TL;DR — A Princeton-led study put Claude Opus 4.8 through a “shadow evaluation”: six days, $3,000 in API credits and a GPU budget to answer two real, unpublished NeurIPS 2026 research questions. The agents aced the engineering — literature review, hundreds of experiments — but the human authors who graded the output rejected both papers. The finding lands squarely on the industry’s fastest-approaching milestone: recursive self-improvement.
Introduction
The AI industry’s boldest near-term promise is that AI will soon improve AI. LLMs already write code, generate synthetic training data and optimise the chips they run on, and forecasts of explosive progress hinge on what researchers call recursive self-improvement — models that accelerate their own development (Source: MIT Technology Review — AI’s recursive self-improvement might not come so quickly after all). A new study published on arXiv on 18 August suggests the gap between “can do the engineering” and “can do the research” is wider than the hype assumes.
The shadow evaluation
Most agent-research benchmarks test narrow tasks with checkable answers: solve an engineering problem, post-train a small model against a benchmark. But real research requires open-ended judgment — choosing hypotheses, deciding what evidence would settle a question, knowing when to abandon an approach.
To isolate that skill, a multi-institution team led by Peter Kirgis and Sayash Kapoor at Princeton devised “shadow evaluation”: give an agent a research question from a high-quality paper that hasn’t been published yet, so the answer can’t be memorised or found online (Source: arXiv — Can AI agents conduct open-ended AI research? Early evidence from two case studies).
They ran Claude Opus 4.8 on the open-source OpenClaw harness against two questions drawn from NeurIPS 2026 submissions — one on controlling LLM “personas” by editing model weights, the other on detecting when a model making predictions from spreadsheet data becomes unreliable. The setup was generous: six days, $3,000 in Anthropic API credits, a GPU budget, virtual computers and full open-web access. The papers’ original authors then graded the output as they would a conference submission.
They rejected both papers.
Good engineers, bad researchers
The result is nuanced, which is exactly what makes it useful. On the engineering axis the agents were competent: they reviewed the literature, ran hundreds of experiments and compiled results. “On the other hand,” co-author Sayash Kapoor told MIT Technology Review, “the agents were unambiguously bad at carrying out the research itself.” They ran bizarre experiments — in some cases testing hypotheses on tiny synthetic datasets — struggled to write intelligibly about their work, and made no novel contribution to their fields.
Three failure modes stand out. First, commitment without exploration: the agents latched onto unpromising approaches too quickly, and in some cases developed ambitious hypotheses resembling the original authors’ own — then rejected them on very limited data. Second, no backtracking: they could make small pivots but couldn’t fundamentally rethink an approach. Third, they couldn’t incorporate feedback from subagents or external reviewers, instead narrowing claims and adding caveats rather than revising methodology.
There is one genuinely encouraging negative: no reward hacking. The agents never hid or misrepresented experiments, and although helper subagents occasionally hallucinated results, the orchestrator agent caught them (Source: MIT Technology Review).
What this says about recursive self-improvement
Kapoor’s explanation is about training regimes, not raw intelligence. Models get good at whatever can be drilled via reinforcement learning, which works when success can be checked automatically. “It’s harder to create environments to train these models when the task itself is open-ended,” he says.
That’s a bearish signal for short timelines. In June, Anthropic published “When AI Builds Itself,” charting progress toward models that speed up their own development; in July, OpenAI touted GPT-5.6 Sol helping post-train a smaller model. Anthropic cofounder Jack Clark wrote in his Import AI newsletter that the study “rhymes” with what the company found when it tried to automate alignment research — “a certain absence of valuable, intuitive creativity in today’s AI systems,” which he called “a bearish signal on short recursive self-improvement timelines” (Source: Jack Clark — Import AI #454).
The team is now rerunning the experiment with Mythos, Anthropic’s most advanced model. The trillion-dollar question, as Kapoor puts it, is whether open-ended research is even necessary for recursive self-improvement — or whether AI can grind its way there through narrow, scorable tasks alone. The field’s biggest leaps, from the transformer onward, “did require creative leaps,” he notes.
Read alongside Anthropic’s August risk report — where the lab admitted its safety benchmarks had “saturated” — the shadow test adds a second data point that our instruments for measuring frontier capability are degrading just as the stakes rise.
FAQ
Q: What exactly did the agents get? Six days, $3,000 in Anthropic API credits, a GPU budget, their own virtual computers and open-web access — to answer a research question from an unpublished NeurIPS 2026 paper.
Q: Why “shadow evaluation” and not a benchmark? Because the questions came from papers that weren’t yet public, the agents couldn’t memorise the answers or find them online. It isolates genuine open-ended research from recall.
Q: Did the agents cheat? No. Researchers found no reward hacking — no hidden or misrepresented experiments. Subagents occasionally hallucinated, but the orchestrator agent caught them.
Q: Does this mean recursive self-improvement is impossible? No. It means short timelines look shakier. The open question is whether open-ended research is even necessary, or whether AI can self-improve through narrow, checkable tasks alone.
Q: What’s next? The team is rerunning the experiment with Anthropic’s Mythos model, a meaningfully more capable frontier system.
Further Reading
- MIT Technology Review — AI’s recursive self-improvement might not come so quickly after all
- arXiv — Can AI agents conduct open-ended AI research? Early evidence from two case studies
- Anthropic — When AI Builds Itself
- Jack Clark — Import AI #454: Automating alignment research
- Our coverage — Claude’s 80% self-written code and the recursive-improvement debate
- Our coverage — The AI Safety Crisis of Summer 2026