1: UNRESOLVED

REASON: My main concern was whether the improvement comes from structured LLM reasoning or simply from providing additional strategic information, especially the strong MCCFR prior. Although the authors added an unstructured control and showed higher payoff, they still did not directly measure whether the reasoning instability identified in Figure 1 was actually reduced. I don't think higher payoff alone is enough to support this claim.

W2: UNRESOLVED

REASON: I wanted to know whether traditional algorithms could achieve similar results when given the same equilibrium policy and opponent belief. The authors added Bayes' Bluff and RNR, which is useful, but this is not the controlled comparison I asked for. Also, RNR even achieves higher average payoff than SAGE in Leduc (1.200 vs. 1.139), so I am still not convinced about the additional value of LLM reasoning here.

W3: RESOLVED

REASON: I wanted to see whether SAGE still helps when the LLM already has thinking capabilities. The authors added the experiments and showed that SAGE still outperforms Vanilla with thinking enabled, which addresses my concern.

W4: RESOLVED

REASON: My concern was about the readability of Figure 1. The authors acknowledged this issue and said they revised the presentation, which is acceptable to me.

Q4: PARTIALLY RESOLVED

REASON: The experiments with previously unseen agents are meaningful, and I appreciate this clarification. However, I was asking about OOD opponent behaviors, for example, a player using a state-dependent mixed strategy that cannot be well represented by the predefined prototypes. Unseen agents and OOD strategies are not necessarily the same thing, so my concern is only partially addressed.

Q5: RESOLVED

REASON: I wanted to know whether counterfactual hypotheses are actually better than simply extracting lessons from history. The authors added a history-lessons baseline and showed improvements across all three games, which I think is acceptable.

Updated Decision:

Therefore, after carefully reconsidering the rebuttal, I will not lower my score (obviously not a 1). However, I now have more confidence that this paper is on the rejection side, since my main concerns in W1 and W2 remain unresolved and might seem hard to be solved.

Detailed Key Reason:

My biggest concern after reading the rebuttal is that I am still not sure about the paper's core contribution. Section 4.2 emphasizes better game payoff and inference efficiency, but when I questioned whether traditional algorithms could achieve similar results, the authors responded that their main goal is to improve LLM reasoning stability. However, although Figure 1 identifies three reasoning failures, the paper does not directly measure whether SAGE actually reduces them. I don't think higher payoff or win rate alone can demonstrate better reasoning stability. If this is the main contribution, I would expect the paper to focus more on this problem and provide corresponding metrics. On the other hand, if the goal is better game-playing performance, then a fair comparison with traditional algorithms using the same strategic information is still missing. I appreciate the additional experiments, but the paper seems to move between these two claims without fully supporting either one, which makes the core contribution kinda unclear to me.

For AC After the rebuttal, the decision goes from a borderline to a clear 2.

Thank you for your follow-up. We would like to clarify three points central to your assessment.

1. We directly evaluated reasoning, not only payoff. The revised paper includes a reasoning-trace analysis in Figure 7. We used Qwen3-32B to assess 100 sampled decisions per method and game for rule violations, contradictions, and fabrications. Across Leduc, Liar's Dice, and Goofspiel, SAGE's combined flagged-category counts are 60, 22, and 23, compared with 85, 32, and 29 for LLM(Eq), respectively. These are direct measurements of unsupported reasoning, rather than inferences from win rate or payoff. Categories can overlap within a decision. Figure 7 provides evidence on unsupported reasoning; the payoff results and unstructured ablation provide complementary evidence on game performance.

2. We added and evaluated the requested non-LLM references. Tables 1–2 report results for Bayes' Bluff and RNR against common opponents and in direct matches with SAGE. In Leduc, RNR achieves a higher aggregate payoff (1.200 versus 1.139), while SAGE obtains +0.10 against RNR in their direct matchup. We present both results as part of the comparison.

Our contribution should be assessed against the problem it addresses. SAGE is designed to mitigate the reasoning instabilities we identify in LLM agents, rather than to propose a new game solver. We have added Bayes' Bluff and RNR as non-LLM comparisons, as requested, to provide useful strategic reference points. Their performance, including cases where a traditional method achieves a higher payoff, does not negate SAGE's contribution to understanding and improving LLM reasoning. To evaluate that contribution directly, Figure 7 analyzes unsupported reasoning in the agents' traces, while the unstructured control tests the effect of SAGE's reasoning organization. We respectfully ask that these results be considered alongside the game-payoff comparisons.

3. Direct matches already evaluate unseen agents and out-of-distribution behaviors. All direct-match opponents in Table 2 are previously unseen agents, and their behaviors are out-of-distribution relative to the common opponent styles. They include diverse LLM agents as well as the non-LLM methods Bayes' Bluff and RNR, each following its own decision procedure. SAGE interacts with these opponents and adapts online across all three games. The direct-match results therefore already provide evidence on the requested generalization to unseen opponents and out-of-distribution behavior; the style-shift experiments further evaluate adaptation when opponent behavior changes over time. We ask that these existing results be considered in the assessment.

Reviewer_XatNOpenReview ↗

Final Assessment for SAGE by Reviewer XatN

1. Regarding Figure 7 and Reasoning Stability

I appreciate the authors adding Figure 7. However, its metrics (rule violations, contradictions, and fabrications) seem relatively superficial and do not directly address the three reasoning problems originally identified in Figure 1 (now Figure 2):

Original Reasoning ProblemAddressed by Figure 7?
Unsupported Planning AssumptionsPartially, through fabrications and rule violations
Inconsistent Opponent EstimatesNot directly measured
Sensitivity to Irrelevant HistoryNot directly measured

I am also concerned about using a 32B model as the judge without further validation, as smaller judges may misclassify nuanced strategic reasoning. The reported 60 error counts across 100 decisions in Leduc also seem surprisingly high, and it is unclear whether these counts overlap or what exactly they represent. I originally did not want to focus on these kinds of reasoning errors, but since the authors now emphasize reasoning stability as their main contribution, I would expect more direct and reliable evidence addressing the specific problems they identified. Therefore, my original W1 concern remains unresolved.

2. Regarding the Paper's Core Contribution and Traditional Algorithm Comparisons

I would like to clarify that my concern was not asking SAGE to propose a new game solver or outperform every traditional algorithm. I was actually following the paper's own claims. For example, the Abstract, Introduction, Section 4.2, and Conclusion all emphasize improvements in game performance and inference efficiency, and Section 4.2 explicitly states that SAGE outperforms most agents across the three games. So I think it is reasonable to ask whether traditional algorithms could achieve similar results with the same strategic information. I appreciate the addition of Bayes' Bluff and RNR, but it is still unclear whether they use exactly the same equilibrium reference and opponent belief as SAGE, which was the main point of my original question. That said, I understand if the authors consider improving LLM reasoning stability as their main contribution, and I don't think they necessarily need to outperform traditional solvers to demonstrate that. However, in that case, game payoff should be supporting evidence rather than the main evidence for reasoning stability. I also appreciate the new Figure 7, which is a meaningful addition, but it only evaluates part of the reasoning failures identified earlier and does not fully isolate the effect of structured reasoning from the additional strategic information. This is why I still find the paper's core contribution kinda unclear, rather than simply disagreeing with the choice of baselines.

3. Regarding OOD Opponents and the Revised Limitations

I appreciate the authors' clarification and the experiments with previously unseen agents. However, I noticed that the original paper did not mention this limitation, while the revised Section 5 now acknowledges that the predefined behavioral prototypes may poorly represent unseen opponents. This is actually the concern I raised in Q4. The authors still have not quantified how different these agents' behaviors are from the predefined prototypes, and using different algorithms does not necessarily mean their strategies are OOD. I appreciate the authors acknowledging this limitation, but I don't think moving this concern into the Limitations section is enough to resolve it, especially when the rebuttal claims that OOD behaviors have already been evaluated. This makes it a bit unclear to me whether the authors consider this issue addressed or still an open limitation.

Final Decision

This is my final decision. The authors did not conduct all the experiments I suggested, which is totally fine. However, my biggest concern is that the newly added experiments still do not provide the crucial evidence needed to support the paper's main claims. I still find the core contribution and its claims rather vague. Therefore, after considering all the rebuttal responses and the revised manuscript, I have decided to lower my Soundness score to 1. I encourage the authors to make the central claim clearer for future submission.

For AC

I have lowered my Soundness score to 1. I also think the writing quality has slightly decreased in the revised manuscript, but I would like to keep the Presentation score at 2, as the authors have added quite a lot of material, and I understand the rebuttal period might have been hectic. To sum up, I suggest rejection. I no longer consider this a borderline case.

Reviewer_kuBfOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This manuscript proposes SAGE, an inference-time framework for LLM-based game-playing agents in repeated imperfect-information games. SAGE structures strategic reasoning around three signals: (1) an equilibrium-guided lookahead anchor, (2) a soft opponent-belief contextual guidance signal, and (3) counterfactual historical recalibration via hypotheses generated from similar past interactions. The paper evaluates SAGE on Leduc Hold’em, Liar’s Dice, and Goofspiel against common opponent styles and several LLM/non-LLM agent baselines. It reports improved payoff over vanilla LLM agents, competitive performance against reasoning-intensive baselines, and lower token/API usage. Ablations, token-usage breakdowns, scaling with Qwen3 model sizes, style-shift experiments, and qualitative case studies are included.

Critical strengths
  1. Clear and intuitive problem framing. The paper identifies a real issue in free-form LLM strategic reasoning: heterogeneous evidence—equilibrium priors, opponent tendencies, and historical experience—is often mixed without explicit roles. The “anchor, adapt, recalibrate” principle is easy to understand and practically motivated.
  2. Modular and practical method. SAGE does not require LLM parameter updates and combines equilibrium guidance, opponent beliefs, and counterfactual hypotheses in a structured prompting/reasoning pipeline. The design is engineering-friendly and potentially applicable beyond the three games studied.
  3. Reasonable empirical coverage for a first study. The paper evaluates on three repeated imperfect-information games, compares against multiple LLM and non-LLM baselines, and includes ablations, token-efficiency analysis, API-call analysis, model-size scaling, and qualitative case studies.
  4. Efficiency claims are potentially valuable. The token-usage breakdown suggests SAGE can achieve competitive payoff with substantially fewer tokens than reasoning-intensive methods such as ReTA, EMO, Suspicion-Agent, and Hypothetical Minds. If verified, this is a meaningful contribution.
  5. Useful appendix and prompt transparency. The appendix provides detailed prompts, opponent settings, belief-update details, and case studies. This is helpful for reproducibility and for understanding the method’s behavior.
Critical weaknesses
  1. Limited novelty as an integration/prompting framework. The individual components—equilibrium policy, opponent modeling, and reflection/history-based adaptation—are not new. The contribution is mainly their orchestration into an inference-time LLM reasoning pipeline. The paper needs a deeper conceptual or theoretical argument for why this particular factorization is principled, rather than mostly empirical prompt engineering.
  2. Limited experimental scope and external validity. The three games are relatively small and have compact action spaces. SAGE relies on game-specific MCCFR equilibrium policies and game-specific behavioral prototypes/features. It is unclear whether the approach scales to larger games, long-horizon interactions, unknown opponent types, or domains without a readily computable equilibrium prior. The conclusion acknowledges this only briefly.
  3. Insufficient statistical rigor. Results are reported without confidence intervals, standard errors, or significance tests. The evaluation uses 30 consecutive games as P0 and 30 as P1, but the number of independent runs/seeds is unclear. Given the variance in poker-like games, claims such as “127.6% improvement” and “positive mean payoff against 5/10, 8/10, 8/10 opponents” need error bars and significance testing. Some results are mixed—e.g., SAGE’s Leduc win rate is lower than ReTA’s—and this should be analyzed rather than glossed over.
  4. Fairness and reproducibility of comparisons are not fully established. SAGE receives privileged game-specific equilibrium guidance from MCCFR, while some baselines may not. Baseline tuning, prompt budgets, and inference budgets are not clearly matched. The LLM backend is described as “DeepSeek v4 flash,” and several references/API versions appear future-dated or unusual in the provided text. These need verification. The paper promises code release, but no code is available in the draft; LLM API nondeterminism and version changes also threaten reproducibility.
  5. Opponent-belief module is game-specific and may not generalize. SAGE maintains beliefs over predefined behavioral prototypes such as LAG, TAG, Nit, Bluffer, etc., using game-specific observable features. This is reasonable for the chosen benchmarks but may not transfer to real or open-ended opponents. The paper itself shows misclassification cases. The benefit of the belief module may be confounded with the equilibrium anchor and prompt design.
  6. Counterfactual recalibration is under-analyzed. The counterfactual hypothesis mechanism depends on retrieval similarity, LLM-generated hypotheses, and a validation step. The ablation shows a smaller contribution than the other components. Risks of hallucinated hypotheses, retrieval bias, and sensitivity to the number/quality of hypotheses are not sufficiently analyzed. More sensitivity studies and failure analysis are needed.
  7. Presentation and formatting issues in the provided PDF. The extracted text begins with a full page of numeric tokens, and there are missing or malformed figures/tables, missing “F” in “Figure,” and inconsistent table formatting. If these are present in the actual submission, they are serious presentation problems. The appendix is very long and detailed, but the main paper should more clearly summarize the method and evidence. Some claims need more precise qualification.
Questions for authors

(a) Beyond combining existing ingredients, what is the core technical contribution that distinguishes SAGE from prior work on equilibrium-guided prompting, opponent modeling, and reflection-based LLM agents?

(b) SAGE relies on game-specific MCCFR equilibrium policies and predefined behavioral prototypes. How well does it scale to games where an accurate equilibrium policy is unavailable or expensive to compute?

(c) How many independent runs/seeds are used for each reported result? Please report confidence intervals or standard errors for payoff, win rate, and token/API comparisons.

(d) Are baselines given comparable inference budgets, prompt budgets, and LLM API access? How were baseline prompts tuned? The paper describes the LLM backend as “DeepSeek v4 flash.” Can the authors clarify the exact model/version, API settings, temperature, and date of experiments?

(e) The belief module uses predefined behavioral prototypes and game-specific features. How were these prototypes chosen, and how sensitive is performance to their number and definition?

(f) The ablation shows counterfactual recalibration has the smallest effect. Can the authors provide a more detailed analysis of when it helps and when it fails?

(g) Several figures/tables in the provided PDF appear malformed or missing labels. Will the final version fix these presentation issues? What are the main failure modes of SAGE, and in which settings would the authors advise against using it?

3 comments

We thank the reviewer for the detailed assessment.

(a) Contribution. The novelty of our work is to identify three sources of instability in free-form LLM strategic reasoning—unsupported assumptions, inconsistent opponent estimates, and interference from irrelevant historical interactions—and to propose a simple framework that mitigates them by organizing strategic reasoning around anchoring, adaptation, and recalibration. We agree that equilibrium strategies, opponent modeling and historical reasoning are established. The revised Related Work now explicitly discusses Ganzfried and Sandholm (2011), Kang (2026), Hypothetical Minds and other relevant LLM-agent methods.

(b) Strategic references and scale. In Leduc, a weaker MCCFR reference and code-based search yield payoffs of 0.825 and 0.842, compared with 1.139 using the original reference and 0.778 for Vanilla. Thus reference quality matters. As discussed in our Limitations section, extending SAGE to larger games requires more scalable strategic solvers. This requirement does not change our contribution of mitigating instability in LLM strategic reasoning, and we believe developing solvers for substantially larger games is a separate research direction.

(c) Statistical evidence. Following prior LLM game-agent work (such as Suspicion-Agent, Hypothetical Minds, and EMO), our main evaluation uses 30 consecutive games per player position. The appendix also reports dynamic payoff trajectories from five independent runs with different seeds for selected comparisons, with shaded regions showing cross-run variability. The 127.6% gain is relative to Vanilla.

(d) Fairness, compute and reproducibility. We consider a strict identical token cap impractical across methods: forcibly truncating an agent's reasoning (e.g., during a <think> trace) could prevent it from completing its intended procedure and make the comparison unfair. We therefore compare payoff at similar scales of token consumption, and token consumption at comparable payoff levels. The efficiency figures report online token and API usage; the MCCFR references for these small games take a few seconds to construct and are reused. We added Bayes' Bluff and RNR and the unstructured control. Appendix E identifies the backend as DeepSeek v4 flash (API identifier deepseek-v4-flash), the experiment period as May–September 2026, and the default context length as 1M tokens. The main experiments use thinking mode disabled; Appendix E also reports an enabled-versus-disabled comparison. We verified that the manuscript provides an anonymous code link that can be opened.

(e) Opponent prototypes. The predefined behavioral prototypes serve as anchors for interpreting opponent behavior, not as a list of previously encountered opponents. All opponents in the direct-match evaluation are previously unseen agents, including Bayes' Bluff and RNR. SAGE uses the same prototype anchors and updates its beliefs from interactions with each agent. The direct-match results therefore show how SAGE adapts to unseen opponents beyond the common-style evaluation.

(f) Historical recalibration. Removing recalibration changes mean payoff by +0.0333, +0.0778 and +0.2722 when calculated as full SAGE minus control. The added history-lessons alternative gives differences of +0.1283, +0.0846 and +0.1071. These results demonstrate the contribution of historical recalibration and its advantage over the tested history-lessons alternative.

(g) Presentation and failure cases. We revised the preliminary-study explanation and added limitations, including SAGE's Leduc loss against LLM(Eq). We could not reproduce the reported missing initial letters or a page of numeric content in the rendered manuscript or that the numbers issue.

Reviewer_kuBfOpenReview ↗

Comments for Author Rebuttal

Thank you for the detailed rebuttal and for adding new baselines, seeds, and ablation data. I appreciate the clarifications regarding the model version, the comparison with Bayes' Bluff/RNR, and the additional evaluation of the counterfactual recalibration module. However, my primary concerns remain unresolved. The paper's heavy reliance on exact MCCFR priors significantly limits its generalizability; as your own ablations show, performance degrades sharply with weaker references. Additionally, the core contribution remains an orchestration of existing prompting techniques without a principled theoretical argument for why this specific factorization should generalize. I therefore maintain my score of 2 (Weak Rejection).

Thank you for your follow-up. We respectfully disagree with the characterization of SAGE as requiring an exact MCCFR prior or as merely orchestrating existing prompting techniques.

SAGE improves LLM reasoning with different-quality strategic references. The weaker-reference experiment shows that the quality of the reference affects payoff, but it does not imply that SAGE requires an exact equilibrium. In Leduc, SAGE with weakly trained MCCFR achieves 0.825, and SAGE with code-based Monte Carlo search achieves 0.842; both exceed the 0.778 payoff of Vanilla. These results show that SAGE can improve an LLM agent using weaker or search-based guidance, without requiring an exact MCCFR prior. Our goal is to improve the LLM's strategic reasoning, not to introduce a new game solver or claim state-of-the-art performance for a non-LLM game-playing algorithm.

Our contribution starts from identifying and decomposing reasoning instability. We identify three distinct problems in free-form LLM strategic reasoning: unsupported assumptions, inconsistent opponent estimates, and interference from irrelevant historical interactions. SAGE addresses them through a targeted reconstruction of the reasoning process: anchoring to strategic guidance, adapting using opponent beliefs, and recalibrating with relevant historical hypotheses. The contribution is this problem-driven decomposition and structured design for mitigating instability, supported by the unstructured ablation across all three games (Table 3) and the reasoning-trace analysis in Figure 7. The individual ingredients have precedents, which we discuss in Related Work; the framework's contribution lies in how its structure addresses the identified LLM reasoning problems.

We hope these results and the problem-driven rationale help clarify SAGE's contribution and the role of the strategic reference.

Reviewer_H6juOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SAGE, a framework for improving LLM agents' strategic decision-making in repeated imperfect-information games. SAGE consists of three components: 1) anchoring decisions to a precomputed approximate Nash-equilibrium policy, 2) adapting decisions using opponent beliefs inferred from past and recent interactions, and 3) recalibrating decisions using counterfactual hypotheses generated from historical experiences.

Experiments on Leduc Hold'em, Liar's Dice, and Goofspiel show that SAGE generally achieves competitive or superior payoffs while using substantially fewer tokens than reasoning-intensive baselines. Additional experiments examine component ablations, weaker equilibrium references, opponent-style shifts, and different LLM sizes.

Critical strengths

Organizing strategic reasoning around equilibrium guidance, opponent adaptation, and historical experience is a useful approach to improving LLM decision-making without additional training.The paper compares SAGE against several relevant LLM-agent methods and includes ablations, alternative strategic references, and opponent-style shifts.SAGE achieves competitive performance with substantially lower token usage than several reasoning-intensive methods, suggesting practical benefits.

Critical weaknesses
  1. (Major weakness) The methodological novelty is limited, and the experiments do not establish the incremental value of SAGE's reasoning structure.

The main ingredients of SAGE have substantial precedents. Ganzfried and Sandholm (2011) already combine equilibrium strategies with opponent modeling for exploitation, while Hypothetical Minds uses historical interactions to generate and refine hypotheses about opponents. Kang (2026, arXiv:2603.18563) studies inference-time Bayesian opponent learning and best-response reasoning, establishing conditional asymptotic guarantees for repeated games. Although its theoretical setting differs, this work demonstrates that opponent inference and strategic adaptation without additional LLM training are not new ideas. Therefore, SAGE's contribution primarily lies in organizing existing ingredients. However, the current ablations remove information sources along with their reasoning components, making it unclear whether the proposed structure itself provides additional value. The authors should compare SAGE against 1) an LLM receiving the same strategic information without the proposed structure, and 2) a simpler Bayesian opponent-modeling and best-response baseline.

  1. (Major weakness) The equilibrium anchor does not guarantee strategically reliable adaptation.

SAGE motivates equilibrium anchoring as a way to prevent unreliable strategic reasoning. However, its subsequent opponent adaptation and historical recalibration can move decisions away from equilibrium, without any guarantee that these deviations are beneficial. This concern is supported by Table 2: SAGE achieves only an 8.3% win rate and −2.967 average payoff against LLM(Eq) in Leduc Hold'em. Thus, the additional reasoning components can substantially worsen performance relative to a simpler equilibrium-guided agent. Kang (2026) provides a useful theoretical contrast by identifying conditions under which opponent learning and approximate best responses lead to asymptotically stable strategic outcomes. SAGE does not establish analogous properties, although such a theorem is not necessarily required for its different setting. The authors should investigate when SAGE's deviations from equilibrium improve performance versus increase vulnerability, ideally through exploitability or best-response-gap measurements.

  1. (Major weakness) The efficiency comparisons do not fully account for offline computation or matched computational budgets.

SAGE relies on precomputed MCCFR equilibrium policies and opponent-style prototypes. However, the efficiency comparisons primarily report online LLM tokens and API calls, excluding the costs of constructing these resources. Moreover, Table 5 shows that Leduc payoff decreases from 1.139 to approximately 0.83 when weaker strategic references are used, suggesting substantial dependence on precomputed policy quality. Finally, Figure 6 compares methods at their tested computational budgets rather than demonstrating superiority when budgets are matched. The authors should account for offline computation and evaluate stronger baselines under comparable computational resources.

  1. (Major weakness) The statistical evidence is insufficient to establish superiority over the closest competitors.

Tables 1–3 do not provide uncertainty estimates for the reported differences. Some improvements are small: in Liar's Dice, SAGE achieves 0.494 payoff compared with 0.472 for Agent-Pro and Hypothetical Minds, a relative improvement of only 4.7%. The prominently reported 127.6% improvement is instead relative to Vanilla. While Appendix B provides variability across runs for selected comparisons, uncertainty estimates for the full baseline comparisons and ablations are missing. The authors should report confidence intervals across independent runs, particularly for small performance differences and the substantial negative direct-match results.

  1. (Major weakness) The statistical evidence is insufficient to establish superiority over the closest competitors.

Tables 1–3 do not provide uncertainty estimates for the reported differences. Some improvements are small: in Liar's Dice, SAGE achieves 0.494 payoff compared with 0.472 for Agent-Pro and Hypothetical Minds, a relative improvement of only 4.7%. The prominently reported 127.6% improvement is instead relative to Vanilla. While Appendix B provides variability across runs for selected comparisons, uncertainty estimates for the full baseline comparisons and ablations are missing. The authors should report confidence intervals across independent runs, particularly for small performance differences and the substantial negative direct-match results.

  1. (Major weakness) The proposed reasoning improvements are not directly validated, and some generated hypotheses appear inconsistent with the game rules.

Figure 1 motivates SAGE through unsupported planning, inconsistent opponent estimates, and irrelevant historical reasoning. However, the experiments do not directly measure whether SAGE reduces these failures. More concerningly, the Leduc counterfactual examples in Appendix F.1 (p. 26) discuss a "turn," subsequent board changes, and "later streets," although the described Leduc game has only one public card and two betting rounds. This raises questions about the reliability of generated hypotheses and the claimed consistency filtering. The authors should clarify whether these hypotheses were used during evaluation and systematically measure the frequency of invalid reasoning and its effect on decisions.

Questions for authors

What is the incremental contribution of SAGE's reasoning structure? Can the authors compare against an LLM receiving identical information without the proposed organization, and a Bayesian opponent-modeling and best-response baseline motivated by Kang (2026)? How does SAGE prevent harmful deviations from equilibrium? In particular, what explains its substantial loss against LLM(Eq) in Leduc Hold'em? Can the authors evaluate exploitability or approximate best-response quality? Does SAGE remain more efficient after accounting for offline computation and matching inference budgets? How sensitive is its advantage to the quality of the precomputed equilibrium policy? Are the reported performance improvements statistically reliable? Please provide uncertainty estimates for the main comparisons, especially the small payoff differences and negative head-to-head results. How are invalid counterfactual hypotheses filtered? Were the rule-inconsistent Leduc hypotheses in Appendix F.1 used during evaluation, and can the authors directly evaluate reductions in the reasoning failures motivating SAGE? How are opponent prototypes and belief updates constructed? How sensitive are the results to misspecified opponent styles and the weighting between recent and historical observations?

1 comment

Thank you for the detailed feedback. We have added the requested comparisons, a reasoning audit, and discussion of related work.

1. Novelty and structure. The novelty lies in identifying instability in free-form LLM strategic reasoning—unsupported assumptions, inconsistent opponent estimates, and interference from irrelevant history, and propose SAGE to mitigate it. SAGE anchors, adapts, and recalibrates the reasoning process. The new unstructured ablation confirms the value of this organization across all three games (Table 3). We expanded Related Work to discuss Ganzfried and Sandholm (2011), Kang (2026), and Hypothetical Minds, and distinguish their contributions from our focus on LLM reasoning instability.

2. Non-LLM adaptive baselines. We added Bayes' Bluff and RNR to Tables 1–2. Against common Leduc opponents, their mean payoffs are 1.05 and 1.20 versus 1.139 for SAGE; in direct matches, SAGE obtains +0.25 and +0.10 against them. These provide algorithmic reference points. We want to highlight that our main goal is to mitigate instability in LLM reasoning by structuring the use of strategic and historical information.

3. Equilibrium deviations and LLM(Eq). Against non-equilibrium opponents, adapting to observed tendencies can improve on following the equilibrium policy alone. SAGE therefore uses equilibrium as an anchor. To investigate its Leduc loss against LLM(Eq) (8.3% win rate; −2.967 payoff), we examined the logs: SAGE repeatedly classified LLM(Eq) as conservative when it held strong cards, contributing to several large losses. Nevertheless, SAGE achieves a better token–payoff trade-off than most evaluated LLM agents.

4. Efficiency and reference quality. A strict identical token cap could interrupt a method before it completes its reasoning, making comparisons unfair. We instead compare payoff at similar token scales and token use at comparable payoff levels. Our token/API measurements concern online inference; the MCCFR reference for each small game takes a few seconds to construct and is reused. In Leduc, weaker MCCFR and code-based search yield payoffs of 0.825 and 0.842 versus 1.139 with the original reference, while both remain above Vanilla (0.778).

5. Statistical evidence. Following prior LLM game-agent work (Suspicion-Agent, Hypothetical Minds, and EMO), our main evaluation uses 30 consecutive games per player position. The appendix also reports dynamic payoff trajectories from five independent runs with different seeds for selected comparisons, with shaded regions showing cross-run variability. The 127.6% gain is relative to Vanilla.

6. Unsupported reasoning and invalid hypotheses. We re-analyzed reasoning traces of all compared LLM agents using Qwen3-32B, identifying rule violations, contradictions, and fabrications in 100 sampled decisions per method and game. Figure 7 now reports the results: across these evaluations, SAGE has lower combined flagged-category counts than most compared methods. The Leduc examples identified by the reviewer illustrate the problem SAGE aims to mitigate; our goal is to reduce reasoning instability, rather than eliminate it entirely.

7. Opponent beliefs and unseen agents. Behavioral prototypes anchor the interpretation of observed actions; they are not a list of previously encountered opponents. All direct-match opponents, including Bayes' Bluff and RNR, are unseen agents. SAGE adapts using its existing anchors and interaction history. The appendix details the features and style-shift experiments.

Comments on the paper

1 comment

We thank all reviewers for their constructive feedback. We have revised the manuscript accordingly and highlighted the added or modified text in blue in the revised manuscript.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.