Pith. sign in

REVIEW 4 major objections 5 minor 120 references

A specialist BPMN-to-agent compiler outperforms general-purpose coding agents on structured workflow generation, cutting token use by over 95% and eliminating repair loops.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 02:01 UTC pith:GK32OE2R

load-bearing objection The BPMN-to-ReAct compiler is a real idea, but the evaluation bench is tilted: the headline TUE gap mostly measures whether the baseline happened to match the specialist's one-tool-per-node decomposition. the 4 major comments →

arxiv 2607.14456 v1 pith:GK32OE2R submitted 2026-07-16 cs.SE cs.AI

Beyond Generalist LLMs: Specialist Agentic Systems for Structured Code Workflow Execution

Harris Borman , Herman Wandabwa , Fusun Yu , Sandeepa Kannangara , Justin Liu , Anna Leontjeva , Ritchie Ng This is my paper
classification cs.SE cs.AI
keywords BPMNagentic workflowsspecialist agentsgeneralist LLM agentsReActtool-use exactnesscode generationtoken efficiency
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tries to establish that a specialist system—a compiler that turns BPMN process diagrams into executable ReAct agents—can beat general-purpose LLM coding agents on structured workflow generation, and do so much more cheaply and reliably. For ten deterministic business workflows, the specialist approach scored 9–20 percentage points higher on tool-use exactness, ran 2.6–3.6 times faster on a penalty-adjusted latency metric, made about 2.5 times fewer tool-call errors, consumed over 95% fewer tokens, and required zero repair iterations. A sympathetic reading: the gains come from externalising the workflow structure—BPMN supplies the plan, so the LLM only fills in local tool logic under tight, per-node context. If correct, this points to a design pattern that could make agentic code generation viable at industrial scale, where consistency and cost matter more than flexibility.

Core claim

The central claim is that a 'specifications-to-agent compiler'—which parses a BPMN 2.0 diagram into typed tool contracts and a control graph, generates API client code, builds each node's tool, validates tools by execution, and then assembles a ReAct agent with a scoped natural-language prompt—produces agents that outperform general-purpose coding agents (Roo and Cline) on every measured axis. The specialist system obtained 57.69% tool-use exactness vs 48.62% for Cline and 38.11% for Roo; averaged 1.27 tool-call errors per run vs ~3.2 for both baselines; achieved a penalty-adjusted latency of 2.49 seconds per effective step vs 6.59 and 9.08 seconds; and generated each agent in about 55k toke

What carries the argument

The load-bearing mechanism is the externalisation of planning: BPMN's explicit control-flow semantics are compiled into a ReAct-style control graph, so the LLM never has to discover the plan or hold the whole workflow in context. Each BPMN node becomes a typed tool contract plus a node-local policy, the control plane enforces branches and joins, and a per-node context-scoping step limits what the model sees. An internal tool-verification loop executes generated tool code and refines it until it passes—this replaces the generalist agents' implicit trial-and-error repair iterated via external error feedback.

Load-bearing premise

The comparison's load-bearing premise is that running the generalist agents in their default configuration, with a prompt that forbids them from testing their own code, is a fair and representative deployment of these tools.

What would settle it

Run Roo and Cline without the 'do not execute your code' instruction—let them run, test, and repair their own output—on the same ten workflows and the same evaluation, and check whether the 9–20 point tool-use exactness gap and the 95% token savings persist; if the gap collapses, the specialist advantage is an artifact of the constrained baseline.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Teams adopting a specialist compiler for a well-specified task class can generate agents in a single pass at roughly 5% of the token cost, with no repair iterations.
  • On structured, deterministic workflows, the resulting agents will call the right tools in the right order 9–20 percentage points more often than agents written by generalist coders.
  • The 2.6–3.6x latency advantage means end-to-end workflow execution is faster even before accounting for fewer errors.
  • The consistency results (coefficient of variation 0.63 vs 0.86 and 1.05) suggest specialist systems reduce run-to-run variance, which matters for maintainability in industrial settings.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The fairness of the comparison depends on the baseline prompt forbidding the generalists from executing their own code; permitting self-testing could narrow the gap, so the paper's quantitative claims should be read as 'given that constraint'.
  • The context-management principle—keeping the model's active context to a single node's contract—is portable: any structured specification (state machines, decision tables, DSLs) could be compiled to an agent the same way, and the token savings would likely scale with workflow size.
  • The tool-use exactness metric, a strict binary per-run check on tool-call sequence, could serve as a reusable benchmark for structured workflow generation beyond BPMN, though it counts a run with a single missed call as a total failure.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes a specialist agentic system that compiles BPMN workflow diagrams into ReAct-style agents, and compares it against two generalist coding agents (Roo and Cline) on ten author-constructed deterministic workflows. The headline claims are that the specialist improves Tool-Use Exactness by 9–20 percentage points, reduces penalty-adjusted latency by 2.6–3.6×, cuts tool-call errors by about 2.5×, lowers generation token cost by over 95%, and eliminates repair iterations. The evaluation uses 300 successful agent runs and 30,543 test cases, with metrics computed by an LLM judge.

Significance. If the result holds, the paper offers a convincing, engineering-oriented demonstration that domain-structured compilation of process models can beat generic coding agents on reliability, speed, and cost for structured deterministic workflows. The authors are transparent about the absence of an established benchmark, the author-constructed workflow set, and the deterministic scope. They also contribute a substantial evaluation corpus (371 generation attempts, 300 successful agents, 30,543 test cases). However, the central comparison is weakened by a scoring rubric that is aligned with the specialist's own decomposition, by a baseline configuration that forbids execution, and by non-comparable token accounting. These issues affect the headline quantitative claims and the paper's industrial-relevance conclusion.

major comments (4)
  1. [§4.2.2, §3.3 Step 3] Tool-Use Exactness, tool-call errors, and penalty-adjusted latency are scored against the expected tool-call sequence from the workflow specification, but the only instruction to baselines (Figure 5) does not define tool granularity. The specialist, by construction, generates one tool per BPMN node (§3.3 Step 3). A generalist that produces a semantically equivalent agent with a different tool decomposition will be penalized as missed/excess. Since TUE is binary and requires zero missed/excess/out-of-sequence calls, the reported 9.1–19.6 pp TUE gap and 2.5× error gap may be partly a rubric-alignment artifact. Please validate the judge against human annotations and provide example traces showing that generalists' tool calls were actually one-to-one with BPMN nodes, or use a functional-outcome metric as the primary comparison.
  2. [§4.1, Figure 5, §3.3 Step 4] The baselines are explicitly instructed: 'Do not attempt to execute your code after completing it; manual testing will be performed and any errors will be passed directly to you.' The specialist, in contrast, includes an internal tool-verification loop (§3.3 Step 4). This differentially handicaps the generalists and makes the repair-iteration and token-cost comparison in §4.3.1 a comparison of the specialist against a deliberately constrained deployment of the baselines. Industrial users can let agents execute tests. Please run a condition in which Roo and Cline may execute their code and self-repair, or explicitly frame and justify the no-execution setting as a 'low-configuration default' rather than a general conclusion about generalist agents.
  3. [§4.2.1] Token usage is measured with Langfuse tracing for the specialist but self-reported by Roo and Cline, with no evidence that the measurements are equivalent. The text also states that baseline token counts are reported 'before accounting for additional tokens consumed in repair cycles', which suggests the reported 1,500k/1,044k figures may exclude repair tokens. Since the >95% token reduction is a headline economic claim, the counts should be measured under a unified protocol that includes all repair iterations.
  4. [§4.2.2] The Process Adherence and TUE metrics rely on GPT-4.1 as an LLM judge, but the judge is not validated against human annotations, and no example execution traces are shown. Given that the judge scores exactly the tool-decomposition constructs the specialist is architecturally aligned with, this is a further risk to the process-adherence and TUE claims. Please include a small human-annotation study or at least a randomly sampled set of trace-level judgments with inter-annotator agreement.
minor comments (5)
  1. [§4.1] The text says baselines were 'allowed to operate unconstrained with their default approaches', yet Figure 5 explicitly forbids code execution. Clarify that this is a deliberate constrained configuration and state the rationale.
  2. [Eq. (1)] The choice of ε = 1 is arbitrary. Report sensitivity of the penalty-adjusted latency results to this parameter, even over a small range.
  3. [Figures 3 and 4] Captions are terse and plots do not show error bars or statistical significance across the 10 runs. Add per-workflow variability information to support the cross-workflow consistency claims.
  4. [§4.3.1] 'Eliminating repair iterations' should be qualified as applying to successful generations only, since 71 of 371 attempts failed outright.
  5. [§4.1] The sentence listing 'Social, Spam, and News' as completed without failures by 'all three systems' should be checked for consistency with the reported attempt counts for the specialist, Cline, and Roo.

Circularity Check

1 steps flagged

TUE and tool-call-error metrics are scored against the specialist's own one-tool-per-BPMN-node decomposition, so the headline 9–20 pp advantage is partially constructed by the rubric.

specific steps
  1. self definitional [Section 4.2.2 (Agent Run Evaluations) vs. Section 3.3 Step 3 (Tool/Context Creation)]
    "The judge received (i) the workflow specification defining the expected tool-call sequence ... Each tool call was categorised as correct, missed (required but not called), excess (called but not required), or out-of-sequence. ... TUE is a binary per-run indicator equal to 1 when the agent invokes exactly the prescribed tools with zero missed, excess, or out-of-sequence calls, and 0 otherwise. ... Using the parsed workflow steps and the API service, the system generates code for each tool corresponding to a workflow node."

    The 'prescribed tools' in the TUE rubric are the BPMN node-level tools; the specialist pipeline is constructed to emit exactly one tool per BPMN node. TUE therefore scores every agent against the specialist's own tool decomposition. A generalist that satisfies the same workflow with a coarser or finer tool granularity is recorded as missed/excess/out-of-sequence even if functionally equivalent, inflating the specialist's 9.1–19.6 pp TUE and 2.5–3× error advantages. The paper neither validates the LLM judge against human annotations nor shows traces demonstrating that the baselines' tool decompositions were actually one-to-one with BPMN nodes, so the advantage is partly an artifact of defining the metric in terms of the specialist's design.

full rationale

This is an empirical benchmark paper rather than a formal derivation; there are no equation-level loops and no load-bearing self-citations. The token-cost, latency, and repair-iteration results are measurements with independent content (although the Figure 5 instruction 'Do not attempt to execute your code after completing it' creates a fairness asymmetry with the specialist's internal verification loop—a validity concern, not a circular one). However, the central run-evaluation claim is partially self-definitional: the specialist's Step 3 tool creation ('code for each tool corresponding to a workflow node') is the same decomposition used by the Section 4.2.2 judge ('the expected tool-call sequence'), so the headline TUE/error advantage is partly guaranteed by the rubric rather than by superior execution. Score 6 reflects partial circularity in the primary comparison; the claim is not wholly circular because TUE is still empirically measured (57.69%, not 100%) and the other metrics retain independent value.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical or mathematical entities. Its load-bearing assumptions are evaluative: the workflows, the baselines, the judge, the ground-truth spec, and the token-accounting comparability are all assumed rather than independently validated.

free parameters (2)
  • epsilon in penalty-adjusted latency (Eq. 1) = 1
    Hand-chosen constant to avoid division by zero in P = max(epsilon, C-(M+E+O)); it affects latency values for runs with net non-positive effective steps.
  • Equal error penalty weights in latency denominator = 1 per missed/excess/out-of-sequence call
    The metric treats M, E, O as equally costly in P = C - (M+E+O). This is a modeling choice, not fitted to data; unequal weights could change reported latency ratios.
axioms (5)
  • domain assumption The ten author-constructed BPMN workflows are representative of real deterministic business processes
    Section 3.1 admits the workflows were manually constructed by the authors and that no established benchmark exists; all conclusions depend on this representativeness.
  • domain assumption Roo and Cline in default configurations are representative of general-purpose coding agents
    Section 4.1 states they were 'allowed to operate unconstrained with their default approaches'; this is the only comparison used for the central claim.
  • domain assumption GPT-4.1 as LLM-as-a-judge reliably identifies missed/excess/out-of-sequence tool calls and process deviations
    Section 4.2.2 uses an LLM judge for Process Adherence and TUE; no human validation or judge accuracy is reported.
  • domain assumption The BPMN 2.0 XML fully determines the expected tool-call sequence and outcome
    Section 3.3 compiles BPMN nodes into typed tool contracts; if BPMN is ambiguous or incomplete, the ground truth for TUE is ill-defined.
  • domain assumption Token accounting from Roo/Cline built-in reporting is comparable to Langfuse tracing for the specialist
    Section 4.2.1 states specialist tokens were measured via Langfuse while Roo/Cline reported their own consumption; different counting scopes could explain part of the reported 95% gap.

pith-pipeline@v1.3.0-alltime-deepseek · 13789 in / 13476 out tokens · 123502 ms · 2026-08-02T02:01:14.281661+00:00 · methodology

0 comments
read the original abstract

Large Language Models (LLMs) have accelerated the adoption of software development agents, now widely available as Integrated Development Environment (IDE) extensions and standalone applications. While these agents are typically general-purpose, it remains unclear whether specialist agents justify their additional development effort. We investigate this question in the context of business process automation, focusing on the transformation of Business Process Model and Notation (BPMN) diagrams into executable agentic workflows. Since BPMN specifies explicit control-flow semantics, we focus on deterministic workflows in which a fixed process model and inputs uniquely determine the executed path. We introduce a specialist workflow for this task and compare it against generalist agents such as Roo and Cline. Our results show that the specialist solution produces agents that outperform generalist baselines by approximately 9-20 percentage points in tool-use exactness, 2-4x in penalty-adjusted latency, and 3x fewer tool-call errors, while reducing generation token cost by over 95% and eliminating repair iterations. We also find that generalist agents generate code inconsistently in both functionality and quality, limiting their suitability for industrial settings where reliability and maintainability are essential.

Figures

Figures reproduced from arXiv: 2607.14456 by Anna Leontjeva, Fusun Yu, Harris Borman, Herman Wandabwa, Justin Liu, Ritchie Ng, Sandeepa Kannangara.

Figure 1
Figure 1. Figure 1: Proposed System Architecture with its GraphFlow controller, failed to autonomously gen￾erate the required files without manual agent wiring. FLOW produced high-level task decompositions but not complete, executable code. As none completed even a single work￾flow end-to-end, we excluded them from the comparative evaluation, focusing on Roo and Cline as the only baselines that successfully completed the task… view at source ↗
Figure 2
Figure 2. Figure 2: Agent Generation Metrics: Repair Iterations, Token Usage (Input and Output) for the Specialist System(Custom), Roo, and Cline. the design principles outlined earlier in the paper: con￾strained execution via externally specified workflow struc￾ture, targeted context management, and modular decompo￾sition. Together, these reduce search space and planning overhead, which in turn contributes to fewer tool-call… view at source ↗
Figure 3
Figure 3. Figure 3: Per Workflow Agent Run Evaluation Metrics: Tool-Use Exactness, Tool-Call Errors, Latency and Process Adherence for the Specialist System 11 [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Key metrics by approach. Comparison of key metrics across different approaches in agent run evaluation. 13 [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Base prompt provided to Roo and Cline for each agent generation attempt. The {num} placeholder was incremented per attempt, and the <API SPEC> and <BPMN> blocks were populated with the actual API specification and BPMN workflow for each experiment. 14 [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Online retail cart/order fulfillment workflow. The process validates the cart, authorises payment, reserves inventory, and routes to either fulfillment/shipping or cancellation/backorder handling depending on payment and stock decisions. After shipping, it monitors delivery and branches into delayed/lost handling, ending in either delivery confirmation, refund, or reorder. 15 [PITH_FULL_IMAGE:figures/full… view at source ↗
Figure 7
Figure 7. Figure 7: Costing/estimation workflow. A deterministic sequence of data-gathering and calculation tasks with decision gateways that select which cost components to apply, followed by aggregation into a final cost/price output. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Deal assessment workflow. The workflow evaluates a candidate deal through staged checks and scoring steps, using decision gateways to route to accept/reject/escalate outcomes based on thresholds and validation checks. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Labelling/classification workflow. A rules/criteria-driven labelling pipeline that applies sequential checks and uses gateways to assign labels and/or exclude items, terminating once a label decision is reached. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: News processing workflow. The process ingests an item, performs validation and enrichment steps, then routes through conditional branches (e.g., eligibility/quality thresholds) to produce a final categorisation/output or a discard/escalation outcome. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Risk & opportunity classifier workflow. The workflow computes risk/opportunity signals, derives a final label (risk/opportunity/mixed/neutral), then routes by a decision gateway into different priority/handling paths (e.g., high/medium vs low) before completing. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Social media workflow. A multi-step pipeline that performs content/user/context checks and then uses decision gateways to choose an action path (e.g., allow, flag, or escalate), ending once an action outcome is produced. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Customer promotion workflow. The process fetches and validates customer profiles, retrieves order history and loyalty points, checks promotion eligibility, segments customers into value tiers (high/low), calculates tier-appropriate discounts, generates recommendations, and routes notifications based on promotion urgency. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Tournament/bracket workflow. The workflow advances items/participants through staged rounds, using decision gateways to determine advancement and termination once a final winner/outcome is determined. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Weather workflow. A deterministic sequence of data retrieval and transformation steps with conditional branches (e.g., data availability/threshold checks), culminating in a final forecast/decision output. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_15.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

120 extracted references · 7 canonical work pages

  1. [1]

    Scaling Learning Algorithms Towards

    Bengio, Yoshua and LeCun, Yann , booktitle =. Scaling Learning Algorithms Towards

  2. [2]

    and Osindero, Simon and Teh, Yee Whye , journal =

    Hinton, Geoffrey E. and Osindero, Simon and Teh, Yee Whye , journal =. A Fast Learning Algorithm for Deep Belief Nets , volume =

  3. [3]

    2016 , publisher =

    Deep learning , author =. 2016 , publisher =

    2016

  4. [4]

    Ibm Cooperation , volume =

    Introduction to BPMN , author =. Ibm Cooperation , volume =

  5. [5]

    Computer Standards & Interfaces , volume =

    BPMN: An introduction to the standard , author =. Computer Standards & Interfaces , volume =

  6. [6]

    Business Process Management , year =

    Concepts, languages, architectures , author =. Business Process Management , year =

  7. [7]

    2018 , publisher =

    Fundamentals of business process management , author =. 2018 , publisher =

    2018

  8. [8]

    2019 , edition =

    Business Process Management: Concepts, Languages, Architectures , author =. 2019 , edition =

    2019

  9. [9]

    Business & Information Systems Engineering , volume =

    Challenges in business process intelligence , author =. Business & Information Systems Engineering , volume =

  10. [10]

    The knowledge engineering review , volume =

    Intelligent agents: Theory and practice , author =. The knowledge engineering review , volume =

  11. [11]

    Handbook on business process management 1: introduction, methods, and information systems , pages =

    The six core elements of business process management , author =. Handbook on business process management 1: introduction, methods, and information systems , pages =

  12. [12]

    2014 IEEE international conference on bioinformatics and biomedicine (BIBM) , pages =

    BPMN4CP: Design and implementation of a BPMN extension for clinical pathways , author =. 2014 IEEE international conference on bioinformatics and biomedicine (BIBM) , pages =

    2014

  13. [13]

    ACM Transactions on Management Information Systems (TMIS) , volume =

    Blockchains for business process management-challenges and opportunities , author =. ACM Transactions on Management Information Systems (TMIS) , volume =

  14. [14]

    Journal on Data Semantics , volume =

    Knowledge-intensive processes: characteristics, requirements and analysis of contemporary approaches , author =. Journal on Data Semantics , volume =

  15. [15]

    2012 , publisher =

    Enabling flexibility in process-aware information systems: challenges, methods, technologies , author =. 2012 , publisher =

    2012

  16. [16]

    International Journal of Business Process Integration and Management , volume =

    Contextualisation of business processes , author =. International Journal of Business Process Integration and Management , volume =

  17. [17]

    Visual Studio Code , author =

  18. [18]

    International conference on advanced information systems engineering , pages =

    Extracting decision logic from process models , author =. International conference on advanced information systems engineering , pages =

  19. [19]

    2025 , url =

    Jiayi Zhang and Jinyu Xiang and Zhaoyang Yu and Fengwei Teng and Xiong-Hui Chen and Jiaqi Chen and Mingchen Zhuge and Xin Cheng and Sirui Hong and Jinlin Wang and Bingnan Zheng and Bang Liu and Yuyu Luo and Chenglin Wu , booktitle =. 2025 , url =

    2025

  20. [20]

    CoRR , volume =

    Guibin Zhang and Kaijie Chen and Guancheng Wan and Heng Chang and Hong Cheng and Kun Wang and Shuyue Hu and Lei Bai , title =. CoRR , volume =. 2025 , month =

    2025

  21. [21]

    Forty-second International Conference on Machine Learning , year =

    Multi-agent Architecture Search via Agentic Supernet , author =. Forty-second International Conference on Machine Learning , year =

  22. [22]

    Journal on data semantics , volume =

    Automated planning for business process management , author =. Journal on data semantics , volume =

  23. [23]

    Advances in neural information processing systems , volume =

    Chain-of-thought prompting elicits reasoning in large language models , author =. Advances in neural information processing systems , volume =

  24. [24]

    Communications of the ACM , volume =

    Process modeling , author =. Communications of the ACM , volume =. 1992 , publisher =

    1992

  25. [25]

    arXiv preprint arXiv:2303.04129 , year =

    Foundation models for decision making: Problems, methods, and opportunities , author =. arXiv preprint arXiv:2303.04129 , year =

    Pith/arXiv arXiv

  26. [26]

    First Conference on Language Modeling , year =

    Autogen: Enabling next-gen LLM applications via multi-agent conversations , author =. First Conference on Language Modeling , year =

  27. [27]

    Nature , volume =

    Large language models encode clinical knowledge , author =. Nature , volume =

  28. [28]

    arXiv preprint arXiv:2302.07842 , year =

    Augmented language models: a survey , author =. arXiv preprint arXiv:2302.07842 , year =

    Pith/arXiv arXiv

  29. [29]

    Business Process Management Journal , volume =

    Opportunities and constraints: the current struggle with BPMN , author =. Business Process Management Journal , volume =. 2010 , publisher =

    2010

  30. [30]

    East European conference on advances in databases and information systems , pages =

    Towards automated performance optimization of BPMN business processes , author =. East European conference on advances in databases and information systems , pages =. 2016 , organization =

    2016

  31. [31]

    Information Systems , volume =

    BPMN Miner: Automated discovery of BPMN process models with hierarchical structure , author =. Information Systems , volume =. 2016 , publisher =

    2016

  32. [32]

    ACM Transactions on Autonomous and Adaptive Systems (TAAS) , volume =

    Supporting dynamic workflows with automatic extraction of goals from BPMN , author =. ACM Transactions on Autonomous and Adaptive Systems (TAAS) , volume =. 2019 , publisher =

    2019

  33. [33]

    arXiv preprint arXiv:2505.06120 , year =

    Llms get lost in multi-turn conversation , author =. arXiv preprint arXiv:2505.06120 , year =

    Pith/arXiv arXiv

  34. [34]

    International Conference on Business Process Modeling, Development and Support , pages =

    Process modeling with large language models , author =. International Conference on Business Process Modeling, Development and Support , pages =. 2024 , organization =

    2024

  35. [35]

    arXiv preprint arXiv:2408.01916 , year =

    Mao: A framework for process model generation with multi-agent orchestration , author =. arXiv preprint arXiv:2408.01916 , year =

    Pith/arXiv arXiv

  36. [36]

    Business Process Model and Notation (BPMN) , url =

  37. [37]

    Communications of the ACM , pages =

    Van Der Aalst, Wil , title =. Communications of the ACM , pages =. 2012 , volume =

    2012

  38. [38]

    Nature Machine Intelligence , volume =

    Explaining machine learning models with interactive natural language conversations using TalkToModel , author =. Nature Machine Intelligence , volume =

  39. [39]

    International Conference on Learning Representations (ICLR) , year =

    ReAct: Synergizing Reasoning and Acting in Language Models , author =. International Conference on Learning Representations (ICLR) , year =

  40. [40]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Toolformer: Language Models Can Teach Themselves to Use Tools , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  41. [41]

    International Conference on Learning Representations (ICLR) , year =

    PAL: Program-Aided Language Models , author =. International Conference on Learning Representations (ICLR) , year =

  42. [42]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Gorilla: Large Language Model Connected with Massive APIs , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  43. [43]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    HuggingGPT: Solving AI Tasks with ChatGPT and its Friends in Hugging Face , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  44. [44]

    arXiv preprint arXiv:2405.04215 , year =

    NL2Plan: Robust LLM-Driven Planning from Minimal Text Descriptions , author =. arXiv preprint arXiv:2405.04215 , year =

    arXiv

  45. [45]

    International Conference on Learning Representations (ICLR) , year =

    Query-Efficient Planning with Language Models , author =. International Conference on Learning Representations (ICLR) , year =

  46. [46]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  47. [47]

    Business & Information Systems Engineering (BISE) , year =

    Retrieval-Augmented Generation (RAG) , author =. Business & Information Systems Engineering (BISE) , year =. doi:10.1007/s12599-025-00945-3 , url =

    work page doi:10.1007/s12599-025-00945-3

  48. [48]

    International Conference on Learning Representations (ICLR) , year =

    Memorizing Transformers , author =. International Conference on Learning Representations (ICLR) , year =

  49. [49]

    International Conference on Learning Representations (ICLR) , year =

    Ring Attention with Blockwise Transformers for Near-Infinite Context , author =. International Conference on Learning Representations (ICLR) , year =

  50. [50]

    Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL) , year =

    Longformer: The Long-Document Transformer , author =. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL) , year =

  51. [51]

    International Conference on Machine Learning (ICML) , year =

    Stabilizing Transformers for Reinforcement Learning (GTrXL) , author =. International Conference on Machine Learning (ICML) , year =

  52. [52]

    NeurIPS Datasets and Benchmarks Track , year =

    Mind2Web: Towards a Generalist Agent for the Web , author =. NeurIPS Datasets and Benchmarks Track , year =

  53. [54]

    International Conference on Learning Representations (ICLR) , year =

    AgentBench: Evaluating LLMs as Agents , author =. International Conference on Learning Representations (ICLR) , year =

  54. [55]

    ACM Transactions on Management Information Systems (TMIS) , year =

    Process Mining: Overview and Opportunities , author =. ACM Transactions on Management Information Systems (TMIS) , year =

  55. [56]

    2016 , url =

    Process Mining: Discovery, Conformance and Enhancement of Business Processes , author =. 2016 , url =

    2016

  56. [57]

    Business Process Management (BPM) , year =

    BPMN 2.0 Execution Semantics Formalized as Graph Rewrite Rules , author =. Business Process Management (BPM) , year =

  57. [58]

    International Journal of Computer Science (IAENG) , volume =

    A Precise Execution Semantics for BPMN , author =. International Journal of Computer Science (IAENG) , volume =. 2012 , url =

    2012

  58. [59]

    Business Process Management (BPM) , year =

    BPEL to BPMN: The Myth of a Straight-Forward Mapping , author =. Business Process Management (BPM) , year =

  59. [60]

    Business Process Management Journal , year =

    Creating AI Business Value through BPM Capabilities , author =. Business Process Management Journal , year =. doi:10.1108/BPMJ-07-2023-0566 , url =

    work page doi:10.1108/bpmj-07-2023-0566 2023

  60. [61]

    Information Systems , year =

    Business Process Management in the Age of AI—Three Complementarities , author =. Information Systems , year =. doi:10.1007/s10257-024-00689-9 , url =

    work page doi:10.1007/s10257-024-00689-9

  61. [62]

    Decision Support Systems , year =

    Large Language Models present new questions for Decision Support , author =. Decision Support Systems , year =. doi:10.1016/j.dss.2024.114156 , url =

    arXiv 2024

  62. [63]

    arXiv preprint arXiv:2403.07541 , year =

    Process Modeling with Large Language Models , author =. arXiv preprint arXiv:2403.07541 , year =

    Pith/arXiv arXiv

  63. [64]

    arXiv preprint arXiv:2412.00023 , year =

    Evaluating Large Language Models on Business Process Modeling: Framework, Benchmark, and Self-Improvement Analysis , author =. arXiv preprint arXiv:2412.00023 , year =

    Pith/arXiv arXiv

  64. [65]

    Introducing the

    K. Introducing the. Proceedings of the Best BPM Dissertation Award, Doctoral Consortium, and Demonstrations & Resources Forum co-located with the 22nd International Conference on Business Process Management (BPM 2024) , series =. 2024 , address =

    2024

  65. [66]

    Cooperative Information Systems , editor =

    Nour Eldin, Ali and Assy, Nour and Anesini, Olan and Dalmas, Benjamin and Gaaloul, Walid , title =. Cooperative Information Systems , editor =. 2025 , publisher =. doi:10.1007/978-3-031-81375-7_27 , url =

    work page doi:10.1007/978-3-031-81375-7_27 2025

  66. [67]

    Computational Science and Computational Intelligence , year =

    Toxtli, Carlos and Li, Wangfan , title =. Computational Science and Computational Intelligence , year =. doi:10.1007/978-3-031-86623-4_18 , url =

    work page doi:10.1007/978-3-031-86623-4_18

  67. [68]

    Berti, Alessandro and Kourani, Humam and van der Aalst, Wil M. P. , title =. arXiv preprint arXiv:2407.13244 , year =

    Pith/arXiv arXiv

  68. [69]

    arXiv preprint arXiv:2405.12842 , year =

    Jain, Arushi and Paliwal, Shubham and Sharma, Monika and Vig, Lovekesh and Shroff, Gautam , title =. arXiv preprint arXiv:2405.12842 , year =

    Pith/arXiv arXiv

  69. [70]

    Proceedings of the 4th ACM International Conference on AI in Finance (ICAIF '23) , year =

    Zeng, Zhen and Watson, William and Cho, Nicole and Rahimi, Saba and Reynolds, Shayleen and Balch, Tucker and Veloso, Manuela , title =. Proceedings of the 4th ACM International Conference on AI in Finance (ICAIF '23) , year =. doi:10.1145/3604237.3626908 , url =

    arXiv

  70. [71]

    arXiv preprint arXiv:2311.10751 , year =

    Ye, Yining and Cong, Xin and Tian, Shizuo and Cao, Jiannan and Wang, Hao and Qin, Yujia and Lu, Yaxi and Yu, Heyang and Wang, Huadong and Lin, Yankai and Liu, Zhiyuan and Sun, Maosong , title =. arXiv preprint arXiv:2311.10751 , year =

    Pith/arXiv arXiv

  71. [72]

    and Lin, Stephanie and Lee, Philip and Hendrycks, Dan and Hsu, Kenny and Ramaswamy, Omkar Vishwanath and Bhatia, Karthik and others , title =

    Patil, Shishir G. and Lin, Stephanie and Lee, Philip and Hendrycks, Dan and Hsu, Kenny and Ramaswamy, Omkar Vishwanath and Bhatia, Karthik and others , title =. Advances in Neural Information Processing Systems (NeurIPS 2024) , year =

    2024

  72. [73]

    Proceedings of the 41st International Conference on Machine Learning , series =

    Zhou, Andy and Yan, Kai and Shlapentokh-Rothman, Michal and Wang, Haohan and Wang, Yu-Xiong , title =. Proceedings of the 41st International Conference on Machine Learning , series =. 2024 , publisher =

    2024

  73. [74]

    arXiv preprint arXiv:2409.07429 , year =

    Wang, Zora Zhiruo and Mao, Jiayuan and Fried, Daniel and Neubig, Graham , title =. arXiv preprint arXiv:2409.07429 , year =

    Pith/arXiv arXiv

  74. [75]

    Advances in Neural Information Processing Systems (NeurIPS 2023) , year =

    Shinn, Noah and Cassano, Federico and Berman, Edward and Gopinath, Ashwin and Narasimhan, Karthik and Yao, Shunyu , title =. Advances in Neural Information Processing Systems (NeurIPS 2023) , year =

    2023

  75. [76]

    Findings of the Association for Computational Linguistics: EMNLP 2024 , pages =

    Xiao, Hangyu and Huang, Buqing and Zhu, Jiaqi and Wang, Zhiheng and Fan, Liang and Tang, Junyou and Zhou, Xin and Li, Jiwei and Jiang, Yong and Sun, Hongzhi and others , title =. Findings of the Association for Computational Linguistics: EMNLP 2024 , pages =. 2024 , month =. doi:10.18653/v1/2024.findings-emnlp.615 , url =

    work page doi:10.18653/v1/2024.findings-emnlp.615 2024

  76. [77]

    NeurIPS 2023 Datasets and Benchmarks Track , year =

    Deng, Xiang and Gu, Yu and Zheng, Boyuan and Chen, Shijie and Stevens, Samuel and Wang, Boshi and Sun, Huan and Su, Yu , title =. NeurIPS 2023 Datasets and Benchmarks Track , year =

    2023

  77. [78]

    and Zhu, Hao and Zhou, Xuhui and Lo, Robert and Sridhar, Abishek and Cheng, Xianyi and Ou, Tianyue and Bisk, Yonatan and Fried, Daniel and Alon, Uri and Neubig, Graham , title =

    Zhou, Shuyan and Xu, Frank F. and Zhu, Hao and Zhou, Xuhui and Lo, Robert and Sridhar, Abishek and Cheng, Xianyi and Ou, Tianyue and Bisk, Yonatan and Fried, Daniel and Alon, Uri and Neubig, Graham , title =. arXiv preprint arXiv:2307.13854 , year =

    Pith/arXiv arXiv

  78. [79]

    and Cao, Yuan and Narasimhan, Karthik , title =

    Yao, Shunyu and Yu, Dian and Zhao, Jeffrey and Shafran, Izhak and Griffiths, Thomas L. and Cao, Yuan and Narasimhan, Karthik , title =. Advances in Neural Information Processing Systems (NeurIPS 2023) , year =

    2023

  79. [80]

    Proceedings of the 40th International Conference on Machine Learning (ICML) , series =

    Gao, Luyu and Madaan, Aman and Zhou, Shuyan and Alon, Uri and Liu, Pengfei and Yang, Yiming and Callan, Jamie and Neubig, Graham , title =. Proceedings of the 40th International Conference on Machine Learning (ICML) , series =. 2023 , publisher =

    2023

  80. [81]

    and Zhang, Tianjun and Wang, Xin and Gonzalez, Joseph E

    Patil, Shishir G. and Zhang, Tianjun and Wang, Xin and Gonzalez, Joseph E. , title =. Advances in Neural Information Processing Systems (NeurIPS 2024) , year =

    2024

Showing first 80 references.