acceptodds
Under review as a conference paper at ICLR 2027

When Interaction Is the Bottleneck: Evaluating and Scaling Agents Under Costly Feedback

Abstract

Large language model (LLM) agents rely on external feedback to test hypotheses, retrieve information, and verify answers. When feedback is costly, capabilities depend on how they acquire information as well as how they reason over it. We introduce ExpGYM, an evaluation suite built on pre-cached feedback and deterministic surrogates for hyperparameter tuning, multi-hop search, and evidence audit. ExPGYM enables repeated evaluation under different feedback budgets without rerunning costly external work, while preserving feedback charges. Moving from Free Feedback to Tight Budget reduces overall performance across all 6 models and changes task-level model preferences. Gemini 3.8 Flash remains the strongest model overall, yet its multi-hop search AVG falls from 71.9% to 17.6% between Free Feedback and Tight Budget. On ParamNet's Adult task, selecting the Free Feedback leader yields a Tight Budget Gap Score 15.22 below the best available model. Trajectory analyses show that limited feedback restricts exploration and evidence completion, while additional requests do not always provide new information. We introduce POOLACT to coordinate acquisition through shared observations and exploration state. With 4 agents under matched per-agent budgets, POOLACT raises Qwen3.8's Moderate Budget evidence audit exact evidence-set accuracy from 68.8% with independent rollouts to 93.7%. These findings highlight feedback acquisition in evaluating agents and coordinating groups under limited budgets.

Primary area:
datasets and benchmarks
Keywords:
Costly Feedback, Feedback Acquisition, Budget-Constrained Evaluation, Parallel Scaling, Test-Time Scaling

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_z32GOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces ExpGym to evaluate LLM agents under limited feedback budgets in hyperparameter tuning, multi-hop search, and evidence audit. It proposes PoolAct to coordinate multiple agents through shared feedback and exploration state. Experiments on six models show that feedback budgets affect performance and model rankings, while PoolAct reduces redundant requests and improves results.

Critical strengths
  1. expgym provides a framework for evaluating agents under limited feedback budgets across three different tasks.
  2. poolact consistently improves evidence-audit performance across all six models under both budget settings.
  3. The component ablations provide useful support for the benefits of feedback sharing and coordinated action selection.
Critical weaknesses
  1. The paper does not compare poolact with a single agent given the same total feedback budget, leaving unclear when multi-agent coordination is preferable to increasing a single agent’s budget.
  2. Agents are not told the budget ceiling or per-call costs before their first action, limiting budget-aware planning under tight constraints.
  3. The decision lock serializes reasoning and action selection across agents, allowing only tool execution to run in parallel. This may increase latency and limit parallel speedups.
Questions for authors
  1. How does PoolAct compare with a single agent given the same total feedback budget and a comparable overall step limit?
  2. What fraction of search instances are solvable by an ideal retrieval strategy within the 900-second Tight Budget under the current withholding rule?
  3. How do the main results change when the initial prompt explicitly states the budget ceiling and estimated per-call costs?
2 comments
AuthorsOpenReview ↗

Reply to reviewer z32G (Part 1/2)

We thank the reviewer for the thoughtful comments on budget allocation and deployment costs. We provide detailed responses to your specific concerns below.

W1: The paper does not compare poolact with a single agent given the same total feedback budget, leaving unclear when multi-agent coordination is preferable to increasing a single agent's budget.

Q1: How does PoolAct compare with a single agent given the same total feedback budget and a comparable overall step limit?

R1: Thank you for asking for this comparison. We added a direct comparison with Long N1: one GLM-5.3 agent receives four workers' combined feedback allowance and step cap (4B, 120 steps), while each PoolAct member receives B and 30 steps. Both methods use a common deadline fixed before execution:

MetricLong N1PoolAct N4
Best visible Gap by shared deadline89.2797.90
Terminal best visible Gap98.4198.19
Feedback-inclusive simulated completion (h)27.648.63

Both methods are scored by their best completed, visible configuration, and the simulation includes recorded training durations and measured inference and coordination time. Over the full runs, Long N1 and PoolAct use 27.44 and 30.07 hours of charged feedback. PoolAct gains 8.63 points at the deadline and wins all nine pairs, while terminal quality is similar. Multi-agent coordination is therefore preferable when results are needed within a practical completion window. In workflows such as scientific discovery and system optimization, giving one agent several times longer for serial experiments can take days, and that time cannot be assumed available. We have added the detailed comparison to Appendix F.5, with a main-text summary in Section 5.5 of the revised manuscript.

Q2: What fraction of search instances are solvable by an ideal retrieval strategy within the 900-second Tight Budget under the current withholding rule?

R2: Thank you for this question. We verified complete-answer certificates for 22 of 73 questions (30.1%) within the 900-second Tight Budget. This is a constructive lower bound, not a solvability ceiling. We constructed evidence certificates offline and replayed their article sets through the released tool, verifying total costs strictly below 900 s. We have added the certificate construction and results to Appendix F.7 of the revised manuscript. Since set F1 also credits correct partial answers, the certificates give F1 lower bounds of 74.55 (Whois) and 45.85 (Whatis), well above current agent scores.

W2: Agents are not told the budget ceiling or per-call costs before their first action, limiting budget-aware planning under tight constraints.

Q3: How do the main results change when the initial prompt explicitly states the budget ceiling and estimated per-call costs?

R3: We added the 900 s ceiling and the 280–320 s cost per new article to GLM-5.3's initial prompt under Tight Budget:

Search (Whois, Tight)Control F1 (%)Upfront F1 (%)Difference
All 39 questions20.0120.96+0.95

F1 increases by only 0.95 points across all 39 questions, suggesting that upfront cost information has little effect on performance in this experiment. As detailed in Appendix B.2, agents already learn the ceiling after the first paid call, since each paid observation reports the incurred cost and remaining budget. For HPO, exact per-call costs are not given in advance and depend on the configuration; as detailed in Appendix E.1, three configurations in one NASBench-201 task cost 3,641 s, 13,525 s and 9,682 s. We have added the prompts and results to Appendix F.8 of the revised manuscript.

AuthorsOpenReview ↗

Reply to reviewer z32G (Part 2/2)

W3: The decision lock serializes reasoning and action selection across agents, allowing only tool execution to run in parallel. This may increase latency and limit parallel speedups.

R4: Thank you for this question. The lock covers only graph reading, action selection and request registration, and it is released before tool execution. In the matched NASBench-101 ablation in Section 5.5, both variants retain the shared cache and exploration graph:

Matched lock ablationGap ScoreTokens (K)Feedback charge (s)
Graph, no decision lock87.68464.88120,312.8
PoolAct, with decision lock96.04410.95110,056.0

Adding the decision lock gains 8.36 Gap points, uses 11.60% fewer tokens and requires 2.85 fewer hours of charged training work across the four members (8.53% less).

To measure parallel speedup with training waits included, we compared PoolAct with a single agent given the same total allowance (see W1/Q1). The four PoolAct members complete 30.07 hours of charged feedback in 8.63 hours, about 3.5 concurrent jobs on average, and finish 3.2× sooner than the single agent; all serialized model calls take 7.5% of that time. We have detailed the matched ablation in Appendix F.2 and the completion-time comparison in Appendix F.5 of the revised manuscript.

Thank you for helping us clarify these comparisons.

Reviewer_62ksOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces ExpGym, a benchmark for evaluating agents when external feedback is costly, across hyperparameter tuning, multi-hop search, and evidence audit. The authors show that limiting feedback substantially affects performance and can change which model performs best. They also introduce PoolAct, which allows multiple agents to share observations and exploration state to coordinate tool calls and reduce redundant work. Experiments across six models show promising improvements, particularly in evidence verification and multi-hop search.

Critical strengths

Important problem and useful benchmark: The paper highlights an important limitation of current agent evaluations: performance can depend heavily on access to external feedback. Evaluating the same models under different feedback budgets is useful, and the analysis of exploration and evidence completeness provides insights beyond final task accuracy.

Promising multi-agent coordination results: PoolAct shows substantial improvements, particularly in evidence audit. I appreciate the comparison against shared caching and the ablations showing that coordinating pending requests can provide additional benefits beyond simply reusing completed observations.

Critical weaknesses

Limited cost-efficiency analysis. PoolAct shows promising improvements over independent rollouts and shared caching, but the evaluation does not fully account for total feedback expenditure, inference tokens, coordination overhead, and wall-clock latency. Since PoolAct adds a shared exploration mechanism and serializes decision-making, a cost-matched comparison is important to establish whether the improvements justify the additional complexity.

Baselines and generalization. While the experiments show benefits from coordination, simpler approaches such as upfront task partitioning or priority-based verification are not evaluated. The three environments are also relatively structured and use deterministic feedback, leaving open how well the findings generalize to more realistic agent settings.

Questions for authors

How realistic are the simulated feedback costs? Would the conclusions hold with actual execution costs and latency?

Since Free Feedback does not expose costs while budgeted regimes do, could some performance differences come from how cost information is presented?

How would PoolAct compare against simpler approaches, such as dividing tasks or hypotheses among agents upfront and sharing the results?

Do the improvements remain when accounting for total inference tokens, coordination overhead, and wall-clock latency? How would PoolAct compare with a single agent given the same total resources?

How statistically reliable are the reported improvements? Could the authors provide confidence intervals or additional independent repetitions, particularly for the multi-agent experiments?

How does PoolAct compare with stronger task-specific baselines, such as Bayesian optimization for tuning or priority-based evidence verification?

How would PoolAct behave with noisy or incorrect feedback, and how does the serialized decision-making scale beyond eight agents?

3 comments
AuthorsOpenReview ↗

Reply to reviewer 62ks (Part 1/3)

We appreciate the reviewer's careful evaluation and helpful suggestions. We provide detailed responses to your specific concerns below.

Q1: How realistic are the simulated feedback costs? Would the conclusions hold with actual execution costs and latency?

R1: Thank you for this question. Our HPO feedback costs reflect real training work. We checked the benchmark cost sources and added their construction details:

HPO groupTasksFeedback-time source
NASBench-101 / NASBench-2016Recorded training times
ParamNet Adult / Higgs / Letter3Runtime RF fitted to measured training data

The ParamNet models use 10,000 real configurations per dataset, each trained for 50 epochs [1]. Separate forests predict performance and runtime, and with step = 50 we charge the predicted full training time, following HPOBench's recorded/predicted-time accounting [2]. As detailed in Section 2.2, budgets scale with each task's reference cost.

The feedback bottleneck also persists with measured latency. As detailed in Appendix C.6 (Figure 7), feedback accounts for 73.4–95.5% of the accounted time on ParamNet and over 99% on all six NASBench tasks, with measured model response times included. We have detailed the cost sources in Appendix F.1 of the revised manuscript.

Q2: Since Free Feedback does not expose costs while budgeted regimes do, could some performance differences come from how cost information is presented?

R2: To check whether presentation drives the gap, we reran GLM-5.3 on all 39 Whois questions under Free Feedback, with costs shown or hidden in the tool observations.

Search questionsFree, cost hidden F1 (%)Free, cost visible F1 (%)Tight F1 (%)
All 39 Whois questions66.1163.3120.01

Displaying costs changes Free F1 by −2.80 percentage points, while Free with costs remains 43.30 points above the archived Tight reference. As detailed in Appendix B.2 and Table 1, Moderate and Tight present costs identically and differ only in the ceiling, yet Search AVG falls by 24.7–37.2 percentage points from Moderate to Tight for all six models. Cost display therefore does not explain the large Free–Tight gap. We have added this ablation to Appendix F.3 (Table 11) of the revised manuscript.

W2: Baselines and generalization. While the experiments show benefits from coordination, simpler approaches such as upfront task partitioning or priority-based verification are not evaluated. The three environments are also relatively structured and use deterministic feedback, leaving open how well the findings generalize to more realistic agent settings.

Q3: How would PoolAct compare against simpler approaches, such as dividing tasks or hypotheses among agents upfront and sharing the results?

R3: We added the suggested baseline: divide the work at the start, execute independently, and combine the final answers. We compare both methods on the same 12 Whois questions with GLM-5.3:

MethodWhois F1 (%)Feedback charge (s)
PoolAct44.183,851.3
Upfront partitioning + final answer aggregation31.684,119.6

PoolAct gains 12.50 F1 points while using 6.51% less feedback charge. In multi-hop search, newly found entities determine which queries are useful next, so coordinating during execution adds value beyond an initial division of work. On generalization, the deterministic environments keep the feedback source controlled; extending reuse to stochastic or stateful tools would require recording observation uncertainty or environment state. We have added the baseline comparison to Appendix F.4 (Table 12) and the extension paths to Appendix G.1 of the revised manuscript.

AuthorsOpenReview ↗

Reply to reviewer 62ks (Part 2/3)

W1: Limited cost-efficiency analysis. PoolAct shows promising improvements over independent rollouts and shared caching, but the evaluation does not fully account for total feedback expenditure, inference tokens, coordination overhead, and wall-clock latency. Since PoolAct adds a shared exploration mechanism and serializes decision-making, a cost-matched comparison is important to establish whether the improvements justify the additional complexity.

Q4: Do the improvements remain when accounting for total inference tokens, coordination overhead, and wall-clock latency? How would PoolAct compare with a single agent given the same total resources?

R4: The improvements remain once these costs are counted. For the GLM-5.3 Tight-budget runs in Section 5.3 (Table 5), we report per-pool inference tokens (including coordination context) and feedback charge:

TaskMethodScoreInference (K tokens)Feedback charge (s)
NASBench-101Naive84.81418.12105,976.4
NASBench-101Cached85.26507.06115,909.8
NASBench-101PoolAct96.04410.95110,056.0
AuditNaive63.13711.503,642.3
AuditCached66.97752.373,522.5
AuditPoolAct81.00648.583,693.2
WhoisNaive19.5892.324,181.6
WhoisCached18.76107.844,127.0
WhoisPoolAct31.16299.953,975.7

PoolAct improves all three scores. It uses fewer tokens than both baselines on NASBench-101 and Audit, and more tokens but the least feedback on Whois. The injected graph text, which is the direct coordination overhead, accounts for 2.58% of input tokens on Audit and 17.03% on Whois.

We also compared PoolAct with a single agent given the same total allowance: Long N1 receives 4B and 120 steps, and each of four PoolAct members receives B and 30 steps, on NASBench-101 A/B/C with three repeats. Completion time includes training, inference and coordination:

MetricLong N1PoolAct N4
Best visible Gap by shared deadline89.2797.90
Terminal best visible Gap98.4198.19
Feedback-inclusive simulated completion (h)27.648.63

PoolAct wins all nine comparisons at a deadline fixed before execution and finishes 3.2× sooner. At full-budget completion, final quality is similar, and Long N1 uses fewer tokens (397.43K vs. 645.70K) and less feedback (27.44 vs. 30.07 h). PoolAct is therefore most useful when costly feedback must arrive within a practical time window. In workflows such as scientific discovery and system optimization, giving one agent several times longer for serial experiments can take days, and that time cannot be assumed available. We have added these analyses to Appendix F.2 and Appendix F.5 of the revised manuscript.

Q5: How statistically reliable are the reported improvements? Could the authors provide confidence intervals or additional independent repetitions, particularly for the multi-agent experiments?

R5: Paired confidence intervals support PoolAct's performance gains across the three task families. We computed 95% paired bootstrap intervals from the existing Tight-budget results:

Metric (0–100 scale)PoolAct − NaivePoolAct − Cached
NASBench-101 Gap+11.23 [4.35, 21.66]+10.78 [5.98, 14.91]
Whois F1+11.59 [3.96, 20.95]+12.40 [3.76, 22.39]
Audit evidence accuracy+17.87 [11.29, 25.00]+14.03 [8.60, 20.36]

All displayed intervals are positive. Audit improves over both baselines across all six models and both budget settings; 19 of 24 paired intervals are entirely above zero. We use 10,000 paired resamples of questions, documents or HPO tasks, and HPO's three recorded repetitions are averaged within each task before resampling. We have added the complete intervals and existing repetition statistics to Appendix F.6 of the revised manuscript.

AuthorsOpenReview ↗

Reply to reviewer 62ks (Part 3/3)

Q6: How does PoolAct compare with stronger task-specific baselines, such as Bayesian optimization for tuning or priority-based evidence verification?

R6: PoolAct is complementary to these methods rather than an alternative to them. Bayesian optimization and priority rules decide what a single worker should request next, while PoolAct coordinates requests across multiple workers, so either could serve as the per-agent policy inside PoolAct. We therefore keep the underlying LLM identical across Naive, Cached and PoolAct to isolate the coordination gain.

For tuning, prior work supports LLM acquisition policies when paid evaluations are scarce: on 32 HPOBench tasks with ten evaluations each, Zhang et al. report mean ranks of 2.42 for GPT-4 Turbo, 3.45 for RF-based BO and 3.80 for GP-based BO [3]. These published results motivate our choice of an LLM policy; they are not a direct BO comparison on ExpGym.

For verification, PoolAct already spreads requests across hypotheses without handcrafted priorities. As detailed in Section 5.4, under Tight Budget averaged over six models:

MethodVisible repliesDistinct hypothesesRepeated requests
Naive9.103.6444.4%
PoolAct9.095.4024.8%

With almost the same number of replies, PoolAct verifies 48% more hypotheses. We have added this scope distinction and the prior-work context to Appendix G.3 of the revised manuscript.

Q7: How would PoolAct behave with noisy or incorrect feedback, and how does the serialized decision-making scale beyond eight agents?

R7: As detailed in Section 5.3, the scaling results at N = 8 are:

Method, N=8Moderate Audit accuracy (%)Tight Audit accuracy (%)
Naive84.263.3
Cached91.966.1
PoolAct100.092.3

Under Tight Budget, PoolAct rises from 64.3% at N=2 to 92.3% at N=8; under Moderate, it reaches 100% by N=6. Eight agents therefore suffice for the Audit tasks studied. Beyond eight, more workers would compete for the same decision queue, so speedups need not grow proportionally. In the added HPO timing study, serialized model time is small relative to feedback-inclusive completion time. In an added HPO timing experiment with four GLM-5.3 agents on NASBench-101 A/B/C, all serialized model calls take 38.70 minutes, 7.5% of the 8.63-hour feedback-inclusive completion time, and expensive tool jobs keep running because the lock is released before tool execution. Finer-grained locks, for example per hypothesis, are a natural extension for larger pools.

Incorrect shared feedback could mislead several agents. For noisy tools, we would retain independent observations with uncertainty and recheck conflicting feedback before reuse. Stateful tools would additionally require state-aware cache and graph entries, or independent environment replicas. We have expanded these extension paths and the scaling discussion in Appendices G.1–G.2 of the revised manuscript.

[1] BOHB: Robust and Efficient Hyperparameter Optimization at Scale

[2] HPOBench: A Collection of Reproducible Multi-Fidelity Benchmark Problems for HPO

[3] Using Large Language Models for Hyperparameter Optimization

Thank you again for these helpful suggestions.

Reviewer_6RbrOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces EXPGYM, an evaluation suite designed to assess LLM agents under explicit feedback budgets. By utilizing pre-cached feedback and deterministic surrogates for tasks like hyperparameter tuning, multi-hop search, and evidence audit, EXPGYM enables reproducible evaluation of agents under Free, Moderate, and Tight feedback constraints without incurring real-world execution costs. The authors evaluate six recent LLMs, revealing that tight budgets severely degrade performance and alter model rankings, leading to high "deployment regret" if models are selected under free feedback conditions. To mitigate this, they propose POOLACT, a multi-agent coordination strategy that shares an exploration graph and pending requests among parallel agents. POOLACT prevents redundant costly queries and significantly improves evidence completion and search performance under constrained budgets compared to independent rollouts or simple result caching.

Critical strengths
  • EXPGYM's use of pre-cached deterministic surrogates to simulate costly feedback is a highly effective way to benchmark budget-constrained agents. It allows for rigorous, reproducible, and computationally efficient evaluation of how feedback limits impact agent capabilities. The introduction of "deployment regret" to quantify the performance cost of selecting a model under the wrong feedback regime is a particularly insightful and practical contribution for real-world deployment.
  • POOLACT addresses a genuine bottleneck in scaling agents: the redundancy of expensive queries in parallel rollouts. The ablation study (comparing Naive, Cached, Peer, Graph, and Full POOLACT) is excellent. It convincingly demonstrates that coordinating future actions via a shared exploration graph and decision lock provides substantial gains beyond merely caching past results, effectively broadening hypothesis coverage without increasing the feedback budget.
Critical weaknesses
  • The EXPGYM framework and POOLACT's exact-match cache reuse rely heavily on the assumption that tools are deterministic and stateless (i.e., the same action always yields the same observation). In many real-world agentic settings (e.g., web navigation, software engineering, or interacting with stateful APIs), tool outputs depend on the environment's state. The paper does not discuss how EXPGYM or POOLACT would generalize to stochastic or stateful environments where simple exact-match caching is invalid.
  • The paper models feedback cost purely as an abstract "simulated seconds" budget, decoupling it from actual wall-clock latency, API rate limits, or the LLM's internal reasoning time. While this cleanly isolates the feedback acquisition bottleneck, real-world agents must balance the latency of a tool call against the time spent reasoning. Ignoring wall-clock time and concurrent execution limits might oversimplify the optimal stopping and exploration behaviors of cost-aware agents.
Questions for authors

See weak points above

1 comment
AuthorsOpenReview ↗

Reply to reviewer 6Rbr

We thank the reviewer for the positive assessment and the practical questions. We provide detailed responses to your specific concerns below.

W1: The EXPGYM framework and POOLACT's exact-match cache reuse rely heavily on the assumption that tools are deterministic and stateless (i.e., the same action always yields the same observation). In many real-world agentic settings (e.g., web navigation, software engineering, or interacting with stateful APIs), tool outputs depend on the environment's state. The paper does not discuss how EXPGYM or POOLACT would generalize to stochastic or stateful environments where simple exact-match caching is invalid.

R1: Thank you for raising this question. We designed ExpGym to keep the feedback source fixed when comparing models and budgets. As detailed in Section 2.3, the same task, action and interaction state produce the same result and cost. Fixed-seed evaluations, fixed-corpus retrieval and annotation-based verification fit this setting.

PoolAct's gains also extend beyond caching. As detailed in Section 5.5 (Figure 4), our ablations retain the same shared cache and exploration graph, isolating what decision coordination adds:

MetricGraph: cache + graph, no decision lockPoolAct: decision coordination added
Whois F1 (%)20.731.2
Audit evidence accuracy (%)76.581.0
NASBench-101 mean-agent Gap Score87.796.0

These gains come from coordinating future requests rather than from reusing past results. For stochastic tools, exact reuse could be disabled or observations retained as samples while preserving pending-request coordination. Stateful tools would require state-aware cache and graph entries, or independent environment replicas. We agree that many real-world tools are not deterministic or stateless. We have added these extension paths in Appendix G.1 of the revised manuscript.

W2: The paper models feedback cost purely as an abstract "simulated seconds" budget, decoupling it from actual wall-clock latency, API rate limits, or the LLM's internal reasoning time. While this cleanly isolates the feedback acquisition bottleneck, real-world agents must balance the latency of a tool call against the time spent reasoning. Ignoring wall-clock time and concurrent execution limits might oversimplify the optimal stopping and exploration behaviors of cost-aware agents.

R2: Thank you for this point. Costly feedback remains the main bottleneck even if inference time is counted. As detailed in Appendix C.6 (Figure 7), feedback accounts for 73.4–95.5% of the accounted time on ParamNet and over 99% on all six NASBench tasks, with measured model response times included.

Request-rate caps and concurrency limits are distinct from the feedback budget; they constrain when tool jobs can start and can be modeled by bounding the number of concurrent jobs. We have detailed these timing and execution assumptions in Appendix F.5 of the revised manuscript.

Thank you again for your constructive feedback.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summarization of paper revision

We thank the reviewers for their constructive feedback. The reviewers find ExpGym useful and effective (6Rbr, 62ks), deployment regret insightful and practical (6Rbr), and PoolAct’s evidence-audit gains substantial and consistent (62ks, z32G).

The revised PDF marks all changes in blue. Appendix F adds resource and timing analyses, equal-allowance single-agent comparisons, additional controls, and confidence intervals. Appendix G discusses generalization and scalability. The main text summarizes the new findings and clarifies scope and limitations.

Thank you for your interesting suggestion for future work, and thank you for your recognition of our work!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.