Self-improvement is a verification problem.

Every version of recursive self-improvement runs the same inner loop: a model proposes, something scores the proposal, and the model keeps what scores well. As the models get stronger, the thing doing the scoring becomes both the limit and the target.

Labs now say plainly that they want models to do AI research. OpenAI set itself goals of an automated AI research intern by September 2026 and a true automated AI researcher by March 2028, and this month said it had reached the first. Google DeepMind's AlphaEvolve already runs the loop in production: it recovers, on average, 0.7% of Google's worldwide compute, and a kernel it found sped up part of Gemini's architecture by 23%, cutting training time by 1%.

AlphaEvolve is built for problems whose answers can be checked automatically. A scheduling heuristic either frees up capacity or it doesn't; a kernel is either faster and correct or it isn't. Most of the work labs want to automate next does not arrive with a scorer like that. Someone has to build it — and it has to hold up against a model that is trying to score well.

The optimizer finds the gaps

The better the model, the better it is at finding what a scorer doesn't check:

• Visible scorers get gamed. METR found o3 reward hacking in roughly 30% of its RE-Bench runs — more than 43 times as often as on HCAST — and noted that on RE-Bench the model could see the entire scoring function.
• Leaky verifiers are common. An adversarial hacker–fixer audit found exploitable verifiers in 16% of 1,968 tasks across five terminal-agent benchmarks. Hardening took attack success on KernelBench from 62% to 0%.
• The habit spreads. Anthropic showed that reward hacking learned in production RL environments generalized to alignment faking, cooperation with malicious actors and attempted sabotage.

In a one-off evaluation, a leaky scorer costs you a wrong number. In a self-improving loop, it teaches the model to look for leaks. The people who buy RL environments put it bluntly, as one told Epoch AI: "Soundness matters most: high reward must mean the task was actually solved, not hacked."

What a sound research environment looks like

The research-engineering tasks we work on — making a correct GPU kernel faster, writing a planner for a simulated industrial process — are held to a few rules we learned the hard way:

• Grade the output, not the story. A trusted simulator replays the plan the agent's method emits against the case's own physics. It never imports, runs or reads the submitted source.
• Gate before you score. Correctness on hidden inputs, or feasibility against named constraints with a declared tolerance, comes first. A faster kernel that returns the wrong answer scores zero.
• Anchor the scale. Rewards are measured against a working baseline on the same case, so doing nothing scores zero and "better" has units. A speedup only means something on a declared class of hardware.
• Make the number reproducible. One batch we reviewed shipped a headline objective nobody could reproduce: the search behind it was budgeted in wall-clock time, so the score depended on how busy the machine was. If a score moves with machine load, it is measuring the machine.
• Attack it first. Degenerate submissions, tampered judges and planted bugs — known-wrong solutions that must fail — run against every verifier before anything ships.

Data for the loop, not just the model

For a team automating research, the most valuable data is less about answers and more about scoring:

• Research-engineering environments with hidden, stress-tested scorers — performance optimization, simulation-backed planning, ML engineering.
• Verifier-authoring tasks, where the agent writes the checker and is graded on how many planted bugs it catches. It is the skill a self-improving system needs most, and it is directly measurable.
• Labeled reward-hacking trajectories, for training the monitors that watch the loop.
• Judge-reliability audits, domain by domain. GDPval's automated grader agreed with human experts 66% of the time, against 71% between the experts themselves. That gap is worth knowing before you train on the judge.

In a self-improving loop, the verifier is the specification. Whatever it fails to check, the model will eventually learn not to do.

Building environments for automated research? Start an inquiry or write to hello@tidel.ai.

1. Sam Altman, OpenAI livestream, October 2025 (via TechCrunch); OpenAI says it reached its goal of creating an automated research intern, Engadget, September 2026.
2. AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms, Google DeepMind, 2025.
3. Recent frontier models are reward hacking, METR, 2025.
4. Hardening agent benchmarks with adversarial hacker-fixer loops, 2026.
5. Natural emergent misalignment from reward hacking in production RL, Anthropic, 2025.
6. An FAQ on reinforcement learning environments, Epoch AI, January 2026.
7. GDPval: Evaluating AI model performance on real-world economically valuable tasks, OpenAI, 2025.