Researchers let Claude Code discover AI scaling algorithms that humans probably wouldn't have designed

This work represents a pivotal demonstration of algorithmic discovery through AI agents, moving beyond using AI to implement human designs and into a paradigm where it invents them. The core innovation is not just the specific algorithm discovered, but the AutoTTS (Automated Test-Time Scaling) framework itself—a meta-algorithm for finding better algorithms.

The AutoTTS Paradigm Shift

The traditional research cycle for improving AI reasoning at inference time—like self-consistency or tree-of-thought—relies on human intuition to design prompting strategies and verification heuristics. AutoTTS flips this script. It treats the design space of test-time compute algorithms as a searchable domain. A coding agent, guided by a high-level objective and evaluation protocol, iteratively proposes, implements, and benchmarks candidate strategies. This transforms algorithm design from a creative, manual process into an automated, empirical optimization loop.

The significance lies in scaling human research effort. The agent can explore a vast, potentially non-intuitive design space at machine speed, evaluating thousands of variations that a human team would find tedious or impossible to consider. This suggests a future where the "ideas" behind core AI techniques are increasingly machine-generated.

Analyzing the Technical Breakthrough

The specific algorithm found is described as cutting compute by ~70% for matching accuracy. This is a substantial efficiency gain. Standard self-consistency works by generating multiple independent answers and taking a majority vote—a method that scales compute linearly with desired reliability. The discovered algorithm likely introduces a more sophisticated dynamic allocation of compute.

It may implement an early-exit or verification mechanism, where the agent doesn't naively sample N full reasoning chains. Instead, it might sample a few initial paths, assess their confidence or divergence, and then allocate the remaining compute budget only to the most promising branches or to resolve specific uncertainties. This moves from a static, parallel search to a sequential, adaptive one. Such a strategy is exactly the kind of complex, conditional logic that is easy to describe conceptually but notoriously difficult to hand-engineer optimally for diverse problem types. AutoTTS found an effective variant empirically.

Practical Implications: Cost and Speed

The efficiency of the search process itself is a headline result. A cost of $40 and 160 minutes to discover a compute-saving algorithm is remarkably low. This democratizes advanced AI research. Previously, finding such optimizations might require extensive human researcher time and multiple expensive experimental runs. Now, a well-architected search system can achieve this overnight with minimal cost.

This creates a powerful feedback loop. The discovered algorithm, once deployed, saves 70% of compute on inference tasks. Those savings can be reinvested to run AutoTTS again to find even better algorithms. The barrier to entry for optimizing inference efficiency plummets, potentially accelerating progress across the field. It also provides a clear economic argument for investing in automated research tools.

Broader Impact on AI Research

This study signals a methodological evolution. It highlights the growing role of AI as a research tool and partner. The focus shifts from crafting the "perfect" algorithm to crafting the perfect search and evaluation environment for the agent. The skill becomes designing the right objective functions, constraints, and test harnesses—the "meta-skill" of AI research.

There's a deeper implication for reproducibility and understanding. If the most effective algorithms are discovered by machines through opaque search processes, does human interpretability matter as long as the outcomes are verifiable? This work argues for a pragmatic view: the algorithm's origin is less important than its empirically proven performance and cost profile. It pushes the field toward performance-driven engineering over theoretically elegant but less efficient solutions.

The most profound judgment here is that AutoTTS is not a one-off trick but a template for a new kind of research methodology. It turns algorithmic innovation into a scalable, compute-bound problem. The near-term future may see similar systems optimizing not just inference, but training procedures, data filtering, and architecture design. The human researcher's role evolves to that of a guide, setting goals and boundaries for a tireless computational explorer. The $40 price tag for a major efficiency gain is the clearest possible proof of concept for this new model of discovery.