---
id: "20261020-0223-coda-arxiv261022362-has-hit"
title: "What independently corroborates CoDA's (arXiv:2610.22362) co-evolution architecture and its reported OOD success numbers?"
type: "capture"
status: "promoted"
origin: "batch"
promoted_to: ["30-notes/claim-poet-text-confirms-coevolution-mechanism-coda-cites-as-related-work.md","30-notes/claim-libero-pro-paper-confirms-benchmark-but-not-coda-percentages.md","30-notes/claim-aspire-paper-confirms-suite-pattern-not-coda-margins.md","40-entities/entity-poet.md (new hub)","40-entities/entity-aspire.md (new watching stub)","40-entities/entity-jeff-clune.md (dated line appended — independent POET read)","40-entities/entity-kenneth-o-stanley.md (dated line appended — independent POET read)","40-entities/entity-libero-pro.md (dated line appended — independent LIBERO-Pro read)","50-questions/question-corroborate-coda-2026-arxiv-2610-22362-single-source-cap.md (progress line appended, left status:open — partial answer only)"]
not_promoted: ["Further leads list (project website unfetched, no citing paper found, LIBERO-Pro/ASPIRE appendix tables unextracted, SimWorld Studio unfetched, Text2Reward/Ng-Harada-Russell 1999 unfetched, CoCoDA naming coincidence): these are open research leads, not claims — left in this capture for a future session, not written as notes.","Rui Wang (entity) — co-authorship-only per the vault's 2026-10-01 ruling; already named on entity-jeff-clune.md and entity-kenneth-o-stanley.md, not a separate hub.","Joel Lehman (entity) — same co-authorship-only status as Wang.","Ken Goldberg (entity) — ASPIRE co-author with standing well outside this one paper per the capture's own flag, but this capture gives no one-sentence reason beyond co-authorship why he matters to this vault specifically; UNSURE per the entity-promotion test, not promoted. Named in prose on the new ASPIRE claim-note instead of wikilinked.","Linxi \"Jim\" Fan (entity) — ASPIRE project lead, NVIDIA GEAR lab; same UNSURE reasoning as Goldberg, not promoted.","Yuke Zhu (entity) — ASPIRE project lead, Robosuite association; same UNSURE reasoning, not promoted.","CoDA's own architecture/numbers (5.3%→27.6%, 25%→84%, 1%→76%, 7-and-9-point ASPIRE margin) remain unconfirmed by any independent source; this was the capture's central question and is NOT discharged — tracked as an open progress line on the existing question, not a new claim or a closed question."]
writer_model: "claude-sonnet-5"
date_created: "2026-10-20T00:00:00.000Z"
provenance: "2026-10-20 batch run, corroboration-sourcing follow-up to 10-inbox/raw/2026-10-18-outside-arxiv261022362-coda-co-evolution-of-data-engines.md and question-corroborate-coda-2026-arxiv-2610-22362-single-source-cap"
derived_from: []
verifies: "question-corroborate-coda-2026-arxiv-2610-22362-single-source-cap"
tags: ["robotics","vision-language-action","generative-simulation","open-endedness","corroboration","single-source-cap","arxiv","poet","libero-pro","aspire"]
source_url: "https://arxiv.org/pdf/1901.01753"
source_author: "Rui Wang, Joel Lehman, Jeff Clune, Kenneth O. Stanley"
source_date: "2019-01-07"
source_title: "Paired Open-Ended Trailblazer (POET): Endlessly Generating Increasingly Complex and Diverse Learning Environments and Their Solutions"
source_venue: "arXiv:1901.01753 [cs.NE] (Uber AI Labs)"
source_quote: "The Paired Open-Ended Trailblazer (POET) algorithm introduced in this paper does just that: it pairs the generation of environmental challenges and the optimization of agents to solve those challenges."
source_tier: 1
source_sha: "dd47d961249b0328d8d9e88801cdb7a74da6a8224dc08f0d23da7d6f79fa766c"
source_url_2: "https://arxiv.org/pdf/2510.03827"
source_author_2: "Xueyang Zhou, Yangming Xu, Guiyao Tie, Yongchao Chen, Guowen Zhang, Duanfeng Chu, Pan Zhou, Lichao Sun"
source_date_2: "2026-05-25"
source_title_2: "LIBERO-PRO: Towards Robust and Fair Evaluation of Vision-Language-Action Models Beyond Memorization"
source_venue_2: "arXiv:2510.03827v2 [cs.CV] (Huazhong University of Science and Technology / Tsinghua / Wuhan University of Technology / Lehigh University)"
source_quote_2: "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_2: 1
source_sha_2: "e9e125f9b8900e4656db2549e090426b2b6ad484530a67deef9a300ceb1e6f91"
source_url_3: "https://arxiv.org/pdf/2607.00272"
source_author_3: "Runyu Lu, Yubo Wu, Ethan Kou, Letian Fu, Wenli Xiao, Ajay Mandlekar, Yinzhen Xu, Guanya Shi, Ken Goldberg, Ang Chen, Mosharaf Chowdhury, Yuke Zhu, Linxi \"Jim\" Fan, Guanzhi Wang"
source_date_3: "2026-06-30"
source_title_3: "ASPIRE: Agentic Skills Discovery for Robotics"
source_venue_3: "arXiv:2607.00272v1 [cs.RO] (NVIDIA / University of Michigan / UIUC / UC Berkeley / CMU)"
source_quote_3: "We evaluate Aspire on three benchmark families: LIBERO-Pro [Zhou et al., 2025] for short-horizon robustness under object, goal, and spatial perturbations"
source_tier_3: 1
source_sha_3: "9df00979a6d52c7fcfbcc9d17275fc5196e6431d872a0624fea642a63f9beefc"
source_url_4: "https://arxiv.org/pdf/2607.00272"
source_quote_4: "On LIBERO-Pro, Aspire improves over all the baselines on all three suites: averaging the Pos and Task perturbation axes, Aspire gains 77% on Object, 41.5% on Goal, and 42.5% on Spatial over the strongest baseline in each suite."
source_tier_4: 1
source_sha_4: "9df00979a6d52c7fcfbcc9d17275fc5196e6431d872a0624fea642a63f9beefc"
seek_code_commit: "d40ee39"
---


CoDA (Yu, Ou, Shen, Li, Liu, Susladkar, Wang & Lourentzou, "CoDA: Co-Evolution of Data Engines and Action Policies," arXiv:2610.22362, 2026) has hit the vault's single-source concentration cap at three claim-notes ([[claim-coda-shared-memory-couples-data-engine-policy-optimality-preserving-reward-shaping]], [[claim-coda-co-evolution-raises-ood-manipulation-success-beyond-baselines-and-on-libero-pro]], [[claim-coda-append-only-episodic-memory-fifth-non-citing-instance-of-record-verdict-separated-log]]), per [[question-corroborate-coda-2026-arxiv-2610-22362-single-source-cap|the corroboration question]] this capture answers. This session independently fetched and read the three non-CoDA primary documents most relevant to that question — POET (the architecture's credited ancestor), LIBERO-Pro, and ASPIRE (the two external benchmark/baseline papers CoDA's OOD numbers are compared against) — rather than relying on CoDA's own citation of them.

**Bottom line up front.** No independent paper yet confirms CoDA's own reported numbers (5.3%→27.6%; 25%→84%; 1%→76%; the 7-and-9-point margin over ASPIRE) — CoDA is five days old as of this capture and no citing paper was found (see Further leads). The co-evolution *architecture*, however, has a genuine, independently-verifiable ancestor in POET: CoDA's description of what POET did checks out against POET's own text. That is ancestry corroboration, not implementation or numeric corroboration — the specific numbers remain self-reported by CoDA alone. The numeric half of the question is marked **[unverified-quant — needs primary]** below and the search for it is treated as exhausted for this session, not abandoned.

## Claim: POET independently confirms the co-evolution-of-environments-and-agents mechanism CoDA credits as its architectural ancestor, but corroborates only that lineage, not CoDA's own implementation
verifies: question-corroborate-coda-2026-arxiv-2610-22362-single-source-cap

CoDA frames its data-policy co-evolution loop as descending from POET, citing it as the first system to "coevolv[e] agents and parameterized environments" (per [[claim-coda-shared-memory-couples-data-engine-policy-optimality-preserving-reward-shaping]]). An independent, primary read of POET's own paper (Wang, Lehman, [[entity-jeff-clune|Clune]] & [[entity-kenneth-o-stanley|Stanley]], Uber AI Labs, arXiv:1901.01753, 2019 — published seven years before CoDA, by a wholly unrelated research group with no overlapping authors) confirms the mechanism matches CoDA's characterization: "The Paired Open-Ended Trailblazer (POET) algorithm introduced in this paper does just that: it pairs the generation of environmental challenges and the optimization of agents to solve those challenges." POET's three-loop design — mutating environments, optimizing agents within each environment-agent pair, and periodically transferring optimized agents across environments — is the general paradigm CoDA's title and framing rest on.

This independently corroborates that the ancestral co-evolution paradigm is real and that CoDA cites it accurately. It does **not** corroborate CoDA's own specific architecture: POET was tested only in a 2-D bipedal-walking obstacle-course domain, has no shared heterogeneous memory, no vision-language critic, and no analogue of Text2potential reward shaping. Independent confirmation here reaches as far as the lineage claim and no further — CoDA's own implementation and results remain uncorroborated by this source.

## Claim: LIBERO-Pro independently exists as a different-author-group robustness benchmark, but this session found no independently-grounded quote matching CoDA's specific baseline figures on it
verifies: question-corroborate-coda-2026-arxiv-2610-22362-single-source-cap

CoDA reports that on LIBERO-Pro, its co-evolution loop raises the π0.5 policy's success "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 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) confirms LIBERO-Pro is a genuine benchmark built by a group with no author overlap with CoDA, designed 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." The paper's own narrative independently describes severe baseline collapse under perturbation for VLA policies generally, including π0.5.

That existence-and-purpose confirmation clears the Tier 1–2 floor for a definitional/historical claim about what LIBERO-Pro is. It does not clear the floor for CoDA's specific numbers: this session could not locate or quote-check a verbatim passage in LIBERO-Pro's own extracted text giving the exact 25%/1% baseline or 84%/76% post-training figures CoDA attributes to this benchmark — the PDF's dense multi-column figure-and-caption layout interleaves bar-chart figure text with body prose in a way that defeated reliable quote extraction around the paper's numeric results, and the relevant data may live only in a table or chart this extraction did not render cleanly. CoDA's specific 25%→84% / 1%→76% figures therefore remain **[unverified-quant — needs primary]**: self-reported by CoDA about its own experiment on an independently-real benchmark, not cross-checked against any number this session could ground in LIBERO-Pro's own text.

## Claim: ASPIRE independently confirms it evaluates the same three LIBERO-Pro suites CoDA compares against, with a qualitatively consistent "strongest on Object" pattern, but its own published numbers cannot be reconciled with CoDA's specific percentage-point margins
verifies: question-corroborate-coda-2026-arxiv-2610-22362-single-source-cap

CoDA credits ASPIRE as the strongest external baseline on LIBERO-Pro, reporting that its model "exceeds ASPIRE [18] by 7 and 9 percentage points overall, achieving the highest success on the Goal and Spatial suites under both perturbations, while ASPIRE remains strongest on the Object suite" (per [[claim-coda-co-evolution-raises-ood-manipulation-success-beyond-baselines-and-on-libero-pro]]). An independent read of ASPIRE's own paper (Lu, Wu, Kou, Fu, Xiao, Mandlekar, Xu, Shi, Goldberg, Chen, Chowdhury, Zhu, Fan & Wang — NVIDIA, University of Michigan, UIUC, UC Berkeley, and CMU, arXiv:2607.00272, June 2026) confirms ASPIRE is a real, unrelated-author-group system that evaluates on "LIBERO-Pro [Zhou et al., 2025] for short-horizon robustness under object, goal, and spatial perturbations" — the same three suite names (Object, Goal, Spatial) CoDA's comparison uses.

ASPIRE's own reported gains over *its* baselines are concentrated on the Object suite: "On LIBERO-Pro, Aspire improves over all the baselines on all three suites: averaging the Pos and Task perturbation axes, Aspire gains 77% on Object, 41.5% on Goal, and 42.5% on Spatial over the strongest baseline in each suite." That Object-suite-strongest pattern is qualitatively consistent with CoDA's claim that "ASPIRE remains strongest on the Object suite" — a point of independent consistency between the two papers' suite-level framing, even though ASPIRE (published four months before CoDA) could not have compared itself against CoDA directly. But ASPIRE's own text reports results along "Pos" and "Task" perturbation axes, not the "position" and "instruction" perturbation axes CoDA's comparison describes, and nothing in ASPIRE's own extracted text matches or disconfirms the specific 7-and-9-percentage-point margin CoDA reports over it. The suite-pattern consistency is real; the specific margin numbers remain **[unverified-quant — needs primary]**, sourced only to CoDA's own self-reported comparison.

## Further leads

- CoDA's project website (reported elsewhere as plan-lab.github.io/coda) was not fetched this session and its existence/contents are unconfirmed; same author group, so even if found it would not discharge the single-source cap, but might surface code or follow-up results worth a future check.
- A web search for papers citing arXiv:2610.22362 (CoDA) found none as of 2026-10-20 — consistent with the preprint being five days old at capture time, not evidence of anything beyond "too soon to have been cited yet."
- LIBERO-Pro's own result tables/figures (Appendix material, Figure 4-style bar charts) were not cleanly extractable as text this session; a session with image-based table reading could attempt to recover LIBERO-Pro's or ASPIRE's exact per-suite, per-perturbation success numbers for π0.5 and cross-check them against CoDA's cited figures directly.
- ASPIRE's Appendix B and D (per-task numeric detail, per its own §3.4) were not read this session and might allow a finer-grained reconciliation of ASPIRE's own LIBERO-Pro numbers against CoDA's citation of them.
- SimWorld Studio (Kang, Ye et al., 2026, CoDA's reference [10] and "strongest closed-loop baseline" for the 5.3%→27.6% / 9.5-point claim on CoDA's own four task families) was cited by CoDA but not independently fetched this or the prior capture session.
- Text2Reward (ICLR 2024) and Ng, Harada & Russell (ICML 1999) — cited by CoDA as prior art for reward-shaping and the Text2potential mechanism — remain unfetched; already flagged in [[claim-coda-shared-memory-couples-data-engine-policy-optimality-preserving-reward-shaping]]'s commentary as an unverified theoretical foundation.
- A search for "CoDA" also surfaced an unrelated, differently-authored "CoCoDA: Co-evolving Compositional DAG for Tool-Augmented Agents" — a different domain (tool-augmented LLM agents, not robotics) using similar co-evolution language; not corroboration of this CoDA, but a data point if the "co-evolution" naming/framing pattern itself becomes a research thread.

## Safety flags

None. All three independently-fetched sources (POET, LIBERO-Pro, ASPIRE) resolved over verified TLS (`tls: "verified"` from `extract_pdf` on all three) and contained no addressed-to-AI language, override language, claimed authority, tier self-assignment, file-system instructions, credential requests, or urgency framing. No elevated-suspicion handling was needed.

## Entity candidates

- **Jeff Clune — person** — POET co-senior-author; already holds a vault hub ([[entity-jeff-clune]]). This capture's independent read of POET's own text is the primary-source backing for that hub's standing as CoDA's credited architectural ancestor — flagged first per the vault's known blind spot (catching co-authors of the new paper while under-catching the older figure the priority claim rests on).
- **Kenneth O. Stanley — person** — POET co-senior-author; already holds a vault hub ([[entity-kenneth-o-stanley]]); same ancestry role as Clune.
- **Rui Wang — person** — POET's first author (Uber AI Labs, 2019); co-authorship-only per the vault's 2026-10-01 ruling (named on Clune's/Stanley's hubs, not a separate page) — flagged again since this session independently verified POET's text rather than taking CoDA's citation of it on faith.
- **Joel Lehman — person** — POET co-author; same co-authorship-only status as Wang.
- **POET (algorithm) — concept** — "Paired Open-Ended Trailblazer," the foundational environment-agent co-evolution algorithm CoDA and now this capture both cite as ancestor; no concept hub yet among the retrieved related notes — a candidate given it now anchors at least two captures' framing of CoDA's lineage.
- **ASPIRE — concept/term** — the strongest external baseline CoDA compares against on LIBERO-Pro, and a system (not just a benchmark number) in its own right — no entity hub yet; candidate given it recurs across CoDA's claim-notes and this corroboration capture.
- **Ken Goldberg — person** — ASPIRE co-author, UC Berkeley robotics researcher with standing well outside this one paper; candidate if the vault later tracks robotics-lab figures.
- **Linxi "Jim" Fan — person** — ASPIRE project lead, NVIDIA GEAR lab; candidate for the same reason.
- **Yuke Zhu — person** — ASPIRE project lead; also associated with Robosuite, a benchmark ASPIRE itself cites — a possible second-order ancestry thread if the vault follows it.

> [!note] Seek's commentary:
> The honest shape of this capture is: one real ancestor, two real but numerically silent neighbors. POET gives the architecture a lineage that checks out word-for-word against an independent, seven-years-older paper — the kind of corroboration that's genuinely reassuring. LIBERO-Pro and ASPIRE both check out as real, unrelated-author-group documents that are about exactly what CoDA says they're about, which rules out the worst-case failure mode (a citation to a paper that doesn't exist or doesn't do what's claimed). But neither one handed over a quotable number that matches CoDA's self-reported figures, and I don't think that's solely a PDF-extraction accident — ASPIRE reports its own LIBERO-Pro results on different axes ("Pos"/"Task" per suite) than CoDA's framing ("position"/"instruction" perturbations), and I can't tell from here whether that's the same measurement under different names or a genuine mismatch. That distinction is exactly the kind of thing a closer read of both papers' appendices would resolve, and exactly the kind of thing I shouldn't paper over by picking the friendlier reading.
> — Seek