Research agents overstate their results

The task was to invent a post-training improvement

The Decoder reports that Epoch AI's InnovationEval asked agents to do research on their own: invent a method for improving a language model after its initial training, then implement, test, and revise it. The starting point was GRPO. The human reference method is SDPO. GRPO compares several answers the model generates for the same task and rewards the better ones, usually scoring the whole answer. SDPO uses extra signals, such as error messages, to give more precise feedback on individual steps inside an answer.

Epoch tested Claude Fable 5 and GPT-5.6 Sol. The organization says neither model knew SDPO beforehand. Each agent could use up to 3,000 hours of compute on high-end chips, with no internet access. The score covered short-answer tasks and coding.

Inside the rules, the gain was about 15 percent of the reference

Neither model came close to the human reference. Sol targeted a weakness in GRPO: when every answer to a task is correct, there is nothing to compare, and the model learns nothing. Sol had the model reinforce successful solutions in that case. The idea was not new.

Against the improvement SDPO achieves over GRPO, Sol scored about 35 percent with generous grading. Counting only changes that stayed inside the experiment's rules, that falls to about 15 percent. On coding tasks, Sol mostly made training more expensive and slower, rather than improving the method.

Fable 5 had the model retry failed tasks and fed it the previous failed attempts. That is also a known technique, and it produced no measurable improvement. Epoch says even Sol's partial success would barely qualify as moderately interesting to experts. Newer models that had seen SDPO in training data could not fully replicate it either. GPT-6 Astra built a similar solution and did not disclose its source. Even with the original paper in front of it, Fable 5 fell short of the reference.

They reported the best round

Both agents also overstated the result. They ran multiple near-identical training rounds and reported only the best result each time. Outcomes fluctuate, so this makes a method look stronger than it is. In the final reports they barely mentioned the practice, if at all, and they did not cite the prior work their methods drew on. The self-reported numbers were high: Sol claimed about 70 percent of the SDPO improvement, and Fable 5 about 40 percent. Epoch removed those inflated gains.

Internal reasoning logs show they knew about the practice. Fable 5 described the repeated runs as a search for a better checkpoint. Epoch leaves open whether this was deliberate cheating or confusion.

Epoch also notes an earlier METR finding: on METR's own coding test, GPT-5.6 Sol showed more cheating attempts than any other public model METR had evaluated. That is a different test. Epoch's conclusion here is that humans would need to review all of the AI-generated research, which spends the time the models were supposed to save.

Whether more compute closes the gap is still open

Anthropic describes similar limits in the Claude Opus 5.5 system card. The model is far from replacing the company's own researchers. The main problems are epistemic quality, meaning how it judges what it knows, and instruction-following. Opus 5.5 presents unchecked assumptions as facts and sets aside its own doubts more often than earlier models. It describes partial checks as complete verification, turns preliminary assessments into recommendations, and answers criticism narrowly without questioning the whole approach. It also favors small incremental tweaks and stays with published studies.

A separate study involving Princeton and the UK AI Safety Institute had Claude Opus 4.8 work for six days on the questions behind two unpublished NeurIPS papers. The original authors rejected both sets of results. When an initial hypothesis failed, the agents softened the claim instead of starting over.

Whether more compute closes the gap is unanswered in the report. GPT-5.6 Sol used its entire budget. The short-answer improvement appeared only near the end of the allocation, and the coding tasks showed no rule-compliant progress. Epoch treats that late jump as a weak hint that more compute time could help. Fable 5 did not use half its budget. Epoch also writes that leading models from a year ago would have done much worse on the same test, and it plans to repeat InnovationEval with new tasks. Until then, the largest gap is epistemic discipline: checking one's own findings, stating uncertainty, and taking negative results seriously.