Counterfactual Sensitivity Is Not Repairability: Auditing Replay Probes for Video Evidence

2026-08-16

AI Summary

This study addresses the lack of retrieval evidence dependency verification in video agent responses by proposing CARVE, a black-box auditing probe. The method innovatively employs a counterfactual replay mechanism that leverages answer sensitivity differences under SHAM and DESTROY pseudo-operations as routing signals rather than direct classification criteria. Experimental results demonstrate that the answer alteration rate under DESTROY operations is significantly higher than under SHAM. Following policy-based filtering, model accuracy improves by 3.26 percentage points, with fallback performance surpassing random baselines. These findings effectively validate the extent to which agents genuinely rely on retrieved evidence, offering a robust framework for assessing evidential grounding in multimodal reasoning systems without requiring internal model access or architectural modifications.

Abstract

Tool-using video agents retrieve visual evidence before answering, but the final answer is not forced to depend on what was retrieved. The natural black box test is counterfactual: destroy the semantic content of the frames the agent retrieved and check whether the answer changes, against a matched sham that re-executes the identical pipeline on those same frames. We introduce CARVE, a black-box counterfactual probe that compares answer changes under matched SHAM and DESTROY replays. Across three independent k=3 runs on a frozen VideoExplorer-style agent, DESTROY changes the answer 29.3 percentage points more often than SHAM, yielding a large and reproducible aggregate effect. Question-level scores are less stable, and increasing the replay budget from k=3 to k=10 reduces ties but weakens the original zero-threshold routing policy. At k=3, CARVE selects 538 of 1,258 LVBench questions and improves accuracy by 3.26 points, with higher fallback yield than most matched random subsets. The score shows only a weak association with annotated temporal coverage, so CARVE is best understood as a routing signal rather than a direct grounding classifier. Our implementation is available at https://github.com/KurbanIntelligenceLab/CARVE.

Problem

• Video Agents
• Visual Evidence Grounding
• Counterfactual Sensitivity
• Black-box Auditing

Innovation

• Counterfactual Probe
• Video Agents
• Evidence Grounding
• Routing Signal
• Black-box Auditing