Open-source large language models (LLMs) are increasingly competitive with closed-source models while offering transparency and the ability to run inference without exposing user inputs to a service p...
arXiv:2609. 01945v1 Announce Type: cross Abstract: Federated Learning enables multiple clients to train a shared model while keeping their local datasets isolated.
arXiv:2606. 09551v1 Announce Type: cross Abstract: Two-server secure inference allows a client to query a hosted large language model (LLM) without revealing prompts or embeddings.
The paper introduces Rewindable Interactive Proofs (RIPs), a new model for scalable oversight where a weak verifier can rewind a powerful AI prover to an earlier state during interaction. RIPs enable verification of oracle‑aided computations that are robust to small changes in oracle answers, achieving near‑linear verifier time without relying on cryptography. The authors provide a construction, implement it on MNIST, and demonstrate that RIPs can outperform existing proof systems in practice.
arXiv:2608. 02774v1 Announce Type: cross Abstract: AI verification crosses a trust boundary: a verifier must learn enough to establish an authorized claim, yet the same evidence can reveal sensitive details about the model, workload, or hardware.
The paper introduces Rewindable Interactive Proofs (RIPs), a new model for scalable oversight where a weak verifier can rewind a powerful AI prover to an earlier state during interaction. RIPs enable efficient verification of computations whose outputs are robust to small changes in oracle answers, achieving near-linear verifier time without relying on cryptography. The authors implement the protocol, demonstrate a preliminary MNIST experiment, and show that an RIP can outperform a state‑of‑the‑art deployed proof system.
arXiv:2607. 21162v1 Announce Type: new Abstract: Outsourced Transformer inference exposes clients to model substitution and incomplete execution, while direct replay removes the computational benefit of delegation.
PCM is a protocol that attaches verifiable evidence to every claim sent between AI agents, using immutable Merkle logs and signed attestations that the sender cannot tamper with. Receivers evaluate messages offline and classify each claim as proven, unproven, or failed, ensuring that self‑reported claims can never be proven and that authority is traceable through recorded human approvals. In a benchmark of 926 real messages, PCM flagged only 10 stale‑evidence lies (1.4%) while other protocols accepted many more, and verification runs in about 1 ms with perfect agreement between independent verifiers.
The paper introduces TLA-Verus, a system that translates TLA+ specifications into the Verus verifier using both deterministic and agentic (LLM-based) translators. It demonstrates that the LLM pipeline can generate proof obligations at scale and produce fully machine‑checkable proofs for safety and liveness properties. The work also provides a new evaluation dataset for assessing LLM capabilities in Verus theorem proving.
The paper investigates whether interactive arguments for oracle‑aided AI computations can be zero‑knowledge, meaning the verifier learns nothing beyond the correctness of the output. It proves that, in general, zero‑knowledge proofs for all oracle‑aided computations are impossible, even in the random oracle model, and this impossibility extends to debate protocols. However, if the oracle signs each answer with a cryptographic signature, then every oracle‑aided computation can be verified in zero‑knowledge with efficient provers and verifiers, assuming only collision‑resistant hash functions.
TensorCommitments (TCs) is a lightweight, tensor-native proof‑of‑inference scheme that enables verifiable inference for large language models (LLMs) without requiring the verifier to rerun the model or possess a powerful GPU. By binding each inference to a commitment stored in multivariate Terkle Trees, TCs detect tampering with only a 0.97% overhead for the prover and 0.12% for the verifier on LLaMA2. The approach improves robustness against tailored LLM attacks by up to 48% compared to previous methods that needed a verifier GPU.
SEVERUS is a context‑aware, gamified framework that generates Verus benchmarks by organizing verification tasks into quests and progressively removing auxiliary lemmas to create levels of increasing difficulty. It constructs 819 quests with 10,095 levels from six open‑source Rust projects, and uses an independent checker to enforce soundness while a noisy binary search estimates agent capabilities. Experiments with GPT‑6 Sol, GPT‑6 Luna, and Claude Code show that as proof context is withdrawn, success rates drop and costs rise, with the strongest agents solving only 35% of levels reliably.