PyTorchKR Papers

최신 논문

총 75편
태그·기간 등 상세 필터

카테고리 탐색

arXiv:2611.05892v1cs.DB

When Is One Query per Template Enough? Adapting Learned Cost Models to New Hardware

A learned cost model that predicts query runtimes is trained on measurements from one machine. When the database moves to other hardware, its plans stay the same but their runtimes change, each by its own factor. We ask how to adapt such a model from k queries measured on the new machine: which queries to run, how to use their runtimes, and how many are enough. On PostgreSQL we timed the same plans of 1,162 TPC-H and TPC-DS queries in six environments: a laptop and a server, each with all its processors or pinned to one, the server also pinned to two, and the laptop also with a small buffer pool. No single factor describes the slowdown between two environments: most of its variance lies between query templates, so a speed factor stops improving after a few queries. A correction that fits separate lines for serial and parallel plans and adds an offset for every measured template brings the moved model, with one measured query per template, to an error (geometric mean over templates of the median Q-error) of 1.10 on TPC-H and 1.19 on TPC-DS, close to the same correction on a model trained on the new machine; fine-tuning a DACE-style neural cost model on the same queries gains at most 2%. On templates never measured, no correction we tried is on average more than 2% better than the speed factor. We then fixed hypotheses and timed new data. On TPC-H at ten times the scale the errors follow these results, but two of five repeated tests miss. On the Join Order Benchmark, whose queries differ widely within a template, the model errs by 2.36 even when retrained on the new machine, no correction gains much, and only the test on new templates holds; boosted trees and a neural model fit these queries better, and correcting the neural model works as well as fine-tuning it. We report every test as registered.

arXiv:2507.22186v3cs.LG

Scalable Source Selection for Machine Learning Tasks

Given a query table, existing data lake discovery techniques seek to identify data sources related to the query table through schema similarity and joinability. The discovered sources, however, are not always useful for downstream analyses; while some improve task generalization, some have the potential to inadvertently introduce harmful noise or distribution shifts. We formalize source selection as the problem of finding a subset of the discovered sources that maximizes downstream task utility. We propose Splice, a target- and cost-aware source selection system that adopts the idea of gene splicing to identify sources which provide additional signals that the union of all sources misses. To speed up the process, we develop two surrogate variants, Splice-DP and Splice-Grad, that estimate the marginal contribution of subsets to task utility by leveraging data profiles in a meta-learning approach and gradient alignment, respectively. On classification tasks over two real-world benchmarks derived from SOTA table union search tools, Splice shows 3-5 percentage points improvement in binary F1 scores compared to the case when all discovered tables are considered. On five additional real-world ML tasks spanning classification and regression, Splice improves task utility over the all-sources baseline by 1.8-13.7 points, while Splice-DP and Splice-Grad match this quality with 3-7x fewer model evaluations (up to 77x at scale).

arXiv:2611.04805v1cs.AI

Where Is the Truth When AI Agents Disagree? Dependence-Aware Belief Propagation for Unsupervised Truth Discovery (Extended Version)

LLM agents often give conflicting answers to the same question, and many deployed agents are variants of one base model, so their errors are correlated. We study truth discovery among such agents: inferring the true answer and each agent's reliability without labels. Classical truth-discovery and crowdsourcing estimators reward an agreeing bloc and can fall below majority vote. We propose BP-TD, which adds per-agent decorrelation weights to mean-field variational inference on a one-coin Dawid-Skene model, so that agents sharing a base model count together as about one vote. Its first iteration is exactly a weighted vote; later iterations estimate agent reliability. With 13 local agents on MMLU, BP-TD beats majority vote without any labels (0.716 vs. 0.624 on 160 questions; 0.707 vs. 0.674 on 10,000) and ranks agents by their true accuracy, which classical estimators invert. The gain depends on the pool: over 2,852 sub-pools of four benchmarks, decorrelation helps when the largest correlated bloc is weaker than the other agents and hurts when it is the strongest, and we found no label-free statistic that tells the two cases apart. Labels change the picture quickly: with 20 labelled questions, selecting the single best agent is more accurate than every aggregator we tested on MMLU, ARC and TruthfulQA.

arXiv:2611.04318v1cs.DB

Text-to-SQL Models Make Choices Without Saying So

A text-to-SQL answer can run, return a plausible number and still compute something other than what the analyst meant. This can happen when the request leaves a calculation choice open, such as whether to count rows or distinct entities, and the model settles it without saying so. Standard benchmarks cannot detect this, because they score each question against a single reference interpretation. We introduce SpecBridge, which tests such a choice by varying it alone: it poses three otherwise identical requests, one leaving the choice open and one per interpretation, and scores each answer by running it. The databases are real, and the choices are ones that made models fail on existing benchmarks. On 204 hand-validated tests, four language models usually compute each interpretation when asked for it, yet for the open request return both in only 15-58% of the tests where they can, often without mentioning the choice. The gap lies in reporting, not in computing. Benchmarks even penalize the alternative: on BIRD questions where such a choice changes the answer, over two in five of the answers marked wrong are right under the other interpretation. Because the gap lies in reporting, a system can narrow it without a stronger model: it first decides whether a request leaves a choice open, then answers each interpretation in its own call. On our tests and on AMBROSIA, an existing benchmark of ambiguous questions, this approach raises the share of open requests that get every interpretation by 29-63 points over the single answer that production assistants give, without lowering precision on clear requests. It also needs fewer user turns than asking a clarifying question.

arXiv:2611.04248v1cs.DB

Navigating Runtime Dynamism in Semantic Query Serving

LLM-powered semantic queries increasingly combine relational processing with multiple LLM-based operators over unstructured data. Efficiently serving such queries is challenging due to runtime dynamism at two coupled levels. Across queries, dynamic arrivals and heterogeneous resource demands continuously change scheduling priorities. Within a query, heterogeneous characteristics across fine-grained input batches cause operator costs and selectivities to vary throughout execution, making a statically optimized plan unreliable. These two forms of dynamism create a circular dependency: inter-query scheduling requires estimates of remaining LLM work, while the remaining work depends on runtime plan choices whose costs become observable only during execution. We present SeqoFlow, a runtime framework that jointly adapts query plans and schedules concurrent semantic queries. SeqoFlow enumerates equivalence-preserving plan alternatives under operator dependencies and decomposes queries into fine-grained units, enabling safe plan adaptation and query preemption at batch boundaries. To support both decisions without exhaustively profiling candidate plans on the target LLM, SeqoFlow uses lightweight speculative estimation to capture runtime costs. These estimates jointly guide plan adaptation and inter-query scheduling, closing the loop between query optimization and LLM inference. Experiments across diverse query structures, data distributions, and arrival rates show that SeqoFlow substantially reduces end-to-end latency compared to existing semantic query systems, adaptive query processing techniques, and conventional scheduling policies.

arXiv:2611.03737v1cs.LG

SSPE: Robust Sequential Staypoint Extraction in Degraded Trajectory Data

Extracting semantically meaningful staypoints from raw trajectory streams is a foundational prerequisite for mobility analytics. However, real-world trajectory data is frequently degraded by severe spatial noise and protracted temporal dropouts. Classic threshold-based extraction methods, such as the widely adopted sliding-window algorithm, are highly sensitive to these anomalies, routinely fracturing continuous stationary episodes into disjointed fragments. To overcome this limitation, we propose the Sequential Staypoint Extraction (SSPE) algorithm. SSPE employs a resilient, state-driven architecture that replaces static spatial anchors with a dynamic centroid and introduces an $n$-step lookahead heuristic to structurally tolerate transient coordinate drift. Concurrently, it utilizes temporal bridging mechanisms to prevent artificial fragmentation during prolonged signal loss. Since optimal extraction parameters vary significantly across datasets, we further introduce a fully unsupervised, consensus-based parameter estimation framework. This framework derives baseline statistical scales directly from the raw trajectory data and identifies the optimal parameter configuration by maximizing structural agreement, thereby eliminating the reliance on annotated ground truth. We evaluate SSPE on the ExtraSensory dataset, a real-world mobility corpus with manually verified ground-truth staypoints, against the sliding-window implementation of Trackintel, one of the most widely adopted extraction baselines in the field. Empirical results demonstrate that SSPE achieves superior accuracy, yielding higher F1-scores while simultaneously improving computational efficiency over this standard.

arXiv:2611.02936v2cs.AI

Why Tabular Foundation Models Excel on Relational Data They Have Never Seen

Given a relational database (RDB) storing heterogeneous tabular information, how can we predict missing (or future) values in some target column of interest? As candidate targets are numerous in enterprise settings, it is preferable to avoid learning a new model from scratch each time there is a new prediction task. Frozen foundation models (FMs) based on in-context learning (ICL) and RDB-specific pre-training provide a viable solution, but the ideal design still remains largely unresolved. Perhaps unexpectedly, preliminary evidence has recently shown that existing single-table FMs (aka TFMs), when combined with suitable parameter-free relational encoders, can be quite competitive in practice despite zero exposure to RDB pre-training or relational data of any kind. In this work we first empirically confirm this phenomenon with a more diverse array of experiments. We then turn to our main contribution, namely, surfacing multiple complementary explanations for why TFMs behave so effectively. These explanations ultimately hinge on innate synergies whereby parameter-free encoders map relational effects to fixed-length embeddings that loosely mimic distributions for which TFMs were originally designed; in so doing, desirable single-table properties transfer to relational settings (e.g., convergent posterior predictive distributions). Collectively, our analysis forges new pathways for improving RDB FMs, particularly when we prioritize interpretability and inevitable integration within agentic systems.

arXiv:2611.03327v1cs.AI

From FAIR to AI-FAIR Metadata-First Infrastructure for AI-Enabled Science

Artificial intelligence (AI) technologies and their capabilities have advanced rapidly over the past few years, leading to widespread interest and adoption in the biological and ecological sciences. However, the data infrastructure and computational resources that provide the foundation for AI workflows in ecology and biodiversity science remain fragmented. Moreover, the flow of value remains largely unidirectional: data move from collectors, through repositories and aggregators into models, and finally to research products or management actions. The derived value, such as corrected taxonomies, annotations, quality flags, and insights, rarely returns to source communities. We propose an extension of the original FAIR principles to address the needs of AI-ready infrastructure: the AI-FAIR metadata criteria, which place AI-enabled science in the context of a bidirectional Question-Data-Model triangle. This framework is designed to facilitate large-scale content-based analysis across diverse collections and ensure that the scientists generating the data benefit from the AI-enabled advances built on their data. Alongside these criteria, we introduce an agentic evaluation workflow that both rates existing datasets on the AI-FAIR and traditional FAIR criteria and provides actionable feedback where critical metadata is missing or incomplete. This modular dataset evaluation tool, paired with a domain-appropriate checklist, translates these principles into a reusable assessment of machine-learning readiness, AI-readiness for specified tasks, traceability, and ethical and CARE-aligned stewardship. Through this approach we aim to create a community-informed system to close the model data loop, so that AI also enriches data to further enable scientific discovery.

arXiv:2611.02695v1cs.CV

3L-D: A Low-Light Level Dataset from a Sensor-Agnostic, Alignment-Preserving Acquisition Framework

Supervised low-light image and video enhancement has achieved remarkable progress, yet current models remain fundamentally constrained by the limitations of existing datasets. Prevailing methods typically rely on uncontrolled ambient captures, camera parameter manipulation (e.g., exposure or ISO adjustments), or neutral density filters, failing to replicate authentic night-sky near-infrared spectral content and genuine low-photon noise distributions. Furthermore, capturing dynamic paired sequences without introducing warping artifacts or sensor-specific biases remains an open challenge. In this paper, we introduce 3L-Dataset, a physically calibrated low-light image and video dataset generated using a modular, sensor-agnostic acquisition framework. Our test bench faithfully reproduces the standardized Midavaine night level scale from Level 1 (full moon, approximately 1000 mlx) down to Level 5 (overcast moonless night, approximately 0.1 mlx), spanning the visible to near-infrared spectrum across diverse realistic environments, including countryside, forest, urban, and desert scale models. Using high-precision programmable mechanical actuation, our acquisition framework guarantees sub-pixel spatial alignment across synchronized sextuplets, each composed of one reference image and five corresponding images acquired at the five standardized night levels, without post-processing image warping. We evaluate multiple representative deep learning baselines across varying night levels, revealing critical trade-offs between reconstruction fidelity, perceptual quality, and inference latency under severe photon starvation. The modular acquisition framework can be adapted to different cameras, optical systems, sensor technologies, and resolutions, enabling the dataset to be continuously expanded.

arXiv:2611.02132v1cs.LG

TSSDA: An Autonomous Agent for Constructing a Time-Series Text-to-SQL Benchmark

As agents increasingly undertake data querying, Text-to-SQL serves as a foundational capability for database access. However, for text-to-SQL in time-series databases, high-quality benchmarks and large-scale training corpora remain scarce, limiting the systematic evaluation and domain adaptation of existing models and agents. Existing automated SQL generation approaches lack sufficient autonomy, leaving key challenges unresolved. First, numerous combinations of temporal syntax and functions complicate comprehensive coverage of valid queries. Second, as the corpus grows, generating valuable and diverse new data becomes increasingly challenging. Third, schema and data heterogeneity hinder reuse of temporal query structures, limiting generation efficiency. To address these challenges, we propose TSSDA, a Time-Series SQL (TSSQL) Generation Data Agent framework, and construct TERN, a comprehensive Time-sERies text-to-SQL beNchmark. First, to achieve broad coverage of valid queries, we systematically model TSSQL syntax and semantics to construct a data generation space. Next, to explore the space and improve diversity, we develop a Data Agent that autonomously orchestrates data generation. The agent comprises a planning and decision-making module and a query generation module, with online adaptation for continuous self-improvement. Finally, we introduce a temporal-aware memory that enables adaptive reuse of experience across heterogeneous data instances, improving generation efficiency. Comprehensive evaluations on TERN reveal the limitations of existing methods in performing Text-to-TSSQL tasks. Furthermore, extensive experiments demonstrate that TSSDA substantially improves coverage, diversity, and complexity.

arXiv:2611.00960v1cs.CL

Self-Refining Prompt Induction for Streaming Text-to-SQL

Text-to-SQL assistants for streaming engines must handle dialect-specific types, windowing rules, and temporal semantics, including constructs poorly handled by generic prompts. We present Self-Refining Prompt Induction (SRPI), an offline method that converts generation failures into reusable dialect instructions while keeping model weights fixed. A Refiner revises the Generator's prompt using reference-conditioned feedback from three LLM judges; a sampled regression guard checks earlier successes, and a separate Compressor reduces prompt length. Inference uses the frozen prompt without retrieval or query repair. We evaluate cross-domain and cross-model transfer on Confluent Cloud Flink SQL using a controlled synthetic workload seeded by 30 expert-written examples, with 335 induction examples and 257 questions from held-out domains. The same induced prompt improves mean judge scores across seven Generators by 7.1-19.2 percentage points, including an increase from 85.3% to 97.5% for the induction model. A separate compression study reduces prompt length by 52.7% with a 0.19-percentage-point score decrease. An assistant using the induced prompt is in early access, with initial user feedback to inform subsequent refinement.

arXiv:2611.00638v1cs.CL

ElectroChemLLM: A Multimodal LLM Pipeline for High-Fidelity Electrochemical Informatics

Electrochemistry is central to solving many modern-day challenges, including energy storage systems for a sustainable future, and its advancement relies heavily on the data from past findings. Existing automated pipelines are limited by text-only architectures, context fragmentation, and numerical hallucinations - collectively preventing accurate, large-scale curation of property databases. We introduce ElectrochemLLM, a multimodal information extraction pipeline that jointly identifies structured electrochemical records encompassing electrolyte compositions, electrode materials, cell configurations, and their measured physicochemical properties from unstructured scientific PDFs at scale. By combining hybrid document parsing with vision-capable LLM orchestration, the pipeline extracts verifiable, structured data across text, tables, and figures. Evaluated on the battery-electrolyte subset of our manually annotated benchmark of 232 papers, ElectrochemLLM achieves an F-1 score of 0.97, and has been applied to curate ElectroChemDB, a structured database spanning 15,000 electrochemical research articles. We demonstrate the utility of this resource by training machine learning models for electrolyte property prediction, illustrating a complete pathway from unstructured literature to predictive materials discovery.

arXiv:2611.00449v1cs.DB

SpBench: An Agent-Native Framework for Cross-Engine Lakehouse Benchmarking

SpBench is an agent-native framework for cross-engine lakehouse benchmarking, in production use for benchmarking Snowpark, Snowpark Connect for Apache Spark, and other execution environments across standardized suites -- such as TPCx-BB and TPC-DI -- and custom benchmarks distilled from production workload patterns. It runs continuously across a broad portfolio of workloads, guiding the performance roadmap and detecting regressions for production-scale data engines. Benchmarking has historically been an expensive investment requiring continuous cross-functional collaboration and domain expertise, making it difficult to obtain actionable performance insights. SpBench provides (i) a typed, code-generated benchmark substrate that compiles platform-agnostic benchmark definitions into fair, comparable platform-specific implementations; (ii) a structured-telemetry contract recording every run as queryable fact and dimension data, letting LLM agents query results deterministically; and (iii) an agent-native harness of three composable skills to expedite building benchmarks and analyzing results. This agent-native architecture enabled the rapid development of more than 400 benchmark queries that run on a regular cadence to detect performance regressions; catching these has directly led to improvements in production.

arXiv:2611.00128v1cs.DB

SPARQLcirc: Probabilistic Query Evaluation with Provenance Circuits

Probabilistic query evaluation (PQE) computes the probability of every answer of a query over a tuple-independent probabilistic knowledge graph. We study exact PQE for SPARQL over graphs accessed through SPARQL 1.1 endpoints. Shared triples among the derivations of an answer make exact PQE #P-hard in general. Knowledge compilation of the Boolean provenance of every answer yields exact probabilities, and the provenance formulas of different answers can share subformulas. We propose SPARQLcirc, which rewrites a SPARQL query into CONSTRUCT queries that name every gate by its content, so that the results of independent queries form one Boolean provenance circuit for all answers. The rewriting covers joins, unions, projection, filters, OPTIONAL, difference, and transitive closures. SPARQLcirc compiles the circuit and computes the exact probability of every answer by weighted model counting. For the evaluation, we draw random probabilities for the triples of WatDiv and gMark and the statements of Wikidata and evaluate the CONSTRUCT queries on GraphDB, Fuseki, and Oxigraph. The ratio of the end-to-end time of PQE with SPARQLcirc to that of PQE by compiling the provenance expressions of the SPARQL provenance rewriting NPCS decreases with the number of derivations.

arXiv:2611.00111v1cs.CV

SpatialVec: A Unified Representation Learning Framework for Geospatial Objects with Boundary-Aware Sampling

Spatial representation learning is essential for Geospatial Artificial Intelligence (GeoAI) applications, enabling machine learning models to capture the geometry, location, and spatial relationships of geographic entities such as points, polylines, and polygons. Existing spatial encoding methods often struggle to preserve fine-grained topological information, particularly for interactions involving complex geometries such as polylines and polygons. These approaches inadequately capture boundary structure and local spatial context, causing topology-sensitive relationships near geometry boundaries to be underrepresented in the learned embedding space. Consequently, accurately modeling complex spatial interactions remains a significant challenge for downstream spatial reasoning tasks. To fill the gap, we introduce SpatialVec, a boundary-aware spatial representation learning framework that adaptively samples geometry boundaries and their surrounding regions, capturing fine-grained geometric structures critical for topology-aware spatial reasoning. A transformer-based encoder-decoder trained through geometry reconstruction produces compact, geometry-aware, and unified representations for points, polylines, and polygons. Empirical evaluation on two real-world datasets shows that SpatialVec substantially improves the preservation of challenging topological relationships involving polygonal interactions and outperforms state-of-the-art baselines significantly on downstream prediction and classification tasks.

arXiv:2610.22061v2cs.AI

FAOPs: Normal Forms and an Algebra for AI Agent Workflows

Organizations increasingly encode operational procedures as multi-step agent skills, here called Agent Operating Procedures (AOPs), that must be traceable, auditable, reproducible, access-controlled, and bounded in time and cost, as regulated sectors require. Yet AOPs are typically kept as coarse-grained files, in which steps bundle actions, copied procedures drift apart, any step may reach any tool, and refactoring goes unchecked. We present Functional Agent Operating Procedures (FAOP), a formal theory that brings database-grade discipline to AI agent workflows. FAOP recasts each workflow as a typed function over scoped state and subjects an entire library to four progressive normal forms: Decomposed (FNF-1) splits a workflow into atomic steps, Typed (FNF-2) makes its data flow explicit and type-checked, Factored (FNF-3) gives each shared sub-workflow a single canonical definition so a change propagates instead of drifting, and Scoped (FNF-4) checks that each FAOP lists resources used by its own actions and the FAOPs it calls, while guardrails declared by it and by FAOPs that call it remain attached to their sources. A FAOP algebra of provably sound equalities turns refactoring and parallelization into compiler rewrites; orthogonal extensions add optimization (+O), bounded completion with time and cost limits (+B), and reachability checks (+R); and an architectural-level classifier (0-3) determines whether a workflow needs AI at all. Together, these make integrity, isolation, typing, stopping limits, cost, and optimization compile-time concerns across the whole library, not runtime hopes. We implement the theory in SkillBase, an AOP management system (what PostgreSQL is to data, SkillBase is to AOPs), and a separate executor that loads FAOPs from SkillBase and enforces these guarantees at run time. A worked financial workflow illustrates the approach end to end.

arXiv:2609.28761v2cs.DB

KathDB-FAO: Optimizing Multimodal Queries through Operator Grouping and Code Reuse

Advances in multimodal large language models (multimodal LLMs) allow data systems to answer analytical queries over text and images alongside relational tables, but at a significant cost for large-scale data. Existing systems reduce this cost through rewrite rules written in advance, which cannot use the logic of an individual query or the semantics of the data it processes. To address this challenge, we design, implement, and evaluate KathDB-FAO, a query evaluation system over multimodal data. KathDB-FAO decomposes a complex query into an execution plan of fine-grained steps, which we call Function-as-Operators (FAOs). To reduce execution cost, it searches for subsets of these FAOs to implement together as one function, enabling query-specific optimizations that make fewer model calls. Synthesizing these functions with a frontier model has its own cost, yet we observe that synthesized FAOs often recur across queries. To further reduce plan generation cost, KathDB-FAO uses pre-built functions and reuses functions synthesized for earlier queries. On SemBench and an additional workload, KathDB-FAO achieves quality comparable to the strongest competing system while reducing end-to-end cost per query by 76% on average.

arXiv:2609.15779v2cs.AI

EvoOntology: A Self-Evolving Ontology Layer for Data Agents

Data agents aim to fulfill natural-language instructions over heterogeneous data, including tables, files, and databases. However, data agents face a challenging agent-data gap: heterogeneous data resides outside the agent, while the agent can access it (e.g., column names and file paths) only through generic tools. Existing approaches either let agents directly explore raw data sources or inject manually constructed semantic layers into prompts. However, neither scales well to large heterogeneous data sources nor adapts to different agent behaviors. In this paper, we introduce EvoOntology, a self-evolving ontology layer for data agents. EvoOntology encapsulates the ontology as an MCP server comprising a schema layer, a content layer, and a tool layer, enabling agents to actively query and interact with the ontology at runtime. To this end, we introduce a builder agent for autonomous ontology construction and a self-evolution loop that continuously refines the ontology through attribution-guided typed edits that are accepted only after a backbone-conditional paired evaluation. Experiments on three well-adopted data-agent benchmarks with four LLM backbones demonstrate that EvoOntology consistently outperforms strong baselines and existing semantic-layer approaches, effectively bridging the agent-data gap and enabling more effective interaction with heterogeneous data. Code: https://github.com/ruc-datalab/EvoOntology

arXiv:2609.08869v2cs.ET

OntoKG-EQ: Evidence-boundary auditing for analytical knowledge graphs

An analytical knowledge-graph result can be supported in its ambient graph while its delivered explanation omits deciding observations. OntoKG-EQ connects competency-question requirements to an executable audit of ambient observations, declared finding support and delivered evidence. A bounded declarative interpreter distinguishes positive, false and unknown numerical status, with same-witness unit readiness reported separately. A public-data application reconstructs England's August 2026 provider-level referral-to-treatment statistical screen. Independent source recomputation reconciles 532 providers with the publisher workbook: 91 meet the inclusive 92% threshold and 441 do not. One additional roster entry remains unavailable rather than zero-filled. All 533 application frames agree with an independently implemented standards-based validation and availability baseline. Transfer requires a disclosed source adapter and arithmetic compilation, but no change to the core interpreter. Twenty-six constructed analytical cases yield 78 status/readiness agreements. Actual financial release replay reproduces two omitted-volume defects: support and delivery suffice for 19 of 21 original findings and all 21 repaired findings after interface adaptation. A separate replay agrees on 487 overlapping financial records. Source audits expose scaling errors and unresolved comparator and temporal assumptions, so financial examples do not establish authenticated market effects. The contribution is a reproducible analytical audit interface and source-reconciled application, not new sufficiency logic or first evidence-boundary isolation. Historical ratings measure preference only; objective human audit benefits and independent production deployment remain untested.

arXiv:2608.19741v5cs.CL

One Success Isn't Reliability: Thinkingbox, a Sandbox and Benchmark for Agents in Stateful Business Workflows

Recent agent benchmarks increasingly ground evaluation in executable environments, from code repair to web navigation, app APIs, and function calling. Yet completing consequential work beyond code requires more than producing a plausible response or valid tool call: agents must gather missing information over multiple turns, follow domain policies, coordinate dependent tools, and realize the correct persistent state transition without collateral effects. In this paper, we introduce Thinkingbox, a sandbox for tool-agent-user interaction that provides isolated MCP-compatible tool sessions, complete execution traces, and outcome evaluation over terminal backend state. Built on this sandbox, Thinkingbox-bench contains 507 policy-conditioned workflows across business scenarios, including retail, hospitality, auto insurance, neobank internal IT, and consulting IT/HR support. Each attempt is evaluated by task-specific executable checks that accept valid trajectories while rejecting wrong, missing, or extra effects; designated tasks additionally check required properties of the final response. Our experiments reveal that even the strongest proprietary and open-weight models show steep reliability drops: Claude Opus 5 falls from 66.50% pass@1 to 47.53% pass^20, and Kimi-K3 from 57.37% pass@1 to 17.60% pass^20. Moreover, many failed trials terminate cleanly after valid state-changing actions, so response- or tool-call-level signals poorly proxy end-to-end completion. Thinkingbox-bench reveals a large gap between occasionally finding a successful trajectory and reliably completing stateful business tasks. We release Thinkingbox, Thinkingbox-Bench, and Thinkingbox-integrated reinforcement learning framework.