acceptodds
Under review as a conference paper at ICLR 2027

Evidence Acquisition And Stopping Via Lattice Distillation

Abstract

Sequential evidence acquisition requires deciding what to observe and when to stop from partial information. We introduce a complete-record view of policy supervision: each fully observed training instance defines a lattice of evidence subsets, with a stop/query decision at every node. For a modest number of evidence groups, this structure can be enumerated even when individual groups contain rich representations. Our method, Subset Lattice Distillation (SLD), scores subsets with a shared masked predictor, constructs multi-step decision targets by cost-sensitive backward planning, and distills them into an observed-history controller. Uniformly resampled masks train the student across the subset space, while combinatorial planning remains offline. Separating subset scores from acquisition prices allows the predictor to be reused when the policy is replanned for new costs. Our analysis identifies action-margin conditions under which privileged supervision preserves optimal online decisions. On retrospective influenza monitoring, SLD achieves higher phase F1 at lower acquisition cost than both the closest-cost adaptive baselines and the fixed full-evidence model. A geographic and seasonal holdout, price perturbations, and ablation studies provide further evaluation.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Evidence acquisition and stopping

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_LpSvOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies sequential evidence acquisition from partial information aiming to balance the cost of acquisition and prediction accuracy. They proposed Subset Lattice Distillation (SLD), which build a teacher model using complete information from an offline dataset to learn cost-sensitive exploration and stopping decision. Then distill its knowledge into a student model that operates under partial (observation so far) information online.

Critical strengths
  1. This paper provided theoretical guarantee for when distilling from teacher model under full information into a student model, which act under partial information so far, preserves good action making. This contribution to the novelty of the proposed algorithm.
Critical weaknesses
  1. The paper lacks a discussion of the best-arm identification and active hypothesis testing literature. The authors should compare their approach with relevant methods and include appropriate baselines from these areas to better contextualize their contributions.
  2. The evaluation does not sufficiently compare SLD with other baselines under diverse acquisition-cost profiles. Table 4 reports SLD's performance under cost perturbations that primarily modify one evidence source at a time, which is insufficient to demonstrate its robustness or comparative advantages across different cost structures.
Questions for authors
  1. Could authors evaluate relevant baselines under a broader range of cost profiles, including scenarios that vary multiple acquisition costs simultaneously and alter their relative rankings.
  2. MLP (WW+ED) is a highly competitive system, as shown in Table 1. How would the author interpret its result in comparison with SLD? Could WW+ED alone be sufficient to achieve comparable performance?
3 comments

Thank you for recognizing the theoretical contribution of transferring decisions from a teacher model with complete information to a student model with partial observations. We address the connections to the literature and provide additional comparisons below.

Q1a: "The paper lacks a discussion of the best-arm identification and active hypothesis testing literature."

Answer: Thank you for pointing out these connections. Existing work in best-arm identification [1,2] and active hypothesis testing [3] does not directly handle our problem setting. Specifically, best-arm identification repeatedly samples arms to identify the highest mean rewards; our task predicts a case's label and reveals each evidence group only once (Section 3). Classical active sequential hypothesis testing requires known observation distributions, which our setting lacks. SLD instead scores evidence subsets in complete records, uses backward planning to construct multi-step queries and stop targets, and distills them into a policy that sees only acquired evidence (Sections 4.2-4.3, Equations (5)-(7)). We will cite these papers and add these connections and distinctions to Section 2.

References

[1] Garivier, A., and Kaufmann, E. Optimal Best Arm Identification with Fixed Confidence. COLT, PMLR 49:998-1027, 2016.

[2] Kaufmann, E., Cappé, O., and Garivier, A. On the Complexity of Best-Arm Identification in Multi-Armed Bandit Models. Journal of Machine Learning Research, 17(1):1-42, 2016.

[3] Naghshvar, M., and Javidi, T. Active sequential hypothesis testing. The Annals of Statistics, 41(6):2703-2738, 2013. DOI: 10.1214/13-AOS1144.

Q1b: "The authors should compare their approach with relevant methods and include appropriate baselines from these areas to better contextualize their contributions."

Answer: These are good suggstions. Following your suggestions, we evaluate the adapted active-testing baseline on a synthetic benchmark (also following Reviewer ZYko's suggestion) because it allows us to compute the optimal acquisition and stopping decisions exactly. Against this reference, SLD achieves lower mean objective regret and higher agreement with optimal actions, as shown in Table R1 below. The methods share the predictor and acquisition costs, and their online decisions rely solely on acquired evidence. We will include these results and the adaptation's query and stopping rules in the revision.

Table R1. Comparison with adapted active hypothesis testing

MethodOptimal-action agreement (%) ↑Objective regret ↓
SLD90.11 ± 3.810.0267 ± 0.0115
Active testing (adapted)71.43 ± 0.000.0400 ± 0.0000
ACO73.25 ± 1.790.1330 ± 0.0400

Results averaged over three runs. Regret is excess expected prediction loss plus weighted acquisition cost relative to the exact online optimum, which uses only acquired evidence.

Q2: "The evaluation does not sufficiently compare SLD with other baselines under diverse acquisition-cost profiles. Table 4 reports SLD's performance under cost perturbations that primarily modify one evidence source at a time, which is insufficient to demonstrate its robustness or comparative advantages across different cost structures." "Could authors evaluate relevant baselines under a broader range of cost profiles, including scenarios that vary multiple acquisition costs simultaneously and alter their relative rankings."

Answer: Good suggestion. Following this suggestion, we ran additional experiments and showed that when multiple acquisition costs and their relative rankings change together, SLD achieves higher mean phase F1 at lower mean cost than both ACO and GDFS in each new setting (Table R2).

Table R2. Adaptive acquisition under joint changes in costs and their rankings

Cost profileMethodPhase macro-F1 ↑Mean acquisition cost ↓
Joint ASLD0.975 ± 0.0041.700 ± 0.100
Joint AACO0.865 ± 0.0102.263 ± 0.008
Joint AGDFS0.870 ± 0.0071.817 ± 0.257
Joint BSLD0.976 ± 0.0042.100 ± 0.100
Joint BACO0.865 ± 0.0092.333 ± 0.145
Joint BGDFS0.868 ± 0.0112.285 ± 0.170

Results averaged over three runs with a shared predictor and prices within each profile. SLD replans and retrains its policy under each profile (Section 6.4). Cost vectors (emergency-department visits, hospital admissions, laboratory positivity, source context): default (1.0, 1.5, 1.2, 0.4); Joint A (0.4, 1.2, 1.5, 1.0); Joint B (1.5, 0.4, 1.0, 1.2). The full acquisition cost is 4.1 across all profiles.

Q3: "MLP (WW+ED) is a highly competitive system, as shown in Table 1. How would the author interpret its result in comparison with SLD? Could WW+ED alone be sufficient to achieve comparable performance?"

Answer: This is a good question. MLP (WW+ED) is a highly competitive low-cost option; SLD improves epidemic-phase prediction and calibration in Table 1. Emergency-department visits are highly informative in this task. SLD improves phase F1 from 0.964 to 0.983 and reduces expected calibration error from 0.073 to 0.014 (Table 1). WW+ED remains cheaper and has stronger binary-alert results (Section 6.2; Appendix C.2, Table 5).

One major contribution is turning complete training records into supervision for acquisition and stopping. SLD achieves higher phase F1 than GDFS at essentially the same acquisition cost and than ACO at lower cost (Section 6.1, Table 1), demonstrating more effective acquisition and stopping.

I appreciate the additional experiment that you have added adapted hypothesis testing and experiments on more diverse cost profile. My concern has been addressed, so I will raise my score.

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

The paper proposes a novel training method for learning active feature acquisition policies called lattice distillation. There are three components. The first is a standard predictor for target variables given any feature subset or partial information, which is parameterized using an encoder-decoder approach. The second is a teacher policy which is constructed via a sample-wise value table, and computing a soft distribution over the optimal actions for the entire trajectory. The third is a student policy that trains to distill the teacher policy's expected action distribution. Experimental validation shows that this approach outperforms baselines on the influenza benchmark.

Critical strengths
  • The motivation for improving algorithms for non-myopic AFA policy learning is clear and well-substantiated, as traditional methods face limitations such as difficulty in value function approximation and training instability.

  • The method of distilling the optimal cost-sensitive feature-action trajectory into the policy represents a significant conceptual contribution, and the empirical results on the influenza dataset are also very strong.

Critical weaknesses
  • The experimental validation is insufficient to demonstrate that non-myopic optimal actions are recovered in a manner that is (1) generalizable across settings and (2) independent of predictor performance. A more rigorous evaluation would make a more convincing argument that distillation approaches should be preferred for AFA. For example, the paper lacks synthetic benchmarks where the ground truth is known to verify whether learned policies successfully recover optimal trajectories. Demonstrating this is important because while non-myopic approaches are theoretically optimal, their primary practical bottleneck lies in training stability across dataset scales and settings. Evaluating on only a single real-world dataset leaves significant questions regarding empirical robustness, given it is a specific setting with only four feature groups and training data is also very limited (n≈5000n \approx 5000).

  • The manuscript is somewhat difficult to parse, even for domain-familiar readers, particularly in terms of how the theoretical results serve to explain and justify the underlying methodology. While the overarching approach is intuitive, the paper would be strengthened by offering clearer intuition alongside each theoretical result, explicitly detailing how each result provides insight into the proposed method.

Questions for authors
  • In Table 1, it is surprising that the baselines perform this poorly, given that this is an experimental setting with only four feature groups and the same predictor is used across methods. Why does the fixed MLP with WW+ED already outperform all baselines, when these baselines should theoretically improve over the fixed MLP by collecting more information? What is the stopping criterion used for the greedy approaches (EDDI, GDFS)? As a sanity check, what is the optimal upper bound performance of SLD in this setting if (1) all features were used or (2) if the optimal features based on the computed value table were used? And how well does SLD recover the optimal trajectory as defined by the value table?

  • Is the theoretical objective different from standard approaches that perform value-function approximation, or does the optimization procedure recover the same oracle (5.1 Proposition 1)? Theorem 2 is also challenging to parse, what is the object that is being characterized in MT(H)M_T(H)? What is the takeaway and how should this quantity be interpreted based on the quantities (such as action-relative mean distortion and ΔH\Delta H)? Some intuition would be helpful and this would also help understand why each quantity was defined in 5.2.

7 comments

Thank you for recognizing our "significant conceptual contribution" and "very strong" influenza results. Following your suggestions, we added a synthetic benchmark in which the optimal acquisition policy can be computed exactly, tested training stability across data sizes and settings, and extended the evaluation to NHANES. These experiments provide direct evidence for SLD's acquisition decisions and the benefit of uniform-mask supervision. We also clarify the baseline comparisons and how each theoretical result supports the method.

Q1a: "The experimental validation is insufficient to demonstrate that non-myopic optimal actions are recovered in a manner that is (1) generalizable across settings and (2) independent of predictor performance."

Answer: Thank you for raising these two points.

  • (1) Generalizability across settings. The submission already evaluates transfer to an unseen region and season with all systems frozen (Section 6.5, Table 3). HHS Region 3 is excluded from fitting and model selection. On its held-out 2024/25 test set, SLD achieves phase F1 0.958 at mean acquisition cost 2.700, compared with 0.928 at 2.975 for ACO. It also outperforms the full-evidence model on both metrics. The higher-F1, lower-cost advantage therefore persists under the tested geographic and seasonal shift.

  • (2) Independence from predictor performance. Our experimental setup uses identical trained predictor weights for SLD, ACO and GDFS (Section 6.1). SLD achieves phase F1 0.983 versus 0.873 for GDFS at essentially the same mean acquisition cost, and also outperforms ACO at lower cost (Table 1). A stronger predictor therefore cannot explain these gains.

Q1b: "Evaluating on only a single real-world dataset leaves significant questions regarding empirical robustness, given it is a specific setting with only four feature groups and training data is also very limited (n ≈ 5000)."

Answer: We assembled the largest set of jointly aligned state–week records we could obtain from public sources under our alignment and temporal-availability criteria: 5,933 state-week episodes in total (Section 6.1). Each episode combines a state's wastewater and clinical surveillance evidence for one week. CDC aggregates influenza surveillance by week, and its reports document uneven wastewater coverage and data completeness [1,2]. This scale supports testing our central hypothesis through shared-predictor comparisons, chronological evaluation and an unseen region and season (Sections 6.1 and 6.5).

Following your suggestion, we extended our evaluation to a second application: diabetes classification using NHANES. SLD achieves higher mean macro-F1 at lower mean acquisition cost than all three acquisition baselines, and matches the reported full-evidence mean macro-F1 while reducing cost by 60.7% (Table R1).

Table R1. NHANES diabetes classification

MethodMacro-F1 ↑Mean acquisition cost ↓
SLD0.671 ± 0.0281.571 ± 0.091
ACO0.658 ± 0.0261.901 ± 0.474
GDFS0.658 ± 0.0271.704 ± 0.219
Fixed subset0.654 ± 0.0242.000
Full evidence0.671 ± 0.0284.000

Results averaged over three runs. All methods share the same frozen predictor within each run. ACO and GDFS are adapted to this task. Each evidence group has unit acquisition cost; the fixed subset was selected using development data.

References

[1] Centers for Disease Control and Prevention. U.S. Influenza Surveillance: Purpose and Methods. FluView, 2025.

[2] Louis S, et al. Wastewater Surveillance for Influenza A Virus and H5 Subtype .... MMWR, 2024;73:804-809.

Q2a: "A more rigorous evaluation would make a more convincing argument that distillation approaches should be preferred for AFA." "For example, the paper lacks synthetic benchmarks where the ground truth is known to verify whether learned policies successfully recover optimal trajectories."

Answer: Thank you for this valuable suggestion, which helped us directly assess recovery of optimal decisions. Following the suggestion, we constructed a synthetic benchmark in which the optimal acquisition policy can be computed exactly. The reference policy uses only acquired evidence. SLD achieves the highest agreement with both optimal actions and complete optimal trajectories, together with the lowest mean regret (Table R2).

Table R2. Recovery of optimal actions and complete trajectories

MethodOptimal-action agreement (%) ↑Optimal-trajectory agreement (%) ↑Objective regret ↓
SLD90.11 ± 3.8166.67 ± 14.430.0267 ± 0.0115
SLD (teacher trajectories only)57.83 ± 16.1416.67 ± 14.430.1900 ± 0.0455
ACO73.25 ± 1.7950.00 ± 0.000.1330 ± 0.0400
Greedy acquisition0.00 ± 0.000.00 ± 0.000.2200 ± 0.0000
Active testing (adapted)71.43 ± 0.000.00 ± 0.000.0400 ± 0.0000

Results averaged over three runs. SLD uses uniform mask sampling; the ablation trains the same model only on teacher trajectories. Trajectory agreement requires recovery of the complete optimal action sequence. Regret is excess expected prediction loss plus weighted acquisition cost relative to the exact online optimum.

With the predictor and training budget fixed, SLD also outperforms its ablation, which is trained only on teacher trajectories. Uniformly sampled masks provide useful supervision beyond the teacher's own paths.

Q2b: "Demonstrating this is important because while non-myopic approaches are theoretically optimal, their primary practical bottleneck lies in training stability across dataset scales and settings."

Answer: Thank you for highlighting the importance of training stability across data sizes and settings. Accordingly, we tested SLD with 500 and 5,000 training records and with four and eight evidence groups, using the same training budget for both SLD variants. SLD achieves the lowest mean regret across all four settings, with run-to-run variation equal to or lower than that of training only on teacher trajectories (Table R3). Its mean regret decreases with more training data at both group counts.

Table R3. Objective regret across training sizes and evidence sets

GroupsTraining recordsSLD ↓SLD (teacher trajectories only) ↓ACO ↓
45000.0400 ± 0.00000.2200 ± 0.00000.1300 ± 0.0400
45,0000.0267 ± 0.01150.1900 ± 0.04550.1330 ± 0.0400
85000.0400 ± 0.00000.3519 ± 0.16180.1333 ± 0.0231
85,0000.0133 ± 0.02310.3495 ± 0.16070.1333 ± 0.0231

Mean ± standard deviation over three runs. Both SLD variants use 1,000 training updates. The eight-group task adds four irrelevant groups while preserving the original label rule.

We will add these experiments and their protocols to the revision.

Q3: "In Table 1, it is surprising that the baselines perform this poorly, given that this is an experimental setting with only four feature groups and the same predictor is used across methods. Why does the fixed MLP with WW+ED already outperform all baselines, when these baselines should theoretically improve over the fixed MLP by collecting more information?"

Answer: WW+ED is a strong baseline because emergency-department visits are highly informative in this task. ACO and GDFS select different subsets for different cases, so higher spending does not guarantee that every case receives WW+ED plus additional inputs. Moreover, additional inputs do not guarantee higher macro-F1 for a fitted predictor (Appendix B.2). SLD achieves the highest phase macro-F1 in Table 1, outperforming GDFS at essentially the same cost and ACO at lower cost. These three methods share the same trained predictor (Section 6.1).

Q4: "What is the stopping criterion used for the greedy approaches (EDDI, GDFS)?"

Answer: EDDI and GDFS stop when no remaining query has an estimated gain larger than its acquisition-cost penalty. Otherwise, they select the query with the largest difference between estimated gain and cost penalty. The penalty is selected on development data under the acquisition-budget constraint. We will state these stopping and parameter-selection rules explicitly in the experimental description.

Q5a: "As a sanity check, what is the optimal upper bound performance of SLD in this setting if (1) all features were used or (2) if the optimal features based on the computed value table were used?"

Answer: Thank you for the suggestion. We address both configurations below:

  • (1) All features. When all features are used, SLD achieves phase macro-F1 0.983 at acquisition cost 4.100 (Table R4). The learned SLD policy matches this reported F1 while reducing acquisition cost by 46.9% (manuscript Table 1), showing the benefit of learning which evidence to acquire and when to stop.

  • (2) Optimal features from the value table. Using the optimal features selected by the value table, SLD (value-table teacher) achieves phase macro-F1 0.988 at mean acquisition cost 1.070 (Table R4). The lattice thus identifies highly predictive subsets at substantially lower cost than acquiring every feature, providing useful supervision for the acquisition policy. The teacher uses complete records to minimize the stopping score plus acquisition cost (Equation (5)).

Table R4. SLD under the two requested acquisition configurations

ConfigurationPhase macro-F1 ↑Acquisition cost ↓
SLD (all features)0.9834.100
SLD (value-table teacher)0.9881.070

Reported point estimates using the same frozen SLD predictor. The first configuration acquires all features; the second follows the complete-record teacher in place of the learned policy. F1 is calculated on answered episodes: 1,796/1,796 and 1,650/1,796, respectively.

Q5b: "And how well does SLD recover the optimal trajectory as defined by the value table?"

Answer: Thank you for suggesting this trajectory check. SLD recovers 37.14% of the optimal trajectories defined by the value table, the highest rate among the compared policies (Table R5). This is a strict measure: every query and the final stopping decision must minimize the value-table cost; tied optimal actions count as matches.

Table R5. Exact recovery of complete teacher trajectories

MethodExact trajectory agreement (%) ↑
SLD37.14
EDDI27.06
GDFS4.18
ACO3.51

Reported point estimates on the same 1,796 test episodes, including forced stops. The reference is the complete-record value-table teacher.

Q6: "Is the theoretical objective different from standard approaches that perform value-function approximation, or does the optimization procedure recover the same oracle (5.1 Proposition 1)?"

Answer: Thank you for asking about the objective and the oracle. SLD uses the standard objective of expected prediction or deferral loss plus weighted acquisition cost; its contribution is how the acquisition policy is trained. The lattice computes multi-step query and stop targets from complete records, and the student learns action probabilities from partial observations rather than regressing an online value function (Equations (4)–(7)).

Proposition 1 identifies the average teacher action probabilities conditional on the student's input (Equation (8)). It does not, by itself, establish recovery of the optimal online policy. Theorem 2 gives sufficient conditions for these learned probabilities to preserve the optimal online action for the same predictor and terminal rule (Equation (10)).

Reviewer_p69WOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors present a way to create an adaptive approach for sampling sets of information when adaptively sampling and halting. They use retrospective influenza monitoring as an interesting test example.

Critical strengths

The problem seems interesting but I’m not yet fully clear on how the authors approached it.

The experimental setting is interesting, there are are theoretical results (proofs not checked) and the setting seems well motivated.

Critical weaknesses

I didn’t fully yet understand how SLD uses structure to avoid the combinatorial subset problem, and what assumptions are needed.

The authors mention that “Appendix A.5 gives the construction cost and a theorem relating pruning error to decision stability.” but it would be nice to see these at least briefly discussed in the main text.

What is the action-relative mean distortion? Please expand on this. In thm 2 it’s stated that if the gap between the best and second action is above this mean distortion, and another ocndition, then we can get optimal action selection, so it’s important I understand.

It would be stronger if a second domain was included.

Questions for authors

See above.

1 comment

Thank you for recognizing the well-motivated problem and interesting experimental setting. Following your suggestion, we added NHANES as a second application, in which SLD matches the reported full-evidence mean macro-F1 while reducing acquisition costs by 60.7%.

Q1: "I didn't fully yet understand how SLD uses structure to avoid the combinatorial subset problem, and what assumptions are needed."

Answer: SLD avoids repeated planning for the same evidence and learns a policy that needs no subset search at test time. For example, acquiring A then B gives the same evidence and remaining choices as acquiring B then A. The lattice treats these as one state and computes its best continuation only once (Equation (5)). We perform this planning over all subsets of the complete labeled training records, then teach the controller to choose the next query or stop based solely on the evidence already observed (Section 4.3).

This relies on evidence staying fixed within each case, each group being acquired at most once, and acquisition costs adding up independently of order (Section 3).

Q2: "The authors mention that "Appendix A.5 gives the construction cost and a theorem relating pruning error to decision stability." but it would be nice to see these at least briefly discussed in the main text."

Answer: Thank you for this suggestion. We will summarize both the construction cost and the decision-stability guarantee in Section 4.3. During construction, the search budget limits how many candidate subsets are evaluated. The retained graph keeps every possible first query but may prune deeper continuations; backward planning then takes time linear in its nodes and edges. Under the stated conditions, Corollary 7 preserves the optimal online action when its guaranteed probability lead exceeds the error bounds from pruning and student approximation (Appendix A.5).

Q3: "What is the action-relative mean distortion? Please expand on this. In Thm 2, it's stated that if the gap between the best and second action is above this mean distortion, and another condition, then we can get optimal action selection, so it's important I understand."

Answer: Action-relative mean distortion measures how much the teacher misstates the cost advantage of one action over another. We average the teacher's costs over records that give the controller the same input. Equation (9) compares the resulting cost differences with those under the best policy that uses only acquired evidence, keeping the predictor and final prediction procedure fixed. If the best action's true advantage exceeds this distortion, the average costs still favor that action.

Distillation converts each record's costs into probabilities and then averages them. This average can favor a different action from the average costs. The quantity MTM_T gives a guaranteed minimum for the best action's lead in these averaged probabilities. That lead must exceed twice the student's largest error in matching these probabilities: the correct action's probability can fall while a competitor's rises. Under Thm 2's assumptions, these two conditions ensure that choosing the student's highest-probability action gives the optimal decision. We will add this explanation beside the theorem.

Q4: "It would be stronger if a second domain was included."

Answer: Thank you for mentioning this. Following your suggestion, we added a second domain. Specifically, for the added NHANES diabetes task, we show that our SLD outperforms all three acquisition baselines, achieving higher mean macro-F1 at lower mean acquisition cost (Table R1). It also matches the reported mean macro-F1 of using all evidence while reducing mean acquisition cost by 60.7%. All methods share the same frozen predictor within each run.

Table R1. NHANES diabetes classification

MethodMacro-F1 ↑Mean acquisition cost ↓
SLD0.671 ± 0.0281.571 ± 0.091
ACO0.658 ± 0.0261.901 ± 0.474
GDFS0.658 ± 0.0271.704 ± 0.219
Fixed subset0.654 ± 0.0242.000
Full evidence0.671 ± 0.0284.000

Results averaged over three runs. Each evidence group has unit acquisition cost. The fixed subset was selected using development data.

These results demonstrate that SLD's advantage in choosing useful evidence and stopping acquisition extends to a second application domain. We will include this experiment and its protocol in the revision.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.