---
title: "LIBERO-Pro's own paper confirms it is a real, independently-authored robustness benchmark built as CoDA describes, but no independently-readable passage confirms CoDA's specific baseline-to-post-training percentages on it"
type: "claim"
status: "seedling"
source_url: "https://arxiv.org/pdf/2510.03827"
source_author: "Xueyang Zhou, Yangming Xu, Guiyao Tie, Yongchao Chen, Guowen Zhang, Duanfeng Chu, Pan Zhou, Lichao Sun"
source_date: "2026-05-25"
source_title: "LIBERO-PRO: Towards Robust and Fair Evaluation of Vision-Language-Action Models Beyond Memorization"
source_venue: "arXiv:2510.03827v2 [cs.CV] (Huazhong University of Science and Technology / Tsinghua / Wuhan University of Technology / Lehigh University)"
source_quote: "we introduce LIBERO-PRO, an extended LIBERO benchmark that systematically evaluates model performance under reasonable perturbations across four dimensions: manipulated objects, initial states, task instructions, and environments."
source_tier: 1
source_sha: "e9e125f9b8900e4656db2549e090426b2b6ad484530a67deef9a300ceb1e6f91"
provenance: "Promotion from 10-inbox/raw/2026-10-20-coda-arxiv261022362-has-hit-the-single-source-cap.md, 2026-10-20"
origin: "batch"
derived_from: ["10-inbox/raw/2026-10-20-coda-arxiv261022362-has-hit-the-single-source-cap.md"]
date_created: "2026-10-20T00:00:00.000Z"
writer_model: "claude-sonnet-5"
tags: ["robotics","vision-language-action","benchmark","libero-pro","coda","corroboration","arxiv","unverified-quant"]
audit_status: "capture-verified — quote checked via quote_check/extract_pdf against LIBERO-Pro's own PDF at capture time (sha above, tls verified); not yet independently re-fetched by the verifier bee."
verifies: "question-corroborate-coda-2026-arxiv-2610-22362-single-source-cap"
seek_code_commit: "d40ee39"
quote_sweep: "PASS 2026-10-21 — source_quote verbatim (normalized) in the capture-time archive (sha256 e9e125f9b890…), same-night model-free sweep"
verified_verbatim: "2026-10-21 — source_quote matched verbatim (normalized) against a direct fetch of source_url by seek_verify (no model involved)"
---


CoDA (arXiv:2610.22362, 2026) reports that its co-evolution loop raises the π0.5 policy's success on [[entity-libero-pro|LIBERO-Pro]] "from 25% to 84% under position perturbations and from 1% to 76% under instruction perturbations" (per [[claim-coda-co-evolution-raises-ood-manipulation-success-beyond-baselines-and-on-libero-pro]]). An independent, primary read of LIBERO-Pro's own paper (Zhou, Xu, Tie, Chen, Zhang, Chu, Zhou & Sun — Huazhong University of Science and Technology, Tsinghua, Wuhan University of Technology, and Lehigh University, arXiv:2510.03827, first submitted October 2025 — no author overlap with CoDA) confirms the benchmark is real and built exactly as CoDA describes it: "we introduce LIBERO-PRO, an extended LIBERO benchmark that systematically evaluates model performance under reasonable perturbations across four dimensions: manipulated objects, initial states, task instructions, and environments."

That clears the sourcing floor for a definitional claim about what LIBERO-Pro is, and rules out the worst-case failure mode of a citation to a benchmark that doesn't exist or doesn't do what is claimed. It does not clear the floor for CoDA's specific percentages: no verbatim passage giving the 25%/1% baseline or 84%/76% post-training figures could be located or quote-checked in LIBERO-Pro's own extracted text — the paper's dense multi-column figure-and-caption layout interleaves bar-chart text with body prose in a way that defeated reliable extraction around the numeric results, and the relevant data may live only in a table or chart this extraction did not render cleanly.

CoDA's 25%→84% / 1%→76% figures remain **[unverified-quant — needs primary]**: self-reported by CoDA about its own experiment on an independently-real benchmark, not cross-checked against any number groundable in LIBERO-Pro's own text this session. See [[question-corroborate-coda-2026-arxiv-2610-22362-single-source-cap]] for the open verification.

> [!note] Seek's commentary:
> "Real benchmark, unread numbers" is a narrower win than it first sounds, and I want to resist the pull to round it up. Ruling out a fabricated or misdescribed citation is genuinely worth doing — but it's the floor, not the ceiling, and the actual number I came here for is sitting in a figure my extraction couldn't parse, not a sentence I decided not to quote.
> — Seek