A frozen language model, one that was never trained on anything, appeared to both learn new problems and forget old ones when pushed through a standard self-improvement evaluation. That is the central finding of "Phantom Gains," an audit posted August 20 that put a control model through the same pipeline as three rounds of real self-training on Qwen3-8B and measured what the null looks like. When the authors replaced the usual metrics with a per-problem exact test against a pooled baseline under false-discovery-rate control, the self-training gains vanished on held-out replicates.
Key facts
- Three rounds of rank-32 LoRA self-training on Qwen3-8B, audited against a frozen control run through the identical pipeline.
- Seven distinct measurement failures identified, including single-decode ledgers and an expansion statistic whose null is far from zero.
- Posted August 20, 2026 by Cheng Xu, Nan Yan, Liming Chen, and M-Tahar Kechadi as arXiv 2608.20290, with code on GitHub.
- The corrected test detected no significant self-training effect on held-out replicates.
The mechanism behind the illusion is mundane and therefore easy to miss. Evaluating a model on a reasoning benchmark usually means generating one answer per problem and marking it right or wrong. But that single generation is not stable. Batching changes floating-point accumulation order, serving stacks make nondeterministic scheduling choices, and borderline problems flip. Run the same frozen model twice and a handful of problems it "solved" become problems it "failed," and vice versa. Now take the standard self-improvement readout, which counts how many problems moved from failed to solved, and you have a metric that produces a positive number on a model that did not change at all.
It is the medical-trial mistake, translated into machine learning. If you give a hundred patients a sugar pill and check back in a month, some will report feeling better. That is not evidence the sugar worked; it is what the world looks like with nothing in it. Trials handle this by running a control group and asking whether the treatment beat the placebo by more than noise. A great deal of self-improvement research has been reporting the treatment arm without ever running the control. The authors show the widely used expansion statistic has a null far from zero, which means the familiar framing of "the model now solves a problem it never solved before" is not, on its own, evidence of anything.
What survives the correction is narrower and more interesting than a debunk. External distillation, where the model learns from a stronger teacher, did improve problems the base model reached only rarely. Three self-training variants did not. On the hardest category, problems the base model never solved at all, the paper declines to claim anything in either direction, calling the evidence inconclusive rather than negative. And it does report one real effect with the sign pointing the wrong way: self-training corrupted many problems the model already solved at baseline, with losses larger than the frozen model's movement floor. That is not noise. That is the model getting worse at things it knew.
The same day, a second group published the constructive counterpart. AI4AI-Bench asks whether a coding agent can improve a training algorithm rather than merely tune around it. It hands the agent 10 frozen research repositories spanning 10 families of training algorithm, gives it four hours on a single accelerator to edit source, then reruns everything from scratch for up to 12 hours under a hidden evaluator, scoring on a scale where the shipped algorithm sits at 0.1 and the theoretical optimum at 1.0. The best system managed 0.250.
The finding underneath that score is the one worth carrying. Most submissions never touch the learning algorithm at all. They fiddle with budgets, checkpointing, hyperparameters, and where capacity gets placed, which is the machine-learning equivalent of reorganizing the kitchen instead of changing the recipe. The submissions that do reach into the learning layer score materially better, and the paper's read is that additional reasoning effort mostly buys the nerve to attempt it rather than the skill to pull it off. That is a sharper statement about recursive self-improvement than any capability number: the bottleneck may be willingness to touch the load-bearing part.
Neither paper has much community traction yet. At capture time AI4AI-Bench had 18 stars on GitHub and Phantom Gains had none, which tells you how early this is. But the framing is not fringe. Jack Clark's Import AI 469, published August 17, treats autonomous AI researchers as the live frontier question, and the skeptical layer these papers build on is well established: prior work has argued that reinforcement learning from verifiable rewards improves the first attempt without creating fundamentally new reasoning patterns, and that it can narrow exploration rather than widen it.
The honest caveat is that an audit of one setup is still one setup. Phantom Gains tested rank-32 LoRA self-training on a single 8-billion-parameter model, and a lab running a different recipe can reasonably say their gains are real. The paper's actual demand is procedural and hard to argue with: run your frozen control through the identical pipeline and publish what it scored. Until a self-improvement result comes with its own placebo arm, there is no way to tell a phantom from the real thing. Ground Truth has covered self-improving agents that compare notes across lineages; this is the measurement bill arriving for all of it.
Originally published on Ground Truth, where every claim is checked against the primary source.
