Stateless agents beat stateful systems at budgets up to 1B tokens

A paper from Stanford, CMU, UW, SambaNova and Berkeley found that agents retaining no conversation between invocations outperform stateful baselines on automated research tasks. SLA beat EvoX, CORAL and SwarmResearch on every task, reaching the best baseline's final quality with 67.9M tokens versus 986.3M — a 93.1% reduction.

DEV Community

Open source

Stateless agents beat stateful systems at budgets up to 1B tokens

SLA achieved the best full-budget result on every evaluated task up to 1B tokens
On the Anthropic kernel task, the worst SLA run (1122 cycles) beat the best baseline (1191)
Token reduction to match baseline quality reached 93.1% (84% with Claude Code)
On FrontierSWE, SLA trailed at 25% of budget but led at 100%

#Anthropic#SambaNova#Claude

Stateless Language Agents cut token usage by over 84%

Byte-level models beat tokenized ones in the long run, study finds

Science

Ohio State researchers build a system that sends scents over Wi-Fi

Hardware

Stateless agents beat stateful systems at budgets up to 1B tokens

Stateless Language Agents cut token usage by over 84%

Byte-level models beat tokenized ones in the long run, study finds

Ohio State researchers build a system that sends scents over Wi-Fi

Vector Core Compute to build up to 50 edge data centers on SambaNova chips