acceptodds
Under review as a conference paper at ICLR 2027

Enabling Timely Guidance before Skill Retrieval: Retaining Helpful Warm Tips in Agent Context

Abstract

Reusable skills help LLM-based agents solve complex tasks, but the agent must receive guidance before it commits to an ineffective approach. Existing skill mechanisms often expose only metadata and load full content on demand, leaving useful guidance unavailable until the agent decides to retrieve it. General memory methods can incur substantial maintenance overhead, while keeping guidance in conversation context risks repeatedly exposing the agent to irrelevant or harmful advice. We propose TipsWarm, a mechanism that complements existing skill mechanisms by maintaining a budgeted pool of skill-derived keypoints, or warm tips, for selective injection into the context of every message turn. By separating event-triggered LLM assessment from inexpensive per-turn screening, it makes transferable skill guidance readily available while controlling maintenance costs. In three coding and iterative task-execution benchmarks, TipsWarm achieves the highest task success rate while remaining time-efficient, compared to recent skill and general memory baselines.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Agent skill, agent memory, context engineering

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_QCxSOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TipsWarm complements on-demand skill retrieval with a small persistent pool of reminders extracted from existing skills. LLM judgments govern tip admission and eviction upon skill events and assess tip utility after task completion, while lightweight scope matching and utility counters select tips for injection at each turn. The method is integrated into Hermes and evaluated on SkillsBench, AppWorld, and text-based ALFWorld, using a warmup phase followed by continued memory updates during testing. TipsWarm improves final success over HermesSkill across the reported settings, with a more pronounced gain in area under the learning curve than in final success on SkillsBench. Ablations examine utility attribution and screening, scope matching, a larger tip pool, and the augmentation of an existing skill bundle.

Critical strengths
  1. The method addresses a concrete problem: useful skill guidance may arrive after an agent has committed to an ineffective approach. Its bounded reminder pool reuses existing skills with minimal changes to the underlying workflow. The experiment augmenting an existing skill bundle further supports its practicality.

  2. The evaluation spans coding and interactive environments within a common agent framework and includes useful component ablations. The results suggest that utility attribution and screening contribute to the SkillsBench learning-curve gains, although improvements vary across configurations. The paper also reports mixed time and token-cost results, providing a more balanced assessment of efficiency.

Critical weaknesses

The benefit of early guidance could be better isolated from that of adding concise context. Comparisons with matched context budgets, initial-only reminders, and earlier skill retrieval would help clarify the advantage of the proposed scheduling mechanism.

The reliability of utility-based scheduling and its efficiency trade-offs remain to be clarified. The attribution prompt disfavors helpful labels for injected tips in failed episodes, which may not reflect their actual contributions. The joint ablation also leaves attribution’s individual effect unclear. Additionally, runtime is averaged over successful tasks only, and reported agent-token costs exclude internal operations.

Questions for authors
  1. How much of the gain comes from early guidance rather than additional context? Comparisons with simpler reminder and early-retrieval baselines would help clarify this.

  2. How sensitive are the results to the attribution rules for failed tasks and non-injected tips? A brief attribution audit and more complete runtime and token-cost reporting would help clarify the reliability and efficiency trade-offs.

2 comments

Thanks very much for your valuable comments. Our responses and revisions are provided as follows. We hope that we have addressed all your concerns and welcome further discussion if necessary. We will be grateful if you would kindly raise the rating if you are satisfied.

W1/Q1. About performance of the early skill-retrieval strategy: In the revision, we added an ablation study of proactive early skill-retrieval strategies (in Section 5.3.Ablation Study->Proactive skill retrieval strategy). The results on the SkillsBench are shown in the following table.

SettingOn-Demand Skill RetrievalEarly Skill RetrievalActive Tip RetrievalSuccess@60 (%)AUC
HermesSkill✓82.3 ±\pm 0.50.456 ±\pm 0.034
HermesSkill-ER✓✓82.6 ±\pm 4.40.472 ±\pm 0.030
TipsWarm✓✓84.1 ± 2.50.533 ±\pm 0.023

As we know, HermesSkill retrieves full skills on demand during the conversation. The Early-Retrieval (HermesSkill-ER) strategy additionally prompts the agent to proactively retrieve a full task-relevant skill into the context at the beginning of the conversation. TipsWarm retrieves tips from the maintained warm-tip pool at every turn. Results show that merely retrieving a full skill early does not obviously improve accuracy when the agent can already retrieve skills on-demand (HermesSkill-ER vs. HermesSkill). In contrast, the warm-tip approach can improve accuracy further by injecting compact guidance that could help avoid ineffective exploration (TipsWarm vs. HermesSkill-ER and HermesSkill).

W2/Q2. About sensitivity of utility attribution rules: Thanks for your suggestion of a sensitivity study of the utility attribution rules. In the revision, we have added a sensitivity study of utility rules in Section 5.3 “Ablation Study->Sensitivity of utility rules”. Table 3 shows the accuracy of different combinations of two utility-attribution rules on SkillsBench. F and N can match the best default seed at 87.0% final success, while their AUCs are lower (0.491 and 0.440, against 0.533) than the default setting, which represents a better attribution for early guidance. As we observe from the trajectories, the default setting opens with a grader constraint and receives some harmful counts on failed tasks. F exempts a path-specific tip from harmful labels in failed tasks, and N leaves two unexposed evaluator-specific tips free of an irrelevant count. A combination of F and N can have both negative effects and drops the final success rate to 78.3% and AUC to 0.420. These early-guidance benefits of the default utility attribution rules can also be inferred from their lower token costs and fewer user message turns than the other variants. In the revision, for brevity, we include only the accuracy results in Table 3.

Table 3. Accuracy and cost results of utility-attribution sensitivity on SkillsBench. The default setting of TipsWarm prohibits helpful labels for injected tips in failed tasks and counts non-injected tips as irrelevant. F allows helpful labels in failed tasks when execution evidence supports a useful contribution, while retaining harmful and irrelevant labels when appropriate. N leaves the utility counters of non-injected tips unchanged. F+N combines both changes. F, N, and F+N each have one run.

SettingAllow Helpful on Failed TasksSkip Updates for Non-injected TipsSuccess@60 (%)AUCToken cost (M)Turn cost
Default84.1 ± 2.50.533 ± 0.0231.316 ± 0.07023.96 ± 1.44
F✓87.00.4911.71127.15
N✓87.00.4402.49433.10
F+N✓✓78.30.4201.92029.39

Thank you for the detailed responses and additional experiments. After considering the responses to other reviewers as well, I am willing to raise my rating to 3. However, several weaknesses remain, and my overall recommendation remains below acceptance. I leave the final decision to the AC.

Reviewer_nBYdOpenReview ↗
Rating
2
Confidence
4(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents TipsWarm, a lightweight mechanism that provides timely, short guidance to LLM agents through tips. Tips are extracted points from the full skill context and can be regarded as an intermediate knowledge representation compared to full skill retrieval. The method extracts these tips from existing skills, maintains a tip pool with a budget, and score the goodness of the tips by attributing the execution results back to the tips. The paper evaluated their method on three different benchmarks: SkillsBench, Alfworld and AppWorld. Results show that their method achieve better success rate compared to existing baselines and has good efficiency.

Critical strengths
  1. The problem studied is interesting. Exploring what information should be provided to the agent beyond full skills and when this information should be provided is timely and important.

  2. The design is lightweight and easy to integrate with existing approaches. The approach also shows good efficiency.

Critical weaknesses
  1. The paper does not establish why their TipsWarm approach is necessary compared to proactively give full skill context to the agent. The central motivation of the paper is that the agent may fail to retrieve relevant skills before making important decisions. However, this problem is more about when this information (e.g. full skill context) should be given to the agent, not what actually needs to be given. A straightforward alternative is to directly retrieve relevant skills before execution when a new task arrives. The current evaluation mainly compares TipWarm with agent-based/memory-based retrieval. However, this does not clearly demonstrate that the benefit comes from the TipsWarm framework design instead of a more proactive knowledge injection point.

  2. The method lacks scalability and may not be robust when the tasks and skills become diverse. Currently, the number of tips in a pool is set to mainly 12. However, a modern skill library may have hundreds of, or even thousands of skills. I believe it is not feasible to maintain a 12-tip pool for a skill library of this size, as the average number of tips each skill can give is very limited. Moreover, lexical overlap is a relatively weak signal. Semantically related tasks can require different routes to solve, and tasks sharing keywords may require different guidance. This limitation becomes more important as the skill library grows and the diversity of tasks increases. It is also unclear whether the historical utility of a tip transfers across different tasks. A tip may be useful for some tasks but harmful or unnecessary for others, but its utility is estimated through aggregated historical counts.

Questions for authors

See weaknesses, moreover,

  1. The improvement of the method compared to other methods seems marginal. On two benchmarks, SkillsBench and Alfworld, the improvement compared to the best baseline is around 1 percent. Also, the authors may consider presenting a comprehensive table for results beyond the current figures.

  2. What is the advantage of retaining only 12 tips globally instead of maintaining a larger tip pool and retrieving K relevant tips per turn?

4 comments

Thanks very much for your valuable comments. Our responses and revisions are provided as follows. We hope that we have addressed all your concerns and welcome further discussion if necessary. We will be grateful if you would kindly raise the rating if you are satisfied.

W1. About comparison with proactive methods: In the revision, we added an ablation study of proactive early skill-retrieval strategies (in Section 5.3.Ablation Study->Proactive skill retrieval strategy). The results on the SkillsBench are shown in the following table.

SettingOn-Demand Skill RetrievalEarly Skill RetrievalActive Tip RetrievalSuccess@60 (%)AUC
HermesSkill✓82.3 ±\pm 0.50.456 ±\pm 0.034
HermesSkill-ER✓✓82.6 ±\pm 4.40.472 ±\pm 0.030
TipsWarm✓✓84.1 ± 2.50.533 ±\pm 0.023

As we know, HermesSkill retrieves full skills on demand during the conversation. The Early-Retrieval (HermesSkill-ER) strategy additionally prompts the agent to proactively retrieve a full task-relevant skill into the context at the beginning of the conversation. TipsWarm retrieves tips from the maintained warm-tip pool at every turn. Results show that merely retrieving a full skill early does not obviously improve accuracy when the agent can already retrieve skills on-demand (HermesSkill-ER vs. HermesSkill). In contrast, the warm-tip approach can improve accuracy further by injecting compact guidance that could help avoid ineffective exploration (TipsWarm vs. HermesSkill-ER and HermesSkill).

W2. About scalability of the method: Thanks for this really practical concern of our method. I would like to address it in multiple aspects.

  1. About the scope of transferability: our intended setting is transfer across similar tasks, as stated in the introduction. It corresponds to our experimental setting, where the agent warms up with some tasks in a benchmark and then tests on other (similar) tasks in the same benchmark, but not across benchmarks.

  2. About the small pool with 12 tips: The tip pool complements the full skill library with compact, reusable guidance, while detailed skills remain available through on-demand retrieval. Moreover, the pool size, which is set to 12 for this experiment, is just a configurable hyperparameter, and a larger pool size may increase accuracy. We will discuss this in detail in our response to your question Q2.

  3. About tip selection: The assessment of relevance is mainly done by LLM judgment during admission and reconciliation of the tip pool during skill visit events. The lexical overlap is just a subsequent, inexpensive scope filter during per-turn injection. This is actually what we emphasize as a practical, novel design for both tip relevance and runtime efficiency, as discussed in the Design Principle paragraph. While task relevance has been somehow assured by the LLM judgment during skill events, in the later scope matching using lexical overlap, tips only need to compete with other tips in the pool, which is less prone to scope bias than competing with all tips across different skills. As a result, the “best of the best” (most relevant and guiding) tips are expected to be selected to inject into the context. Actually, as shown in the ablation study, the AUC of TipsWarm without scope (NS: 0.513) matching is slightly lower than the original TipsWarm (ALL: 0.533), indicating that this lexical-based scope matching has an effect on refining tip selection for early guidance.

  4. About utility transferability: Historical utility is aim for similar tasks, as we have scoped, and incorporates task-specific helpful, harmful, and irrelevant judgments based on execution evidence, rather than success counts alone.

To further study the effect of different utility attribution rules, we have added a sensitivity study of utility rules in Section 5.3 “Ablation Study->Sensitivity of utility rules” in the revision. Table 3 shows the accuracy of different combinations of two utility-attribution rules on SkillsBench. F and N can match the best default seed at 87.0% final success, while their AUCs are lower (0.491 and 0.440, against 0.533) than the default setting, which represents a better attribution for early guidance. As we observe from the trajectories, the default setting opens with a grader constraint and receives some harmful counts on failed tasks. F exempts a path-specific tip from harmful labels in failed tasks, and N leaves two unexposed evaluator-specific tips free of an irrelevant count. A combination of F and N can have both negative effects and drops the final success rate to 78.3% and AUC to 0.420. These early-guidance benefits of the default utility attribution rules can also be inferred from their lower token costs and fewer user message turns than the other variants. In the revision, for brevity, we include only the accuracy results in Table 3.

Table 3. Accuracy and cost results of utility-attribution sensitivity on SkillsBench. The default setting of TipsWarm prohibits helpful labels for injected tips in failed tasks and counts non-injected tips as irrelevant. F allows helpful labels in failed tasks when execution evidence supports a useful contribution, while retaining harmful and irrelevant labels when appropriate. N leaves the utility counters of non-injected tips unchanged. F+N combines both changes. F, N, and F+N each run once.

SettingAllow Helpful on Failed TasksSkip Updates for Non-injected TipsSuccess@60 (%)AUCToken cost (M)Turn cost
Default84.1 ± 2.50.533 ± 0.0231.316 ± 0.07023.96 ± 1.44
F✓87.00.4911.71127.15
N✓87.00.4402.49433.10
F+N✓✓78.30.4201.92029.39

Q1. 1) About performance gain interpretation: We got your concern about the modest final success-rate gains on SkillsBench and ALFWorld. We claim to compare the performance benefit of TipsWarm over other methods in a broader evaluation aspects, including various metrics and scenarios, instead of focusing on a single metric in specific scenarios.

Considering multiple metrics, when TipsWarm’s final accuracy is only 1.7 pp higher than the second best HermesSkill on SkillsBench (84.1% vs. 82.3%), its success rates at earlier turns are also consistently higher than those of HermesSkill, as indicated by the curve of Fig. 4(a), resulting in an AUC gain of 0.078 (0.533 vs. 0.456), as shown in Fig 4a. Its completion time is also only 69% of that of HermesSkills, as shown in Fig. 4b.

Considering more task scenarios, TipsWarm improves the best final final success by 4.9 pp in AppWorld (Fig. 5a).

Considering more LLM backbones, increases SkillsBench AUC by 0.102 from 0.557 to 0.659, demonstrating faster task completion even when final success saturates at 100% (Fig. 7a).

Furthermore, as shown in the ablation study of the pool size in our revision, increasing the tip-pool budget yields a final success rate of 94.2% success and an AUC of 0.546, exceeding HermesSkill by 11.9 pp and 0.090 AUC, respectively (Table 2). These results show that a simple budget adjustment can further improve TipsWarm’s performance.

These results show comprehensive benefits beyond the final success rates of the default configuration.

2) About result presentation: As for the table presentation of the main results, our initial consideration is that the curve figure over different turns may better reflect the change in accuracy as the agent conversation proceeds and better illustrate and compare the AUCs of different methods. We agree that a comprehensive table can present exact values of the results for detailed comparison. Therefore, in the revision, we have added a comprehensive table containing values of Success@k, AUC, and time to success across methods and benchmarks in Appendix C and referred to this appendix in Section 5.2.Main Results.

Q2. About design considerations and the advantage of maintaining a specific number of tips in the pool: As described in the method sections 4.1 and 4.2, ** the current method already maintains a pool of a specific size and uses a lightweight scope matching to retrieve K relevant tips per turn for per-turn efficiency. The pool size and K are just configurable hyperparameters**, which are default to 12 and 4, respectively, as specified in the experiment hyperparameter setting in Appendix B.1. When you are concerned about the advantage of a larger pool size, we have already preliminarily explored it on SkillsBench in the ablation study in Section 5.3 in the initial submission. The bigger pool (BP) setting has a pool size of 16 and K of 8.

For your interest about the performance of a larger pool, we further about evaluate TipsWarm with an even bigger pool. The default setting of TipsWarm has a pool size budget Cp=12C_p=12, per-skill tip budget Cs=4C_s=4, and injection quota Ci=4C_i=4. The bigger-pool (TipsWarm-BP) setting increases these sizes, with Cp=16C_p=16, Cs=4C_s=4, and Ci=8C_i=8, and the even-bigger-pool (TipsWarm-EBP) setting has Cp=32C_p=32, Cs=8C_s=8, and Ci=16C_i=16. The results on the SkillsBench are shown in the following table. Results show that a sufficiently big pool can further improve performance substantially, reaching a success rate of 94.2% and an AUC of 0.546 within the turn budget (TipsWarm-EBP and TipsWarm-BP vs. TipsWarm).

SettingBigger PoolEven Bigger PoolSuccess@60 (%)AUC
TipsWarm84.1 ± 2.50.533 ±\pm 0.023
TipsWarm-BP✓88.4 ±\pm 2.50.501 ±\pm 0.028
TipsWarm-EBP✓94.2 ±\pm 2.50.546 ±\pm 0.043

In the revision, we compare these results and analysis together with the proactive method in Section 5.3.”Ablation Study->Proactive skill retrieval strategy”.

Thank you for your response. While it addresses some of my concerns, the following concerns remain:

  • The gain of the method seems marginal compared to other methods. The figures in the main text, after revision, still make it hard for people to understand the gains, and after a careful read, one would find the improvement is marginal.

  • The scalability of this method when the number of skills is large (hundreds or thousands) is uncertain considering the small set of number of skills.

  • The 'when the information should be given to the agent' problem is only investigated for one baseline, which makes the conclusion not convincing enough

  • The BP & EBP experiment raises the concern that the chosen hyperparameter of the method is sub-optimal for the method. The EBP results are much more better than the original results. The author may want to figure out what the optimal configuration is for their own method so that the community can follow it.

Considering these, plus that the technical contribution(overall method presented and the insights discussed) of this paper sounds incremental, I believe this paper is below the bar of ICLR, and it is appropriate to maintain my current score.

Reviewer_x9BMOpenReview ↗
Rating
3
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The work is related to increasing accuracy and efficient of multi-turn agents via better utilizing skills mechanism. A new approach called "tipswarm" is proposed which aims at solving the problem of cold-start of skills and instead provide warmed up compressed tips to model. Due to readily availability of these tips at each user-turn, the LLM can better get timely hints instead of getting late to realize skills were needed. The paper is well written, easy to follow, and the idea is sound. I have some questions/missing evaluations which I list in below sections.

Critical strengths
  1. The idea of using some warmed-up tips (short compressed skills) to LLM at the time of user-turns which can avoid the situation of progressing forward without realizing skills could help is nice. Its also visible from results that with turns budget, the tipsWarm has lead compared to baselines
  2. The paper is well written and methodology is also explained properly.
Critical weaknesses
  1. Adding more data to stochastic systems input like LLM is not always useful. Due to additional info, now agents can take diverse paths as well. I am not seeing in this work it being quantified or studied. For example, does adding more/less tips confuse model or not, some ablation is needed.
  2. Another missing results are impact to input tokens which directly would lead to cost. Having tips as additional mechanism can save/increase tokens, which is not present in the main text results.
  3. In Sec-5.2, SkillBench, the accuracy is computed over how many tasks? is it full 88 or 23 tests samples (as mentioned in line 355-356)? Same need to be clarified for other benchmarks as well.
Questions for authors
  1. Line 190-191, the n must be a func of "t" right?
  2. Fig-3 typo. "Alwasy". I suggest do a thorough proof-read of entire paper
1 comment

Thanks very much for your valuable comments. Our responses and revisions are provided as follows. We hope that we have addressed all your concerns and that you would kindly raise your rating if you are satisfied. We also welcome further discussion if necessary.

W1. About varied pool size and tip injection: In our initial submission, we did try to explore the influence of pool size in the mechanism ablation study. We explored the performance of a bigger pool and larger injection quota (BP, with the pool size of 16 and injection quota of 8) than the default setting of TipsWarm (ALL, with the pool size of 12 and the injection quota of 4) on SkillsBench.

To fully address your concern about the influence of varied pool sizes and additional information in the context, we add an ablation study of different proactive skill retrieval strategies in the revision (in Section 5.3.Ablation Study->Proactive skill retrieval strategy). The results on the SkillsBench are shown in the following table.

SettingOn-Demand Skill RetrievalEarly Skill RetrievalActive Tip RetrievalBigger PoolEven Bigger PoolSuccess@60 (%)AUC
HermesSkill✓82.3 ±\pm 0.50.456 ±\pm 0.034
HermesSkill-ER✓✓82.6 ±\pm 4.40.472 ±\pm 0.030
TipsWarm✓✓84.1 ± 2.50.533 ±\pm 0.023
TipsWarm-BP✓✓✓88.4 ±\pm 2.50.501 ±\pm 0.028
TipsWarm-EBP✓✓✓94.2 ±\pm 2.50.546 ±\pm 0.043

As we know, HermesSkill retrieves full skills on demand during the conversation. The Early-Retrieval (HermesSkill-ER) strategy additionally prompts the agent to proactively retrieve a full task-relevant skill into the context at the beginning of the conversation. TipsWarm retrieves tips from the maintained warm-tip pool at every turn, using the default setting of a pool size budget Cp=12C_p=12, per-skill tip budget Cs=4C_s=4, and injection quota Ci=4C_i=4. As this setting already represents a small pool, we do not evaluate on a smaller pool, but focus on bigger pools. The bigger-pool (TipsWarm-BP) setting increases these sizes, with Cp=16C_p=16, Cs=4C_s=4, and Ci=8C_i=8, and the even-bigger-pool (TipsWarm-EBP) setting has Cp=32C_p=32, Cs=8C_s=8, and Ci=16C_i=16. Results show that merely retrieving a full skill early does not obviously improve accuracy when the agent can already retrieve skills on-demand (HermesSkill-ER vs. HermesSkill). In contrast, the warm-tip approach can improve accuracy further by injecting compact guidance that could help avoid ineffective exploration (TipsWarm vs. HermesSkill-ER and HermesSkill). A sufficiently big pool can further improve performance substantially, reaching a success rate of 94.2% and an AUC of 0.546 within the turn budget (TipsWarm-EBP and TipsWarm-BP vs. TipsWarm).

W2. About cost results: We agree that tokens are an important cost factor. In the revision, we present and analyze the total token cost efficiency results, which were originally in the appendix, along with accuracy and time cost efficiency results in Section 5.2."Main Results" in the main text. The input token cost consistently constitutes about 98%-99% of the total token cost across different methods and benchmarks. We present the detailed decomposition of the agent's total tokens, including input and output tokens, along with the mechanism-internal tokens in Appendix C.

W3. About metrics: In all benchmarks, the accuracy of one experiment run is computed on the held-out test split only. “Section 5.1.Benchmarks and Splits” provides the test split information. That is, SkillsBench uses 23 test tasks, AlfWorld uses 134 unseen test tasks, and AppWorld uses 168 test-normal tasks. Please also note that the average accuracy reported in the figures and text is further averaged over three experiment runs, as already specified in “Section 5.1.Running Details”.

To avoid ambiguity you are concerned about, we explicitly clarify that “accuracy results are computed only on the test tasks” in Section 5.1.Metrics in the revision.

Q1: Yes, “n” is a function of “t” in the skill-evolving setting. We have changed “n” to “ntn_t” to reflect this setting.

Q2: Thanks for pointing out the typos. We have done a thorough proofreading of the entire paper for typos and grammar errors.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the rebuttal

Dear SACs, ACs, and reviewers,

Thanks for your comments, which are very valuable to us in improving the final paper. As the end of the rebuttal phase approaches, we would like to briefly summarize what we have done in the revision to address the concerns raised in the comments.

1) We adde results for comparison with the early skill retrieval method. Results show that TipsWarm performs better than the early skill retrieval approach in both final success rate and success rate in early turns. TipsWarm better maintains compact, cross-skill guidance than early retrieval of full skills.

2) We extend the ablation study of the pool size and utility attribution rules. Although we already had the component-level ablation study of the pool size and utility attribution in our initial manuscript, we further evaluate larger pool sizes and various utility attribution rules to study the mechanism effects more deeply in TipsWarm.

3) We also clarify the interpretation of the performance gain, the description of metrics, and the presentation of cost results.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.