acceptodds
Under review as a conference paper at ICLR 2027

Generate Multiple Actions, Execute One: Experience-Guided Action Selection for LLM Agent Reinforcement Learning

Abstract

Group-in-Group Policy Optimization (GiGPO) improves credit assignment in reinforcement learning for LLM agents by comparing the returns of actions already executed at the same state. Comparing more candidate actions could improve action selection and provide additional learning signals. However, directly obtaining terminal feedback for each candidate requires completing a separate downstream trajectory, which involves many additional environment interactions. To avoid this cost, we introduce Experience-Guided Action Selection (EGAS), a method that uses an experience pool of historical state–action outcomes to guide both action selection and policy learning. At the current state, EGAS uses the experience pool to estimate whether an alternative action could be more likely to lead to task success than the initially generated action. When this is the case, EGAS generates additional candidates from the current policy and selects the action with the highest estimated probability of success among the initial action and candidates with sufficient historical support. Only the selected action is executed. EGAS then uses the experience pool to construct pairwise preferences between the selected action and unexecuted candidates. An auxiliary objective incorporates these preferences into the base RL objective, enabling the policy to learn from unexecuted candidates without requiring additional downstream trajectories. EGAS improves held-out task success over GiGPO by 1.7–7.9 percentage points across ALFWorld, WebShop, and Sokoban. It also reaches GiGPO's best-observed performance while using 7.3–25.0% fewer completed training trajectories. These results demonstrate that comparing and learning from more actions need not require executing more trajectories.

Primary area:
reinforcement learning
Keywords:
LLM agents, reinforcement learning, action selection, experience reuse, preference learning, sample efficiency

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_s8XYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Experience-Guided Action Selection (EGAS), an extension of GiGPO for training LLM agents. EGAS maintains an experience pool of executed (task, state, action) tuples with time-decayed success and failure counts, modeled as Beta posteriors over task success. During a subset of rollouts, the policy samples an initial action. If the pool suggests a historically supported alternative may do better, the policy samples K−1 additional candidates and executes the one with the highest posterior mean. Only one action is executed per state, so no extra environment branches are needed.

The pool also shapes learning in two ways. GiGPO advantages are reweighted per state group by the estimated potential improvement. An auxiliary PPO-style objective trains unexecuted candidates using pairwise preferences, estimated from the pool, between each candidate and the executed action.

Across ALFWorld, WebShop, and Sokoban with Qwen3 1.7B/4B models, EGAS reports gains of 1.7–7.9pp in held-out success over GiGPO. It also reaches GiGPO’s best-observed performance with 7.3–25% fewer training trajectories.

Critical strengths

The core idea is simple and well motivated.

The ablation suite is fairly thorough. It covers component ablations, intervention allocation, candidate-set size, and auxiliary weight. The retrieved-action ablation in Appendix F is a nice design-justifying experiment.

The paper is clearly written and the method is easy to follow.

Critical weaknesses

No confidence intervals or significance testing. None of the main results, efficiency metrics, or ablations report confidence intervals. I understand these experiments are expensive, but the gains are small enough that some evidence of statistical significance is necessary. Each checkpoint is evaluated on 128 episodes (256 for WebShop). At a success rate of about 55%, the binomial standard error of a single run is about 4.4pp. A difference between two independent runs therefore has a 95% interval of roughly ±12pp from evaluation sampling alone. Even the largest headline gain (+7.9pp on ALFWorld-1.7B) is not distinguishable from zero on that basis, and seed variance adds to it: Appendix G reports EGAS seed SDs of 2.9pp on ALFWorld-1.7B and 4.7pp on Sokoban. Multi-seed results are given only for EGAS, not for GiGPO or other baselines, and the main table and ablations appear to be single runs. Multi-seed means with confidence intervals for both EGAS and GiGPO would substantially strengthen the paper. This is the main thing that would change my assessment.

Reliance on exact matching limits the scope of the method. As I understand Appendix C.4, a candidate is eligible for selection only if its normalized (task, observation string, action string) key exactly matches a pool entry with sufficient support. There is no semantic matching. This is plausible for ALFWorld and Sokoban, whose action spaces are small, enumerable sets of commands or moves. It seems much less likely for free-text actions such as WebShop’s search[...] queries, where selection may effectively operate only over click[...] actions. Keys also include the task identifier, so the pool helps only when the same task is resampled in a later iteration. The paper should also state explicitly that the method targets deterministic environments with discrete, string-matchable observations and a finite set of repeated training tasks.

The “learn without executing” framing is somewhat overstated. Every pool entry with a success estimate exists because a trajectory containing that exact state-action pair was executed in an earlier iteration. The method avoids executing a candidate again now. It does not evaluate truly unexecuted actions: unsupported candidates receive only a prior-based signal. Framing the pool explicitly as a tabular Monte Carlo critic estimated from past rollouts would make both the contribution and its limits clearer.

Overall, I think this is a good paper with a sensible, practical idea, and I lean toward weak reject in its current form. My main concern is the lack of statistical evidence for gains of this size. I would be happy to raise my score if the authors report multi-seed results with confidence intervals for both EGAS and the GiGPO baseline.

Questions for authors

Can you report results over multiple seeds, with confidence intervals, for both GiGPO and EGAS? This applies to the main results (Table 1, Figures 3–4) and ideally the key ablations (Figure 7).

3 comments
AuthorsOpenReview ↗

Response to Reviewer s8XY

Q: Can you report results over multiple seeds, with confidence intervals, for both GiGPO and EGAS? This applies to the main results (Table 1, Figures 3–4) and ideally the key ablations (Figure 7).

As suggested, we added main results for both GiGPO and EGAS using seeds 0, 1, and 2, together with three-seed component ablations on ALFWorld with Qwen3-1.7B.

We report individual-seed results, Mean ± SD, and approximate 95% confidence intervals computed using Student’s tt method, with training seeds as the statistical units. All metrics use a 0–100 scale.

  1. Main results

    Setting / MetricMethodSeed 0Seed 1Seed 2Mean ± SD95% CI
    ALFWorld 1.7B / SuccessGiGPO50.0060.1655.4755.21 ± 5.08[42.58, 67.84]
    EGAS60.1664.8059.3861.45 ± 2.93[54.17, 68.73]
    ALFWorld 4B / SuccessGiGPO64.8462.5063.2863.54 ± 1.19[60.58, 66.50]
    EGAS71.1070.3073.4071.60 ± 1.61[67.60, 75.60]
    WebShop 1.7B / ScoreGiGPO80.6676.3977.6878.24 ± 2.19[72.80, 83.68]
    EGAS80.2080.9077.1079.40 ± 2.02[74.38, 84.42]
    WebShop 1.7B / SuccessGiGPO62.8949.6151.5654.69 ± 7.17[36.87, 72.50]
    EGAS64.5065.2061.8063.83 ± 1.80[59.37, 68.29]
    WebShop 4B / ScoreGiGPO86.1986.9187.7286.94 ± 0.77[85.04, 88.84]
    EGAS89.6088.4088.2088.73 ± 0.76[86.85, 90.61]
    WebShop 4B / SuccessGiGPO78.9173.8380.0877.61 ± 3.32[69.35, 85.86]
    EGAS80.5079.3079.7079.83 ± 0.61[78.32, 81.35]
    Sokoban VL-4B / SuccessGiGPO85.9476.5681.2581.25 ± 4.69[69.60, 92.90]
    EGAS80.5085.9089.8085.40 ± 4.67[73.80, 97.00]

    EGAS achieves higher cross-seed means on all reported metrics, with the SDs and confidence intervals showing the associated variability and uncertainty.

  2. Key component ablations

    ConfigurationSeed 0Seed 1Seed 2Mean ± SD95% CI
    EGAS60.1664.8059.3861.45 ± 2.93[54.17, 68.73]
    Random states54.6955.4758.5956.25 ± 2.06[51.12, 61.38]
    Random selection43.7545.3142.9744.01 ± 1.19[41.05, 46.97]
    No time decay53.1350.7853.9152.61 ± 1.63[48.56, 56.65]
    No potential weighting52.3454.6950.7852.60 ± 1.97[47.71, 57.49]
    No auxiliary preference49.2253.9152.3451.82 ± 2.39[45.89, 57.75]

    Full EGAS outperforms all five ablation variants in each seed, providing further support for the consistent direction of the component contributions.

Clarity issues

We thank the reviewer for pointing out these clarity issues. We will revise the relevant sections to make the methodological descriptions more precise and clearly state the scope of our claims and the method’s limitations.

Thank you for running the additional experiments. I appreciate the effort.

The paper’s contribution is primarily empirical. For a contribution of this kind, we need evidence that the gains are statistically distinguishable (or at least plausibly so). Its really hard for me to justify raising my score for an empirical paper at ICLR without that evidence.

From your new results, EGAS only outperforms GiGPO in one setting (ALFWorld-4B). For the remaining settings, the intervals on the difference are wide and include zero. The consistent positive direction across settings is encouraging, but I find it hard to recommend acceptance without stronger evidence. I’d encourage running additional seeds.

For now I’m keeping my score, but I’d be glad to revisit it if additional seeds establish the gains clearly.

Thank you for pointing out the limitations of our statistical evidence. We agree that, although EGAS achieved higher cross-seed mean performance in all reported settings, the confidence intervals for the performance differences were still wide with only three training seeds, providing insufficient evidence to establish statistically significant improvements across settings.

To address this concern, we added three training seeds for both GiGPO and EGAS, increasing the number of runs per method in each setting from three to six. We now report the complete per-seed results, means, standard deviations, and the mean performance differences between EGAS and GiGPO with their 95% confidence intervals. The updated analysis shows that the confidence intervals for the differences are entirely above zero for all seven reported metrics. All seven comparisons also remain statistically significant at the 0.05 level after applying two-sided Welch t-tests and Holm correction for multiple comparisons. These additional results provide stronger statistical support for the mean performance improvements of EGAS in the reported settings. The detailed results are provided below.

Table 1: Complete results across six training seeds

Setting / MetricMethodSeed 0Seed 1Seed 2Seed 3Seed 4Seed 5
ALFWorld 1.7B / SuccessGiGPO50.0060.1655.4749.2149.2152.34
ALFWorld 1.7B / SuccessEGAS60.1664.8459.3861.7160.1664.84
ALFWorld 4B / SuccessGiGPO64.8462.5063.2860.1661.7160.93
ALFWorld 4B / SuccessEGAS71.1070.3073.4070.3071.1071.88
WebShop 1.7B / ScoreGiGPO80.6676.3977.6873.4074.0573.82
WebShop 1.7B / ScoreEGAS80.2080.9077.1081.2081.6082.01
WebShop 1.7B / SuccessGiGPO62.8949.6151.5654.2952.7354.68
WebShop 1.7B / SuccessEGAS64.5065.2061.8066.4065.2363.28
WebShop 4B / ScoreGiGPO86.1986.9187.7288.7684.3384.91
WebShop 4B / ScoreEGAS89.6088.4088.2090.3492.3192.10
WebShop 4B / SuccessGiGPO78.9173.8380.0875.0071.4871.48
WebShop 4B / SuccessEGAS80.5079.3079.7082.4282.8183.98
Sokoban VL-4B / SuccessGiGPO85.9476.5681.2578.6878.2578.03
Sokoban VL-4B / SuccessEGAS80.5085.9089.8089.0688.7189.73

Table 2: Cross-seed means, standard deviations, and 95% confidence intervals for each method

Each method has six training runs in each setting. Mean ± SD denotes the cross-seed mean and sample standard deviation. The two-sided 95% confidence interval for each method's mean is computed using Student's t method with five degrees of freedom.

Setting / MetricGiGPO Mean ± SDGiGPO 95% CIEGAS Mean ± SDEGAS 95% CI
ALFWorld 1.7B / Success52.73 ± 4.36[48.15, 57.31]61.85 ± 2.44[59.29, 64.41]
ALFWorld 4B / Success62.24 ± 1.69[60.47, 64.01]71.35 ± 1.17[70.12, 72.57]
WebShop 1.7B / Score76.00 ± 2.83[73.03, 78.97]80.50 ± 1.78[78.64, 82.37]
WebShop 1.7B / Success54.29 ± 4.60[49.46, 59.12]64.40 ± 1.63[62.69, 66.12]
WebShop 4B / Score86.47 ± 1.68[84.71, 88.23]90.16 ± 1.77[88.30, 92.02]
WebShop 4B / Success75.13 ± 3.66[71.28, 78.98]81.45 ± 1.89[79.47, 83.43]
Sokoban VL-4B / Success79.79 ± 3.38[76.24, 83.33]87.28 ± 3.62[83.49, 91.08]

Table 3: Mean differences (EGAS − GiGPO), 95% confidence intervals, and significance tests

We use training runs as the statistical units, with n = 6 per method in each setting, and define the mean difference as Δ = Mean(EGAS) − Mean(GiGPO). We apply independent-samples, two-sided Welch t-tests and compute the approximate 95% confidence intervals for the differences using Welch–Satterthwaite degrees of freedom. We also apply Holm correction to the seven comparisons reported in this table.

Setting / MetricMean Difference Δ95% CI for ΔTwo-sided p-valueHolm-adjusted p-value
ALFWorld 1.7B / Success+9.12[4.40, 13.84]0.002190.01223
ALFWorld 4B / Success+9.11[7.21, 11.01]0.000001950.00001364
WebShop 1.7B / Score+4.50[1.38, 7.62]0.010110.01629
WebShop 1.7B / Success+10.11[5.28, 14.94]0.002040.01223
WebShop 4B / Score+3.69[1.47, 5.91]0.004110.01629
WebShop 4B / Success+6.32[2.39, 10.25]0.006300.01629
Sokoban VL-4B / Success+7.50[2.99, 12.00]0.004070.01629

Thank you again for encouraging us to conduct this additional validation. We will include the complete results and statistical methodology in the revised manuscript. We hope this additional evidence addresses your main concern regarding the statistical reliability of the main results.

Reviewer_R9BTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper explores how to use a replay-buffer during RL training for reasoning to explore more effectively without needing explicit tree-style branching. The paper defines a posterior for each task, state, action and at each state, use a beta posterior that is continually updated to select the most promising action. The model is updated using standard RL (in this case GiGPO) and an aux loss that is advantage-weighted policy gradient, where the advantage comes from the posterior's likelihood of the unexplored actions from the buffer at that state. Overall this is an intriguing idea, but the experiments are rather simplistic (e.g. ALFWorld) and the replay buffer requires explicit state-key match. Moreover, there is no insight into the correctness of the beta posterior and whether the aux loss makes sense (i.e. can we actually trust this posterior since n is small and the beta definition assumes independence across actions despite them coming from same trajectory). Lastly, there is no FLOPs-equivalent baseline which is concerning as the learning process is introducing more FLOPs via the replay-buffer and especially the aux term. For these reasons, I lean toward weak reject.

Critical strengths
  • Method improved performance around ~10% on the provided benchmarks and requires less steps compared to vanilla Gigpo/grpo
  • Ablations clearly isolate the main contributors to the boost.
Critical weaknesses
  • Lack of insight into the aux loss. Can we trust the posterior for this? I see the ablation shows that aux loss is important, but is it sound? The beta posterior is also quite noisy since n is small (and selection criteria is just n>=3).
  • Lack of FLOPS-equivalent baselines. There are additional FLOPs introduced, e.g. the aux loss and there doesn't seem to be a equal-FLOPs comparision with vanilla Gigpo/grpo.
Questions for authors
  • Can you provide analysis into the soundness of the posterior and the advantage computed from it for the aux loss? Does this explain with the method doesn't consistently improve as num_trajectories scale?
  • Can you provide FLOPs-equivalent vanilla baseline?
1 comment
AuthorsOpenReview ↗

Response to Reviewer R9BT

Q1: Can you provide analysis into the soundness of the posterior and the advantage computed from it for the aux loss? Does this explain with the method doesn't consistently improve as num_trajectories scale?

Thank you for this question, which helps us clarify the soundness of the posterior and auxiliary loss.

  1. The posterior is grounded in state overlap across trajectories. Different trajectories for the same task often revisit shared states, where past successes and failures provide evidence for current action choices. The posterior aggregates this evidence, and the derived advantage guides the auxiliary loss toward better-supported actions. Its reliability therefore depends on the coverage and quality of relevant feedback, not trajectory count alone. This offers a plausible explanation for non-monotonic gains: additional trajectories may provide redundant evidence rather than improve estimates at relevant states.

  2. The auxiliary loss remains effective across different trajectory compositions. On ALFWorld with Qwen3-4B-Instruct-2507, we hold the total number of trajectories fixed at eight and increase the experience-guided trajectories from four to eight. As shown in the table below, the gain from the auxiliary loss increases from 4.69 to 6.25 percentage points. This supports its effectiveness across these configurations, rather than establishing independence from total trajectory count.

    Standard + guided trajectoriesWithout aux lossWith aux lossGain
    4 + 466.41%71.09%+4.69 pp
    0 + 858.59%64.84%+6.25 pp

Q2: Can you provide FLOPs-equivalent vanilla baseline?

Good suggestion. We added a budget-matched comparison with vanilla GiGPO on ALFWorld with Qwen3-1.7B. The table compares checkpoints with closely matched cumulative FLOPs, measured in 101810^{18}:

GiGPO / EGAS stepsGiGPO FLOPsEGAS FLOPsBudget differenceGiGPO success rateEGAS success rateSuccess-rate gain
130 / 1259.95459.9131−0.416%52.34%57.81%+5.47 pp
140 / 13510.636110.6576+0.202%53.75%60.34%+6.59 pp
150 / 14511.290111.3573+0.595%53.97%62.34%+8.37 pp

The budget differences are below 0.6% for all three pairs. Under these approximately equal FLOPs budgets, EGAS achieves higher success rates than vanilla GiGPO.

Reviewer_QioQOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

Group-in-Group policy optimization addresses the problem that sparse terminal rewards make it difficult to assign credit to individual actions. However, under sparse terminal feedback, directly obtaining a terminal outcome for each candidate requires completing a separate downstream trajectory. This paper introduces EGAS, a method that uses an experience pool of historical state-action outcomes to guide both action selection and policy learning, avoiding additional environment interactions that come from obtaining feedback for each candidate.

At each state, the policy first generates an initial action. The pool estimates whether an alternative could be more likely to lead to task success than this action. The pool estimates whether an alternative could be more likely to lead to task success than this action. When the estimate indicates potential improvement, EGAS generates additional candidates from the current policy. Among all the candidates, it selects the action with the highest estimated probability of task success. The experience pool also guides policy learning. The executed trajectories are used to update the policy through a weighted objective. The unexecuted candidates contribute through an auxiliary pairwise preference objective.

Critical strengths
  1. This paper studies an interesting topic: how to use the training trajectories more efficiently with an experience pool, built on top of GiGPO, which shows advantages in the credit-assignment problem in long-horizon tasks.

  2. EGRA is combined with GiGPO and GRPO and compared against the corresponding base algorithms across various environments and tasks. In most of the cases EGAs reaches the GiGPO's best performance with fewer training steps.

Critical weaknesses
  1. When kappat>0kappa_t>0, the authors claim that the experience pool only chooses candidates with "sufficient historical support". A clearer quantitative validation is needed here.

  2. In line 229, the RL advantages and updates of the executed experience-pool entries are all based on one trajectory's outcome. Is the outcome going to be noisy? Feel free to correct me if I misunderstood.

Questions for authors
  1. Before equation (1), the authors mentioned that, due to the policy evolution, the earlier collected outcomes are not very representative of the current outcome distribution, and they used exponential time decay on the stored evidence. Can you elaborate on the details of that within a time window?

  2. In (2), when choosing alternative actions ata_t in Atsup⁡\mathcal{A}_t^{\operatorname{sup}}, how can we know that the time-delayed success is caused by actions at a certain step tt, but not a result of consecutive actions?

1 comment
AuthorsOpenReview ↗

Response to Reviewer QioQ

Q1: Before equation (1), the authors mentioned that, due to the policy evolution, the earlier collected outcomes are not very representative of the current outcome distribution, and they used exponential time decay on the stored evidence. Can you elaborate on the details of that within a time window?

Thanks for this valubale question. Following the general principle of gradually reducing the influence of older experience in Discounted Thompson Sampling [1], DEPPO [2], and FreshPER [3], we use exponential decay to form a soft time window. We set the half-life to H=50H=50 training steps: historical evidence weights decrease exponentially during training, halving every 50 steps (Appendix C.4).

Q2: In (2), when choosing alternative actions in Atsup\mathcal{A}_t^{\mathrm{sup}}, how can we know that the time-delayed success is caused by actions at a certain step tt, but not a result of consecutive actions?

We agree that time-delayed success may be caused by an action at a certain step or by the combined effects of consecutive actions, and terminal feedback alone makes these contributions difficult to distinguish.We therefore adopt a practice used in GiGPO [4] and DEPPO [2], treating a single action at the current state as the unit of evaluation: we select an action at step tt, let the policy execute subsequent actions, and use the final outcome to evaluate that choice. Although the modeling unit is a single action, its evaluation includes the combined effects of subsequent actions.

W1: When κt>0\kappa_t>0, the authors claim that the experience pool only chooses candidates with "sufficient historical support". A clearer quantitative validation is needed here.

Good point. As suggested, we added a quantitative experiment on ALFWorld with Qwen3-1.7B. Here, mm denotes the minimum total historical support required for an action, with a default of m=3m=3. Additional candidates are eligible for selection only if they meet this threshold.

Metricm=1m=1m=3m=3 (default)m=5m=5
Success rate (%)53.960.255.46
Candidate-generation events8,4015,1873,011
Actual action replacements2,7551,891852

The default m=3m=3 achieves the highest success rate, suggesting that filtering historical evidence should also preserve sufficient opportunities for experience guidance.

W2: In line 229, the RL advantages and updates of the executed experience-pool entries are all based on one trajectory's outcome. Is the outcome going to be noisy? Feel free to correct me if I misunderstood.

We apologize for the confusion. A single trajectory’s outcome is one sample: RL advantages compare returns across multiple trajectories in the same group (Section 3.3), while experience-pool updates incorporate each new outcome into accumulated historical feedback to refine the action’s estimate (Appendix C.4). We will revise line 229 and the related explanations to clarify these two processes.

References:

[1] Raj and Kalyani. Taming Non-stationary Bandits: A Bayesian Approach. 2017.

[2] Dong et al. DEPPO: Dual Experience Pool Policy Optimization for Long-Horizon Reinforcement Learning. EMNLP, 2026.

[3] Ma et al. Freshness-Aware Prioritized Experience Replay for LLM/VLM Reinforcement Learning. arXiv:2604.16918, 2026.

[4] Feng et al. Group-in-Group Policy Optimization for LLM Agent Training. arXiv:2505.10978, 2025.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Revised manuscript (changes in blue)

We sincerely thank all reviewers for their valuable suggestions, which have helped improve the clarity, rigor, and completeness of our paper. We have revised the manuscript accordingly, with changes marked in blue:

  • Unexecuted candidates. We clarify that candidates unexecuted at the current decision may have historical execution records.
  • Experience pool and support threshold. We explain the pool as a tabular Monte Carlo-style success critic and define the default support threshold of three, excluding prior pseudo-counts.
  • Evidence reuse and matching. We clarify task isolation, exact-matching requirements, coverage limitations, and the semantic observation-matching extension.
  • Six-seed main results. We report complete per-seed results, summary statistics, and 95% confidence intervals. All seven comparisons remain significant under two-sided Welch tests with Holm correction.
  • Component ablations. We report three-seed results and paired confidence intervals, identifying which intervals exclude zero.
AuthorsOpenReview ↗

General Response to the Area Chair and All Reviewers

Dear Area Chair and Reviewers,

We sincerely appreciate the valuable time and effort you have devoted to our submission. We greatly value your recognition: Reviewer QioQ appreciates the research value of improving training-trajectory efficiency through an experience pool; Reviewer R9BT finds the idea intriguing and recognizes that the ablations clearly identify the sources of improvement; and Reviewer s8XY acknowledges the simple, well-motivated core idea and the fairly thorough ablation suite. These assessments are deeply encouraging.

Your constructive suggestions have directly strengthened the paper. To address statistical reliability, we expanded the main experiments to six training seeds for both GiGPO and EGAS. Improvements across all seven reported metrics remain statistically significant after Holm correction. We also added approximately FLOPs-matched comparisons, multi-seed ablations, and historical-support threshold experiments, while clarifying the methodological foundations, scope, and limitations. All revisions are incorporated into the manuscript, with changes marked in blue.

These discussions have strengthened the empirical evidence and helped us articulate the contribution more precisely. We sincerely thank the Area Chair for coordinating the review and all reviewers for their careful, insightful feedback. We greatly appreciate this opportunity for exchange and will thoughtfully address any further questions or suggestions.

Best regards, The Authors of Submission 24472

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.