Beyond Next-Token Prediction: An RLVR Proof of Concept for Tool-Use Agents on Atlassian Workflows

Jul 16, 2026

Upload by Violet

Large language models are trained to predict the next token, yet enterprise SaaS workflows require correct tool calls with the right nested arguments and ordering. This mismatch can cause silent failures such as missing required fields, hallucinated tools, and premature stopping after early reads. The work evaluates Reinforcement Learning with Verifiable Rewards (RLVR) using synthetic, schema-faithful Jira REST v3 and Confluence v2 environments, with rewards computed from tool-call traces only. Prompted Qwen3 models improve average reward from a 4B baseline range of 0.35–0.92 to 0.95–1.00 on non-degenerate scenarios, highlighting a path toward outcome-optimized small models.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages6

Reading Time15 min

AccessFree to download

Introduction

Objective mismatch in next-token training

RLVR and verifiable reward design

Contribution

Synthetic Atlassian environments

Verifiable reward functions

Prompted baselines and RLVR training

Beyond Next-Token Prediction: An RLVR Proof of Concept for Tool-Use Agents on Atlassian Workflows Karthikeya Aditya Vissa∗ Centific [karthikeyaaditya.v@centific.com](karthikeyaaditya.v@centific.com) Sankalp Mane∗ Centific [sankalp.mane@centific.com](sankalp.mane@centific.com) Ananya Mantravadi Centific[ananya.mantravadi@centific.com](ananya.mantravadi@centific.com) Harshit Rajgarhia Centific[harshit.rajgarhia@centific.com](harshit.rajgarhia@centific.com) Abhishek Mukherji Centific [abhishek.mukherji@centific.com](abhishek.mukherji@centific.com) arXiv :2607 .0 1465v 1 [ cs .AI] 1 Jul 2026 Abstract 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 → 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. Keywords reinforcement learning, tool use, verifiable rewards, GRPO, synthetic environments, agent evaluation, Jira, Confluence 1 Introduction LLM-driven agents are moving from chat into workflows: opening tickets, transitioning issues, updating wikis. A growing literature treats tool invocation as a first-class capability to be elicited via supervised fine-tuning or RL on real APIs [9, 11, 15, 17], yet public RL benchmarks for agents target the open web [18], software engineering [4], or broad agentic suites [6, 13] . Niche enterprise SaaS workflows —narrow, schema-heavy, dominated by the “check the resource, then mutate it” pattern — get comparatively little attention. The deeper issue is an objective mismatch: LLMs are optimised against next-token likelihood on internet-scale text, not to act inside a particular API surface. Asked to create a Jira sub-task or a ∗ Equal contribution. Confluence page in our suite, a prompted Qwen3.5-4B knows the rough shape of the call but fills slots wrong, omits required fields, or stops after the first read — behaviours that read as fluent under a token-prediction objective and score poorly under any check that grounds the output back to the environment (our promptedbaseline scores in Section 6 bear this out). Reinforcement Learning with Verifiable Rewards (RLVR) [1, 3, 5, 12] gives up reward modelling in favour of programmatic checkers wherever correctness can be inspected. Tool-use is naturally amenable: an agent’s output is a sequence of structured tool calls whose argument values and ordering can be inspected directly. We exploit that observation to train agents against outcome-level rewards in synthetic copies of the target environment, with no live API, learned judge, or human label in the loop. Contribution. We present a suite of five synthetic, schemafaithful Atlassian environments, the verifiable rewards that

FAQs

Why does next-token prediction underperform in Jira/Confluence tool-use workflows?

The next-token objective optimizes fluency rather than API correctness, leading to wrong slot filling, omitted required fields, hallucinated tools, or stopping after a first read. Reward checks grounded in the environment expose these failures.

How does the RLVR proof of concept compute rewards?

Rewards are computed entirely from the agent’s tool-call trace using hand-designed verifiable reward functions. The setup avoids live APIs, learned judges, and human labels in the loop.

What results are reported for RL-trained policies compared with prompted baselines?

On non-degenerate scenarios, RL-trained policies raise average reward from a 4B baseline range of 0.35–0.92 to 0.95–1.00. The largest single gain is on Confluence page creation (0.35 → 1.00).

What limitations does the work highlight?

Hand-crafting verifiable rewards may not scale beyond the small number of endpoints modeled. One scenario (ticket-transition) has a saturating reward shape where the prompted 4B already reaches the maximum.

Jul 16, 2026

Upload by Violet