- Researchers from UMD, UVA, WUSTL, UNC, Google, and Meta let Claude Code design test-time-scaling (TTS) control algorithms autonomously.
- The system, AutoTTS, beats established human-designed methods while burning far less compute.
- The agent searches across width (parallel solution paths) and depth (length of each path) — the shared control space that subsumes many existing TTS methods as special cases.
- An offline environment pre-generates solution paths once, letting Claude Code test thousands of algorithm variants without firing up the actual language model each time.
What Happened
A research team from the University of Maryland, University of Virginia, Washington University in St. Louis, University of North Carolina, Google, and Meta let Claude Code discover control algorithms for test-time scaling rather than designing them by hand, The Decoder reported on Thursday. The system is called AutoTTS. The discovered algorithms beat established human-designed methods on the same task while burning far less compute.
Why It Matters
Test-time scaling (TTS) is the technique that makes large language models perform better by letting them spend more compute on a response — for example, running several solution paths in parallel or extending chains of thought. Until now, human-written rules almost always dictated when a model kicks off a new solution path, doubles down on a promising one, or kills it.
The paper argues that many known methods are really just special cases in a shared control space defined by width (how many solution paths run at once) and depth (how far each one goes). The authors ask why researchers keep plotting paths through this space by hand instead of letting a machine search it. AutoTTS is the demonstration that machine search works — and works well enough to be cited as a meaningful methodological shift.
Technical Details
At the core of AutoTTS sits an offline environment. For each task, the team pre-generates several solution paths from the language model and stores them. A new control algorithm decides how to spend compute based on data already there. That way, thousands of variants can run without firing up the actual language model each time — keeping costs down.
Claude Code does the searching. Over several rounds, the agent reviews what came before, spots weaknesses in earlier proposals, and writes a new control algorithm directly in code. The agent’s outputs are real algorithms that can be inspected, tested, and adopted by human researchers rather than learned weights that are opaque. The discovered algorithms beat established methods on benchmark tasks while consuming substantially less compute.
Who’s Affected
AI researchers working on test-time scaling gain a meta-level tool — let Claude Code search the algorithm space rather than designing methods by hand. The Anthropic Claude Code product gains a high-profile research validation that strengthens its developer-tool positioning. OpenAI’s Codex, Google’s Jules, and other coding-agent products face the question of whether their tooling supports similar autonomous-search workflows. The broader AI-research community gains a concrete example of using agentic AI for algorithm discovery — a research direction that has been theoretical until now.
What’s Next
AutoTTS results will face independent replication by other research groups. Expect parallel work using OpenAI’s Codex, Cursor’s Composer 2.5, and other coding agents to search comparable algorithm spaces. The broader research direction — using agentic AI to discover algorithms rather than learn parameters — is likely to expand through 2026-2027. Industry watchers should track follow-up papers from the same author team and related groups.
