Research map / Large language models / Reasoning and chain of thought

TensorCommitments: A Lightweight Verifiable Inference for Large Language Models

NeurIPS 2026 · Large language models › Reasoning and chain of thought · approach: Efficient LLM serving

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes TensorCommitments (multivariate polynomial commitments over activation tensors) organized in Terkle trees plus a spectral-score layer selector, enabling lightweight verifiable LLM inference without verifier GPU.

Problem

Verifying that a remote LLM inference was executed correctly without re-running the model or exposing internal activations; existing cryptographic proofs are too slow and non-cryptographic methods need verifier GPUs or retraining.

Approach

Commit to activation tensors as multivariate polynomials (tensor-native commitments) rather than univariate vector polynomials, organize commitments in a high-arity Terkle tree with a single root, and verify only layers chosen by a heavy-tailed spectral score under an interval-selection DP.

Key results

• For LLaMA 2-13B, prover overhead 0.96% and verifier overhead 0.12% of 10.165s base inference, with 2 B commitment size per token.
• Attack detection accuracy 96.02% vs 82.59% for TOPLOC; median detection improved by 12% under noise injection and 48% under prompt tampering.
• Terkle trees verify up to 1416x faster than Merkle trees and 14x faster than Verkle trees at branching factor 64.

Surprising findings

• Last-layer top-k fingerprints (TOPLOC) can be evaded by prompt tampering that preserves the top-k activations while changing the output semantically.

Limitations

• Requires prover-side GPU with full activation access.
• Not full zero-knowledge; public parts of the model are not hidden.
• Terkle trees are slow and sparse at very low arity.

Experimental setup

Metrics: prover time verification time commitment size per token attack detection accuracy attack manipulation coverage GPU memory

Baselines: zkLLM SVIP TOPLOC Raw Activations Merkle tree Verkle tree

Builds on: KZG polynomial commitments Verkle trees Merkle trees TOPLOC alpha-pruning

Closest papers by meaning

• TOPLOC: A Locality Sensitive Hashing Scheme for Trustless Verifiable Inference (ICML 2025 · similarity 0.90)
• Hidden No More: Attacking and Defending Private Third-Party LLM Inference (ICML 2025 · similarity 0.88)
• EncryptedLLM: Privacy-Preserving Large Language Model Inference via GPU-Accelerated Fully Homomorphic Encryption (ICML 2025 · similarity 0.87)
• Puzzle: Distillation-Based NAS for Inference-Optimized LLMs (ICML 2025 · similarity 0.87)
• VIA-SD: Verification via Intra-Model Routing for Speculative Decoding (ICML 2026 · similarity 0.86)
• Exploring the limits of strong membership inference attacks on large language models (NeurIPS 2025 · similarity 0.86)