P051 Sampling procedures and sampling randomness
Can an adversary who controls or chooses sampling seeds encode exfiltration bits or evade detection without raising the false-positive rate?
Sampling procedures and seed control as a verification gap.
Status
Analyzed A written analysis of the seed-control channel and the design requirement that closes it (a trusted cryptographic RNG bound to the request, not choosable or biasable by the adversary); no bit count is measured and no seed-controlling adversary is tested. Rinberg et al.'s measured exfiltration bounds (Llama and Qwen) assume trusted randomness, so they do not bear on this headline. Amodo's Stage 1 prototype records that its seed is set by the prover and should be verifier-defined, a documented gap. Speculative decoding and non-Gumbel samplers are unevaluated (DiFR).
"If the adversary can manipulate randomness, they can transmit information by selecting seeds/samples while leaving per-step likelihoods within normal ranges. In this regime the verifier’s guarantees degrade sharply: many exfiltration strategies become indistinguishable from benign stochasticity, and a malicious provider could even watermark or backdoor outputs via RNG tampering while remaining white-box undetectable [GKVZ24, CG24]."
Posed by
- Rinberg et al. (2025), Verifying LLM Inference to Detect Model Weight Exfiltration:
If the adversary can manipulate randomness, they can transmit information by selecting seeds/samples while leaving per-step likelihoods within normal ranges.
(App. E.5, p.28, L2206–2207)
Also raised in: Karvonen et al. (2025), DiFR: Inference Verification Despite Nondeterminism; Amodo (2026), An Inference Verification Prototype — Stage 1.
Sub-questions
| ID | Sub-question | What the sources report |
|---|---|---|
| P051.1 | Does logging per-token sampling mode let verification cover speculative / constrained decoding, and how many extra exfiltratable bits does a draft model give? | open |
| P051.2 | Does Token-DiFR keep honest/cheat separation under non-Gumbel samplers (e.g. min-p)? | open |
| P051.3 | Does deterministic tie-breaking at branching tokens cut exfiltratable bits by a stated factor at a stated throughput cost? | open |
| P051.4 | How many bits can an adversary who controls or chooses sampling seeds encode without raising the false-positive rate? | open |
Possibly relevant work found by automated search
Found by AI agents searching for this problem's sub-questions, mostly by keyword and citation in the OpenAlex scholarly database and on the web; a second AI pass reviewed each one, and we spot-checked some by hand. They seemed reasonably related, as far as one can tell from outside each paper's field. Some come from neighbouring fields, such as arms control and financial audit, and bear on the problem by analogy. Not reviewed against the status above; a paper listed here does not change that status. 928 of 1052 quotes checked word for word against the paper's text.
For P051.1: Does logging per-token sampling mode let verification cover speculative / constrained decoding, and how many extra exfiltratable bits does a draft model give?
For P051.2: Does Token-DiFR keep honest/cheat separation under non-Gumbel samplers (e.g. min-p)?
For P051.3: Does deterministic tie-breaking at branching tokens cut exfiltratable bits by a stated factor at a stated throughput cost?
For P051.4: How many bits can an adversary who controls or chooses sampling seeds encode without raising the false-positive rate?
Data
- problems.csv — one row per problem
- problem_status.csv — status, scope and evidence per problem
- problem_sources.csv — the posed-by quote and other documents per problem
- leaves.csv — all sub-questions, including those without a checkable form
- leaf_sources.csv — every source row behind each sub-question