Mechanism · Sampled inference recomputation
Sources
On this page
- BR. Rinberg et al. (2025). Verifying LLM Inference to Detect Model Weight Exfiltration. arXiv. Source recordSupports: reference architecture; trust assumptions, including external control of user queries; benign evaluation prompts; adversarial prompts an open question; prefill-only cost; exfiltration results; stand-alone limitation · abstract; §4.2; §5; §5.3; §6.1; §8; Fig. 7
- BA. Karvonen et al. (2025). DiFR: Inference Verification Despite Nondeterminism. ICML 2026 Workshop on Technical AI Governance Research. Source recordSupports: Token-DiFR and Activation-DiFR; benign nondeterminism; detection results; weights and sampling requirements; limitations and speculative-decoding sketch · abstract; §3.3; §5; §5.1; §7.2; §7.4; Appendix F
- CAmodo Design (2026). Example Schemes for Verifying High-Stakes AI Agreements. Amodo Design. Source recordSupports: TOPLOC and Token-DiFR as example schemes; recomputation server in prover's data centre; correctness vs completeness · introduction; inference schemes
- BN. Cankaya (2026). A System Overview for Near-Term, Low-Trust AI Compute Verification. Machine Intelligence Research Institute. Source recordSupports: bit-exact replay metadata; evaluation in auditing environment; sampling statistics; attribution problem · §2b; §3.2.2; §5.2.2; Appendix A1
- BN. Cankaya (2026). Bit-Exact AI Inference Verification Without Performance Tradeoffs. ICML 2026 Workshop on Technical AI Governance Research. Source recordSupports: statistical schemes bound but do not close covert bandwidth; detection probability; bit-exact pass/fail · abstract; §1
- CR. Dean (2026). Verification Plan. AI 2040. Source recordSupports: network taps feeding a recomputation server; partial recomputation of random samples · Concrete inference-only retrofitting proposal
- AJ. M. Ong et al. (2025). TOPLOC: A Locality Sensitive Hashing Scheme for Trustless Verifiable Inference. Proceedings of the 42nd International Conference on Machine Learning (PMLR 267), pp. 47196-47211. Source recordSupports: TOPLOC mechanism, results and stated limitations, including untested KV-cache compression · abstract; §4; §5; §6.1-6.4
- BPrime Intellect (2025). PrimeIntellect-ai/toploc (GitHub repository). GitHub. Source recordSupports: public TOPLOC implementation · README; release v0.1.6
- BPrime Intellect Team et al. (2025). INTELLECT-2: A Reasoning Model Trained Through Globally Decentralized Reinforcement Learning. arXiv. Source recordSupports: developer use of TOPLOC to validate untrusted workers · §2.3; §2.4.2
- BA. Karvonen (2025). adamkarvonen/difr (GitHub repository). GitHub. Source recordSupports: public DiFR implementation with vLLM integration · README
- CAmodo Design (2026). Scaling Recomputation Inference Verification. Amodo Design. Source recordSupports: prototype architecture, scale and verifier advantage · whole note; Fig. 3
- BAmodo Design (2026). Amodo-Design/Inference-Recomputation-Prototype (GitHub repository). GitHub. Source recordSupports: prototype code and workflow; DiFR module is a modified copy of the difr library · README
- CPrime Intellect (2025). SYNTHETIC-2 Release: Four Million Collaboratively Generated Reasoning Traces. Prime Intellect blog. Source recordSupports: provider-reported TOPLOC v2 use in SYNTHETIC-2; 1,253 GPUs; false-positive rate over 4 million samples · verification section; GPU section
- CAmodo Design (2026). An Inference Verification Prototype — Stage 1. Amodo Design. Source recordSupports: Amodo's first prototype reuses the DiFR library; tests limited to a fraction of possible attacks · setup; limitations
- CAmodo Design (2026). AI 2040 Plan A — Verification SITREP. Amodo Design. Source recordSupports: status of recomputation algorithms, capture, server security and red-teaming · status items
- BN. Kezins (2026). Adversarial Entropy Inflation Against Gumbel-Based Inference Verification. arXiv. Source recordSupports: independent prompt-control attack on the weight-exfiltration detector; scope limited to the exfiltration bound · abstract; results