P082 Statistical inference verification (TOPLOC / DiFR family)

Do statistical inference-verification methods reliably catch cheating with few false failures on large models, and do adaptive attacks defeat them?

Statistical inference verification (TOPLOC / Token-DiFR family).

Terms: TOPLOC, Token-DiFR: named published methods that statistically check whether a model really produced the outputs it claims, without re-running the whole computation. k: in top-k sampling, the number of most likely tokens the sampler chooses among; the "simulated k = 2 bug" makes 1% of tokens a uniform pick from the top two.

Status

Started Token-DiFR re-implemented by a third party and run on a tray of H100s against non-adaptive attacks (substituted models including gpt-oss-120b versus 20b, LoRA append and append-on-trigger on Qwen2.5-7B); the best honest-versus-cheating pass-rate gap was about 25%, the threshold must be set per model, and most misses were outputs under 5 tokens. Adaptive-attacks clause: the DiFR paper (Karvonen et al. 2025) shows Token-DiFR robust to one simple adaptive attack (temperature tuned to match mean cross-entropy, which defeats cross-entropy checks); stronger adaptive attacks untested, and Amodo's own note says its testing 'only covers a fraction of possible attacks'. Models tested are well below frontier size; the smallest deviations on Qwen3-30B-A3B 'remain difficult to separate' in the DiFR paper.

Our results add further evidence that Token DiFR can identify different incorrect-model attacks reliably across a small range of prompts.
Amodo Design (2026), An Inference Verification Prototype, Stage 1 (design note) (2026-06-29), L235

Posed by

Also raised in: Rinberg et al. (2025), Verifying LLM Inference to Detect Model Weight Exfiltration; AI 2040 (2026), Get Involved in Verification; Amodo (2026), AI 2040 Plan A — Verification SITREP; Amodo (2026), An Inference Verification Prototype — Stage 1; Amodo (2026), Example Schemes for Verifying High-Stakes AI Agreements.

Sub-questions

IDSub-questionWhat the sources report
P082.1Have TOPLOC / Token-DiFR been validated on frontier-scale models (a stated parameter count) and relevant production hardware?partly answered
P082.2At what per-model threshold does Token-DiFR give a stated honest-failure rate with a stated cheat-detection rate, per response?partly answered
P082.3Can thresholds or short outputs (<5 tokens) be exploited by prompt-based attacks to evade detection?open
P082.4Does any adaptive attack pass Token-DiFR at a stated honest-failure rate while changing a stated fraction of tokens (red-teaming)?open
P082.5Can a provider pass TOPLOC/LOGIC with arbitrary text (reconstructability vs generation authenticity)?a source suggests an answer (not checked)
P082.6Do temperature-tuning or temperature-zero selective cheating defeat cross-entropy or spot-check verification?a source suggests an answer (not checked)
P082.7Can small deviations (sampling temperature 1.1, top-k bugs), MoE batch-dependent routing and single sequences within a batch be verified at a stated AUC?partly answered

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 P082.2: At what per-model threshold does Token-DiFR give a stated honest-failure rate with a stated cheat-detection rate, per response?

Data