Don't Solve, Just Compare: Tiny Advisors for Runtime Intervention in LLM Agents

TL;DR - COTA improves LLM agents at runtime using a small comparator that selects promising alternative actions rather than solving tasks or generating corrections itself. It improved performance across all nine tested actor-environment settings, suggesting weaker auxiliary models can still provide effective recovery guidance.

• The comparator evaluates same-prefix alternative continuations against the agent’s proposed action using pairwise judgments.
• Repeated comparisons determine whether intervention is needed; preferred alternatives are offered as non-binding advice so the original agent can replan.
• Training uses pairwise supervision from counterfactual branches sharing the same task prefix.
• COTA outperformed the compared baselines across WebShop, ALFWorld, and tau³-Retail with three different actors.