- 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
-
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.
-
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.
-
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
-
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? -
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. -
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. -
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? -
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
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.
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 Problem | Addressed by Figure 7? |
|---|---|
| Unsupported Planning Assumptions | Partially, through fabrications and rule violations |
| Inconsistent Opponent Estimates | Not directly measured |
| Sensitivity to Irrelevant History | Not 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.
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.