PaperlayerView exact revision on arXiv

Source-linked AI summary

Learned Enterprise Data Comprehension: Compression and Routing for Data Agents

Ethan Torres, Eric Mills

arXiv:2609.25286v1cs.AI

TL;DR

Enterprise data agents must expose relevant internal evidence, yet agents are often tasked with discovering data organization as well as reasoning over it. The paper introduces latent equivalence learning to route recurring query requirements through persistent identities to dataset-specific evidence realizations. On DAB, the architecture reaches 94.67% dataset-macro stratified Pass@1, versus 55.51% for the Claude Opus 4.6 reference agent.

  • Problem

    Enterprise data agents must expose relevant internal evidence, while agents are commonly asked both to discover an unfamiliar system's organization and reason over it.

    Source: §1 · p. 1

  • Method

    Latent equivalence learning separates persistent task-relevant identities from dataset-relative realizations and maps recurring query requirements to those identities for evidence routing.

    Source: §1 · p. 2, §3.4 · p. 7

  • Results

    94.67% dataset-macro stratified Pass@1, versus 55.51% for the Claude Opus 4.6 reference agent, on the Data Agent Benchmark.

    Source: §4.2 · p. 8

  • Takeaways & Limitations

    Across the heterogeneous DAB tasks studied, learning reusable structure upstream can improve data-agent performance while preserving separation between representation and query-specific reasoning.

    Source: §5 · p. 9

    The framework assumes a bounded identity inventory covers the deployment domain and sufficient evidence exists to realize and distinguish those identities.

    Source: §5 · p. 9

Abstract

from arXiv · show

Structured-data agents in enterprise settings must reason over complex data environments whose relevant evidence is distributed across schemas, relationships, policies, and recurring business roles. Modern agentic systems often address this burden through reusable markdown-style memory or skill files that preserve previously discovered information for later queries, reducing the need to rediscover the same structure repeatedly. This is useful, but it obscures a natural division of labor: agents are well suited to semantic reasoning, while learned systems are well suited to predicting and organizing recurring structure. We introduce latent equivalence learning to bridge this gap. The framework separates persistent task-relevant identities from their dataset-relative realizations. In our realization, supporting and opposing evidence shape support-realized Gaussian prototypes that learn how those identities are expressed in a particular data environment, while soft-membership profiles retain distinctions lost under a hard assignment. A separate learned query-prototype system represents recurring evidential requirements and maps them through a learned compatibility function into the same persistent identity structure. This identity-factorized, query-conditioned routing materializes the relevant dataset-specific evidence for downstream reasoning, allowing the agent to operate over an already organized evidential state rather than reconstructing cross-schema structure at every query. On the Data Agent Benchmark, spanning 54 queries across 12 heterogeneous datasets, our full implementation achieves 94.67% dataset-macro stratified Pass@1 over five complete trials and 258/270 successful raw query attempts, compared with 55.51% for the benchmark's Claude Opus 4.6 reference agent, ranking first among 40 leaderboard entries at submission.

1 INTRODUCTION

Enterprise data agents face the challenge of exposing relevant internal evidence, while neither agentic decomposition alone nor assigning all semantics to a learned predictor resolves the division-of-labor problem. The paper introduces latent equivalence learning to organize evidence around persistent task-relevant identities, then route recurring query requirements to dataset-specific realizations.

1 INTRODUCTION

  • Agents face a bottleneck in exposing relevant internal evidence, and adding more agents or reasoning steps can expand exploration without improving evidence organization.The paper frames the challenge in terms of computation, cost, and context.Source: §1 · p. 1
  • Latent equivalence learning organizes enterprise evidence around persistent task-relevant identities and routes recurring query requirements to their dataset-specific realizations.The framework separates persistent identities from dataset-relative realizations; Gaussian prototypes are one concrete realization, not a prescribed prototype family.Source: §1 · p. 2
  • A customer identifier can remain one persistent business role when represented as buyer id in one system and customer id in another.Observed columns, records, and relations provide supporting or opposing evidence from which each dataset's realization is learned.Source: §1 · p. 2
  • The modeled identity inventory is finite and scoped to a deployment domain, so genuinely new roles require extending it.The representation aims to absorb variation consistent with a shared identity while preserving distinctions the evidence indicates remain task-relevant.Source: §1 · p. 2
  • Support-realized Gaussian prototypes learn dataset-specific identity realizations, while separate query prototypes map recurring evidential requirements to persistent identity activations.A bounded view of supporting, opposing, and competing evidence is passed to the Arbiter and data agent for interpretation, verification, and computation.Source: §1 · p. 2
  • Soft-membership profiles preserve secondary distinctions lost under hard identity assignments, and the Gaussian realization supports a local assignment-stability bound.The bound characterizes when embedding perturbations preserve hard identity under a fixed prototype bank.Source: §1 · p. 2

2 RELATED WORK

Related work covers learned representations and prototypes, structured correspondence and weak supervision, and evidence routing for data agents. The paper distinguishes its focus as a learned interface that activates reusable evidence identities before downstream agent interpretation and execution.

2 RELATED WORK

  • Prototype learning provides ingredients for realizing persistent evidence identities, while the paper's local assignment certificate specializes an established margin-sensitivity relation rather than claiming that principle as new.The cited ingredients include support-derived prototypes and Gaussian prototypes with precision-weighted support and class-dependent distances.Source: §2 · p. 3
  • Structured correspondence and weak-supervision methods can contribute evidence about an environment, but they do not define the paper's persistent identities.The distinction is between a persistent identity and its dataset-relative realization, whose memberships and associated evidence remain available for query-conditioned use.Source: §2 · p. 3
  • Unlike retrieval and agent-workflow approaches described here, the paper focuses on a learned interface that maps query-prototype memberships to persistent evidence identities before agent execution.The downstream agent retains responsibility for semantic interpretation and execution.Source: §2 · pp. 3–4

3 LATENT EQUIVALENCE LEARNING

The framework separates shared task-relevant identities from their dataset-specific evidence, retaining soft memberships while learning Gaussian realizations and routing query requirements into identity coordinates.

3.1 PROBLEM SETUP AND SIGNED EVIDENCE

  • A finite identity set represents shared task-relevant roles, while evidence objects represent the dataset-specific structures through which those roles are expressed.Supporting and opposing evidence form an application-specific interface for assigning evidence to identities.Source: §3.1 · p. 4

3.2 SOFT MEMBERSHIP AND ASSIGNMENT FIBERS

  • Hard identity readouts partition evidence into dataset-specific fibers, but full soft-membership profiles preserve secondary memberships that hard assignments discard.The profile comparison is a pseudometric: zero distance means identical profiles and the same hard identity, but the converse need not hold.Source: §3.2 · p. 5

3.3 SUPPORT-REALIZED GAUSSIAN PROTOTYPES

  • A feature map and learned encoder place evidence in a metric space where support-derived Gaussian prototypes realize identities for each dataset.Precision-weighted support contributes to prototype construction, with interpolation retaining the persistent prior when dataset support varies.Source: §3.3 · pp. 5–6
  • Supporting evidence attracts embeddings toward corresponding prototypes, while opposing evidence penalizes proximity within a prescribed squared-distance margin.The training objective also includes an episodic classification loss and terms for soft-target consistency, calibration, and geometric regularization.Source: §3.3 · p. 6
  • With a fixed prototype bank, embeddings within the local stability radius retain their hard identity; schema changes or rebuilt prototypes require additional control.The certificate is specifically about embedding perturbations with the bank fixed.Source: §3.3 · p. 6

3.4 IDENTITY-FACTORIZED ROUTING AND AGENTIC COMPOSITION

  • A separate query-prototype space represents recurring evidential requirements and maps their soft profiles through learned compatibility into activations over persistent identities.Evidence membership and query demand use the same identity coordinates but have different meanings.Source: §3.4 · p. 7
  • Query-conditioned materialization selects a bounded, source-preserving evidence view, separating reusable query requirements from dataset-specific evidence realization.The Arbiter operates only on this bounded view and does not search the full static state.Source: §3.4 · p. 7

4 EXPERIMENTS

The experiments evaluate the learned representation and routing against an agentic Core system and the DAB reference across heterogeneous datasets. The full architecture improves the primary dataset-macro result over both comparisons.

4.1 EXPERIMENTAL SETUP

  • DAB spans 54 queries across 12 datasets, nine application domains, and four database systems, with each system evaluated over five complete trials.Each trial includes one attempt at each query; the primary metric weights datasets equally, while raw query-attempt accuracy weights all attempts equally.Source: §4.1 · p. 8
  • Core retains the Opus 5 agentic harness but omits learned equivalence representations; the proposed architecture adds prototype realizations and query-conditioned evidence construction.The corresponding PERMUTE EQ and PERMUTE CORE labels distinguish the full architecture and Core in Figure 2.Source: §4.1 · p. 8, Figure 2 · p. 9

4.2 RESULTS AND ABLATIONS

  • 94.67% dataset-macro stratified Pass@1, versus 84.13% for Core and 55.51% for the Claude Opus 4.6 reference agent.The architecture also succeeds on 258/270 raw query attempts (95.56%), compared with 236/270 (87.41%) for Core.Source: §4.2 · p. 8, Figure 2 · p. 9
  • The gain over Core is attributed to learned identity-based representation and query-conditioned evidence routing upstream of the downstream reasoner.The comparison tests the learned representation in a strong agentic setting.Source: §4.2 · p. 9

5 CONCLUSION

The conclusion presents upstream evidence organization as a practical way to improve data-agent performance across the heterogeneous DAB tasks studied. It bounds that conclusion by finite identity coverage and evidence constraints, and the paper reports AI-use and reproducibility details.

5 CONCLUSION

  • The framework assumes a finite identity inventory and sufficient evidence; unsupported identities, relationships, or distinctions may fall outside its modeled structure.Performance is also limited by evidence quality, routing, and the materialization budget, while substantial realization changes require separate control.Source: §5 · p. 9
  • Across heterogeneous DAB tasks, learning reusable evidence structure upstream can improve performance while leaving request-specific semantic interpretation and execution to language-model reasoning.The conclusion frames this as a design principle within the bounds of the studied tasks.Source: §5 · p. 9

5.1 AI USE STATEMENT

  • AI tools assisted with prose, notation, code development and debugging, ablation design, and presentation of graphs, figures, and diagrams.The authors state that proofs, implementation details, algorithmic constructions, code and diagram design, and architectural decisions are their own.Source: §5.1 · p. 10

5.2 REPRODUCIBILITY STATEMENT

  • The paper provides main-text architectural definitions, training objectives, and inference procedures, but does not disclose some implementation details or release proprietary source code.Complete mathematical proofs and additional derivations are in the appendix.Source: §5.2 · p. 10
  • DAB evaluations followed official rules and were externally validated without access to gold answers or benchmark-specific prompt cues.Training and evaluation used two local consumer-grade workstations without external compute clusters or cloud infrastructure.Source: §5.2 · p. 10

A.1 PER-DATASET PERFORMANCE

The per-dataset breakdown shows where aggregate DAB differences arise across configurations. The full Opus 5 realization is perfect on nine datasets and exceeds Core on five.

A.1 PER-DATASET PERFORMANCE

  • The full Opus 5 realization attains perfect Pass@1 on nine datasets, exceeds Core on five, and matches it on the other seven.The five datasets with higher Pass@1 are AG News, CRM Arena, Deps.dev, GitHub, and PanCancer.Source: Figure 3 · p. 12
  • Figure 3 reports Pass@1 percentages in cells for each of the 12 benchmark datasets and evaluated configurations.The breakdown makes visible where aggregate differences arise rather than treating DAB as a single homogeneous task family.Source: Figure 3 · p. 12, §A.1 · p. 12
  • The largest gaps over Core are GitHub (50% →100%), PanCancer (67% →100%), and Deps.dev (60% →90%).These per-dataset differences are descriptive; the aggregate evaluation remains the primary comparison in the main text.Source: §A.1 · p. 13

A.2 COST, RESOURCES, AND CONSISTENCY

The section characterizes the evaluated systems’ quality–cost operating points and resource use under conservative verification, while clarifying that these configurations were not jointly optimized. It also defines how repeated-trial consistency and fixed-prototype assumptions should be interpreted.

A.2 COST, RESOURCES, AND CONSISTENCY

  • In manually inspected traces, correct answers appeared after approximately 3–5 tool calls, followed by about 15 additional verification calls; this observation is qualitative, not benchmarkwide.The evaluation deliberately measured a high-reliability operating point rather than minimizing token use.Source: §A.2 · p. 13
  • 94.7% stratified Pass@1 at $82.82 is the highest measured quality, while Sonnet 5 reaches 82.9% at $41.10 and Core reports 84.1% at approximately $56.The quality–cost comparison covers one complete 54-query trial; the Core cost is approximate.Source: §A.2 · p. 13, Figure 4 · p. 13
  • The reported operating points were not jointly optimized over reasoner choice, verification depth, stopping policy, or token budget.Core’s approximate cost reflects missing cost fields in its accepted artifacts.Source: §A.2 · p. 13
  • Core averages 24.110 million total tokens per complete trial, comprising 23.455 million input and cache tokens and 0.655 million output tokens.The usage totals include conservative post-answer verification.Source: Figure 5 · p. 14
  • Figure 6 compares dataset-macro capability with consistency, where Pass3 requires success in all three sampled trials and is distinct from Pass@3.The diagonal represents no consistency gap; the supplied passages define the comparison but do not report the configurations’ outcomes.Source: Figure 6 · p. 15

B ASSIGNMENT STRUCTURE, ROUTING, AND STABILITY

B.1 ASSIGNMENT FIBERS AND SOFT MEMBERSHIP

The section defines hard assignment fibers as equivalence classes of evidence objects, then retains soft-membership profiles to preserve distinctions that hard assignments discard.

B.1 ASSIGNMENT FIBERS AND SOFT MEMBERSHIP

  • Hard assignment by identity partitions the evidence objects into nonempty fibers, with each fiber corresponding to one identity label.The quotient correspondence follows for any single-valued assignment map; the passage does not claim novelty for that construction.Source: §B.1 · p. 15
  • Prototype decision cells are obtained by comparing prototypes, and their evidence-space preimages contain the objects assigned to each identity.The passage distinguishes the identity label and an individual prototype from the set-valued equivalence class.Source: §B.1 · pp. 15–16
  • Soft membership profiles preserve differences between evidence objects that share a hard assignment, so later evidence materialization can distinguish their secondary memberships.The zero-distance classes of the membership pseudometric refine the hard-assignment fibers, and that refinement can be strict.Source: §B.1 · p. 16
  • The assignment and membership constructions are well-defined, but these properties do not guarantee that optimization attains every desired separation.An opposing-evidence term vanishes at its squared-distance margin and is positive below it only when its weight is positive.Source: §B.1 · p. 16

B.2 IDENTITY-FACTORIZED ROUTING

Query-side soft profiles map through a compatibility function to persistent identity activations, which a dataset-specific map uses to materialize evidence. Under fixed query-side objects, changing datasets changes materialization but not the query activation profile.

B.2 IDENTITY-FACTORIZED ROUTING

  • For the same query and fixed query-side model, replacing dataset D with D′ over the same identity inventory leaves γ(q) unchanged; only the dataset-specific materialization map changes.The factorization composes the compatibility map ΓM with query membership β, then applies the dataset-specific evidence-selection map.Source: §B.2 · p. 16
  • Dataset-invariant routing is an architectural separation property, not a guarantee that requested evidence exists or that answers are statistically sufficient.The conclusion also depends on the query encoder, compatibility map, identity inventory, and query-side inputs remaining dataset-independent.Source: §B.2 · p. 16
  • Routing uses the full query membership profile because queries with the same hard maximizer can produce different identity activations under the compatibility map.The supplied example has output odds exp(1/2) and exp(1/3) for two profiles sharing a hard maximizer.Source: §B.2 · p. 17

B.3 LOCAL STABILITY OF PROTOTYPE ASSIGNMENTS

The section gives a local certificate for preserving a prototype’s winning identity under bounded embedding perturbations, conditional on a fixed prototype bank. It distinguishes identity stability from semantic correctness and arbitrary schema invariance.

B.3 LOCAL STABILITY OF PROTOTYPE ASSIGNMENTS

  • An embedding perturbation smaller than the positive radius rD(e) preserves the unique winning identity when the prototype bank is fixed.The radius is based on the winning identity’s separation from every competitor.Source: §B.3 · p. 17
  • At a tie the argument gives no positive stability radius, and a runner-up-only sensitivity bound is insufficient unless all other competitors are controlled.With only one identity, the hard assignment is constant everywhere.Source: §B.3 · p. 17
  • The certificate preserves the selected identity, not semantic correctness or the full soft-membership vector, and does not establish invariance to arbitrary raw-schema transformations.Feature perturbations are covered only within a neighborhood where the embedding component has the stated Lipschitz bound; the entire prototype bank remains fixed.Source: §B.3 · p. 17

B.4 FROM SIGNED-EVIDENCE LOSSES TO A STABILITY MARGIN

The signed-evidence losses yield a conditional stability guarantee for a candidate identity under positive opposing coverage. When the prototype realization changes, preserving the winner additionally requires controlling embedding and prototype drift.

B.4 FROM SIGNED-EVIDENCE LOSSES TO A STABILITY MARGIN

  • If b_e > a_e, c_0 is the unique nearest identity with a positive distance margin over every competitor.The proof bounds the candidate distance by a_e and competitor distances below by b_e.Source: §B.4 · p. 18
  • Supporting loss controls proximity, opposing loss controls separation, and positive opposing coverage of every competitor is essential to this uniform guarantee.Training need not attain b_e > a_e, and the bound does not establish that the evidence endorses the semantically correct identity.Source: §B.4 · p. 18

B.5 WHEN THE PROTOTYPE REALIZATION ALSO CHANGES

  • When support and the prototype bank are recomputed, the fixed-bank certificate cannot be used unchanged; winner preservation requires bounds on embedding, mean, and precision changes.The extension bounds distance drift between the original and modified realizations.Source: §B.5 · p. 18
  • Arbitrary dataset changes are not assumed to satisfy the bounds required for preserving the original winner.The extension identifies additional quantities that must be controlled when support is recomputed.Source: §B.5 · p. 18

C SAMPLE EXECUTION TRACE

The execution trace contrasts two agents answering an LGG expression query: both perform the requested computation, but they select different histology groupings. Permute EQ reports the required ICD-O-3 groups, while Permute Core returns broader labels and therefore an incorrect-grain answer.

C SAMPLE EXECUTION TRACE

  • Permute Core’s trace profiles clinical categories, sample types, schemas, and candidate histology fields before finalizing a three-label aggregation.Its run is summarized as 20 agent steps and 20 tool calls.Source: C SAMPLE EXECUTION TRACE · p. 19
  • Permute EQ’s trace verifies IGF2 validity, explores diagnosis frequencies and cohort wording, compares histology grains, then computes selected means by code.The trace includes checks of joined records and evaluation of the unshifted log transform.Source: C SAMPLE EXECUTION TRACE · p. 19
  • The Permute EQ answer reports means of 2.7136, 2.6014, 2.5584, 2.6967, and 2.5826 for histology groups 9382/3, 9400/3, 9401/3, 9450/3, and 9451/3, respectively.The query requests average log10-transformed IGF2 expression for LGG patients across histology types, using valid values and unbracketed annotations.Source: C SAMPLE EXECUTION TRACE · p. 19
  • Permute Core’s numerically consistent computation is incorrect because it uses three broader descriptive histology labels instead of the five required ICD-O-3 groups.Both systems identify the LGG cohort, join clinical and expression data, and compute the requested transformation.Source: Figure 7 · p. 20

Related papers

Loading 2609.25286v1…