acceptodds
Under review as a conference paper at ICLR 2027

EviLink: Multi-Path Schema Linking with Uncertainty-Guided Evidence Acquisition for Large-Scale Text-to-SQL

Abstract

Schema linking is a difficult and important step in large-scale Text-to-SQL, where systems must identify a compact yet sufficient schema context from large and ambiguous databases. Existing methods often treat schema linking as deterministic selection around a single SQL path, but complex questions may admit multiple valid realizations with different schema needs. We reframe schema linking as uncertainty-aware schema-need inference over multiple plausible SQL paths, where the system distinguishes required schema items from path-dependent uncertain ones and acquires evidence only where needed. We instantiate this reframing with EviLink, which combines multi-hypothesis schema grounding with uncertainty-guided evidence acquisition. Experiments on BIRD-Dev and Spider2-Snow show that this perspective improves the balance among schema completeness, schema relevance, and token cost. On Spider2-Snow, EviLink achieves 93.04% field-level strict recall rate, uses 116.55K average tokens, and improves downstream SQL generation under a fixed generator.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Schema Linking, Text-to-SQL, Large Language Models

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

The paper presents EviLink, which at its core resembles a two-step pipeline for text-to-SQL that intends to consider multiple SQL candidates for aligning with the underlying schema until committing to one. The first step generates multiple SQL candidates and informs which SQL / schema elements are necessary and which are uncertain, after which the second investigates the uncertain elements further in an agentic manner. The evaluation shows that EviLink holds some promise, but comes with several limitations.

Critical strengths

S1 The idea to consider multiple SQL paths before committing to one SQL, as done in EviLink, is promising and the motivation aligns with earlier findings.

S2 Using candidate overlap to inform common hence more certain schema elements versus uncertain ones that guide the agentic data exploration seems an efficient approach.

S3 The evaluation includes most prior relevant approaches as baselines, relevant metrics and datasets, although the related DisambigParse [1] is missing.

Critical weaknesses

W1 Methodological weaknesses:

  • W1.1 SQL path hypothesis generation isn’t developed in a principled manner, while it is one of the core ideas of the paper. In its current form, SQL hypothesis generation relies fully on the LLM’s capability to generate plausible and diverse SQL queries. Moreover, K is set “adaptively” without explanation, whereas K can also be a single hypothesis.
  • W1.2 The uncertainty assessment is based on an LLM’s own assessment through a dense prompt given the collected evidence elements, rather than a reliable measure.

W2 The performance analysis shows that while EviLink is effective on Spider-Snow, EviLink still has significant shortcomings as its SRR is relatively far off from the best-performing approach and its precision is compromised significantly. On Spider-Snow EviLink indeed performs better but it isn’t shown that the database sizes in Spider-Snow explains this.

W3 The ablation study confirms that while multiple SQL path hypotheses are effective for improving recall, it significantly sinks the precision (hence F1). Meanwhile, the uncertainty-grounded refinement isn’t contributing as much to the final performance.

W4 Altogether, the contribution made in this paper is not solid, novel nor substantial enough for publication in its current form.

[1] Saparina, I., & Lapata, M. Disambiguate first, parse later: Generating interpretations for ambiguity resolution in semantic parsing. In Findings of the Association for Computational Linguistics: ACL 2025.

Questions for authors

Q1 How is K set adaptively in multi-hypothesis SQL generation?

Q2 It would be interesting to explain the evidence-guided agentic exploration with a concrete example.

2 comments
AuthorsOpenReview ↗

Rebuttal - Part I

We sincerely thank you for recognizing multi-path preservation, the potential efficiency of overlap-guided exploration, and the evaluation choices. Your questions help clarify hypothesis generation and its benefits. Below, we address them through methodological explanations, experimental results, and a concrete example.

1. How the number of hypotheses is chosen

W1.1 SQL path hypothesis generation isn’t developed in a principled manner, while it is one of the core ideas of the paper. In its current form, SQL hypothesis generation relies fully on the LLM’s capability to generate plausible and diverse SQL queries. Moreover, K is set “adaptively” without explanation, whereas K can also be a single hypothesis.

Q1 How is K set adaptively in multi-hypothesis SQL generation?

In §3.1, K is the fixed maximum budget (default 4), while M ≤ K is the actual number of hypotheses retained. The prompt encourages coverage of distinct join routes, filters, aggregation, or temporal interpretations within this budget. Hypotheses describe plausible SQL realizations and need not be executable queries. Each drives its own table and field selection; we preserve these grounded paths and their field support for refinement.

The original Spider2 runs produced:

Hypotheses (M)GLM-5.1 (main)Qwen3.8-27B (ablation)
1192107
2122372
39828
410510
Total517517

K = 4 is a reasonable default balancing path coverage and computational cost, rather than an optimal setting. The appendix reports sensitivity to K (Figure 4).

2. Basis of the uncertainty signal and refinement’s contribution

W1.2 The uncertainty assessment is based on an LLM’s own assessment through a dense prompt given the collected evidence elements, rather than a reliable measure.

W3 The ablation study confirms that while multiple SQL path hypotheses are effective for improving recall, it significantly sinks the precision (hence F1). Meanwhile, the uncertainty-grounded refinement isn’t contributing as much to the final performance.

In §3.1, the score is computed from schema support across independently grounded paths using Eqs. (3)–(5), rather than an LLM’s self-reported confidence. Like an SQL engineer comparing implementations, it uses agreement and disagreement over schema requirements to prioritize investigation. It is a path-support heuristic, not a calibrated probability of correctness.

Multi-path grounding prioritizes coverage of alternative SQL implementations, trading precision for recall. With grounding already achieving high recall, UGAR focuses on improving selection quality. The Table 2 ablation shows higher precision and NSF on both datasets after adding UGAR.

In this additional control on all 517 examples, both agents share initial grounding outputs and the paper’s Qwen3.8, tools, and budgets. The generic agent uses the same tools for refinement without uncertainty guidance: it does not partition schema items into certain and uncertain sets based on cross-path support.

AgentComplete recallSRRNSRNSPNSFRefinement tokensLinking tokens
Guided392/51775.8292.6656.5765.4911.07K99.05K
Generic389/51775.2492.4656.4765.3212.73K100.71K

With equal turn limits, guidance reduces refinement tokens by 13.05%, evidence-tool calls from 567 to 424, and model calls from 952 to 750, with small observed gains in all four schema metrics. This shows improved schema quality at lower cost. We will add this experiment and analysis to the revised manuscript.

3. Recall, precision, and database size

W2 The performance analysis shows that while EviLink is effective on Spider-Snow, EviLink still has significant shortcomings as its SRR is relatively far off from the best-performing approach and its precision is compromised significantly. On Spider-Snow EviLink indeed performs better but it isn’t shown that the database sizes in Spider-Snow explains this.

Submitted Table 1 gives EviLink the highest Spider2 field SRR: 93.04% vs. APEX-SQL’s 80.85%, with lower NSP/NSF (26.04%/36.65% vs. 47.50%/58.64%). On BIRD, it has lower SRR than AutoLink (93.22% vs. 98.89%) but higher NSP/NSF (51.58%/63.60% vs. 11.39%/19.28%). Submitted Table 12:

Schema groupNEviLink SRRAPEX-SQL SRRDifference, pp
Small (≤91)17198.2587.72+10.53
Medium (92–505)17892.7084.83+7.87
Large (≥506)16888.1069.64+18.45

EviLink leads APEX-SQL across sizes, most on large schemas (18.45 pp vs. 7.87 on medium), supporting its large-schema coverage advantage.

AuthorsOpenReview ↗

Rebuttal - Part II

4. A concrete evidence-guided refinement example

Q2 It would be interesting to explain the evidence-guided agentic exploration with a concrete example.

Figure 2 illustrates the grounding-to-refinement pipeline on the real Spider2-Snow example sf_bq391. Below, we provide a turn-by-turn trace of another example, sf_local163, to show how the agent uses evidence and makes decisions.

In sf_local163, refinement reduces 11 fields to 6 while retaining all four reference fields. The question asks for faculty salaries closest to their rank's mean, returning rank, first name, last name, and salary. Refinement starts with L2 evidence, including available field descriptions, sample values, and null ratios, to assess field relevance. The current GLM trace records:

StepRecorded actionEffect
GroundingThree paths select one table and 11 fieldsKeeps four reference fields plus seven extras
Turn 1Five successful discard calls remove FacState, FacCity, FacDept, FacHireDate, and FacSupervisorPrunes attributes using default L2 evidence and query scope
Turn 2list_columns checks STAFFHOURS, SALARYDATASET, and UNIVERSITY_STUDENTInspects candidates; adds no fields
Turn 3keep receives an empty item listReturns six fields with all four reference fields retained
Field metricBefore refinementAfter refinement
Selected fields116
Reference fields retained4/44/4
Precision36.36%66.67%
Recall100.00%100.00%
F153.33%80.00%

This trace shows how evidence-guided pruning improves precision while retaining recall in this example; two extra fields remain. It does not imply universal gains. We will add this walkthrough in the revised manuscript.

5. Relation to DisambigParse

S3 The evaluation includes most prior relevant approaches as baselines, relevant metrics and datasets, although the related DisambigParse [1] is missing.

We thank you for highlighting DisambigParse. Our introduction already cites the AmbiQT and AMBROSIA papers (Bhaskar et al., 2023; Saparina and Lapata, 2024) to motivate preserving multiple valid SQL implementations. DisambigParse addresses this related problem by generating natural-language interpretations, filling in missing meanings, and parsing each into SQL. EviLink focuses on schema linking: grounding alternative SQL paths separately and using their schema requirements to guide refinement. We will add a comparison with DisambigParse in the related work of the revised manuscript.

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

This paper studies schema linking for large-scale Text-to-SQL. Its central premise is that a question may admit multiple structurally distinct SQL realizations, so committing early to a single presumed path can remove schema elements required by another valid solution. EviLink adaptively generates up to four structured hypotheses representing different SQL-relevant paths, independently grounds each hypothesis to tables and fields, and consolidates the path-wise selections using a smoothed cross-hypothesis support score. It then partitions schema items into a locked required set, an uncertain set selected by only some hypotheses, and a zero-support candidate set.A refinement agent examines the uncertain items using semantic field profiles and, when needed, richer statistics or lightweight SQL probes. It may also recover overlooked items from the candidate set. Items are discarded only when the available evidence supports doing so, while unresolved items are retained by default to prioritize recall.

The method is evaluated against seven recent schema-linking or Text-to-SQL systems on BIRD-Dev and an expanded Spider2-Snow evaluation set under shared backbones and fixed downstream SQL generators. On Spider2-Snow, EviLink achieves 93.04% field-level strict recall with 116.55K average schema-linking tokens. It also obtains the highest reported execution accuracy on both benchmarks under the fixed generators, although its margins over the strongest baseline are small.

Critical strengths
  1. The multi-path formulation is a meaningful and practically relevant view of schema linking. Treating schema requirements as path-dependent directly addresses a failure mode of early pruning: a field that looks unnecessary under one SQL plan may be essential under another valid join or filtering strategy. The required/uncertain/candidate decomposition is simple, understandable, and can be incorporated into an agentic Text-to-SQL pipeline without changing the downstream generator.

  2. The evaluation is broad. The authors compare against seven recent methods, hold the backbone and downstream generator fixed within each benchmark, report both retrieval quality and token use, and include ablations, schema-size analysis, an official-SQL subset, additional backbones, prompts, and annotation details. The strongest result is on medium and large Spider2-Snow schemas, where EviLink substantially improves strict field recall while using far fewer linking tokens than APEX-SQL.

Critical weaknesses
  1. The experiments do not establish that uncertainty-guided refinement is a major source of the improvement. The ablation indicates that multi-hypothesis grounding accounts for most of the recall gain. On Spider2-Snow with Qwen3.8-27B, adding UGAR changes SRR from 74.47% to 74.66% and NSF from 62.47% to 63.32%, while increasing the token count from 79.43K to 91.36K. In the GLM-5.1 main run, refinement reduces SRR from 94.00% to 93.04%, losing complete recall on six examples and gaining it on only one. Since “uncertainty-guided evidence acquisition” is central to the paper's title and framing, the current evidence supports it mainly as a modest precision-oriented cleanup stage rather than as a consistently beneficial uncertainty-resolution mechanism. Tool-use frequencies, evidence-level transitions, and a cost-matched refinement baseline would make this part of the claim substantially more convincing.

  2. The main downstream and Spider2-Snow results need a stronger robustness check. The execution-accuracy gains over APEX-SQL are 0.46 points on BIRD-Dev (seven examples) and 0.58 points on Spider2-Snow (three examples), and the official-SQL subset differs by only one example. These results come from a stochastic pipeline, but no repeated runs, confidence intervals, or paired significance tests are reported. In addition, 407 of the 517 Spider2-Snow reference queries were constructed with model assistance, while only 44 examples received documented human SQL review. The authors are commendably explicit about this limitation and provide a 95-example unchanged-official-SQL subset, but that subset is small and shows the same narrow downstream margin. This leaves uncertainty about the magnitude and stability of the end-to-end advantage.

  3. The novelty is limited. The core components of EviLink are already well-established. Multi-hypothesis generation is not new: Relying on LLMs to generate multiple SQL drafts to assist in schema extraction has been extensively explored. Utilizing LLM agents equipped with tools to acquire evidence, such as querying data distributions or executing probe SQLs, is already the central mechanism in recent state-of-the-art works like AutoLink. The primary reframing is simply a voting rule that partitions schema elements into three categories and selectively routes only the "uncertain" ones to the agent, which is also a basic routing trick.

Questions for authors
  1. Can the authors report paired confidence intervals or a paired significance test, such as bootstrap resampling or McNemar’s test, for the execution-accuracy differences in Table 3? Since the reported gains correspond to only 7 additional correct examples on BIRD-Dev and 3 on Spider2-Snow, results across multiple generation seeds would also help distinguish test-set uncertainty from stochastic variation in the generation pipeline.

  2. How often does UGAR call get_field_stats, probe_sql, list_tables, and list_columns, and how are its refinement tokens distributed across evidence levels? Could the authors also compare UGAR with a refinement policy that has access to the same tools and a comparable token budget but does not use the cross-hypothesis uncertainty partition? This would help isolate the contribution of uncertainty guidance from the effects of additional evidence and agentic refinement more generally.

  3. In the GLM-5.1 results in Table 15, refinement reduces field-level SRR from 94.00% to 93.04%, losing complete recall on six examples while recovering it on only one. What evidence and discard rationales led the agent to remove the missing gold fields, and what were their cross-hypothesis support counts or credibility scores? Could a threshold adjustment or an additional safeguard have prevented these failures?

  4. How sensitive are the Spider2-Snow conclusions to the choice of reference SQL? For a sample of cases where EviLink retains fields outside the constructed reference SQL, could independent human review determine whether these fields support a valid alternative SQL path or reflect ordinary over-retrieval?

4 comments
AuthorsOpenReview ↗

Rebuttal - Part I

We sincerely thank you for recognizing path-dependent schema needs, evaluation coverage, and recall–cost performance on medium and large schemas. Your questions helped us assess refinement gains and failures. We respond below with experiments and case analyses.

1. UGAR's contribution and the same-tool agent comparison

The experiments do not establish that uncertainty-guided refinement is a major source of the improvement.

Could the authors also compare UGAR with a refinement policy that has access to the same tools and a comparable token budget but does not use the cross-hypothesis uncertainty partition?

We thank you for suggesting this control. Both agents use the paper’s Qwen3.8, tools, and budgets in a single run on all 517 examples, sharing initial grounding outputs; Generic omits uncertainty partitioning and path protection.

AgentComplete recallSRRNSRNSPNSFRefinement tokensLinking tokens
Guided392/51775.8292.6656.5765.4911.07K99.05K
Generic389/51775.2492.4656.4765.3212.73K100.71K

Both agents share the same turn limits. Compared with the generic agent, guidance reduces evidence-tool calls from 567 to 424, model calls from 952 to 750, and refinement tokens by 13.05%, with small observed gains in all four schema metrics. Together, these results show improved schema quality at lower exploration cost. We will add this experiment and analysis to the revised manuscript.

2. Tool-use frequencies and evidence accounting

How often does UGAR call get_field_stats, probe_sql, list_tables, and list_columns, and how are its refinement tokens distributed across evidence levels?

Tool calls across all 1534 BIRD and 517 Spider2 examples:

ToolBIRD Qwen3.6-35B-A3B (1534)Spider2 GLM-5.1 (517)
get_field_stats1251
probe_sql370350
list_tables355
list_columns85126
keep312179
discard9102373

Refinement starts with L2 and requests L3 statistics as needed. We report examples receiving nonempty L3 statistics and token shares of model calls before/after receipt, counting all input and output tokens:

Evidence use and token shareBIRD Qwen3.6-35B-A3BSpider2 GLM-5.1
Acquired L3 / agent-active examples10/60231/261
Model-call tokens before L3 receipt97.32%88.56%
Model-call tokens after L3 receipt2.68%11.44%

3. The six losses of complete recall

In the GLM-5.1 results in Table 15, refinement reduces field-level SRR from 94.00% to 93.04%, losing complete recall on six examples while recovering it on only one. What evidence and discard rationales led the agent to remove the missing gold fields, and what were their cross-hypothesis support counts or credibility scores? Could a threshold adjustment or an additional safeguard have prevented these failures?

In the submitted run, Grounding retrieved all eight fields lost across these six cases. Their scores were below the 0.85 locking threshold, so UGAR could inspect and prune them. The table summarizes logged discard reasons and our error analysis:

ExampleRemoved reference field(s)Support n/MCredibilityDiscard rationale and error analysis
sf_bq023INDIV20.entity_tp3/40.8125Treated as filing metadata rather than a contribution filter
sf_bq349PLANET_FEATURES.osm_timestamp2/40.5000Treated as editing metadata
sf_bq349PLANET_NODES.latitude, longitude3/4 each0.8125 eachPreferred an available geometry representation
sf_bq309POSTS_QUESTIONS.title1/20.5000Focused on body length and omitted the title projection
sf_bq182YEAR._2023.created_at2/30.6875Preferred a date-specific DAY table over the YEAR path
sf_bq159MC3_MAF_V5_ONE_PER_TUMOR_SAMPLE.FILTER1/20.5000Confused mutation classification with quality-control reliability
sf_bq325LOCUS2GENE.pos3/40.8125Considered filtering/ranking needs but missed a join role

The paper uses 0.85 as a configurable default, without an extensive search for an optimal threshold. Lowering it to 0.8125 would protect four of the eight fields. With the recorded scores, reducing it from 0.85 to 0.8125 would lock entity_tp, latitude, longitude, and pos (3/4 support), blocking their deletion; the other four remain below it.

The cases also expose errors in interpreting field roles. For example, FILTER = PASS identifies records passing quality checks; Variant_Classification describes mutation types. The agent mistook mutation type for reliability and deleted FILTER. These cases suggest checking filtering, projection, and join roles across candidate paths before deletion; other available fields or paths alone do not justify removing a field.

AuthorsOpenReview ↗

Rebuttal - Part II

4. Statistical strength of downstream improvements

Can the authors report paired confidence intervals or a paired significance test, such as bootstrap resampling or McNemar’s test, for the execution-accuracy differences in Table 3?

Our main experiments use single runs due to multi-seed cost. For updated results against all seven baselines, we report paired 95% bootstrap CIs (100,000 resamples) and exact McNemar tests with Holm correction across 14 comparisons.

BaselineBIRD ΔEX [95% CI], ppp (Holm)Spider2 ΔEX [95% CI], ppp (Holm)
TA-SQL+3.59 [+1.89, +5.28]0.0005+8.32 [+4.45, +12.19]0.0005
RSL-SQL+0.98 [-0.39, +2.35]0.9183+1.93 [-1.55, +5.42]1.0000
ReFoRCE+37.09 [+34.42, +39.77]<0.0001+6.00 [+2.13, +9.86]0.0287
LinkAlign+1.89 [+0.33, +3.46]0.1236+7.93 [+4.45, +11.41]0.0003
AutoLink+0.91 [-0.52, +2.35]0.9868+5.80 [+2.13, +9.48]0.0287
DSR-SQL+2.22 [+0.59, +3.85]0.0715+5.61 [+2.13, +9.09]0.0279
APEX-SQL+0.65 [-0.65, +1.96]1.0000+0.19 [-3.48, +3.87]1.0000

EX gains remain significant after correction against TA-SQL and ReFoRCE on BIRD, and against TA-SQL, ReFoRCE, LinkAlign, AutoLink, and DSR-SQL on Spider2. Against APEX-SQL, EX gains are not significant, while linking tokens fall by 44.91% and 78.91%, respectively. This observed Pareto improvement matters for enterprise data agents: with downstream quality maintained, lower inference cost and fewer calls may matter more than small schema-metric gains, helping control computation and user waiting time.

5. Constructed references and alternative valid paths

How sensitive are the Spider2-Snow conclusions to the choice of reference SQL? For a sample of cases where EviLink retains fields outside the constructed reference SQL, could independent human review determine whether these fields support a valid alternative SQL path or reflect ordinary over-retrieval?

Valid SQL paths can use different fields: single-reference precision may penalize alternatives, and SRR may fall when reference fields are omitted. We have not quantified sensitivity across equivalent reference SQLs; systematic evaluation is beyond scope.

In Table 3’s downstream evaluation, we fix the SQL generator and compare execution results with benchmark reference CSVs. This scoring does not depend on our constructed gold SQL and accepts different SQL implementations producing correct results.

We examined extra-field uses in EviLink’s correct cases within the official-SQL subset. These diagnostic examples illustrate valid alternatives and over-retrieval; they are not an independent human review:

CaseUse of an extra fieldFinding
sf_bq252reads sample_repo_name, avoiding the official file-table joinvalid alternative
sf_bq263gets cost through inventory_item_id rather than the official product_id pathvalid alternative
sf_local354uses DRIVES.first_round/last_round rather than deriving them from RESULTS + RACESvalid alternative
sf_bq264carries USERS.id into a CTE without using itordinary over-retrieval

These cases show that some extra fields support correctly executed alternative SQL paths, while others are unused and constitute over-retrieval. Thus, lower precision against a single reference SQL cannot be attributed entirely to irrelevant-field retrieval.

6. Contribution relative to multi-sampling and tool agents

The primary reframing is simply a voting rule that partitions schema elements into three categories and selectively routes only the "uncertain" ones to the agent, which is also a basic routing trick.

We respectfully disagree that this is merely a routing trick. The initially most plausible SQL path may not yield a correct solution. Inspired by experienced SQL engineers (Figure 1), EviLink preserves alternatives until evidence supports pruning. Low path support can indicate a useful alternative, not irrelevance.

In Figure 2, H1's status-based path misses authoredOn and reasonReference, which H3 retrieves. Selecting only H1 would exclude H3's required fields; if this alternative is needed, the linked schema would be insufficient. APEX-SQL merges drafts into one master plan before grounding. EviLink instead grounds each path separately and preserves field-to-path membership, so alternative requirements remain identifiable during refinement.

UGAR focuses L2 inspection and on-demand L3 statistics or SQL probes on uncertain needs, retaining unresolved fields (§3.2). In Rebuttal - Part I’s same-tool control, guidance cuts refinement tokens by 13.05%, with small observed gains in all four schema metrics. Rich fixed evidence is costly: Table 2’s Spider2 Static L2/L3 use 522.88K/1997.60K linking tokens versus Full’s 94.58K.

Thank you for the detailed rebuttal and additional experiments. I acknowledge that the authors have addressed several of my experimental concerns, and I appreciate the effort.

However, my main concern regarding the novelty of this work remains unresolved.

Specifically, generating multiple SQL hypotheses and using them to identify relevant schema elements have been extensively explored in prior work, including PET-SQL, SQL-to-Schema, RSL-SQL, and Dynamic-SQL. Similarly, agentic schema exploration through database inspection, evidence acquisition, and SQL-based verification is already well established, particularly in AutoLink. EviLink's main additional contribution is to preserve hypothesis-level schema support, partition schema items into Required, Uncertain, and Candidate sets, and use this partition to guide subsequent agentic refinement.

I acknowledge that this organization is reasonable and that the resulting system provides empirical improvements. However, the distinction from existing approaches, including preserving field-to-path membership rather than consolidating hypotheses into a unified plan, does not by itself establish a sufficiently substantial conceptual or methodological contribution. The proposed support-based partitioning and uncertainty-guided routing are useful design choices, but remain relatively straightforward combinations of existing techniques.

The experiments demonstrate that the proposed combination works; they do not fundamentally change the nature or originality of the underlying ideas. My concern is therefore not primarily about experimental soundness or implementation quality, but about whether the methodological advance is sufficiently substantial for ICLR.

Given that my central concern about the incremental nature of the contribution remains, I will maintain my original rating.

Thank you for your quick response and for acknowledging that our additional experiments addressed several of your concerns. This is encouraging to us.

We see EviLink’s core contribution as making schema relevance conditional on alternative SQL implementations: a field unnecessary for one path may be essential for another. EviLink carries these distinct requirements through evidence-guided refinement, allowing schema pruning to account for viable alternatives.

Thank you again for your time and thoughtful feedback.

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

This paper addresses schema linking for large-scale Text-to-SQL, arguing that single-path schema linkers prematurely prune tables and columns needed by alternative valid SQL realizations. The authors propose EviLink, a two-stage framework (Section 3, Figure 2):

  1. Stage 1 (Multi-Hypothesis Grounding, Section 3.1, Eqs. (1)–(5)): An LLM generates M≤KM \le K (K=4K=4) distinct SQL realization hypotheses H={hm}m=1M\mathcal{H} = \lbrace h _ m \rbrace _ {m=1}^M, grounds table and column selections per hypothesis, aggregates support counts nxn _ x (Eq. (2)) into a Beta–Binomial credibility score cx=Pr⁡(px>0.5∣nx,M)c _ x = \Pr(p _ x \gt 0.5 \mid n _ x, M) (Eqs. (4)–(5)), and partitions elements at τreq=0.85\tau _ {\mathrm{req}} = 0.85 into a locked required set Sreq\mathcal{S} _ {\mathrm{req}}, an uncertain set Sunc\mathcal{S} _ {\mathrm{unc}}, and a remaining candidate pool Scand\mathcal{S} _ {\mathrm{cand}}.
  2. Stage 2 (Uncertainty-Guided Agentic Refinement / UGAR, Section 3.2, Eqs. (6)–(7)): An agentic loop inspects Sunc\mathcal{S} _ {\mathrm{unc}} and Scand\mathcal{S} _ {\mathrm{cand}} using progressive L0–L3 schema profiles (Table 8) and read-only tools (get_evidence, expand_Field, probe_sql, confirm, discard), retaining unresolved items in Sunc\mathcal{S} _ {\mathrm{unc}} by default (Eq. (7)).

EviLink is evaluated on BIRD-Dev (1,534 instances) and a 517-instance Spider2-Snow subset (110 released SQLs and 407 GPT-5.6-constructed SQLs, Appendix C Table 5) against 8 pipeline and 3 agentic baselines (Tables 1–3, Appendices D–I).

Critical strengths
  1. Strong strict-recall (SRR) and linking-token trade-off on large schemas. On the 517-instance Spider2-Snow benchmark (averaging 61.3 tables and 812.9 columns per database, Table 4), EviLink with GLM-5.1 achieves 93.04% field-level SRR using 116.55K linking tokens, outperforming APEX-SQL (84.14% SRR at 231.04K tokens, Table 1) with particularly strong gains on the Large (>200\gt 200 cols) stratum (88.10% vs. 73.81% SRR, Table 12).

  2. Controlled downstream SQL generation protocol and runnable supplementary code. Using a fixed downstream SQL generator (Qwen3.6-27B / GLM-5.1, Section 4.3 Table 3) helps isolate the downstream utility of the linked schema context, and the supplementary package provides clean, runnable Stage 1–2 code alongside all 1,534 BIRD-Dev and 517 Spider2-Snow gold schema/SQL annotations.

Critical weaknesses
  1. Undisclosed "safety_policy": "guarded" pre-confirmation, table-discard blocking, and benchmark-specific hardcoded heuristics in Stage 2 (UGAR), plus degenerate Beta credibility in Stage 1.

    • Unconditional table locking and fast_skip in Stage 2: Both src/bird/config/method.json:85 and src/spider2/config/method.json:85 run UGAR under "safety_policy": "guarded". Before Turn 1 of the UGAR LLM agent (src/bird/agent/ugar/orchestrator.py:581-600, src/spider2/agent/ugar/orchestrator.py:587-589), every uncertain table in Sunc\mathcal{S} _ {\mathrm{unc}} is unconditionally moved to confirmed_keep, and src/bird/agent/ugar/tools/discard.py:198-207 blocks discarding any table ("reason": "table_discard_blocked"). This explains why Table 15 (p. 25) reports 0 (0.00%) tables removed for both GLM-5.1 and Qwen3.8-27B, and why Full has lower Table NSP than w/o UGAR in Table 11 (77.78% vs. 77.85% on BIRD; 78.88% vs. 79.08% on Spider2). Furthermore, every uncertain field matching any of 5 deterministic rule tags (retrieval_unit, connector, surface_aligned, evidence_aligned, operator_aligned in field_role_tagger.py:222-304) is pre-confirmed and blocked from discard (discard.py:208-221). Because this pre-confirmation empties working_set before Turn 1 on most instances, orchestrator.py:720-761 triggers fast_skip=True with zero LLM calls in Stage 2, explaining why 65.76% (340/517) of GLM-5.1 and 74.27% (384/517) of Qwen3.8-27B instances in Table 15 are untouched by UGAR, and why UGAR actually reduces Spider2-Snow field SRR on the primary GLM-5.1 model from 94.00% (486/517) to 93.04% (481/517).
    • Evaluation-set-specific heuristics: src/bird/agent/ugar/field_role_tagger.py (lines 117, 130-136, 483-592, 1605-1872) hardcodes domain- and benchmark-specific rules: e.g., mapping "profit" / "margin" in the question to protect columns matching {"cost", "price", "quantity", ..., "wholesale"}—which directly drives the highlighted sf_bq273 profit-margin case study in Appendix J (p. 25, lines 1346–1348)—alongside rules for "unemployment" + "labor"/"force", genomic "variant"/"allele" →\to {"alt", "chrom", "pos", "ref"}, GitHub _YYYYMM tables, BIRD european_football_2 home_player_1..11 columns, Spider2 NOAA hurricane wind_radii_r34_ne..nw columns, and ChEMBL versioned tables. Moreover, src/bird/pipeline/evidence_anchor.py:1-569 and two_stage_pipeline.py:121-203 explicitly cite specific BIRD-Dev evaluation instance IDs (iid=864, 93, 142, 689, 51, 189) and state that the deterministic FK-closure rule is "recovering the gold join key."
    • Degeneracy of Beta credibility for M≤4M \le 4: Under Eqs. (4)–(5) (cx=1−FBeta(nx+1,M−nx+1)(0.5)c _ x = 1 - F _ {\mathrm{Beta}(n _ x + 1, M - n _ x + 1)}(0.5)) with τreq=0.85\tau _ {\mathrm{req}} = 0.85 and M≤4M \le 4, Sreq\mathcal{S} _ {\mathrm{req}} reduces to a strict unanimity rule (nx=Mn _ x = M): for M=4M=4, nx=4  ⟹  cx=31/32=0.96875≥0.85n _ x = 4 \implies c _ x = 31/32 = 0.96875 \ge 0.85, whereas nx=3  ⟹  cx=26/32=0.8125<0.85n _ x = 3 \implies c _ x = 26/32 = 0.8125 \lt 0.85. No matched-budget repeated single-path union/voting baseline is tested to separate multi-hypothesis prompting from simple 4-sample ensembling.
  2. Confounded ablations (w/o Multi, Static L0–L3), Static L2/L3 outperforming Full, asymmetric model matrices, and fragile downstream EX gains.

    • Confounded ablations & Static L2/L3 superiority: Setting M=1M=1 in w/o Multi (bucketing.py:228-231, Section 3.1 lines 236–237) places all Stage-1 items into Sreq\mathcal{S} _ {\mathrm{req}} and sets Sunc=∅\mathcal{S} _ {\mathrm{unc}} = \emptyset, simultaneously disabling UGAR pruning. In Static L0–L3, stage_executor.py:455-463 hardcodes Stage 1.A table selection to L0, while field_role_tagger.py:322-377 unconditionally reads level=2 (L2) metadata during Stage 2 pre-confirmation even under Static L0/L1. Furthermore, in Table 2 (p. 8) on Spider2-Snow, Static L3 beats Full on field SRR (71.37% vs. 71.18%) and NSR (88.14% vs. 87.98%), while Static L2 beats Full on NSP (71.67% vs. 70.10%) and NSF (74.99% vs. 74.38%); in Table 11 (p. 21), both Static L2 and Static L3 outperform Full on Spider2-Snow table SRR and NSF.
    • Asymmetric model evaluation and omitted Table 3 baselines: In Tables 1 and 10, all baselines on BIRD-Dev are evaluated only with Qwen3.5-27B (none with GLM-5.1), whereas all baselines on Spider2-Snow are evaluated only with GLM-5.1 (none with Qwen3.5-27B). In Table 3 (p. 9) and Table 7 (p. 18), Spider2-Snow downstream SQL evaluation omits all 8 pipeline baselines—even though CHESS achieved 42.55% field SRR in Table 1 (47.37% in Table 6), nearly 3×3\times higher than DB-Copilot (14.51%).
    • Small single-seed EX margins and baseline framing: Despite stochastic sampling (temperature = 1.0 in linking, 0.7 in SQL generation), no multi-seed variance or paired significance tests are reported; EviLink improves over APEX-SQL by only 7 queries on BIRD-Dev (1016 vs. 1009), 3 queries on Spider2-Snow (202 vs. 199), and 1 query on the 95-example unchanged official SQL subset (41 vs. 40, Table 5). Finally, the Abstract's -27.2% and -71.3% generation token reductions are measured against Origin Schema rather than against the best baseline (MCS-SQL uses 1.61K generation tokens on BIRD-Dev vs. 1.63K for EviLink).
Questions for authors
  1. Ablation of undisclosed deterministic heuristics and guarded pre-confirmation (decisive for soundness): What are the results in Tables 1–3 when (a) "safety_policy": "guarded" pre-confirmation and table-discard blocking (orchestrator.py:581-600, discard.py:198-221), (b) the domain- and benchmark-specific rules in field_role_tagger.py (including the "profit" rule used in the sf_bq273 case study, "unemployment", genomic "variant", GitHub _YYYYMM, european_football_2, NOAA hurricane, and ChEMBL patterns), and (c) BIRD-specific evidence_anchor.py and deterministic FK closure (two_stage_pipeline.py:121-203) are disabled? Were these rules developed after inspecting errors on BIRD-Dev (iid=864, 93, 142, 689, 51, 189) and the 517 Spider2-Snow instances?

  2. Matched-budget single-path ensemble control and unconfounded ablations: Since Eqs. (4)–(5) with τreq=0.85\tau _ {\mathrm{req}} = 0.85 and M≤4M \le 4 reduce to unanimity (nx=Mn _ x = M), how does EviLink compare against running the Stage-1 single-path schema linker independently 4 times (M=4M=4) at temperature = 1.0 and applying simple union or unanimity/majority voting? In addition, what is the performance of w/o Multi when Stage-1 selections are placed into Sunc\mathcal{S} _ {\mathrm{unc}} (so UGAR can still refine them) rather than locked into Sreq\mathcal{S} _ {\mathrm{req}}?

  3. Completing the Table 3 pipeline baselines on Spider2-Snow and multi-seed statistical testing: Why were all 8 pipeline baselines (especially CHESS, which achieved 42.55% field SRR in Table 1) omitted from Spider2-Snow downstream SQL generation in Table 3 and Table 7, and why was EviLink (GLM-5.1) omitted from BIRD-Dev in Table 3? Please provide these missing rows alongside multi-seed means, standard deviations, and paired McNemar tests for the 7-query (BIRD) and 3-query (Spider2) EX differences over APEX-SQL.

3 comments
AuthorsOpenReview ↗

Rebuttal - Part I

We sincerely thank you for recognizing EviLink’s combination of high recall and lower linking cost on large schemas, as well as the evaluation design that holds the downstream SQL generator fixed. Your concerns about the implementation and evaluation prompted us to examine the relevant rules and conduct additional experiments, strengthening the empirical support for our conclusions.

Clarifications on the experimental setup and results

1. Coverage of the downstream baseline evaluation

In Table 3 (p. 9) and Table 7 (p. 18), Spider2-Snow downstream SQL evaluation omits all 8 pipeline baselines—even though CHESS achieved 42.55% field SRR in Table 1 (47.37% in Table 6), […] higher than DB-Copilot (14.51%).

The original Tables 3 and 7 already report downstream results for all seven evaluated baselines (§4.1), reproduced below:

MethodTable 3: Spider2-Snow correct / 517Table 7: official-SQL subset correct / 95
TA-SQL15728
RSL-SQL19035
ReFoRCE16936
LinkAlign15932
AutoLink17031
DSR-SQL17134
APEX-SQL19940
EviLink20241

Tables 1 and 6 report no CHESS results. CHESS is cited as related work; neither it nor DB-Copilot is an evaluated baseline.

2. APEX-SQL results in Table 1

outperforming APEX-SQL (84.14% SRR at 231.04K tokens, Table 1)

Table 1 reports 80.85% field SRR and 580.80K linking tokens for APEX-SQL on Spider2-Snow. The corresponding EviLink results are 93.04% and 116.55K.

3. Model configurations across benchmarks

Asymmetric model evaluation and omitted Table 3 baselines: In Tables 1 and 10, all baselines on BIRD-Dev are evaluated only with Qwen3.5-27B (none with GLM-5.1), whereas all baselines on Spider2-Snow are evaluated only with GLM-5.1 (none with Qwen3.5-27B).

The main linking backbone is Qwen3.6-35B-A3B on BIRD-Dev and GLM-5.1 on Spider2-Snow (§4.1). The linking and downstream generation configurations are listed separately in Table 8:

BenchmarkMain schema-linking backboneFixed downstream SQL generator
BIRD-DevQwen3.6-35B-A3BQwen3.8-27B
Spider2-SnowGLM-5.1Qwen3.6-27B

Within each benchmark, EviLink and all seven baselines share the same linking backbone and fixed downstream SQL generator. Different backbones were chosen across benchmarks for cost reasons; each within-benchmark comparison uses common model conditions.

The submitted Table 13 also evaluates EviLink with four additional backbones—GLM-5.2, GLM-5.3 Low, DeepSeek-4 Pro, and Qwen3.8-27B—on the same 517 Spider2-Snow examples. These experiments show that EviLink maintains high schema recall across different backbones, supporting its cross-model generalizability.

4. Static evidence-level control

stage_executor.py:455-463 hardcodes Stage 1.A table selection to L0, while field_role_tagger.py:322-377 unconditionally reads level=2 (L2) metadata during Stage 2 pre-confirmation even under Static L0/L1.

L0 is the table-selection default. Static experiments use the selected level (L0, L1, L2, or L3) for table selection, field selection, and refinement. Pre-confirmation reads schema objects built at that level, without separately loading L2 metadata under Static L0/L1. The original experiment configurations and logs confirm these settings.

5. Reading the ablation comparisons

Static L3 beats Full on field SRR (71.37% vs. 71.18%) and NSR (88.14% vs. 87.98%), while Static L2 beats Full on NSP (71.67% vs. 70.10%) and NSF (74.99% vs. 74.38%)

The submitted Table 2 reports:

Spider2-Snow configurationField SRRField NSRField NSPField NSFLinking tokens
Full74.66%91.54%54.58%63.32%91.36K
Static L278.34%92.47%54.59%63.03%521.07K
Static L374.08%92.01%55.99%63.62%1991.14K

Static L3 has lower SRR than Full; Static L2 has lower NSF. Relative to Full, Static L2 raises SRR, and Static L3 raises NSR and NSF, at substantially higher linking cost (§4.3).

in Table 11 (p. 21), both Static L2 and Static L3 outperform Full on Spider2-Snow table SRR and NSF.

In the original submission, Static L2 and L3 have higher table SRR, but lower table NSF: 81.29% and 79.64%, versus 82.56% for Full (Table 11).

6. Scope of the token-cost comparison

the Abstract's -27.2% and -71.3% generation token reductions are measured against Origin Schema rather than against the best baseline (MCS-SQL uses 1.61K generation tokens on BIRD-Dev vs. 1.63K for EviLink).

The submitted abstract contains neither these percentages nor a comparison with Origin Schema. Table 3 reports 2.83K BIRD generation tokens for EviLink; MCS-SQL is not included in that table. The manuscript explicitly measures schema-linking token reductions relative to APEX-SQL.

AuthorsOpenReview ↗

Rebuttal - Part II

Responses to methodological and experimental concerns

7. Domain rules, recall safeguards, and reevaluation

Undisclosed "safety_policy": "guarded" pre-confirmation, table-discard blocking, and benchmark-specific hardcoded heuristics in Stage 2 (UGAR) […]

We thank you for raising generalization concerns. During development, we examined BIRD and Spider2 cases to understand the limitations of existing schema-linking approaches. Following your feedback, we reviewed the safeguards and their scope.

Removed: domain-specific protection/recovery for profit, unemployment, genomic variants, GitHub partitions, football lineups, and hurricane radii; automatic version/literal recovery including ChEMBL; broad bare-measure and geographic protection. Lexical surface_aligned tags no longer force retention.

Retained: pre-confirmation of keys, identifiers, and query-aligned fields, table-discard blocking, and protection of at least one intact candidate path to reduce the risk of deleting join information or breaking a viable SQL path. Task-provided evidence highlights field clues, and declared FKs guide join-field completion. These are general recall safeguards based on task inputs and schema structure.

Similar recall safeguards appear in baseline code: APEX-SQL retains columns whose names or descriptions contain id/code/name/type when kept_num < 200 [1]; AutoLink adds columns whose names contain id/name/code from candidate tables and retains them in the final merge [2]; LinkAlign merges a pre-sampled reserved subset back after model filtering [3]. These are different implementations of field retention or expansion alongside model decisions.

We reevaluated the affected examples and updated the full-benchmark results for the revised implementation:

ConfigurationNSubmitted SRR → current SRRSubmitted NSR → current NSRSubmitted NSP → current NSPSubmitted NSF → current NSFLinking tokens: submitted → current
BIRD, Qwen3.6-35B-A3B153493.22% → 92.89%98.05% → 97.95%51.58% → 51.78%63.60% → 63.78%25.68K → 28.07K
Spider2, GLM-5.151793.04% → 93.42%97.86% → 97.95%26.04% → 26.23%36.65% → 36.90%116.55K → 122.48K
BIRD, Qwen3.8-27B153492.57% → 92.37%97.82% → 97.75%56.13% → 56.31%67.77% → 67.93%28.16K → 28.80K
Spider2, Qwen3.8-27B51774.66% → 74.27%91.54% → 91.50%54.58% → 54.70%63.32% → 63.34%91.36K → 94.58K
BenchmarkSubmitted correctCurrent correct
BIRD1016/15341019/1534
Spider2202/517200/517

Across all four configurations, SRR changes by at most 0.39 percentage points and NSF improves. Rankings against all seven baselines remain unchanged for Table 1 field metrics and linking cost, and Table 3 downstream EX. EviLink retains the highest Spider2 field SRR and the highest EX on both benchmarks, while using 44.91% and 78.91% fewer linking tokens than APEX-SQL on BIRD and Spider2, respectively. We will update the corresponding values in the revised manuscript.

triggers fast_skip=True with zero LLM calls in Stage 2, explaining why 65.76% (340/517) of GLM-5.1 and 74.27% (384/517) of Qwen3.8-27B instances in Table 15 are untouched by UGAR

fast_skip avoids unnecessary LLM calls: after locking and protective pre-confirmation, it returns the current schema when the refinement working set is empty and candidate recovery is not triggered. This reserves agent computation for cases requiring inspection or recovery.

Table 15 reports whether the final schema changed. Agent execution is a separate quantity, as the original traces show:

Original submitted runUnchanged final schemaRecorded fast_skipUnchanged schema with recorded LLM calls
GLM-5.1340/517259/51781/517
Qwen3.8-27B384/517252/517133/517

In 81 GLM-5.1 and 133 Qwen3.8-27B cases, UGAR made LLM calls but retained the same schema. These cases reflect inspection without a change to the output, rather than skipped refinement.

References

[1] APEX-SQL: field protection.

[2] AutoLink: field expansion; merge and entry point.

[3] LinkAlign: reserved-set retention.

AuthorsOpenReview ↗

Rebuttal - Part III

8. Beta support and the independent single-path ensemble control

reduces to a strict unanimity rule

At M = 2–4, 0.85 requires unanimity for locking, protecting shared schema needs from pruning. Partial support may reflect a valid alternative SQL path, so those fields remain for evidence-guided refinement. This separates protection from deletion: unresolved fields are retained, not dropped by a vote.

No matched-budget repeated single-path union/voting baseline is tested to separate multi-hypothesis prompting from simple 4-sample ensembling.

We evaluated four independent single-path samples with Qwen3.8-27B on all 517 Spider2-Snow examples, applying union, majority, and unanimity to the same outputs. We will add this control in the revised manuscript. The table reports actual costs; budgets are not matched:

Stage 1 methodSRRNSRNSPNSFStage 1 tokens / example
MHSG75.82%92.45%55.93%64.68%87.98K
Four single paths + union84.91%95.22%48.45%59.14%189.27K
Four single paths + majority58.99%84.34%71.92%74.11%189.27K
Four single paths + unanimity43.71%74.81%78.60%73.77%189.27K

MHSG offers a favorable balance of recall, precision, and cost. It uses 53.5% fewer Stage 1 tokens than all three ensembles. Relative to union, NSP rises from 48.45% to 55.93% and NSF from 59.14% to 64.68%, while NSR remains 92.45%. Relative to majority/unanimity, it retains higher recall at lower precision. These results support multi-hypothesis grounding as a cost-efficient way to preserve candidate SQL paths.

9. Single-path control with UGAR pruning enabled

In addition, what is the performance of w/o Multi when Stage-1 selections are placed into […] (so UGAR can still refine them) rather than locked into […]?

We compared locked and refinable selections on all 517 examples using exactly the Stage 1 outputs underlying revised Table 2 w/o Multi, with the same tools and budgets.

ConditionComplete recallSRRNSRNSPNSFStage 2 tokensLinking tokens
Locked (w/o Multi)325/51762.8686.9860.7066.896.14K44.18K
Refinable316/51761.1286.8461.5667.6411.96K50.00K

Enabling UGAR pruning introduces a trade-off rather than an overall improvement: NSP rises from 60.70% to 61.56% and NSF from 66.89% to 67.64%, while SRR falls from 62.86% to 61.12% and linking tokens increase by 13.2%. The refinable single-path SRR remains below Full’s 74.27% in updated Table 2. This supports the recall benefit of multi-path grounding: its advantage is not explained solely by locking single-path selections. We will add this control and analysis to the revised manuscript.

10. Downstream uncertainty, models, and comparison scope

Small single-seed EX margins and baseline framing: Despite stochastic sampling (temperature = 1.0 in linking, 0.7 in SQL generation), no multi-seed variance or paired significance tests are reported

Given the cost of repeating the full pipeline across seeds, our main experiments use a single-run setting. Using the updated evaluation results, we compare all seven baselines with paired 95% CIs from 100,000 question-level bootstrap resamples and exact McNemar tests; p-values use Holm correction across 14 comparisons.

BaselineBIRD ΔEX [95% CI], ppp (Holm)Spider2 ΔEX [95% CI], ppp (Holm)
TA-SQL+3.59 [+1.89, +5.28]0.0005+8.32 [+4.45, +12.19]0.0005
RSL-SQL+0.98 [-0.39, +2.35]0.9183+1.93 [-1.55, +5.42]1.0000
ReFoRCE+37.09 [+34.42, +39.77]<0.0001+6.00 [+2.13, +9.86]0.0287
LinkAlign+1.89 [+0.33, +3.46]0.1236+7.93 [+4.45, +11.41]0.0003
AutoLink+0.91 [-0.52, +2.35]0.9868+5.80 [+2.13, +9.48]0.0287
DSR-SQL+2.22 [+0.59, +3.85]0.0715+5.61 [+2.13, +9.09]0.0279
APEX-SQL+0.65 [-0.65, +1.96]1.0000+0.19 [-3.48, +3.87]1.0000

EX gains remain significant after correction against TA-SQL and ReFoRCE on BIRD, and against TA-SQL, ReFoRCE, LinkAlign, AutoLink, and DSR-SQL on Spider2. Against APEX-SQL, EX gains are not significant, while linking tokens fall by 44.91% and 78.91%, respectively. This observed Pareto improvement supports our emphasis on linking efficiency with strong downstream performance.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.