acceptodds
Under review as a conference paper at ICLR 2027

ELVA: VERIFICATION BEGINS BEFORE THE VERIFIER

Abstract

Web agents can report success even when every observation they rely on is correct. This can occur when observation processing preserves individual facts but loses which entities and actions they refer to, allowing unrelated facts to be combined into an unsupported completion proof. We formalize this failure as *inadmissible conjunction*. Our key idea is that verification must begin when evidence is collected, before these relationships are lost. We introduce ELVA, a runtime that preserves these relationships and checks whether earlier evidence remains valid as execution proceeds. These checks determine whether observations can jointly support completion or whether the agent needs further observation or repair. For the supported task dependencies, we prove a conditional guarantee excluding completion proofs that reuse stale action references or evidence whose supporting links have broken. Across five WebArena-Verified sites, ELVA achieves 47.51% overall task success, exceeding the strongest reproduced baseline by 6.15 percentage points. A controlled diagnostic also shows lower false-completion and missed-valid-evidence rates than a local verifier without lineage keys at the same evidence horizon. The largest ablation effects come from preserving evidence before verification and using verification results to guide subsequent actions.

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.