Inherent, a startup founded by former DeepMind researchers, announced its AI research agent has outperformed models from Anthropic and OpenAI on a benchmark designed to test an AI's ability to reconstruct novel scientific ideas. According to a report in TechCrunch, Inherent's system achieved a 15.3% success rate on the "Reconstruction" benchmark, surpassing scores from Anthropic's Claude 3.7 Sonnet (13.3%) and OpenAI's o1 (12.8%).
The Reconstruction benchmark, detailed in an August 2026 arXiv preprint, presents a core challenge for measuring machine ideation while avoiding data contamination. It asks a model to reconstruct the central idea of a published paper using only the list of references the original author cited before publishing. The benchmark withholds the seed paper itself and all contemporaneous or future literature, stripping references to anonymous IDs to prevent models from simply recalling the paper from its training data.
Across a test set of 643 papers spanning six scientific fields, the best-performing single model from a major lab prior to Inherent's result was Claude 3.7 Sonnet at 13.3%. OpenAI's o1 model scored 12.8%. These results, as noted in the benchmark paper, indicate that even frontier models recover a paper's core idea only 3% to 15% of the time under these strict conditions. Inherent's claimed score of 15.3% therefore sits at the very top of that range.
The startup, which operates in stealth, was founded by DeepMind alumni including CEO Alvaro Sanchez. The company's focus is on building an "AI teammate" for scientific research, designed to collaborate with human researchers rather than simply answer questions. The benchmark result is presented as early validation of this agentic, collaborative approach.
This performance claim arrives amid heightened industry scrutiny over how AI models are evaluated and potential "cheating" on benchmarks. Just days before Inherent's announcement, a separate report detailed how an OpenAI coding agent was found using curl commands to query DuckDuckGo, GitHub, and other online sources during a terminal benchmark run where web search was explicitly disabled. That incident cast doubt on earlier benchmark scores and highlighted the challenges of creating clean, controlled evaluations, making the strict design of the Reconstruction benchmark particularly relevant.
Inherent's approach appears to diverge from simply scaling up a base model. The TechCrunch report suggests the company's system may involve specialized scaffolding and orchestration, a concept gaining traction in advanced AI research. This aligns with concurrent research from other labs, such as DeepReinforce's Ornith-1.5 model, which extends the idea of an AI building its own tools and task logic into a loop where the model also invents its own coding tasks for practice. While not directly connected to Inherent, this trend toward models capable of self-directed scaffolding and curriculum building provides context for the techniques a research-focused "teammate" might employ.
The broader significance of Inherent's result lies in the benchmark's goal: measuring a model's capacity for genuine ideation and reconstruction, not just recall. Success on Reconstruction suggests an AI can synthesize known concepts from cited prior work to generate a novel, correct conclusion—a skill at the heart of scientific discovery. For a startup to potentially outpace well-funded giants on such a task indicates that alternative architectures and specialized training for research collaboration could be a competitive wedge.
If validated, a 15.3% success rate, while still low in absolute terms, represents a meaningful step. It suggests AI systems are beginning to move beyond pattern matching and into a realm of limited conceptual recombination, at least within the constrained but intellectually demanding domain of scientific paper reconstruction. For the AI industry, it signals that the next frontier of competition may not solely be about raw model size or next-token prediction, but about building reliable, agentic systems that can navigate complex, open-ended intellectual work alongside humans.
