“Give yourself time. Ideas'll come. Life'll shake you, roll you, maybe embrace you. The music'll find you.” —— Johnny Silverhand

Phantom Liberty — Dawid Podsiadło & P.T. Adamczyk

Phantom Liberty — Dawid Podsiadło & P.T. Adamczyk

The score is visible. The truth is not. — Generated with GPT Image 2

By Ziyu Li, with some help from ChatGPT

Aug 21, 2026

When I was working on this year’s ICLR/CVPR paper, I came across DDCM, a paper I had first read last summer, back when I barely knew anything about diffusion models. Quite some time had passed since then. But when I read it again, something suddenly felt different. I began to feel that perhaps I needed to stop pushing forward for a moment—to give myself a little more time and space to think about how my research had been going, what research really is, and what we should want it to be.

Real research can involve plenty of tedious work. But at its core, there should be a question whose answer we genuinely want to know. What has troubled me about much of my recent research is that it often feels like the question itself has disappeared, replaced by the pursuit of SOTA performance through combinations of A+B engineering tricks. And this is hardly unique to me. There is now a strong tendency toward this kind of work across the AI research community.

There may be many reasons for it: funding pressure in academia, the astonishing pace of model scaling in industry, the widening gap in resources between academia and industry, or simply the incentive to chase whatever is hot, build on mature directions, make incremental improvements, and produce more papers.

I want to call this kind of SOTA-oriented research “benchmark engineering.”

Benchmark engineering itself is not necessarily bad. But when improving the benchmark stops being evidence for an idea and becomes the objective itself, research begins to resemble reward hacking. I want to call this research hacking.

Goodhart’s law says, “When a measure becomes a target, it ceases to be a good measure.” In reinforcement learning, we often turn evaluation metrics into reward functions. In RL-based image generation, this can lead to familiar patterns of reward hacking: a model discovers shortcuts that exploit the reward function rather than genuinely improving the desired behavior. The reward score goes up, yet the actual output may become obviously worse.

Typical reward hacking patterns in image generation.

Reward hacking happens, but that does not necessarily mean we are optimizing the model in the wrong way. In some sense, it is a natural outcome of optimization. Once we tell the model what it should improve on, it will naturally search for whatever behavior can push that objective higher—even if that behavior is not quite what we originally had in mind.

This is how AI learns, and in some ways, it is not so different from how we learn as humans. A simple example is how students often shape their learning strategies around the goal of getting high scores on exams. This can work remarkably well, especially when the surrounding system is itself built around the same metric. High scores can directly lead to more “rewards”: admission to a better university, exemptions from entrance exams for graduate school, or, more simply, a stronger resume. And if we want, we can always extend this into an even longer “logic chain,” listing all the other advantages that better scores might eventually bring.

But humans are not starting from scratch. Our brains have been shaped by a very long history of evolution, and then further trained by years of education and experience. In the language of AI, I would imagine that we are already extremely well-pretrained. A PhD student, after years of scientific training, may even be seen as having gone through some additional supervised fine-tuning.

Perhaps this is why research can be different from simply “learning” to optimize a reward. Instead of blindly searching for whatever makes a score go up, we can first form an understanding of how something works and ask: “If I understand this correctly, then the world should behave in this way.” We can then design experiments to see whether reality actually agrees with us—or, even better, whether it can prove us wrong. As Richard Feynman put it, “We are trying to prove ourselves wrong as quickly as possible, because only in that way can we find progress.”

A healthy research process.