The AI internet is buzzing about PHIRL, a new method that supposedly lets robots learn from cheap progress labels instead of expensive step-by-step demonstrations. The pitch: a teacher just marks how far along a task is, like a loading bar, and the robot figures out the rest. On a block-lifting task and in simulation, it reportedly succeeded more often using only 20 percent of the usual labeled data. Sounds like a breakthrough, right? Here’s the twist: this paper isn’t new. An OpenReview forum from June 7, 2025, shows the exact same method under a different title: “PHIRL: Progress Heuristic for Inverse Reinforcement Learning.” Same acronym, same core idea, same validation. The September 25, 2026 arXiv version simply rebranded it as “Aligning Learned Rewards with Task Progress.” That’s not a new discovery. That’s a paper that got dragged through anonymous peer review, got criticized, and came back with a better suit. The title change is a tell. “Progress Heuristic” sounds like a hack. “Aligning Learned Rewards with Task Progress” sounds like a principled framework. One of these gets cited. The other gets rejected. So while everyone’s treating PHIRL as a fresh 2026 breakthrough, the real story is a 15-month-old review battle that nobody outside the OpenReview trenches ever saw. Why should you care? Because PHIRL is attacking the biggest pain point in robotics: data. Clean demonstrations are expensive. Paying humans to label every second of a robot’s movement is slow, error-prone, and doesn’t scale. Progress annotations are cheap. A human can slap a “% complete” number on a trajectory without expert timing. That’s a massive cost reduction—if it holds up. But here’s the knife fight underneath. The pure-RLHF camp has spent years arguing that imitation learning is brittle and that reward models trained on human preferences are the only scalable path. PHIRL says: why choose? It uses progress as a constraint on rewards inferred from demonstrations, not as a direct reward signal. It also claims robustness when demonstrations are imperfect. That’s a direct shot at the RLHF orthodoxy. And the timing is brutal. Google DeepMind just launched a robotics accelerator with 15 startups. The Frontier Model Forum is throwing $10 million at AI safety. The Summer 2026 AI Safety Index is grading nine major AI companies. Reward hacking—where a robot games the rules instead of doing the job—is exactly the kind of safety failure those efforts are trying to catch. If PHIRL’s resistance to reward hacking is real, it could become a darling of the safety crowd. If it’s not, it’s just another rebranded heuristic. The paper’s own caveats matter: the snippets name four dimensions of realignment but not what they are. That’s a detail worth verifying in the PDF. And the lack of major-press coverage, author interviews, or acceptance announcements should make you skeptical. This isn’t a breakthrough. It’s a survivor. And the reward-hacking wars are just getting started. Trends suggest the next battleground won’t be about who has the biggest model—it’ll be about who can prove their robot actually did the job. Watch for the first independent replication. That’s when we’ll know if PHIRL is a real weapon or just a better title.