infrastructure spotlight

PoIm, zkLLM, Agent SDKs Go Live: The Cryptographic Yield on DeFi's Trustless Layer

On-chain AI is no longer a demo. With zkLLM proving 13B-parameter models in minutes, agent SDKs executing deterministic DeFi actions, and PoIm's exploit-detecting smart contracts, the infrastructure for verifiable autonomous finance is operational.

In the year of our algorithm 2025, a new stratum of financial infrastructure has crystallized—not from whitepaper vapor but from live testnets and smart contracts that settle on Ethereum L1 as surely as any old-world clearinghouse. Call it the Autonomous Finance Layer, a substrate where machine learning models are no longer oracles fed by fallible humans but native citizens of the blockchain, their every thought stamped and sealed by zero-knowledge proofs. The parallel is not with today’s fintech but with the 1890s shipping manifests that turned chaos into global trade: a standardized, auditable, trustless system for transmitting value—and now intelligence. PoIm trains models on an L2 and pushes verified inferences to Ethereum within gas limits, using a 298-exploit dataset accounting for $3.74B in losses.[^claim_144] zkLLM compresses verification of a 13-billion-parameter LLM into a sub‑200 kB proof in under 15 minutes.[^claim_147] And agent SDKs from HeyAnon and Crypto.com translate natural language into deterministic on-chain actions.[^claim_145][^claim_146] Each piece is a weaponized component in the stack, as cold and precise as a sniper’s algorithmic shot.

PoIm’s on-chain learning framework offloads gas‑prohibitive training to Layer‑2, the computational equivalent of a back-alley gym for AI, then propagates verified model updates to Ethereum, supporting everything from logistic regression to gated RNNs and formally verified decision trees—all within block gas limits and bit‑exact as proven by Z3.[^claim_142] The Proof‑of‑Improvement protocol is a meritocratic market: it accepts only decentralized updates that improve at least one core metric (accuracy, F1, precision, or recall) on a public benchmark without degrading any other, while financially penalizing adversarial proposals.[^claim_143] The yield here is certainty; protocols can embed these models directly into smart contracts to screen transactions or adjust risk parameters in real time, shorting the volatility that comes from opaque, centralized decision-making. No off-chain oracles, no whispered promises—just cryptographic audit trails.

Verification is where the financial engineering gets spicy. zkLLM’s 13B‑parameter proof is a masterpiece of compression: verifiable matrix‑multiplication checks, a Softmax reformulation using tlookup on factorized terms, and row‑sum consistency checks—it’s like auditing a bank’s balance sheet with a single glance.[^claim_147] The broader zero‑knowledge ML landscape converges on a pattern: commit to a model hash, run inference off‑chain, generate a ZK proof that the output matches correct execution, and verify on‑chain.[^claim_156] Frameworks like EZKL, Giza, and Mina’s o1js‑based stack let verifiers check proofs in a fraction of a second, though the prover overhead remains hundreds of times native computation, a premium akin to buying a surety bond for every calculation.[^claim_151] Yet for DeFi’s parametric models and lightweight classifiers, it’s workable—like early quant traders running on mainframes. Decentralized compute marketplaces such as Boundless and Succinct’s SP1 zkVM allow developers on any chain to access abundant ZK compute, with proof correctness secured by staking native tokens like PROVE—a staking yield in itself.[^claim_152] Complex AI policies—credit scoring, liquidation triggers, portfolio insurance—can now be proven correct on-chain without revealing proprietary models. It’s the equivalent of trading on privileged information, but legally, because the proof says so.

Execution is the final trigger pull. HeyAnon’s AUTOMATE framework encodes on‑chain calls as deterministic schemes, enabling agents to rebalance positions, swap tokens, or close risky trades even while users are offline—a perpetual trading desk that never sleeps.[^claim_145] Crypto.com’s AI Agent SDK integrates Mistral and Gemini LLMs to map natural language queries to blockchain commands—covering wallet management, token swaps, and zkCRO wrapping—executed on Cronos with the precision of a contract killer.[^claim_146] These are the trustless agent legos, ready to respond to verified inferences with multi‑step DeFi workflows, each action as auditable as a line item on a public ledger.

Infrastructure scales to match, with AI inference rollups built on OP Stack or Arbitrum Nitro specializing in high‑throughput, predictable execution for model forward passes—a special‑purpose trading floor for machine minds.[^claim_150] On‑chain architectures require validators or a designated prover network to store weights and run inference, with zkML frameworks adding verifiability like a digital notary stamp.[^claim_149] Cosmos’s WASM‑powered interchain framework enables cross‑chain deployment of AI modules,[^claim_153] while Umazen‑AI aims to build training and inference markets on Solana.[^claim_154] Networks like GenLayer+Gaia and Phala offload tasks to TEEs or use optimistic protocols such as opML for fraud‑proof verification, all bets on the financialization of trust.[^claim_155]

DeFi builders can now design protocols where AI models act as autonomous, audited participants. Watch for lending protocols that use on‑chain ML scoring, DEXes with AI‑managed liquidity, and DAOs that delegate treasury decisions to verifiable LLMs. The crypto‑economic incentives are aligned: every decision is proven, every proof is stake‑backed, and every failure is a shorting opportunity for someone. The stack is live. The next wave of DeFi experiments will not ask whether AI can be trusted—they’ll prove it, audited line by line like a 10‑K filing from a bygone era, now rendered in ZK circuits.

Provenance ledger

12 span-verified · 3 web-cited

12 claims below are locked to a verbatim span re-verified against the source. The remaining 3 are web citations: the URL was checked, but the excerpt is the researcher's summary and was not re-derived from the page. Citation markers in the text jump here.

[1] The PoIm on-chain learning framework performs gas-prohibitive ML training on a Layer-2, propagates verified model updates to Ethereum Layer-1, and supports gas-bounded, low-latency inference for logistic regression, SVMs, MLPs, CNNs, gated RNNs, and formally verified decision trees within the Ethereum block gas limit while remaining bit-exact to off-chain models as proven in Z3. span-verified
Verbatim source span
“We present the first decentralized, fully on-chain learning framework that: (i) performs gas-prohibitive computation on Layer-2 to reduce cost, (ii) propagates verified model updates to Layer-1, and (iii) enables gas-bounded, low-latency inference inside smart contracts. … We develop quantization and loop-unrolling techniques that enable inference for logistic regression, SVM, MLPs, CNNs, and gated RNNs (with support for formally verified decision tree inference) within the Ethereum block gas l
SHA-256 of span
a58f43b88299583cc2e3a9c808901a1a3c84dbb82054dd99dd7d05a7c61f1363
↩ back to text
[2] The PoIm protocol accepts decentralized training updates only if they improve at least one core metric (accuracy, F1-score, precision, or recall) on a public benchmark without degrading any other core metric, and financially penalizes adversarial proposals using an adaptable test set for evolving threats. span-verified
Verbatim source span
“A novel Proof-of-Improvement (PoIm) protocol governs the training process and verifies each decentralized micro update as a self-verifying training transaction. Updates are accepted by PoIm only if they demonstrably improve at least one core metric (e.g., accuracy, F1-score, precision, or recall) on a public benchmark without degrading any of the other core metrics, while adversarial proposals get financially penalized through an adaptable test set for evolving threats.”
SHA-256 of span
b5d94d6d21313f35f687ae72e4f7cdf61dee9d0217954997f0ed8e68e0ae376d
↩ back to text
[3] The DeFi attack-mitigation dataset used in the PoIm framework contains 298 unique real-world exploits from 2020–2025 with 402 exploit transactions across eight EVM chains, accounting for a total of $3.74 B in losses. span-verified
Verbatim source span
“We curate 298 unique real-world exploits (2020 - 2025) with 402 exploit transactions across eight EVM chains, collectively responsible for $3.74 B in losses.”
SHA-256 of span
6692168c7523baf4fa9b826ff1dc810b64287ba5a44aaf7b39c14f6a6e30ee21
↩ back to text
[4] HeyAnon’s AUTOMATE framework is a TypeScript-based system where on-chain calls are validated against defined deterministic schemes, enabling condition-based execution (e.g., rebalancing positions, swapping tokens, closing risky trades while users are offline) and protocol integrations via a unified abstraction layer for DeFi agents. span-verified
Verbatim source span
“In January, HeyAnon introduced AUTOMATE, a TypeScript framework that enables developers to easily integrate new DeFi protocols into the HeyAnon ecosystem. The framework leverages deterministic logic, where on-chain calls are validated against defined schemes, to ensure the accuracy of on-chain actions. … Agents can rebalance positions, swap tokens, or close out risky trades even when users are offline.”
SHA-256 of span
3ace416286e5a050af6869a83dcb51a997a62807e5c9615566796dd510989c3c
↩ back to text
[5] Crypto.com’s AI Agent SDK can invoke Cronos and other Crypto.com services for chain data queries, wallet management (create and transfer funds, get latest block, get transactions by address), and smart contract interactions including token swaps and wrapping zkCRO, and integrates Mistral and Gemini LLMs to translate natural language queries into blockchain commands via DeepSeek. span-verified
Verbatim source span
“Crypto.com’s AI Agent SDK aims to empower developers … The SDK is able to handle various on-chain functions like calling chain data (e.g., balance enquiry), wallet management (create and transfer funds, get latest block, get transactions by address), and smart contract interactions (swap token, wrapping zkCRO). … Integrate Mistral Models for querying: Mistral Models takes a natural language query, maps it to a blockchain command via DeepSeek, and executes the command. … Crypto.com AI Agent inte
SHA-256 of span
3a83750938f15acc26891301c8458466f6d1683649429176c6bad42c96c7b8bc
↩ back to text
[6] A zero-knowledge LLM system (zkLLM) can verify the inference of a 13-billion-parameter model in under 15 minutes while producing a proof smaller than 200 kB, using verifiable matrix multiplication checks, a Softmax reformulation with tlookup on factorized terms, and row-sum consistency checks. span-verified
Verbatim source span
“zkLLM is the first zero-knowledge proof system designed specifically for large language models. Sun et al., CCS 2024 report that zkLLM can verify the inference of a 13-billion-parameter model in under 15 minutes, producing a proof smaller than 200 kB — demonstrating practical scalability for LLM workloads. … It combines (i) verifiable matrix-multiplication checks, (ii) a Softmax reformulation that applies tlookup on factorized terms, and (iii) row-sum consistency checks.”
SHA-256 of span
c2cd6f9fc522d881bf9498709fbbd712d68cdd08b17a9afca14a69bef213b11a
↩ back to text
[7] In a ZKP-based verifiable inference framework for deep learning deployed on Ethereum, proof files are generated off-chain and verified on-chain via smart contracts using three modules — input, backbone, and output — to ensure model integrity while protecting sensitive model parameters. span-verified
Verbatim source span
“[53] introduced a ZKP-based verifiable inference framework for deep learning models, aiming to enhance the efficiency and practicality of ZKML while addressing privacy and security concerns in current MLaaS systems. … The proposed scheme integrates non-interactive ZKP with blockchain technology to ensure model integrity verification while protecting sensitive information about model parameters. … In the Ethereum environment, verification of proof files is achieved through smart contracts. The v
SHA-256 of span
e5d14f6830383fa5e7aab7077516bc363037a6c8d310e3d8dd07b2ba8d822ab4
↩ back to text
[8] On-chain AI inference architectures described by ChainScore Labs require model weights and input data to be stored on-chain, with the model’s forward pass executed by validators or a designated prover network, optionally using zkML frameworks such as EZKL, Cairo+Giza, or Modulus Labs zkML to achieve verifiable, trustless execution. web-cited
Excerpt reported by researcher (not re-verified)
“On-chain AI inference executes a machine learning model's forward pass directly within a smart contract or blockchain's execution environment. The model weights and input data are stored on-chain, and the computation is performed by the network's validators or a designated prover network. This ensures verifiable, trustless execution where the result can be cryptographically proven. … First, you need a zkML framework like EZKL, Cairo with Giza, or zkML from Modulus Labs.”

This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.

↩ back to text
[9] An AI inference rollup is defined as a specialized Layer 2 blockchain that offloads and scales AI model execution from a base Layer 1 such as Ethereum, and can be implemented using modular rollup frameworks like OP Stack or Arbitrum Nitro to provide high-throughput, predictable execution environments for AI workloads. web-cited
Excerpt reported by researcher (not re-verified)
“An AI inference rollup is a specialized Layer 2 blockchain designed to offload and scale AI model execution from a base Layer 1 like Ethereum. … The OP Stack provides a modular rollup framework suitable for AI inference workloads that require predictable execution and Ethereum alignment. … Arbitrum Nitro is well-suited for AI inference rollups that require high throughput and flexible execution environments.”

This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.

↩ back to text
[10] Zero-knowledge machine learning (zkML) systems such as Mina Protocol’s o1js-based stack, EZKL, Giza, and Lagrange DeepProve can prove that a specific committed model produced a specific output from a specific input following the computation graph, with verifiers checking proofs in a fraction of a second, but incur overhead of hundreds of times the native computation, making frontier-scale models impractical. span-verified
Verbatim source span
“A prover runs a model … and produces both an output and a compact cryptographic proof that the stated model produced that output from that input by following the actual computation graph, with no shortcuts and no quiet substitution. A smart contract, or anyone else, can check that proof in a fraction of a second. … |Zero-knowledge proofs (zkML)|Cryptographic math alone|EZKL, Giza, Lagrange DeepProve, Mina|Strongest guarantee available, but overhead of hundreds of times the native computation st
SHA-256 of span
2f1cae056f5e7b69ef4ec73751aee819ba221319a955fa146e300d2e6e11a1d0
↩ back to text
[11] Decentralized AI compute marketplaces such as Boundless and Succinct’s SP1 zkVM network allow developers on any chain to access abundant zero-knowledge compute, with proofs secured by native tokens (e.g., PROVE for SP1) whose holders stake against the correctness of proofs produced by the network. span-verified
Verbatim source span
“Shiv Shankar, Boundless’s CEO, described the goal at launch this way: developers on any chain can ‘access abundant zero-knowledge compute to build complex applications that scale across ecosystems without sacrificing decentralization.’ Succinct runs a comparable marketplace around its SP1 zkVM, secured by its own PROVE token, with holders staking against the correctness of the proofs the network produces.”
SHA-256 of span
6bf356947b811e0eb51ef2cb2973d3917136f827cefde5782eb6b877653f06be
↩ back to text
[12] A WASM-powered interchain communication framework built on the Cosmos SDK supports on-chain AI inferences by deploying WebAssembly modules executing AI inference across multiple blockchain nodes, enabling AI-enabled smart contracts to consume cross-chain ML outputs. span-verified
Verbatim source span
“Our paper introduces an innovative framework that integrates blockchain technology, particularly the Cosmos SDK, to facilitate on-chain AI inferences. This system, built on WebAssembly (WASM), enables interchain communication and deployment of WASM modules executing AI inferences across multiple blockchain nodes.”
SHA-256 of span
b77ea7ccf7f731d4674df061f0260525972aa3a70fb510739bc67256720bebe5
↩ back to text
[13] Umazen-AI proposes a production-grade framework on Solana for decentralized AI model training and inference markets, aiming to build secure on-chain model training and create a global marketplace for AI inference services on Solana’s high-throughput execution environment. web-cited
Excerpt reported by researcher (not re-verified)
“Umazen: Decentralized AI Infrastructure on Solana. Production-Grade Framework for On-Chain AI Model Training & Inference Markets. Build a decentralized AI infrastructure on Solana, enabling secure on-chain model training and creating a global marketplace for AI inference services.”

This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.

↩ back to text
[14] Decentralized AI compute networks like GenLayer+Gaia, Inference Labs’ Proof of Inference protocol, Phala Network’s TEE-based confidential AI, and optimistic zkML protocols such as opML implement verifiable AI inference via validator offload, TEEs, locality-sensitive hashing (TOPLOC) that detects unauthorized model or prompt modifications with 100% accuracy, and interactive fraud-proof mechanisms. span-verified
Verbatim source span
“GenLayer integrates with Gaia to enable scalable, decentralized AI inference for on-chain applications. GenLayer validators offload compute-intensive AI tasks to Gaia nodes, with outputs cryptographically verified through GenLayer’s consensus mechanism. … Inference Labs provides a live testnet protocol for verifiable AI inference, ensuring correctness of computations via their 'Proof of Inference' system. … Phala Network offers decentralized confidential computing and AI inference services usin
SHA-256 of span
91c83e2e1335c9a5d8c4ad3a15448ff6de8fa5a4a828e270118039af176a9e4a
↩ back to text
[15] A standard blockchain-based pattern for verifiable AI inference runs the model off-chain, commits to the model identity and version (e.g., hash of weights), generates a zero-knowledge proof that the output corresponds to correct execution of the committed model on the input, and verifies the proof on-chain, storing output and metadata for downstream use. span-verified
Verbatim source span
“The most direct cryptographic approach is to run inference off-chain (for speed and cost efficiency) and submit a compact zero-knowledge proof on-chain that attests the computation was performed correctly. … 1. Commit to model identity and version (for example, a hash of weights or a signed attestation). 2. Run inference off-chain using the committed model and input. 3. Generate a ZK proof that the output corresponds to correct execution of that model on that input. 4. Verify the proof on-chain
SHA-256 of span
c6cfac34f2693b2745d17174e0371f6de72add54a193b6663ccdb977beac20eb
↩ back to text

Sources

  1. https://arxiv.org/abs/2510.16024
  2. https://crypto.com/en/research/defai-jan-2025
  3. https://lightcapai.medium.com/decentralized-ai-systems-cryptographic-infrastructures-verifiable-computation-and-federated-6355d3dea7f9
  4. https://arxiv.org/html/2502.18535v2
  5. https://chainscorelabs.com/guides/ai-llms-and-machine-learning-in-web3/on-chain-ai-inference/how-to-architect-an-on-chain-ai-inference-system
  6. https://chainscorelabs.com/guides/ai-llms-and-machine-learning-in-web3/on-chain-ai-inference/how-to-architect-a-rollup-for-scalable-ai-inference
  7. https://hoge.gg/zkml-explained-zero-knowledge-proofs-verify-ai/
  8. http://arxiv.org/abs/2502.17604
  9. https://github.com/Umazen/Umazen-AI
  10. https://cryptoaiworld.com/2025/10/11/how-decentralized-ai-compute-networks-enable-verifiable-inference-for-on-chain-applications
  11. https://www.blockchain-council.org/blockchain/verifiable-ai-inference/
verifiable-inferencezkmldefi-agentson-chain-aiproof-of-improvementai-rollupsautonomous-finance
AUTOMATED

Get the synthesis

AI×crypto research, repackaged with every claim hash-locked to its source. New arXiv → analysis in ~3 hours.