Beyond Next-Token Prediction: An RLVR Proof of Concept for Tool-Use Agents on Atlassian Workflows
06:00 · July 3, 2026 · arXiv cs.AI RSS
Large language models are trained to predict the next token, not to act inside a specific API. In niche enterprise SaaS workflows -- where success means hitting the right endpoint with the right nested arguments in the right order -- this objective mismatch shows up as silent failures: dropped required fields, hallucinated tools, or early stops after a single read. We ask whether Reinforcement Learning with Verifiable Rewards (RLVR), applied directly in the target environment, closes the gap. As a proof of concept we build a suite of five synthetic environments emulating the Jira REST v3 and Confluence v2 APIs at schema fidelity; rewards are computed entirely from the tool-call trace, with no live API, no learned judge, and no human label in the loop. Scoring prompted Qwen3-1.7B and Qwen3.5-4B on the same checkers that drive GRPO training, we find that on the four scenarios whose rewards are non-degenerate the RL-trained policy lifts average reward from a 4B-baseline range of 0.35--0.92 to 0.95--1.00, with the largest single gain on Confluence page creation ($0.35 \rightarrow 1.00$). We position this as a preliminary step toward outcome-optimised small models for niche enterprise APIs, and foreground two limitations a workshop reader should weigh: hand-crafting verifiable rewards does not scale beyond the handful of endpoints reported here, and one of our five scenarios (ticket-transition) has a saturating reward shape that the prompted 4B already maxes out.
Summary
Large language models are trained primarily for next-token prediction on broad text corpora, which leaves them poorly aligned with the precise, schema-driven requirements of enterprise SaaS tool use. In workflows such as creating Jira issues or Confluence pages, this mismatch produces characteristic failures: omitted required arguments, invented tool names, or premature termination after an initial read operation. The authors investigate whether Reinforcement Learning with Verifiable Rewards (RLVR) can close the gap when applied directly inside synthetic replicas of the target APIs.
They constructed five schema-faithful environments that emulate selected Jira REST v3 and Confluence v2 endpoints, including issue creation, sub-task linking, page creation, and ticket transitions. Each environment maintains stateful resource pools with parent–child constraints yet resets deterministically before every rollout, allowing reward computation solely from the sequence of tool calls and their argument values. No live services, learned reward models, or human annotations are required. Training employed Group Relative Policy Optimization (GRPO) on the Qwen3-1.7B and Qwen3.5-4B models, with the same programmatic checkers used both for reward signals and for final evaluation.
On the four scenarios whose reward functions are non-degenerate, the RL-trained policies raised average reward from a prompted 4B baseline range of 0.35–0.92 to 0.95–1.00. The largest single improvement occurred on Confluence page creation, where the score rose from 0.35 to 1.00. These gains were achieved in tens to low hundreds of generation batches using only synthetic data.
The work is presented as an early demonstration that outcome-level optimization can produce compact, task-specific agents for narrow enterprise surfaces. At the same time, the authors explicitly note two constraints: manually engineered verifiable rewards do not extend readily beyond the small set of endpoints tested, and one scenario (ticket transition) already saturates under the prompted 4B baseline, limiting observable improvement.
Why it matters
It provides actionable research for Dutch AI practitioners developing autonomous agents for enterprise workflows, a key area for SME AI adoption. Furthermore, the focus on optimizing smaller models aligns with European data privacy preferences and cost-effective local deployments.
