acceptodds
Under review as a conference paper at ICLR 2027

Actionwise Support Control Closes the Global-Veto Gap in Logged Test-Time Compute

Abstract

Conservative logged controllers often decide whether to continue a language-model trajectory by collapsing actionwise support to a single state-level test. That aggregation can reject every continuation when one action is weakly covered, even when another action is supported and improves on a complete fallback. We introduce Actionwise Support Control (ASC), a feasibility operator that preserves the identity of supported actions until value comparison. ASC is defined for any fixed score, support statistic, fallback, and deterministic tie rule; it therefore composes with pessimistic estimators instead of replacing them. Our central result gives an exact gate-equivalence partition: an actionwise gate and a global gate disagree precisely on states with at least one failed support test and at least one support-eligible, fallback-improving action. The resulting global-veto mass is an interpretable diagnostic of where support aggregation can alter a deployed decision. We also give an outcome-isolated audit contract and a matched-comparison accounting contract that make empirical tests of this mechanism reproducible. Together, these results establish support topology as a first-class design choice for logged test-time compute: conservatism can discard unsupported actions without discarding qualified alternatives.

open until 14 Dec 2026

est. 32% chance this paper gets accepted at ICLR 2027.

Reject 68%Accept 32%

What do you think this paper will get?

All positions stay anonymous.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.