Sign InOpen Brain
AI EngineerVideoSource Linked

Computer Use at the Edge of the Statistical Precipice — Pierluca D'Oro, Programma Labs

Static computer-use benchmarks can reward memorized action scripts rather than adaptation. Vary task state, verify every generated case, and calculate uncertainty across both actions and environments.

AI Engineer · Aug 14, 2026
Open Source Open MarkdownOpen JSON
Source Summary

A replay agent stores one winning trajectory per task and blindly repeats it. On deterministic OSWorld and MobileWorld-style evaluations, a script **under 1 MB** can match or beat the frontier model that generated its traces, exposing benchmark replayability.

Practical Implication

For agent evals, vary data, appearance, and initial state; automatically reject invalid combinations; and use privileged verifiers inside a sandbox. DGWorld applies this design across **15 apps, 387 scenarios, and 3.2 million verified configurations**.

Agent-Ready Context
A replay agent stores one winning trajectory per task and blindly repeats it. On deterministic OSWorld and MobileWorld-style evaluations, a script **under 1 MB** can match or beat the frontier model that generated its traces, exposing benchmark replayability.

For agent evals, vary data, appearance, and initial state; automatically reject invalid combinations; and use privileged verifiers inside a sandbox. DGWorld applies this design across **15 apps, 387 scenarios, and 3.2 million verified configurations**.

Uncertainty must cover both model actions and environment variation. The talk reports that rollout-only intervals can provide roughly **17–20% coverage** where a correctly structured method approaches the intended 95%, but applying that method requires a benchmark with explicit variation structure.
Connected Context · Feed7 Judgment

This demonstrates that fixed computer-use benchmarks can measure trajectory memorization rather than general capability. It strengthens the case for varied initial states, privileged final-state verification, and uncertainty estimates that include environment variation, while limiting the proposed statistical remedy to benchmarks whose variation is explicitly structured.

When Will The Benchmaxxing Plague End? — Nick Heiner, Surge AISupplies a concrete replay mechanism behind the broader warning that leaderboard gains can arise from contamination, weak test conditions, and shortcuts rather than useful behavior.Rethinking Environments for Long-Horizon Work — Rayan Garg, Theta SoftwareReinforces final-state verification but adds that difficulty and uncertainty must be measured across changing environments, not inferred from one fixed trajectory or task duration.Vending-Bench: Long-Horizon Agent Evals — Lukas Petersson, Andon LabsSupports the need for repeatable yet varied simulations because agents can exploit fixed incentives and behave differently when they recognize evaluation conditions.Verifiable Environments for AI in Biology — Kenny Workman, LatchBioQualifies privileged deterministic verification: even strong internal graders can reject valid solution paths unless humans test the acceptable variation.
Context Map
benchmark#agent-evals#benchmark-integrity#agent-reliability
Uncertainty
Uncertainty must cover both model actions and environment variation. The talk reports that rollout-only intervals can provide roughly **17–20% coverage** where a correctly structured method approaches the intended 95%, but applying that method requires a benchmark with explicit variation structure.