Stanford's Stateless Language Agents Match Best Kernel Baseline With 84% Fewer Tokens

StanfordUniversity · hf · 2026-10-12

Automated research systems increasingly run LLM agents over long horizons, but more inference does not by itself produce more progress: agents replay growing histories, duplicate work, or stop experimenting while tokens keep burning. Most evaluations use short budgets or benchmarks that saturate early, leaving these failure modes untested. The authors trace failures to two choices: where research state lives and who decides what to try next.

They introduce Stateless Language Agents (SLAs), built on stateful search with stateless agents: no agent carries its conversation across invocations; instead the harness owns the research state (candidate solutions and measured outcomes) and reconstructs a fresh, role-specific context each time.

In the SLA framework, a stateless Advisor reads harness-summarized evidence across search directions and assigns concrete experiments to parallel Workers.
On software engineering, kernel optimization, and algorithm design at budgets up to one billion tokens, SLA achieves the best final result on every task and reaches the strongest kernel baseline's final performance with over 84% fewer tokens.
Ablations show focused contexts and explicit assignments each contribute and can compound over full runs, while the Advisor consumes under 0.6% of tokens.

The results argue for keeping durable research state out of agent conversations, and show short evaluation horizons can misjudge research systems.

Original post →

Stanford's Stateless Language Agents Match Best Kernel Baseline With 84% Fewer Tokens

More from coding & agent