WorkGenesis: Building the Worlds That Teach Agents to Work

1 Introduction

• Scaling credible occupational scenarios is difficult because settings require interconnected materials, coherent business context, defensible requirements, and costly expert review.Unconstrained synthesis can lack grounding or conceal contradictions among evidence, objectives, and evaluation criteria.Source: §1 · p. 2
• Fx-Work-35B leads comparable-scale baselines on all five reported metrics, raising GDPvalAA-v2 Elo from 1053 to 1352 and surpassing DeepSeek-V4-Pro-Preview.APEX-Agents pass@1 rises from 13.4% to 25.2%, and JobBench Main/Easy scores rise from 14.99/56.02 to 25.19/70.06.Source: §1 · pp. 2–3
• The evaluation trains Fx-Work-35B with tool-interaction trajectories collected from 20K units of occupational work and tests it on GDPval, APEX-Agents, and JobBench.The benchmarks represent economically valuable work, long-horizon cross-application work, and work professionals prioritize for delegation, respectively.Source: §1 · p. 2
• WorkGenesis combines evidence-based construction from real files and occupational knowledge with execution-guided checks that expose and repair task or rubric defects.The framework assembles context, materials, objectives, and acceptance criteria around real records; reference execution tests feasibility, professional reasonableness, consistency, and difficulty.Source: §1 · pp. 2–3
• The paper analyzes synthesized task-data properties and model cost-effectiveness, and open-sources Fx-Work-35B.The authors describe the release as intended to advance working-agent development.Source: §1 · p. 3

2 Related Work

• WorkArena and OSWorld (Xie et al., 2024) evaluate browser and computer-use tasks, while GDPval (Patwardhan et al., 2026), APEX-Agents (Vidgen et al., 2026), and JobBench (Li et al., 2026) target professional work.These benchmarks primarily evaluate agents rather than provide a scalable construction method for diverse, file-backed training scenarios with auditable requirements.Source: §2 · p. 3
• Instruction synthesis and environment-generation methods expand training data, whereas WorkGenesis starts from public artifacts and organizes them into workplace scenarios with requests, companion files, and rubrics.The evidence-first construction keeps scenarios connected to real work while allowing surrounding materials to be generated automatically.Source: §2 · p. 3
• WorkGenesis uses reference execution to test scenario consistency: unsatisfied rubric items are attributed to the agent, task, or rubric, and task or rubric defects are repaired before acceptance.Unlike methods that improve an answer or later agent behavior, this process checks whether the prompt, files, and criteria describe the same professional activity.Source: §2 · p. 3

3.1 Overview and Problem Formulation

• Professional-work units combine a request, real and synthesized reference files, and an itemwise rubric; construction must address weak grounding and cross-material inconsistencies.Source: §3.1 · p. 4

3.2 Evidence-Based Work Construction

• O*NET-guided search retrieves occupation-related public documents as anchors, then WorkGenesis synthesizes companion attachments, a workplace request, and rubric into a complete scenario.Source: §3.1 · p. 4, §3.2 · pp. 4–5
• An independent agent sees only the prompt, reference files, and tools; its deliverable tests whether the materials support the request and the rubric can be applied itemwise.Source: §3.2 · pp. 5–6

3.3 Execution-Guided Consistency Verification

• Each criterion is scored against the deliverable and materials, and failures are attributed to Agent, Rubric, or Task; agent capability failures are recorded but left unchanged.Source: §3.3 · p. 6
• Task- or rubric-attributed defects drive revisions to the request, companion attachments, and rubric; verification repeats until none remain, or the scenario is discarded after three failed rounds.Source: §3.3 · p. 6

4.1 Experimental Setup

• From 65 O*NET occupations, WorkGenesis produced 20K tasks and trajectories for SFT; evaluation used GDPval, APEX-Agents, and JobBench.The comparison included frontier models and ≤35B models as comparable-scale baselines.Source: §4.1 · p. 7
• 4 of 5 metrics: Fx-Work-35B surpasses 1.6T DeepSeek-V4-Pro-Preview, while beating all comparable-scale models on all three benchmarks.GDPvalAA-v2 rubric and Elo scores reach 86.66 and 1352; APEX-Agents-AA pass@1 reaches 25.2%; JobBench Main and Easy scores are 25.19 and 70.06.Source: §4.1 · pp. 7–8, Table 1 · p. 8
• 31.00 average score at $0.77 per task; Fx-Work-35B costs 36% less than similarly scoring DeepSeek-V4-Pro-Preview and scores 74% higher than similarly priced Nex-N2-mini.The reported differences are $0.44 per task and 13.26 points, respectively.Source: §4.1 · p. 8, Figure 4 · p. 9

4.2 Discussion and Analysis

• After training, average turns rose to 44.6, 44.8, and 89.2 across the three benchmarks; simple tasks were solved faster while complex long-horizon tasks were handled more meticulously.The authors report relatively balanced overall turn distributions despite variation across tasks.Source: §4.1 · pp. 8–9, Figure 6 · p. 10
• With identical training hyperparameters, Elo rose steadily as WorkGenesis data grew from 5K to 20K scenarios; rubric scores dipped slightly at 5K before trending upward.The authors attribute the initial rubric dip to tasks exceeding the prescribed turn budget, while Elo is less affected because it emphasizes completed-task quality.Source: §4.2 · pp. 9–10, Figure 5 · p. 9
• WorkGenesis tasks mainly use PDFs, spreadsheets, and Word documents, while also including specialist reference files such as .step and .psd.Source: §4.2 · p. 10, Figure 7 · p. 10
• Rubric and Task defects fall from 1.66 and 0.92 after the initial version to zero after three verification rounds, and accepted work has zero defective items.Source: §4.2 · pp. 10–11, Figure 8 · p. 10

5 Conclusion

• Fx-Work-35B leads comparable-scale baselines on every reported metric and surpasses the 1.6T DeepSeek-V4-Pro-Preview on four of five.GDPval Elo rises from 1053 to 1352, APEX-Agents pass@1 from 13.4% to 25.2%, and JobBench Main/Easy from 14.99/56.02 to 25.19/70.06.Source: §5 · p. 11
• WorkGenesis retrieves public anchor files using O*NET occupations, builds requests, companion materials, and itemized rubrics, then repairs task and rubric defects exposed by reference execution.The resulting 20K work units span 65 occupations and provide trajectories for training Fx-Work-35B.Source: §5 · p. 11
• The authors identify expansion to more occupations and reinforcement learning as promising next steps.They conclude that evidence-grounded, executable work is a practical source of training data for working agents.Source: §5 · p. 11

AI use statement

• The authors report that AI-assisted outputs underwent execution-based verification, item-level rubric checking, manual inspection, and reproducible evaluation scripts.They take responsibility for the paper's final content, claims, code, data, and artifacts.Source: AI use statement · p. 12

A Implementation Details

• The section provides details on WorkGenesis synthesis, Fx-Work-35B training, and evaluation.Source: A Implementation Details · p. 16

A.1 Synthesis Details

• GLM-5.2 drives each WorkGenesis module; synthesis uses 65 O*NET occupations, four fixed tools, Serp API for web search and fetch, and Stirrup with GLM-5.2 to render deliverables.The tools are web_search, web_fetch, code_exec, and view_image.Source: §A.1 · p. 16

A.2 Training Details

• Qwen3.6-35B-A3B is trained with standard supervised fine-tuning; the training loss curve and hyperparameters are presented in Figure 9 and Table 3.Source: §A.2 · p. 16

GDPvalAA-v2

• GDPvalAA-v2 uses the official Artificial Analysis criteria and Stirrup harness; GPT-5.5 and Gemini-3.6-flash score each deliverable, and their scores are averaged.For ELO, the authors use the August 4, 2026 Artificial Analysis snapshot for previously evaluated models and compute scores for the others using that snapshot, their results on 15 models, and Artificial Analysis’s ELO model.Source: §A.3 · p. 17

APEX-Agents-AA

• APEX-Agents-AA evaluation covers 452 tasks across Law, Management Consulting, and Investment Banking, with an attempt passing only when every binary rubric criterion is met.The two Investment Banking worlds using real-time external data are excluded; each task is repeated three times and pass@1 is the fraction of passing attempts among all 452 × 3 attempts.Source: §A.3 · p. 17

JobBench

• JobBench deliverables are scored with GPT-5.5 and Gemini-3.6-flash against the task prompt, parsed attachments, and rubric, then averaged.Evaluation uses Stirrup as the agent harness.Source: §A.3 · p. 17

Aggregation Score

• The Aggregation Score averages normalized GDPvalAA-v2 ELO, APEX-Agents-AA pass rate, and the JobBench Main score.Source: §A.3 · p. 17

• The 20K synthesized occupational tasks are evenly distributed across 65 occupations.Figure 10 depicts their distribution.Source: B Analysis and results · p. 17

• The section presents two specific synthesized occupational tasks, but this passage does not describe their details.Source: C Case Study · p. 17
• Figure 10 depicts how the 20K constructed work units are distributed across 65 occupations.Source: Figure 10 · p. 18

Task scope and deliverable

• The task requests an engineer-ready Excel decision workbook for a Line 4 roll-former retrofit, using 12 reference files and five specified working sheets.The files include drawings, data, constraints, and OSHA booklets; the deliverable is for deciding whether a concept can move into detailed design.Source: §C.1 · pp. 18, 19, 21

Decision summary

• The Decision Summary must conditionally advance Concept A, hold B, and reject C, while stating cost, schedule, residual risk, and release conditions.Concept B’s 14.5 in available separation is about 20.3 in short of the preliminary 34.8 in screen; Concept C leaves hazards physically reachable.Source: §C.1 · pp. 19, 22, 24

Risk and job hazard analysis

• Risk & JHA must cover threading, running, debris clearing, service, and sharp-edge handling, with formula-driven initial and residual risk and accountable verification.The guide-roll nip starts at likelihood 4 and severity 5 (risk 20); the exposed chain is a separate hazard, and residual risks must meet the internal target of 6 or below.Source: §C.1 · pp. 19, 23

Guard geometry and preliminary screening

• Guard Geometry calculates the 54.0 in fixed-guard span from a 52.0 in face and two 1.0 in returns, checks the feed opening and clearances, and computes required separation for the hinged option.The 0.25 in feed opening is checked against a 0.25 in maximum, and both 1.25 in clearances against a 1.00 in minimum. With a 0.42 s maximum measured stop time, the supplied equation 63 × (maximum stop time + 0.10 s) + 2.0 gives about 34.76 in required separation.Source: §C.1 · pp. 19, 23

Concept comparison and budget

• Concept Comparison must score all three concepts on six stated criteria using the weighted model 35%, 20%, 15%, 10%, 10%, and 10%.The weighted result must be formula-based, visibly rank A above B and C, and show the technical basis for lower ratings.Source: §C.1 · pp. 19, 24

Action plan and release gate

• Action & Verification must assign owners, dates, and acceptance evidence, including a no-go rule against restart before required checks are accepted.Checks include feed opening at or below 0.25 in, clearance at or above 1.00 in, covered chain drive, secure tool-only hardware, and no new pinch or shear point.Source: §C.1 · pp. 19, 25

Scope and safety caveats

• The workbook must distinguish internal planning screens from regulatory requirements and retain EHS and machine-safety review as the final authority.It must not claim OSHA compliance, certification, or final regulatory approval, or substitute unsupported safety-distance, opening, or clearance rules.Source: §C.1 · pp. 19, 23, 25

Concept comparison and budget

• Concept A’s estimated cost is about $8,878, B’s about $13,582, and C’s about $1,576; A fits the $10,000 ceiling and four-hour outage, while B exceeds both constraints.The workbook must provide cost build-up, contingency, capital variance, lead time, and outage for each concept.Source: §C.1 · p. 24

• The acceptance rubric has 43 items, with 37 positive and 6 penalty rows; 117 required points form the pass threshold within 121 positive points.The threshold means required points cannot be treated as a flat checklist.Source: §C.2 · pp. 28–30
• The task asks for a 24–28-second Acadia feature-page highlight from Shirley Beccue’s interview, aimed at visitors arriving cold and encouraging them to watch the full interview.Only the supplied video is approved for picture and sound.Source: §C.2 · p. 28
• The export must be a playable MP4 at 854×480 and 23.976 fps, with progressive H.264 video, yuv420p, and 48 kHz stereo AAC audio.The cut must preserve the source footage’s proportions and keep dialogue intelligible and in sync.Source: §C.2 · pp. 28, 30
• The deliverable must include the exact opening title, Beccue identification, and closing copy, with accurate burned-in captions kept readable and clear of the identification.The title appears within two seconds, and the closing copy remains legible for at least two seconds.Source: §C.2 · pp. 30–31
• The edit must build around Beccue’s reflection that art touches emotions and preserve her complete line about being invited into a place through painting, music, or dance.The reflection should resolve into a complete thought without reordered or synthetically altered dialogue.Source: §C.2 · pp. 28, 31

D.1 Stage 1: Discovering Real-World Anchors

• Stage 1 uses occupation knowledge and lineage memory to guide searches toward realistic, non-reused files and records accepted anchors in metadata.At least one usable real file is required, and the prompt prefers two to four strong, on-occupation files.Source: §D.1 · p. 33
• Anchor discovery searches and inspects public files, accepting only usable artifacts that meet occupational fit, substance, veracity, novelty, solvability, and safety checks.Accepted files are kept unchanged, with their provenance and suitability recorded; failed searches are reported rather than fabricated.Source: §D.1 · pp. 33–35

D.2 Stage 2a: Synthesizing the Companion Attachments

• Stage 2a treats the parsed real files as authoritative evidence, keeps them unchanged, and documents synthetic companions’ provenance and assumptions in a manifest.The combined real and generated package is placed in the reference set for later stages.Source: §D.2 · pp. 36–38
• Companion attachments fill scenario-specific gaps in real files, adding only the minimum grounded context needed to support realistic professional work.The prompt cautions against padding the package or adding unrelated material.Source: §D.2 · p. 36
• The package should prompt work across multiple files where appropriate, with generated materials reusing real entities and avoiding unfair contradictions.The workflow reopens and parses generated files, then checks internal consistency, cross-file references, dates, totals, and task sufficiency.Source: §D.2 · pp. 36–37

D.3 Stage 2b: Writing the Work Request and the Rubric

• Stage 2b begins with a professional task author preparing the work request and rubric after the reference files for the scenario already exist.The supplied passage ends before describing further instructions for this stage.Source: §D.3 · p. 38
• The work request should pose a demanding, occupation-specific workplace task grounded in the attachments, requiring substantive analysis rather than mere summarization or transcription.It should establish a concrete role, situation, audience, and sufficient context while keeping detailed verification expectations in the rubric.Source: §D.3 · pp. 38–39
• The request’s style and deliverables should fit the scenario, with one strong deliverable usually preferred and public research requested only when the occupation calls for it.When attachments are blank or partial templates, request a concrete product that uses them as a reference structure.Source: §D.3 · p. 39
• The rubric should balance scores by importance and use concrete, decidable criteria covering the content, computation, structure, and compliance required by the specific task.It may include a small number of negative-score prohibitions and should avoid criteria judgeable only from a hidden gold deliverable.Source: §D.3 · pp. 39–40
• Before finalizing, check the request and rubric against parsed files so their facts, deliverable formats, scope, scoring, and item coverage remain mutually consistent.Do not demand unsupported content unless the request calls for professional composition or bounded public research.Source: §D.3 · pp. 39–40

D.4 Stage 3: Rendering the Reference Deliverable

• The execution agent sees the task request and reference files, follows the request, and treats the files as authoritative for facts, using outside knowledge only when permitted.The agent is instructed to make reasonable assumptions when information is missing and record them when submitting.Source: §D.4 · p. 41
• The agent must write deliverable files in the designated deliverable directory and submit their absolute file paths with a brief accomplishment summary using the finish tool.The instructions specify submitting files rather than folders.Source: §D.4 · p. 41
• Render a complete, high-quality deliverable in the requested file types, not an outline or restatement; when ambiguity remains, proceed with a reasonable reading and record assumptions.Record assumptions that materially affect numbers, classifications, or recommendations, as well as evidence that could not be found.Source: §D.4 · p. 41

D.5 Stage 4: Scoring the Deliverable

• Inspect actual output files for rubric-required format and artifact properties, and give no credit when required audio or video evidence is absent from supplied materials.Check items such as filenames, structure, formulas, charts, citations, and links when required.Source: §D.5 · pp. 42–43
• Score each rubric item independently against observable evidence, award partial credit under its positive or negative rule, and give concrete, traceable rationales.Use the rubric rather than a hidden gold deliverable; when files conflict about their contents, the reference file is authoritative.Source: §D.5 · pp. 42–43

D.6 Stage 5: Attributing the Unsatisfied Items

• Classify each unsatisfied item as Agent when a fair, grounded requirement was missed, Rubric when its criterion is defective, or Task when the scenario makes it unsatisfiable or unfair.An item may receive multiple causes when they overlap.Source: §D.6 · p. 43
• Repair only items labeled Rubric or Task, leaving Agent-labeled capability gaps untouched because they are not scenario defects.Misattribution either hides a real defect or changes a scenario that was already working.Source: §D.6 · p. 43
• For attribution, verify the scorer’s rationale against the deliverable and test whether the requirement is decidable from the request or files; when unsupported, favor Rubric or Task over Agent.Combine labels only when causes overlap, and explain any tie-break.Source: §D.6 · pp. 44–45

D.7 Stage 6: Updating and Re-Auditing the Scenario

• Re-audit repairs until no rubric or task defects remain; if the same defect persists for three rounds, discard and rebuild the scenario while recording source URLs and hashes.Repairs may change the request, companion attachments, and rubric, but not the real reference files.Source: §D.7 · p. 45
• Keep the real public reference files immutable during repair; update only the request, companion attachments, or rubric, and leave Agent-only items unchanged.A repair must remove the defect itself because the updated scenario is scored again.Source: §D.7 · p. 45
• Group defects by whether they belong to the rubric, request, companion files, or their interactions, then repair rubric and task defects at their source.Keep rubric coverage and score balance intact, align request and rubric scope, and adapt to fixed real files when they conflict.Source: §D.7 · p. 46
• Re-derive facts affected by each change, re-check every linked criterion and request statement, re-parse regenerated files, recompute SHA-256, and confirm the deliverable contract remains satisfiable.State whether the deliverable must be rendered again; leave defect-free scenario parts unchanged.Source: §D.7 · p. 46
• Choose “accepted” when no defective item will remain, “updated” for repairable defects before the third failed round, or “discarded” when the same unresolved item reaches that round.Use stop reasons “defect_set_empty,” “reaudit,” and “unresolved_after_three_rounds,” respectively.Source: §D.7 · p. 46
• Record the round status, stop reason, addressed defects, changes, files changed and verified, rubric indices, rerender requirement, unresolved defects, and next-pass notes in repair_log.json.If a defect cannot be fixed without contradicting a real file, mark the scenario for reconsideration rather than forcing a repair.Source: §D.7 · pp. 46–47
• Update lineage memory on every outcome, carrying each accepted anchor’s domain, source and download page URLs, canonical URL, and SHA-256 into the update.Include the final scenario summary, outcome, rounds taken, and discard reason when applicable.Source: §D.7 · p. 47
• Leave anchors unchanged, repair only items labeled Rubric or Task, and make the smallest coherent fix without changing Agent-only items.When changing the request, retain its workplace-request style and keep the rubric's mixed-score style and penalty prohibitions.Source: §D.7 · pp. 46–47