Self-improving coding agents show 10x efficiency boost
This research shows a significant efficiency improvement for coding agents, which could make them more accessible for developers and companies, especially those with limited computational resources.
Researchers from MIT and Sakana AI present a new method called SIFT that makes self-improving coding agents 10 times more efficient than previous approaches. The SIFT method reduces CPU hours from 224 to under 50 for a Qwen3-30B model, cutting costs from $34 to $3.4. It achieves 35.1% on Polyglot benchmark after 30 expansions, compared to DGM's 30.7% after 80 nodes. The approach uses an LLM judge to rank candidate modifications before evaluation, focusing on promising paths.
Recommend read. Love this new arc toward more affordable, more efficient self-improving coding agents. DAIR.AI @dair_ai Banger paper from MIT and Sakana AI. They show that self-improving coding agents work. The best part is that their approach, Self-Improvement via Fast Tree-search (SIFT), runs at a tenth of the CPU hours of DGM. They reach 35.1 percent on Polyglot with o3-mini after 30 expansions. DGM reaches 30.7 percent after 80 nodes of tree search. SIFT does it in under 50 CPU hours and under 5 hours of wall clock. The Qwen3-30B configuration runs its full search at 224 CPU hours and $34 of API spend, a tenth of the DGM baseline. The saving comes from where the money goes. Benchmark evaluation is the runtime bottleneck, so an LLM judge ranks candidate self-modifications first and only promising candidates get evaluated. Judge quality decides the run. On TerminalBench, gpt-5.4-high as the pairwise judge finds a 36.7 percent agent against a 29.2 percent starting point. gpt-5 finds 34.5 percent, and its top-ranked candidate is not the best agent its search produced. Paper: academy.dair.ai/papers/self-im… 🔗 View Quoted Tweet 💬 3 🔄 0 ❤️ 3 👀 2043 📊 4 ⚡
