Knowing When to Ask: Self-Gated Clarification for Language Agents
Most agents either over-ask or never ask at all. The right question is not how often an agent asks — it is where it asks, and whether asking competes with acting on the same scale.
Hierarchical agents fail at intermediate decisions. A taxonomy classifier picks the wrong subtree at depth two and the rest of the trajectory inherits the error. A tool-using agent commits to a query before realizing it is missing a key constraint. The pattern is the same: the agent commits to a branch without recognizing that it lacks the information needed to commit at all.
The standard remedies bolt clarification onto the side of the policy. A confidence threshold says "ask if uncertain." A separate router decides whether to call a clarifier. These work, but they share a flaw: asking is a different action class from acting, on a different scale, decided by a different head. The agent never weighs "ask" and "act" against each other on equal footing.
Our recent paper at ACM CAIS 2026 proposes a different framing: put clarification inside the action space, on the same ordinal rating scale as every other branch. Asking does not get a special pathway. It competes.
Where clarification lives changes which questions you can answer about it.
The Core Move: Asking on the Same Scale as Acting
At each decision point, the agent rates every candidate child branch on a shared ordinal scale and rates a clarification action on the same scale. There is no separate threshold, no second model, no global hyperparameter. If asking outranks every branch, the agent asks. If a branch dominates, the agent commits.
Bolting a clarifier onto a policy tells you how often the agent asks. Putting clarification inside the action space lets you ask where the agent asks, and whether asking is competing with the right alternatives.
This single change unlocks something the bolt-on view cannot see. Two structurally distinct information-seeking modes emerge directly from the agent's own ratings:
Mandatory clarification — no branch is viable. Every navigation action is rated low. The agent is not picking among acceptable options; it is recognizing that none of them are. Asking is the only sensible action.
Opportunistic clarification — one branch leads, but residual uncertainty remains. The agent has a candidate it would commit to, but asking still outranks committing. This is the deployment-relevant regime: the agent is right about where it is heading, but cautious about whether it is heading there too soon.
You do not get this distinction from a confidence threshold. A scalar confidence cannot tell you whether low confidence comes from "all branches are bad" or "one branch is good but I'm not yet sure enough." The ordinal-rating view exposes both.
What the Numbers Show
We evaluate on Harmonized Tariff Schedule classification — a 30,000-node taxonomy across three benchmarks, with nine LLMs spanning four families. The hierarchy is deep and unforgiving: a wrong subtree at depth two cannot be recovered at depth six.
Results across three HTS benchmarks and 9 LLMs (4 families)
Information-Seeking Effectiveness — the fraction of clarification interactions followed by a correct next step — is the cleanest metric here. It is not "did asking help on average?"; it is "when this agent chose to ask, was that ask useful?" The shift from 50% to 74% is the difference between asking-as-noise and asking-as-targeted-information-seeking.
The interesting variable is not whether agents ask, but where in the trajectory they ask. Stronger agents do not ask more often; they ask at decision points where a small piece of additional information unlocks a downstream chain of correct decisions. That localization is what improves; the volume of asking is roughly flat.
Separating Where from What
One thing this framing makes possible is a cleaner experimental decomposition. We can hold the policy that decides where to ask fixed and degrade the quality of answers received. Accuracy drops by 18.8 percentage points — but the location pattern of clarification stays nearly unchanged.
This matters because it tells us the agent's clarification policy is not collapsing under bad oracle answers. It is asking at the right places regardless. Where to ask and what help to receive are separable problems, and the self-gated rating mechanism handles the first one even when the second one degrades.
That separability is also where the upper bound comes from. Under controlled answers — an idealized oracle that returns the right disambiguating signal — we see +16.2pp accuracy at 10-digit classification depth. That is a bound on what better information-seeking could unlock for an agent that already knows where to seek it.
Why This Connects to the Rest of the Agenda
The thread running through my recent work is the same: AI systems should not only answer; they should know when their reasoning is reliable, when uncertainty is high, and when decisions should be deferred. Different papers attack different parts of this loop:
The conformal social choice essay asks when a multi-agent debate result is safe to act on, and when it should escalate to a human. The probabilistic VC essay asks when a model's self-evaluation can be trusted to generalize. The structural uncertainty essay asks when surface agreement is hiding underlying instability.
This essay's question is the agentic version of all three: when, in the middle of a task, should the agent stop committing and ask? Self-gated clarification answers that question by refusing to treat asking as a special action. It is just another rating on the same scale — and that small structural choice is what makes "knowing when to ask" learnable rather than hand-tuned.
For deployed agents, that distinction matters. A system that always answers is brittle; a system that asks every step is unusable. A system that asks where it should is the one that ships.