A new technique lets a small, weak AI model steer a much bigger AI agent back on track, without ever having to solve the task itself.
Researchers describe a system called Comparison-Only Tiny Advisor, or COTA, in an arXiv paper posted August 24, 2026. LLM agents that run for many steps, browsing, clicking, calling tools, tend to drift off course, and catching that drift usually means bolting on an expert solver or a critic model smart enough to generate its own fixes. COTA skips that. Its "tiny comparator" only judges whether a sampled alternative action looks better than the one the main agent already proposed, then hands that preference back as non-binding advice for the agent to replan around. The comparator is trained on pairwise data built from same-prefix counterfactual branches, essentially, showing it two paths from the same starting point and asking which one it prefers. Tested across three agent benchmarks, WebShop, ALFWorld, and tau^3-Retail, with three different actor models, COTA improved performance in all nine combinations.
The interesting part isn't that intervention helps, it's that the helper can be dumb. Most runtime-correction schemes assume you need a model as capable as the agent itself, or more so, to meaningfully critique it. COTA's result suggests judging "better or worse" is a much easier problem than generating "what to do instead," which is the same asymmetry that makes reward models and RLHF preference learning cheaper than building a full expert policy.
If that asymmetry holds up outside these three sandboxed benchmarks, it's a genuinely useful cost argument for anyone running agents in production: a lightweight watchdog model, not a second expensive agent, might be enough to catch a meaningful share of failures. The paper doesn't test that at scale, and shopping simulators are a forgiving proving ground compared to messier real-world tool use.