# Teach it to stop, not just to click

arXiv: https://arxiv.org/abs/2607.17136 · HTML: https://mlsanity.com/paper/2607.17136/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- Placed in Software engineering coding agents, one of 16 directions in LLM agents and tool use.
- Builds on 10 of its 13 references that are in the map.
- First posted on arXiv 19 Jul 2026.

## Placed in

- LLM agents and tool use / [Software engineering coding agents](https://mlsanity.com/agents/software-engineering-coding-agents/)

## Builds on (most cited first, 10 of 10)

- [On-Policy Distillation of Language Models: Learning from Self-Generated Mistakes](https://mlsanity.com/paper/2306.13649/) (2023)
- [Deep Reinforcement Learning at the Edge of the Statistical Precipice](https://mlsanity.com/paper/2108.13264/) (2021)
- [Deep Reinforcement Learning that Matters](https://mlsanity.com/paper/1709.06560/) (2017)
- [GUI-Shepherd: Reliable Process Reward and Verification for Long-Sequence GUI Tasks](https://mlsanity.com/paper/2509.23738/) (2025)
- [Refined Policy Distillation: From VLA Generalists to RL Experts](https://mlsanity.com/paper/2503.05833/) (2025)
- [GUI-GENESIS: Automated Synthesis of Efficient Environments with Verifiable Rewards for GUI Agent Post-Training](https://mlsanity.com/paper/2602.14093/) (2026)
- [VLA-OPD: Bridging Offline SFT and Online RL for Vision-Language-Action Models via On-Policy Distillation](https://mlsanity.com/paper/2603.26666/) (2026)
- [Adaptive Milestone Reward for GUI Agents](https://mlsanity.com/paper/2602.11524/) (2026)
- [OS-Themis: A Scalable Critic Framework for Generalist GUI Rewards](https://mlsanity.com/paper/2603.19191/) (2026)
- [MagicGUI-RMS: A Multi-Agent Reward Model System for Self-Evolving GUI Agents via Automated Feedback Reflux](https://mlsanity.com/paper/2601.13060/) (2026)

## Built on by (most cited first, 1 of 1)

- [cua-speedrun: Standardized Benchmarking of the Speed of Computer-Use Agents](https://mlsanity.com/paper/2609.40284/) (2026)

## Abstract (arXiv)

Agentic computer-use RL is reported in single runs, and those numbers mislead. Using verifier-guided repair of a 35B computer-use agent (CUA) across five oracle-graded environments, we show a repaired policy's success rate is dominated by upstream variance: a variance-components decomposition across three cells (crossed data-draw × seed grid, bootstrap CIs) finds evaluation variance negligible (σeval ≈ 0) and the training-seed effect small everywhere (≤ 10%); instead it splits between the data draw and run-to-run nondeterminism, the data draw's share rising to dominant (48%) on the hardest cell. There the run-to-run distribution is bimodal (Hartigan dip p=0.07, k=10), so a single run has roughly a 30% chance of the failure mode and mean±std is the wrong summary. On that footing, two findings hold. First, repairability is two-tier in how constrained the corrective action is: a single fixed token installs reliably (done-detection 0.97±0.06), while open-ended corrections are only partial -- spatial-coordinate clicks (grounding 0.53±0.35) and a generative field-fill (0.14±0.04). Second, the frame-level repair transfers to task success only when the corrective action is the task's sole remaining blocker (LinkedIn 8/20 vs. base 0/15, Fisher p=0.006). We caught two of our own over-claims -- a sample-efficiency curve and a 'grounding cannot be bought' boundary -- only by replicating across seeds; a stress test makes the stakes external: a single-run improvement of the size this field publishes would have the wrong sign roughly one-third of the time in a comparable regime. We release a library (cuareliability) for routine k-seed reporting. The apparatus is, to our knowledge, the first multimodal segment-aggregated on-policy self-distillation (SA-OPSD) update on a real 35B CUA policy.