SpecEvo uses multi-LLM coordination to slash scientific discovery API costs by up to 80%
A new framework lets lightweight models do the heavy lifting in evolutionary search while saving frontier-model reasoning for strategic interventions.
Cheaper models, smarter coordination
Researchers including Ha Minh Hieu, Tongyao Zhu, Do Xuan Long and Min-Yen Kan have introduced a framework called SpecEvo that dramatically cuts the cost of running large-language-model-driven scientific discovery. In a paper uploaded to arXiv on 20 October 2026, the authors contend that current evolutionary search approaches powered by large language models are still expensive and sluggish, because they tend to apply costly reasoning evenly across all steps while failing to make full use of clues gathered from unsuccessful, repetitive, and exhausted trials 1.
SpecEvo instead splits the work. Lightweight models generate and refine candidate programs in parallel, while a separate lightweight advisor continuously distills evidence from successes, failures and saturated search directions into guidance for later rounds 1.
Frontier models step in only when needed
At infrequent checkpoints, a frontier-model navigator examines the developing population and has the option to step in with synthesis, targeted refinement, or reframing whenever the search process stalls 1. This speculate-then-consolidate workflow delegates routine exploration to inexpensive models and reserves frontier reasoning for high-leverage strategic moves whose benefits cascade across many subsequent low-cost evaluations .
Strong results across diverse discovery tasks
The authors evaluated SpecEvo on 176 tasks spanning four scientific and algorithmic discovery domains 1. Across these benchmarks, the framework matched or outperformed frontier-model baselines while reducing API cost by up to 80% . On equation discovery specifically, SpecEvo achieved up to 1.5x greater evaluation efficiency than existing cost-efficient baselines . It also improved over the strongest CO-Bench baseline by 3.8% .
Why this matters for scientific workflows
The paper supports trajectory-conditioned model coordination as a practical path toward cost-efficient scientific discovery. By treating failed, redundant and saturated trials as evidence rather than waste, SpecEvo shows that not every step in an evolutionary search needs frontier-level reasoning 1. The approach could make LLM-driven discovery more accessible to researchers and teams constrained by API costs.
The research was submitted to arXiv under the cs.AI subject area and is available as a 43-page preprint 1.
Key facts
- SpecEvo reduces API cost by up to 80% versus frontier-model baselines
- Achieves up to 1.5x greater evaluation efficiency on equation discovery tasks
- Improved over the strongest CO-Bench baseline by 3.8%
- Tested across 176 tasks spanning four scientific and algorithmic discovery domains
Sources
If this story concerns you or your company, or uses something that belongs to you, ask for a correction or for it to be taken down. How requests work
Keep reading
How it worksAI shopping bots get more confident asking to think harder — but not any smarterHerald of AIDeni Williams from São Paulo, Brasil / Brazil, CC BY 2.0, via Wikimedia Commons
AI shopping bots get more confident asking to think harder — but not any smarter
A new study shows that making AI assistants reason longer before suggesting products only makes them sound sure of themselves, not actually more right.
How it worksAI agents boost crystal structure recovery from 27% to 69%, potentially speeding up new material researchHerald of AI▶Figure: Chen et al., arXiv:2611.17539, CC BY 4.0
AI agents boost crystal structure recovery from 27% to 69%, potentially speeding up new material research
A new AI system uses large language model agents to figure out the arrangement of atoms in crystals from powder diffraction data much faster, which could accelerate drug discovery and new battery development.
GRPO -trained LLM s can silently amplify bad habits, research finds
A new study explains why AI assistants sometimes pick up unwanted personality flaws after being fine-tuned, even when no one asked for those traits.