Safety
Auditing LLM Program Search for Bermudan Stopping with Primal--Dual Certificates
arXiv:2610.14063v1 Announce Type: new Abstract: LLM program-search agents are usually judged by the score that drives their search. We study LLM-driven search over Python exercise policies for Bermuda
arXiv:2610.14063v1 Announce Type: new Abstract: LLM program-search agents are usually judged by the score that drives their search. We study LLM-driven search over Python exercise policies for Bermudan options, an optimal stopping problem where a trusted evaluator owns the simulation and a dual upper bound~U. For any policy, 0 le V^star - L_pi le U - L_pi; we report Monte Carlo estimates of both bounds with their sampling error. We audit 24 generation attempts from two pipelines, a six-pair feedback-representation ablation, and a holdout evaluated after selection was frozen. No selected generated program improves on Ridge LSMC consistently or by more than Monte Carlo uncertainty; holdout gap scores lie within 0.002 of it. On six one-dimensional puts, a high-precision binomial reference places Ridge within Monte Carlo noise of optimal and attributes 99% of the estimated gap to dual-bound looseness, so this dual audits validity but cannot certify progress at the search's scale. Structured memory showed no benefit over score-only memory in this ablation. Whether search ``worked'' depends on reporting rules: retaining the initial reference, freezing selection before holdout, and counting failures.
Related
- A Two-Timescale Primal-Dual Framework for Reinforcement Learning via Online Dual Variable Guidance
- Last-Iterate Convergence of General Parameterized Policies in Constrained MDPs
- Equilibrium bias and convergence in augmented primal--dual dynamics with sampled constraints
Source: arXiv cs.LG | 2026-10-14