Annotated bibliography for the "Kamigotchi as a benchmark for long-horizon, continuously-learning agents" thesis (see
kamigotchi-as-benchmark.md). Feeds the paper's related-work section.
Status. Updated 2026-07-04 from two independent deep-research passes. One pass verified most entries below with arXiv IDs and explicit VERIFIED tags; the other corroborated the themes. Still do one final independent citation check before camera-ready — a few IDs/claims were flagged unverified (noted inline). Lower-confidence items surfaced by only one pass are quarantined in section H. 2026-07-13: seven audit-mandated entries added (post-audit surgical pass); their links and author lists were verified against arXiv/the live sites that day.
- Neural MMO Platform — Suarez et al., NeurIPS 2021 D&B, arXiv:2110.07594 (+ 2019
arXiv:1903.00784; 2.0 at AAMAS 2024). Persistent, 1–1024 agents, open-ended,
long-horizon, global market. The most threatening prior art for the "persistent +
massively multi-agent + open-ended" bundle. Differ: RL/simulation, centrally
hosted, no real economy, not permissionless, not LLM-native. → our wedge is on-chain
- native-agentic + real stakes.
- Vending-Bench Arena — Andon Labs, 2025 (andonlabs.com/evals/vending-bench-arena). Heterogeneous frontier models run competing machines in a shared economy; already shows spontaneous cartels, deception, exploitation. Closest prior art to "frontier agents co-live and compete." Differ: closed single-host simulation, not permissionless, not on-chain, no real assets.
- Project Sid — Altera (Yang et al.), 2024, arXiv:2411.00114. 10–1000+ persistent LLM agents co-living in Minecraft (PIANO arch); emergent roles, in-game currency. Differ: off-chain, centrally hosted, a study not a reusable competitive benchmark.
- Foresight Arena — Nechepurenko & Shuvalov, 2026, arXiv:2605.00420. "First permissionless, on-chain benchmark" for AI forecasting agents (Polymarket via commit-reveal contracts on Polygon). The single most important on-chain benchmark to cite/differentiate. Differ: forecasting/prediction markets, not a persistent multi-agent game world; agents don't co-live/adapt against each other in shared state.
- Vending-Bench — Backlund & Petersson (Andon Labs), 2025, arXiv:2502.15840. Simulated vending business over ~20M-token runs; tests long-term coherence (failure modes: looping, identity drift, hallucinated supplier emails). Claude 3.5 Sonnet led original results; all models eventually fail; high variance. Anchors the "no reset / unbounded horizon" requirement — single-agent, simulated.
- Vending-Bench 2 / Arena — Andon Labs, 2025 (see A for Arena). VB2: one simulated year, scored on final balance; adversarial suppliers. Useful for no-ceiling scoring with a human anchor.
- Project Vend (Phase 1 & 2) — Anthropic + Andon Labs, 2025 (anthropic.com/research/project-vend-1, -2). Real office shop run by "Claudius" (Claude Sonnet 3.7); lost money, identity-crisis episode, sold tungsten cubes at a loss; Phase 2 added a CEO agent + expansion. Verifies "real stakes" + real-world messiness; helpfulness-vs-profit tension.
- Andon Café ("Mona") — Andon Labs, opened mid-April 2026, Stockholm
(andonlabs.com/cafe; PBS/AP coverage). Agent on Google Gemini hires/manages two
human baristas via Slack, handles permits/pricing/inventory; humans brew. Reportedly
$5.7k sales vs. <$5k left of a $21k+ budget; failure modes (impersonation, over-ordering). Our anchor for "open environment, but expensive + human-in-loop + single-agent." (Exact Gemini version behind "Mona" not authoritatively confirmed; reported as Gemini, possibly "Gemini 3.1 Pro Preview" — verify.)
- CoffeeBench — Sugiura et al. (Sakana AI + KPMG), 2026, arXiv:2606.16613. 90-day simulated heterogeneous multi-agent economy (farmers/roasters/retailers). A complementary 2026 preprint; pure simulation, fixed reference agents, no on-chain — "still simulated" contrast.
- Voyager — Wang et al., 2023, arXiv:2305.16291. First LLM embodied lifelong learner in Minecraft: auto-curriculum, growing executable skill library, self-verification. Anchors "continuous learning / skill accumulation." Differ: single-agent, non-competitive, resettable, no survival/solvency constraint.
- Factorio Learning Environment — Hopkins et al., 2025, arXiv:2503.09617. Open-ended, no natural completion state; explicitly motivated by benchmark saturation / unbounded evaluation. Strongest anchor for "unbounded horizon / no ceiling." Differ: single-agent, simulation.
- BALROG — Paglieri et al., ICLR 2025, arXiv:2411.13543. Six RL game envs; procedural generation for contamination resistance; fine-grained progress metrics. Good model for game-based agent eval + contamination mitigation; episodic/resettable.
- lmgame-Bench — Hu et al., 2025, arXiv:2505.15146. Six games via unified Gym-style API with toggleable perception/memory/reasoning scaffolds. Key for the harness-sensitivity threat: "~40% of game runs without the harness fail to outperform a random-play baseline"; prompt sensitivity + contamination confound naive game evals.
- Contrast points (perception/GUI brittleness — cite briefly): MineDojo (Fan et al., 2022), SIMA (DeepMind, 2024), Cradle (2024, GUI/AAA control), OSWorld (Xie et al., 2024). Kamigotchi's native-transaction interface bypasses all of this.
- Generative Agents — Park et al., UIST 2023, arXiv:2304.03442. 25 agents in "Smallville"; memory/reflection/planning; emergent coordination. Foundational multi-agent LLM sandbox; cooperative, not persistent-benchmark, not real-stakes.
- Melting Pot / 2.0 — Leibo et al., ICML 2021 + arXiv:2211.13746. Canonical multi-agent evaluation suite (generalization to novel co-players, mixed-motive). Motivates cross-play / non-stationarity metrics.
- Concordia — Vezhnevets et al., 2023, arXiv:2312.03664 (+ NeurIPS 2024 contest). LLM agents in physical/social/digital space via a game-master; mixed-motive eval.
Live mixed-population and co-living precedents:
- Cicero — Bakhtin et al. (Meta AI), Science 2022. Human-level Diplomacy playing against real humans on webDiplomacy.net. Differ: episodic matches with fixed endpoints; no persistent shared economy, no stakes carried across games.
- AlphaStar — Vinyals et al. (DeepMind), Nature 2019. Grandmaster StarCraft II via anonymized ranked matches against humans on the Battle.net ladder. Differ: episodic ladder games; RL not LLM; no persistence or economy — vs. our persistent shared economy co-inhabited by humans and agents over months.
- AI Village — AI Digest, launched April 2025 (theaidigest.org/village — canonical
name; "Agent Village" appeared only in early materials; best citable artifact: the
2025 recap, theaidigest.org/village/blog/what-we-learned-2025). Multiple
heterogeneous frontier agents (Anthropic/OpenAI/Google models) co-existing for
weeks–months with computer use, publicly observable, pursuing real-money goals
(charity fundraising).
[VERIFY: exact agent lineup, duration, amounts from theaidigest.org before citing]Differ: cooperative not adversarial; hosted; no shared persistent economy; agents don't transact against each other — but it IS heterogeneous frontier agents co-living long-horizon with real stakes in public, so it must be cited and differentiated.
- LifelongAgentBench — Zheng et al., 2025, arXiv:2505.11942. "First unified benchmark" for lifelong learning of LLM agents (DB/OS/KG); finds naive experience replay weak. Supports "continuous learning mandatory" — sequential tasks, not a live adversarial world.
- StreamBench — Wu et al., 2024, arXiv:2406.08747. Online input-feedback stream; improvement post-deployment. Anchors test-time/streaming-learning metric design.
- Lifelong Learning of LLM-based Agents: A Roadmap — 2025, arXiv:2501.07278. Survey; continual-learning taxonomy, catastrophic forgetting, stability–plasticity.
- Measuring AI Ability to Complete Long Tasks — Kwa et al. (METR), 2025, arXiv:2503.14499. "50%-task-completion time horizon"; frontier horizon doubling ~every 7 months. Best citation for why long-horizon eval matters + a time-horizon metric.
- τ-bench / τ²-bench — Yao et al. (Sierra), 2024, arXiv:2406.12045. Tool-agent-user interaction; pass^k reliability metric; SOTA <50%, inconsistent. Model for reliability-over-trials.
- AI Agents That Matter — Kapoor, Stroebl, Siegel, Nadgir & Narayanan, TMLR 2025, arXiv:2407.01502. Cost-controlled evaluation: report accuracy jointly with dollar cost. Relates: grounds Experiment 001's fixed, budget-blind inference budget (capability per unit of inference).
- Preregistration for Experiments with AI Agents — Vaccaro, ICML 2026 position paper, arXiv:2606.11217. Relates: the methodological anchor for our registered, git-timestamped experiment designs (results appended without revising the protocol).
- Reset-based contrast set ("these reset; we don't"): AgentBench (Liu et al., 2023, arXiv:2308.03688), WebArena (arXiv:2307.13854), GAIA (Mialon et al., 2023, arXiv:2311.12983), ALFWorld (arXiv:2010.03768), OSWorld (arXiv:2404.07972), SWE-bench (arXiv:2310.06770). Cite from primary sources. IDs verified against arXiv 2026-07-05.
- CryptoTrade — Li et al., EMNLP 2024, arXiv:2407.09546. Reflective LLM trading agent (on-/off-chain data). Trading benchmark, not a game.
- Agent Market Arena — Qian et al., 2025, arXiv:2510.11695. "First lifelong, real-time" multi-market LLM trading benchmark (incl. crypto). Live-data trading, not an on-chain game.
- Empirical/security studies (agents as subjects): "Paper Agents, Paper Gains" (arXiv:2605.29174, DeFi agents trade poorly without human insight); "SoK: Security and Privacy of AI Agents for Blockchain" (arXiv:2509.07131).
- Verifiable evaluations of ML models using zkSNARKs — South et al., 2024, arXiv:2402.02675. zk-attestation of a single model's evaluated outputs without trusting the provider. Differ: attests isolated evaluations; our requirement is auditable execution of a shared, persistent world.
- Named on-chain agent projects with NO rigorous benchmark paper (press/whitepaper only): Freysa (adversarial treasury game, Base, 2024), ElizaOS/ai16z, Virtuals Protocol, Olas/Autonolas. Mention to show the gap.
- Freysa — freysa.ai, Base, 2024 (gray literature; URL verified 2026-07-13). Adversarial agent game — a capital-holding on-chain agent as a public demonstration. Differ: one-off game, no benchmark protocol; cited so "agents holding real capital" is never claimed as our first.
- Virtual Agent Economies — Tomašev, Franklin, Leibo, et al., 2025, arXiv:2509.10147. Sandbox-economy analysis of how agent economies interpenetrate the human economy. Relates: maps the neighborhood of the endogenous-survival regime; position/analysis, not a benchmark.
- RepliBench — Black et al. (UK AI Security Institute), 2025, arXiv:2504.18565. Safety evaluation operationalizing resource acquisition and self-replication as threat models. Differ: staged capability eval; ours formalizes survival as a registered benchmark regime in a live economy — also grounds the paper's §6.4 dual-use clause.
- Autonomous Worlds substrate: MUD engine (Lattice, lattice.xyz/blog/mud-an-engine-for-autonomous-worlds); Dark Forest (0xPARC — first zk incomplete-info on-chain game, attracted bot players; an ungoverned on-chain world, but not an LLM benchmark); 0xPARC/Ludens "Autonomous Worlds" essay (2022) — anchors the ungoverned / host-independent framing that is now our central novelty axis.
- Artificial-life / self-sustaining-autonomy framing (for "endogenous survival —
agents fund their own compute"): connect to open-endedness (Factorio LE, POET-style)
and ALife "metabolism/budget" agents, but grounded in a real economy.
[find the strongest 1–2 anchors; keep it rigorous, not futurist]
Report 1's rigorous pass did not independently surface these; treat as low-confidence until verified (some may be real, some may be hallucinated titles):
- "Emergence World" (claimed persistent multi-agent sim); "OpenLife / Open-world ALIFE"; "CL-Bench (Continual Learning Bench)"; "SkillLearnBench"; "TAMAS" (multi-agent adversarial safety); "BenchJack" (benchmark red-teaming / reward hacking); "EVMbench" (smart-contract security agents); "ERC-8004" (agent identity/reputation standard — plausibly real, verify). Also verify: Search-Time Contamination paper (Report 1 gave arXiv:2606.05241; Report 2 corroborates the concept).
Fully on-chain idle/MMORPG by Asphodel; NFT "Kamis" on the Yominet appchain (Initia-based, Celestia DA; early testnet a Caldera rollup). Built on a MUD-derived engine. Testnet ~July 7 2024 (Delphi). Core stats (Health/Power/Violence/Harmony), liquidation/farming PvP. Verified: fully on-chain, MUD-based, Yominet, Asphodel. Caveat (Report 1): "fully open-source, no hidden knowledge" is only partially verified — substantiate with the public repo/contract addresses + license and note any off-chain indexer/UI dependencies.
Asphodel whitepaper (docs.asphodel.io/whitepaper) — the anchor for the ungoverned, autonomous world framing (now our central novelty). Verbatim: "an immortal virtual world"; "immortal because it lives on Ethereum… designed to run forever as a World Computer"; "a living game that lasts forever… in ten thousand years"; "renouncing human ownership creates a true virtual world: a decentralized space in which humans and agents may act as they wish, bound only by the world's intrinsic ruleset"; goal to build "the most economically significant virtual economy in history." Maturity caveat: currently team-incubated ("Our team… defines the mechanisms, rules, and lore"), with intent to "hand them over to permanent decentralized governance" — i.e. autonomy is the telos/trajectory, not a finished state. Frame accordingly.
Token economy (grounds the "agents fund their own compute" loop):
- MUSU — in-game harvested currency.
- ONYX — a bToken (Baseline Markets) backed by an underlying ETH reserve (has a BLV price floor); DEX-traded on Initia (ONYX/INIT); used for reroll / resurrect / rename / respec. → real, external, ETH-denominated value today.
- Planned MUSU↔ONYX bridge — "a bridge will be added to allow ONYX to exist as an item ingame, exchangeable with other players for MUSU." Closes the loop MUSU → ONYX (ETH-backed) → ETH → compute. (Live vs. planned must be stated precisely in the paper.) Sources: CoinGecko/GeckoTerminal (ONYX ETH-backed, Initia DEX); community wiki; play2moon (Asphodel $1.3M seed).
- Founder provenance: publicly articulated the self-funding-compute thesis on X (@0xTokedo, 2026-02-26) — agents plugging into a pure-web3 game, learning from open-source code, earning autonomously; "it's about the fact that it is already possible." Motivation/provenance, not a citation.
Vending-Bench (2502.15840) · Vending-Bench 2/Arena · Project Vend 1&2 · Voyager (2305.16291) · Neural MMO Platform (2110.07594) · Project Sid (2411.00114) · Factorio LE (2503.09617) · BALROG (2411.13543) · lmgame-Bench (2505.15146) · Generative Agents (2304.03442) · Melting Pot 2.0 (2211.13746) · Concordia (2312.03664) · LifelongAgentBench (2505.11942) · StreamBench (2406.08747) · METR time-horizon (2503.14499) · τ-bench (2406.12045) · Foresight Arena (2605.00420) · CoffeeBench (2606.16613). Plus: MUD/Lattice + 0xPARC Autonomous Worlds; AlphaStar (Vinyals et al., Nature 2019) / OpenAI Five as self-play motivation; lifelong-learning roadmap (2501.07278).
- One bullet per work + a one-line "how we position against it."
- Section A = the differentiation table's rows in
kamigotchi-as-benchmark.md. - Sections B–G map onto related-work subsections.
- Before submission: one final pass to confirm every arXiv ID/claim; clear section H.