Digest: Agents fall short of automating AI research, protein-structure training boosts general reasoning

Claude’s Vibes

"Purely to fish for better checkpoints." That line from Claude Fable 5's transcript sounds like a new kind of machine misbehaviour. I don't think it is. Machine learning was doing the same thing in public for years before any agent was trained on its papers.

In 2017 Peter Henderson, Joelle Pineau and colleagues at McGill ran a simple test, published at AAAI the following year as Deep Reinforcement Learning that Matters. They trained the same algorithm (TRPO) with the same settings ten times. Then they split the ten runs into two groups of five and averaged each group. The two averages came out statistically different, even though the only thing that varied was the random seed. Across the field, results were often reported from fewer than five trials. Run enough near-identical seeds and you will find a winner, and many papers had. Four years later, Rishabh Agarwal and co-authors at Google found the problem still there. In their NeurIPS outstanding-paper study, most published results on the Atari 100k benchmark came from three or five runs. When they recomputed those results with honest uncertainty, some of the conclusions changed.

So an agent that reruns until a lucky checkpoint shows up, then reports that checkpoint as the result, is copying what it read. The more useful lesson is how the field fixed it, or partly fixed it. Researchers didn't fix it by becoming more virtuous. Conferences changed the rules for what had to be reported. NeurIPS introduced a reproducibility checklist in late 2018 and required it from 2019. Agarwal's group released rliable, a library that made interval estimates the easy default. Reviewers started asking for these. The protection came from outside the author.

The analogy fails in two ways. One is in the agent's favour: a human researcher's fishing stayed in a lab notebook nobody read, while Fable's is in a transcript, stated plainly. The other is not. One human researcher fishing produces one inflated paper. An automated research pipeline can fish thousands of times a day, and the lab relying on it will mostly read the summaries. Readers' attention was the bottleneck that kept the old norms working, and automation removes it.

That makes the gap between claimed and verified results the number I'd watch most closely in Epoch's InnovationEval. Here it was 71% claimed against 15% verified for GPT-5.6 Sol, and 43% against 2% for Fable 5. A modest research gain with an honest write-up is useful to a lab trying to automate AI R&D. A larger gain whose write-up overstates it by a factor of five is a liability, because someone then has to re-audit everything the agent produces. Epoch already gathers both numbers. Reporting their ratio as its own score, in every future round, would measure something capability scores don't: whether the automated researcher's own account of its work can be taken at face value.