acceptodds
Under review as a conference paper at ICLR 2027

Risk-Controlled Evidence Selection for Retrieval-Augmented Question Answering

Abstract

Evidence compression must account for both shared context and the reader's dependence on the removed text. A low-cost solution to a predicted support graph need not preserve an answer. We separate these questions through shared-resource selection and calibration of the complete compression-and-fallback policy. The selector retains the classical additive covering guarantee while a validated flow relaxation supplies a separate, instance-specific bound on graph cost. A common budget-aware search addresses a further mismatch between penalized selection and hard admission budgets. This search preserves already-admitted contexts but must itself be included in calibration. Whole-policy calibration controls paired answer degradation among admitted questions, not absolute factual correctness. On 120 controlled graphs, safeguarded shared-cost selection reduces mean cost by 18.5% relative to additive greedy. Its graph-cost certificates have median ratio 1.132 and maximum 2.076, with no zero lower bounds in this family. A budget-repair study covers 2,400 paired selector configurations and retains a case where finite search misses a feasible budget. Simulations over 1,500 calibration samples quantify the cost of fallback. These studies establish optimization and statistical properties.

Primary area:
optimization
Keywords:
retrieval-augmented generation, evidence selection, context compression, combinatorial optimization, shared-resource optimization, approximation guarantees, risk control, selective prediction, whole-policy calibration, question answering

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_L8jSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies evidence selection for retrieval-augmented question answering with a fixed candidate context and reader. It separates optimization on a predicted support graph from the effect of compression on answer quality. The proposed CRESS-S selector combines additive and shared-cost greedy selection, retaining the additive solution as a safeguard. A validated flow relaxation provides an instance-specific certificate of graph optimization quality. The framework also includes a common budget-aware repair search and calibration over a fixed family of compression-and-fallback policies. The reader-level target is paired answer degradation relative to the full-context baseline, rather than absolute answer correctness. Evaluation consists of controlled graph experiments and calibration simulations. A public-data reader study is specified but has not been executed.

Critical strengths
  1. Careful separation of optimization quality and reader reliability. The paper clearly distinguishes coverage of predicted requirements, near-optimality under the declared graph objective, and preservation of the reader’s answer. The counterexample in Section 4.1 explains why locally calibrated support scores need not remain valid after cost-based selection. Calibrating the complete admission-and-reading policy addresses the relevant selection effect, and the discussion of paired degradation explicitly avoids interpreting the guarantee as absolute answer correctness. These distinctions make the formulation useful for understanding what evidence-selection guarantees do and do not establish.
  2. The controlled experiments make optimization and support-prediction errors distinguishable. The graph comparisons use matched candidates, costs, and corrupted support memberships, together with MILP reference optima and independently validated lower bounds. Evaluation against latent support labels is kept separate from predicted coverage. The paper also reports certificate tails, search failures, and fallback-only calibration outcomes. This experimental design makes the controlled results interpretable and provides useful evidence about the behavior and limitations of the proposed components.
Critical weaknesses
  1. The value of the combined approach for question answering has not been demonstrated. Section 5.4 and Appendix G explicitly state that the public-data reader study has not been run. The completed studies evaluate graph optimization and policy calibration separately. They therefore do not establish whether contexts selected using learned support predictions can be certified at a useful admission rate while reducing measured deployment cost. In particular, the calibration simulation uses 16 threshold policies with a risk target of 0.10, whereas the proposed reader study uses 42 policies with a target of 0.02; the simulation does not establish the operating behavior of that reader pipeline. The authors acknowledge this limitation, and I do not regard it as a contradiction of the stated theorem. Nevertheless, because the paper identifies the interface between selection, certification, and reader calibration as its main contribution, the absence of a joint evaluation leaves a central part of its value unresolved.

  2. The incremental contribution beyond simpler combinations is not sufficiently established. The manuscript appropriately identifies the weighted-cover guarantee and simultaneous finite-family calibration as established ingredients. This makes the additional value of the safeguard, graph certificate, and common repair particularly important. Table A.4 reports an 18.5% mean cost reduction relative to additive CRESS, but a more modest 2.3% reduction relative to union-cost greedy, which already accounts for shared resources. After common repair, the differences narrow further: at budget 48, CRESS-S and union-cost greedy both admit all 120 graphs, with fallback-inclusive normalized graph costs of 0.073 and 0.074, respectively (Table A.7). These results show a benefit, but do not yet establish whether the additional selection machinery provides a consequential advantage once computation is included. Furthermore, Appendix G specifies that the reader policy does not call the graph LP. The certificate is useful as an optimization diagnostic, but its role in the proposed interface needs a clearer account. Taken together, the current evidence does not sufficiently establish the significance of the full combination beyond a simpler shared-cost selector with the same repair and calibration.

Questions for authors
  1. What is the principal additional capability provided by the proposed interface? Beyond combining shared-cost selection with finite-family risk calibration, which methodological or theoretical result do the authors consider the main new contribution? In particular, is the graph certificate intended solely for offline optimization auditing, or does it inform the construction or selection of reader policies? Clarifying this would help assess the contribution independently of the unexecuted reader study.

  2. Why retain omission penalties when admission requires complete predicted coverage? Section 3.3 and Appendix F.3 acknowledge that omission penalties can cause the core selector to return a context that fails the full-coverage admission gate. What motivates retaining these penalties rather than treating all predicted requirements as mandatory before applying the budget gate? Would that alternative change the interpretation of the gains attributed to common repair? This question concerns omission-induced rejections, separately from cases in which greedy selection misses a budget-feasible solution.

  3. What computational overhead accompanies the safeguard’s remaining benefit after repair? For the matched comparisons in Table A.7, what additional selection and repair time does CRESS-S require relative to union-cost greedy? Since their reported graph costs after repair are close, matched computational measurements would materially affect the assessment of whether the safeguard provides a useful net benefit.

1 comment

Thank you for distinguishing the soundness of the analysis from the unresolved application value. Your questions identify where the individual components need to justify their place in the complete policy.

1. What the graph certificate adds. In the implemented reader protocol, the graph certificate is an offline optimization diagnostic. It neither constructs the selected context nor gates reader admission. Appendix G's admission rules use coverage, confidence, and budget, not the LP ratio. It would be incorrect to attribute a reader-quality improvement to that certificate.

Its additional capability is an instance-specific upper bound on suboptimality against the shared-cost graph optimum, rather than only an additive-fragment guarantee. With a validated positive lower bound LL, F(H)/OPTG≤F(H)/LF(H)/\mathrm{OPT}_G\le F(H)/L. This can identify an inefficient selected graph without a uniform fragment-completeness assumption. It cannot identify missing semantic evidence. The rational residual correction makes the numerical lower bound checkable, but we do not claim that primal-dual certification is a new optimization principle.

The broader contribution is the separation of quantities that can otherwise be conflated. Predicted coverage, graph optimality, and reader degradation are different targets. Section 4.1 shows that even calibrated local scores need not remain valid after cost-based selection. C-RAG (Kang et al., ICML 2024) controls generation loss, whereas Principled Context Engineering (Chakraborty et al., ECIR 2026) controls relevant-snippet retention. Our use of finite-family testing follows Learn then Test (Angelopoulos et al., 2025). We are not claiming a new statistical framework or that every component is necessary in every deployment.

2. Omission penalties versus mandatory coverage. We agree that a full-coverage gate makes mandatory coverage a particularly important control. We ran it on the same 120 graphs. At budget 48, all 43 original CRESS-S rejections arise from missing predicted requirements. None is over budget. Removing omission choices raises initial admission from 77 to 113. The seven remaining rejections exceed the budget. Common repair brings both variants to 120 admissions.

The mandatory variant needs seven searches instead of 43. Its aggregate normalized resource cost after repair is 0.079300.07930, versus 0.073140.07314 for the penalized variant, while latent true-support recall is 0.92360.9236 versus 0.92220.9222. Same-host selection-plus-conditional-search time is 1.461.46 ms versus 5.975.97 ms per graph. Thus penalties are not intrinsically preferable. Here they produce more search and somewhat lower retained-resource cost, not a demonstrated reader benefit. This control also shows why the recovered admission cannot be credited solely to better budget search. Much of it repairs omissions introduced by the penalized objective itself.

3. Overhead behind the remaining safeguard gain. For the matched budget-48 comparison, union-cost greedy takes 0.370.37 ms for selection and 5.115.11 ms for conditional search, totaling 5.485.48 ms per graph. CRESS-S takes 0.670.67 ms and 5.295.29 ms, totaling 5.975.97 ms. The approximately 0.490.49 ms overhead accompanies aggregate normalized resource costs of 0.073900.07390 and 0.073140.07314. Selector timings are warmed seven-repeat medians per instance, then averaged. Search is timed once per graph-budget and shared between matched comparisons. These one-host CPU measurements exclude scoring and reader inference. They do not establish net deployment savings.

Checking this comparison exposed a caption error in Table A.7. The displayed costs are ratios of total deployed to total full-context resource cost, not means of per-instance ratios. The corresponding per-instance means are 0.066160.06616 and 0.065310.06531. Both summaries preserve your conclusion that the repaired methods are close. The accompanying results give both aggregations explicitly.

Joint validation and the calibration settings. We agree that the 16-policy, 0.100.10-risk simulation cannot establish behavior of the unrun 42-policy reader study at 0.020.02. Even with no observed paired failures, the latter requires 334 admitted calibration questions at confidence error 0.050.05. The direct-search and mandatory-coverage alternatives are controlled ablations only. Incorporating them into a reader comparison requires fixing the enlarged family and its simultaneous correction before calibration.

The public-reader result remains missing. Your requested joint evaluation must include absolute accuracy, all four paired outcomes, admission, and fallback-inclusive computation. The new controls narrow the empirical claim rather than manufacture a large gain. We respectfully ask that the paper's value be assessed in its validated separation of optimization and statistical claims, together with the documented tradeoffs and failure cases, without treating the modest safeguard improvement as proof of end-to-end efficacy.

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

This paper studies evidence selection for retrieval-augmented QA while distinguishing graph optimization quality from downstream answer reliability. Its CRESS-S selector combines additive and shared-cost greedy selection, retaining a classical covering guarantee. A flow relaxation with rationally validated lower bounds provides instance-specific graph-cost certificates. The framework additionally introduces a common budget-repair search and calibrates the complete compression-and-fallback policy using simultaneous finite-family testing, targeting paired answer degradation among admitted questions. Evaluations are done on controlled graphs, paired selector configurations, and calibration simulations, which illustrate optimization gains and calibration behavior, but the proposed public-data QA study has not been executed.

Critical strengths
  1. The paper addresses an important tension between evidence compression and answer reliability, carefully distinguishing computational guarantees from semantic sufficiency.

  2. The residual-safe, rationally evaluated lower bound makes the graph-cost certificate more credible than relying directly on a numerical solver’s objective.

  3. Experiments performance is good, CRESS-S reduces mean union cost by 18.5% over additive greedy, demonstrating the value of accounting for shared resources in the tested setting.

Critical weaknesses
  1. Section 5.4 and Appendix G explicitly state that the public-reader experiment has not been run, leaving no evidence that the complete method achieves useful admission while preserves answer quality.

  2. The paper’s main contribution is their integration rather than a substantial new algorithmic or statistical result, which is more engineer focus.

  3. The main experiments use a fixed, shallow graph family with 67 vertices and 66 edges, without runtime scaling results for the flow relaxation and rational validation. It's necessary to add extra experiments.

Questions for authors
  1. Can the authors provide results from the prespecified HotpotQA study, including absolute accuracy, paired degradation, and fallback-inclusive computation, etc?

  2. How does CRESS-S with repair compare against running the same budget-aware search directly, followed by identical calibration?

  3. Based on the framework and paper content, which technical result or algorithmic mechanism should be regarded as the principal advance beyond classical covering, flow-relaxation certificates, and finite-family risk control?

1 comment

Thank you for the specific questions and for recognizing the independently validated lower bound as a strength. The blue revision adds the requested direct-search control and a separate timing study of the flow relaxation and rational check.

1. The HotpotQA study and fallback-inclusive cost. The public-reader study remains to be run. Section 5.4 and Appendix G now distinguish deployment computation from calibration and comparative evaluation. A deployed question incurs scoring and selection, including any repair, followed by compressed-context inference on admission or full-context inference otherwise. Absolute accuracy, the four paired correctness outcomes, admission, fallback, and measured component costs are reported separately in that protocol. The graph certificate is an offline optimization diagnostic and is not an input to the reader's admission gate.

2. Direct search versus CRESS-S followed by repair. Section 5.2, Table 4, and Appendix F.4 now contain this comparison. Direct search uses the identical candidate pool, predicted memberships, costs, width-32 search, stopping rule, and budgets. The comparison covers all four budgets on the same 120 graphs, with no reference solution or latent support label supplied to the search.

At budget 48, direct search and CRESS-S plus conditional repair both admit 120/120 graphs. Their aggregate normalized resource costs are 0.07144 and 0.07314. CRESS-S accepts 77 initial contexts and searches only 43, reducing measured selection-and-search time from 15.13 to 5.97 ms per graph, or 60.6%. Thus the initial selector provides a substantial reduction in search work at matched admission, while direct search retains fewer resources. Table A.10 separates selection from search, and Appendix F.4 reports both ratios of total costs and averages of per-instance ratios.

We also added mandatory coverage, which admits 113 graphs before repair and needs seven searches. It reaches 120 admissions in 1.46 ms, with resource ratio 0.07930. This control explains the role of omission-induced rejection rather than attributing every recovered admission to budget search. The new alternatives are controlled graph ablations. Any reader comparison including them must fix the complete policy family before applying the same simultaneous calibration rule.

3. Scaling of the flow relaxation and rational validation. Section 5.1 and Appendix F.5 add constructed trees at depths two and four. The four settings contain 67, 79, 277, and 553 vertices, with 12, 12, 24, and 24 requirements. We use three seeds and three timed repetitions per seed.

Table A.11 reports median total times of 0.032, 0.033, 0.229, and 0.571 seconds for matrix construction, LP solution, and rational validation. Separate saved-witness checks take 0.019, 0.021, 0.148, and 0.332 seconds, including matrix reconstruction but no new LP solve. Those check times are not added again to the totals. The appendix gives seed-level ranges and the recorded Xeon Platinum 8573C host.

The formulation has (m+1)∣E∣+m+∑i∣Ti∣(m+1)|E|+m+\sum_i|T_i| variables. Reporting both dimensions explains the dependence on graph size and requirement count. This study measures the new certificate itself, separately from the archived greedy-kernel timings. Valid bounds are obtained on every recorded workload.

4. The principal advance. Section 1.1 now identifies the method as shared-resource selection with separate computational and statistical guarantees. The safeguard returns the less costly feasible graph from additive and union-cost greedy while retaining F(H^)≤HmOPTAF(\hat H)\le H_m\mathrm{OPT}_{\mathcal A}. The validated flow bound supplies a separate graph-level certificate, F(H^)/OPTG≤F(H^)/LF(\hat H)/\mathrm{OPT}_G\le F(\hat H)/L for L>0L>0. Repair preserves existing admissions, and the complete repaired policy is the unit of reader-risk calibration.

Weighted-cover analysis and finite-family testing are established tools. C-RAG controls generation risks (Kang et al., ICML 2024), and Learn then Test provides the testing framework (Angelopoulos et al., 2025). The contribution here is a precise account of how sharing, repair, and fallback affect the guarantees, supported by counterexamples and matched controls. The within-bin example shows why local calibration is insufficient, while the budget example identifies a distinct failure of penalized selection.

The revision now answers the direct-search, computational-scaling, and mechanism questions with explicit results and locations in the paper. We respectfully ask that these additions be reflected in the assessment of the method and its evidence.

Reviewer_epM5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents a framework for evidence compression in retrieval-augmented QA that combines shared-resource selection, graph-cost certificates, budget repair, and calibration of the complete compression-and-fallback policy. Its statistical guarantee controls paired answer degradation relative to full-context reading. Validation consists of controlled graph experiments and calibration simulations, and the proposed public-data QA evaluation has not been executed.

Critical strengths
  1. The paper separates optimization quality from answer reliability. Its whole-policy calibration accounts for selection and fallback, while its paired-loss definition makes clear that limiting degradation does not ensure absolute correctness. Such distinctions provide a useful structure for evaluating evidence compression.
Critical weaknesses
  1. The paper does not evaluate the complete framework on a real QA task. The graph experiments and calibration simulations assess individual components but do not show whether any policy achieves effective compression and admission rates while controlling reader-level risk. Therefore, the intended application, including its accuracy and computational cost, remains unvalidated.
  2. The main guarantees rely on classical weighted-cover analysis, flow relaxation, and established finite-family risk testing. Although the integration is thorough, its added value is unclear. CRESS-S reduces synthetic graph cost by 2.3% compared to union-cost greedy, and the common repair further reduces differences between selectors.
Questions for authors
  1. Can the complete framework certify a useful compression policy on a real QA benchmark while reducing total deployment cost? Reporting results in absolute accuracy, paired degradation, admission, and costs, including scoring, search, and fallback, could materially change my assessment.
2 comments

Thank you for identifying the separation of optimization quality and reader reliability as a strength. We have revised the paper in blue to make the incremental contribution explicit and to answer the computational questions with matched comparisons.

1. What the method adds beyond its classical ingredients. Section 1.1 now states the contribution in terms of three complementary guarantees. The safeguard chooses the cheaper feasible graph produced by additive and shared-cost greedy, so its core output is never more costly than either candidate under the declared objective. Theorem 1 retains the additive covering guarantee. Theorem 2 independently bounds suboptimality against the shared-cost graph optimum using a validated lower bound. Theorem 3 controls the complete admission-and-reading policy, including fallback. Each guarantee has its own comparator and assumptions.

This separation has operational consequences. Section 4.1 constructs calibrated local scores with population error 0.0890.089, yet cost-based selection raises the selected subgroup's error to 0.80.8. Section 3.3 shows why satisfying a penalized objective and finding a context within a hard budget are different tasks. Together, these results explain why optimizing evidence and calibrating individual scores cannot simply be treated as an answer-level certificate.

The positioning follows the relevant literature. C-RAG controls generation risk (Kang et al., ICML 2024), and Learn then Test supplies finite-family risk testing (Angelopoulos et al., 2025). Our contribution specifies how shared resources, budget repair, and fallback enter the optimization and calibrated policy. Section 1.1 and Appendix E make that distinction explicit.

2. The safeguard's incremental benefit. We agree that union-cost greedy is the appropriate comparator for isolating the safeguard. The abstract and Section 5.1 now report both comparisons. CRESS-S reduces mean graph cost by 18.5% against additive greedy and by 2.3% against union-cost greedy. The latter's seed-cluster interval is [1.1%,3.8%][1.1\%,3.8\%] in Table A.4. Independent true-support recall is reported separately. The contribution is therefore supported by a matched comparison with a selector that already accounts for sharing, rather than only by the larger additive-baseline gap.

3. What common repair contributes and what it costs. Section 5.2, Table 4, and Appendix F.4 now compare CRESS-S plus conditional repair directly with running the same search on every graph. Candidates, memberships, costs, width 32, and budgets are identical.

At budget 48, both admit 120/120 graphs. CRESS-S preserves 77 initially admitted contexts and searches only the remaining 43. Its measured selection-and-search time is 5.97 ms per graph, versus 15.13 ms for direct search, a 60.6% reduction. Their aggregate resource ratios are 0.07314 and 0.07144, respectively. The direct-search control therefore identifies the computational value of conditional repair while also reporting its resource tradeoff.

The revision also includes mandatory coverage, which needs only seven searches and takes 1.46 ms, with resource ratio 0.07930. Against repaired union-cost greedy, the safeguard changes the resource ratio from 0.07390 to 0.07314 at approximately 0.49 ms additional computation. Table A.10 gives the timing components. These are one-host CPU measurements of selection and search, with scoring and reader inference accounted for separately.

4. The requested real-reader result and deployment accounting. The HotpotQA reader study remains to be run. Section 5.4 and Appendix G now state the complete accounting explicitly. Deployment includes scoring, selection and repair, then compressed-context inference on admission or full-context inference on rejection. Calibration is a separate offline expense. The graph certificate is an offline diagnostic in the stated reader protocol, so its LP cost is not charged to an admission decision that does not use it. Absolute accuracy, all four paired outcomes, admission, fallback, and computation are separate reporting targets.

The revision thus adds measured answers to the incremental-value and computational questions while keeping the reader result distinct. We respectfully ask that the assessment be reconsidered in light of the new direct-search and mandatory-coverage controls, the quantified safeguard overhead, and the clarified relationship between the optimization and statistical guarantees.

W1

The concern is that the complete framework had not been evaluated on a real QA task. The authors confirm that the HotpotQA reader study has still not been run. They add an evaluation plan and cost-accounting rules. Without results, it remains unclear whether the framework achieves useful compression.

W2

The concern is the added value. CRESS-S reduces cost by 2.3% relative to union-cost greedy and repair further narrows the gap between selectors. The authors restate the roles of the three guarantees and repeat the 2.3% result. They do not address the convergence of all selectors to similar costs after repair. The incremental value therefore remains limited.

Q1

The question is for total deployment cost on a real benchmark, including scoring, search, and fallback. The authors report millisecond-level selection and search times on synthetic graphs. These measurements exclude scoring and reader inference. As reader inference is the main cost of real deployment, these controls do not answer the question.

The main concerns remain unresolved. I lean to maintain the rating.

Comments on the paper

1 comment

We thank the reviewers and Area Chair for their careful assessment and constructive discussion. The revision incorporates the additional comparisons and clarifications in blue. We appreciate epM5’s recognition of the separation between optimization and reader reliability, TDWJ’s assessment of the validated graph bound, and L8jS’s recognition of the counterexample and controlled experimental design.

Reviewers epM5 and TDWJ, and incremental value and computational cost. The abstract and Section 5.1 now report both relevant comparisons. CRESS-S reduces mean graph cost by 18.5% relative to additive greedy and 2.3% relative to union-cost greedy, with a seed-cluster interval of [1.1%,3.8%][1.1\%,3.8\%] for the latter. This distinguishes the benefit of accounting for shared resources from the safeguard’s additional gain.

Section 5.2, Table 4, and Appendix F.4 add direct budget search and mandatory coverage as controls. The methods use the same candidates, memberships, costs, and width-32 search. At budget 48, all four methods pass the coverage-and-budget gate on 120/120 controlled graphs.

MethodGraphs searchedResource ratioSelection and search, ms
Direct search1200.0714415.13
Mandatory CRESS-S with repair70.079301.46
Penalized union-cost greedy with repair420.073905.48
Penalized CRESS-S with repair430.073145.97

The resource ratio divides total retained resource cost by total full-context resource cost. Times are mean CPU selection-and-search times per graph on the same host, excluding scoring and reader inference.

These comparisons answer two questions.

  • Conditional repair. CRESS-S preserves 77 initially admitted contexts and searches the remaining 43. It matches direct search’s admission with 60.6% less selection-and-search time, at the resource costs shown above.
  • Additional computation. Relative to repaired union-cost greedy, it reduces the resource ratio from 0.07390 to 0.07314 and adds approximately 0.49 ms of selection-and-search time per graph. Table A.10 makes this computational price explicit alongside the incremental saving.

Reviewer L8jS. Omission penalties and mandatory coverage. The new control distinguishes the reasons for rejection. At budget 48, the 43 initial rejections under penalized CRESS-S all arise from omitted requirements. Mandatory coverage admits 113 graphs immediately, leaving seven over-budget selections. Both reach 120 admissions after the same repair.

When admission requires complete predicted coverage, mandatory selection avoids omission-induced rejection and reduces search work. In these experiments, penalized selection followed by repair retains fewer resources. The comparison explains this tradeoff and identifies which admission gains come from the common repair. Appendix F.4 reports all four budgets.

Reviewer TDWJ, and certificate computation and scaling. Section 5.1 and Appendix F.5 add measurements on constructed trees of different sizes and depths, up to 553 vertices and 24 requirements. Median total certificate time ranges from 0.032 to 0.571 seconds across the four settings, including matrix construction, LP solution, and rational validation. Table A.11 separately reports checks of saved witnesses. Section 3.2 also makes clear that the graph certificate is an offline optimization diagnostic in the specified reader protocol.

All reviewers. The contribution and its implications. Sections 1.1 and 3.2 now state the three guarantees together. CRESS-S accounts for shared resources while retaining the additive covering guarantee. The validated lower bound assesses the returned graph against the shared-cost optimum. For a fixed family of complete policies, calibration controls paired reader degradation among admitted questions, with full-context fallback on rejection.

These distinctions matter because cost-based selection can increase error within a subgroup even when local scores are calibrated, and repair can change the population admitted for compression. The counterexamples and new controls show how those effects arise and how to account for them. Together, the analysis and controls establish which guarantees apply at each stage and which decisions require recalibration.

The public-reader evaluation remains outstanding. Section 5.4 and Appendix G keep that study separate from the completed graph experiments and calibration simulations, and specify accuracy, paired degradation, admission, and computation accounting.

We appreciate L8jS’s follow-up acknowledgment that the additions address several methodological questions and quantify the tradeoffs more clearly. The revision provides measured answers on direct search, mandatory coverage, computational overhead, and certificate computation. We respectfully ask that the final assessment give full weight to these additions and the technical strengths identified in the reviews.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.