AI safety & alignment · Verification & proof search · Statistical inference / Open question
Can Finite Audits Certify an Adaptive Conversation?
Multi-turn certificates rely on worst-case transition and persistence properties, while practical audits observe only a finite selection of conversations. Efficiency payoff: Spend audits on safety-relevant states and transitions instead of repeatedly testing whole conversations without a coverage argument.
Seed catalog · Showing the bundled record.
Problem statement
Let conversational states lie in , attacker messages in a known finite set of size , and the fixed model induce unknown kernels that are -Lipschitz in total variation. A known rational polytope is the safe set. With paid reset-and-query samples at chosen , learn a partition into cells and a lower bound on , valid with probability against every history-adaptive attacker . Determine the audit cost of accuracy .
Assumptions and scope
- The queried kernel is the actual fixed conversational system, and the state is sufficient for its dynamics in this benchmark. Reset access to arbitrary states is an explicit strong assumption; a logs-only version needs separately proved coverage or impossibility bounds.
- Choose the partition from a finite family of rational-polytope partitions supplied as input. Include selection complexity and within-cell variation; a representative state's empirical success rate does not bound the worst state in the cell.
- Accuracy means 0<=rho_H-hat_rho_H<=epsilon on the confidence event. Known finite-state robust dynamic programming is a baseline. Charge state queries, partition search and certificate computation, and expose any exponential dependence on state dimension or horizon.
What would settle it?
A high-confidence adaptive audit/abstraction algorithm with tight sample and computational dependence, or an indistinguishability lower bound showing which uniform guarantees finite logs cannot support.
Selected for a theoretical or computational target. A fitted scaling curve or improved benchmark score alone does not establish the stated learning-theoretic guarantee.
Status and evidence
MTCR already gives conditional mode/persistence composition and a dynamic-programming calculation; it explicitly limits formal certification to bounded perturbations. The residual target is learning and validating the required uniform quantities. The continuous-state reset model is an editorial mathematical benchmark, not a claim that deployed LLM hidden states are freely resettable or certified Markov. This is an editorial, source-motivated research target. The preliminary review through 2026-11-20 does not certify novelty or exclude an unlocated resolution.
Preliminary source check; not an expert-verified open status or an independent audit of proposed proofs. The compilation date is not a literature-review cutoff.
Sources and further reading
- Liu et al. — Certified Multi-Turn Robustness for LLM Safety via Compositional Bounds and Safety Persistence ↗partial theory · source checked 2026-11-20
HTML Sections 2–3 checked: conditional mode/persistence bounds, dynamic programming and an explicit bounded-perturbation scope. Proposition 3.7's sufficient conditions and empirical parameter estimates are distinct; finite sampling does not by itself verify infima over every state and adaptive input.
- Kumar et al. — Certifying LLM Safety against Adversarial Prompting ↗partial theory · source checked 2026-11-20
HTML Sections 3–7 and Theorem 4.1 checked: erase-and-check inherits harmful-prompt detection guarantees from a base filter; general infusion enumerates erased subsets. Benign utility is evaluated empirically and the guarantee is not unrestricted semantic output safety.
- Seferis et al. — Randomized Smoothing Meets Vision-Language Models ↗partial theory · source checked 2026-11-20
Primary search abstract checked: generated responses mapped to discrete classes by a bounded-error oracle, with sample/radius laws. Output-level smoothing already has theory conditional on the classifier assumptions.