cd /news/large-language-models/skip-expert-personas-in-prompts-no-a… · home › topics › large-language-models › article
[ARTICLE · art-144440] src=vibeleaderboard.ai ↗ pub= topic=large-language-models verified=true sentiment=· neutral

Skip expert personas in prompts: no accuracy gain, up to 4.5x the cost

A study testing 503 profession-specific system prompts across nine science benchmarks found no clear accuracy gain over minimal prompts, while expert personas produced up to 2.3x more output tokens and raised per-call cost 2.2 to 4.5x. The finding argues against adding expert personas to prompts for accuracy, since the practice mainly increases token usage and cost.

key takeaway

Across 503 profession-specific system prompts on nine science benchmarks, expert personas delivered no clear accuracy gain over minimal prompts while generating up to 2.3x more output tokens and raising per-call cost 2.2 to 4.5x.

read1 min views21 publishedOct 3, 2026

A study tested 503 profession-specific system prompts on nine science benchmarks. Matched expert profiles showed no clear accuracy gain over minimal prompts, yet produced up to 2.3x more output tokens and raised per-call cost 2.2 to 4.5x. Read: A study tested 503 profession-specific system prompts on nine science benchmarks. Matched expert profiles showed no clear accuracy gain over minimal prompts, yet produced up to 2.3x more output tokens and raised per-call cost 2.2 to 4.5x. Read: OpenRouter launched Model Router Benchmarks, scoring seven routers including NVIDIA Switchyard on six benchmarks with a blended Router Index, and explained why routing often loses to a single model: cache rebuilds, weak complexity signals and latency. Read: Ai2 released AstaBrief 8B, a Qwen3-8B fine-tune that turns a research question and literature excerpts into a cited report. Weights and training data are open, and it cut report latency from 178.5s to 51.1s against the Claude-based mode in Asta. Read: A study of reviewer models auditing 411 coding-agent traces found grounding in execution evidence lifted defect catch and cut over-rejection, while reviewer size predicted little. A cascade using generated tests worked without official tests. Read: Uber detailed its MCP Gateway, a proxy that exposes existing HTTP, gRPC and TChannel services as MCP tools, with a registry, API-crawling discovery and a control plane. It now hosts more than 800 MCP servers and 5,000 tools. Read: Shopify described ShopGym, which converts live storefronts into self-contained sandbox shops called ShopArena and generates grounded shopping tasks, so shopping agents can be benchmarked repeatably despite changing prices and bot detection.

── more in #large-language-models 4 stories · sorted by recency
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/skip-expert-personas…] indexed:0 read:1min 2026-10-03 · —