Language-model judges and reward models enable scalable supervision, but finite optimization can exploit evaluator errors rather than improve response quality. We characterize this failure through the covariance geometry of evaluator ensembles. For calibrated judges, the ensemble mean retains common-mode error along the all-ones direction, whereas cross-judge disagreement captures only orthogonal error. Consequently, disagreement can be high despite robust aggregation, or low while shared response-dependent errors persist. We prove that common-mode error is not identifiable from internal judge scores alone. Under a joint sub-Gaussian model, we bound best-of-K selection overstatement and target-quality regret, extending the guarantees to predictably adaptive search under conditional calibration. The resulting search terms scale as the square root of log K and are asymptotically tight for Gaussian projected errors. We further show that noisy quality proxies introduce artificial rank-one covariance without changing disagreement, and propose a bounded two-anchor Bernstein certificate for finite-search error and regret. Fixed-seed Gaussian stress tests over 120 (J, rho, K) configurations and real-model audits validate the theory while revealing the limits of disagreement-based diagnostics under increasing search pressure.
Process reward models (PRMs) score intermediate reasoning steps and are widely used for search, ranking, and training, but optimization can exploit these learned proxies by increasing reward while turning correct reasoning into incorrect reasoning. We formulate PRM stress testing as a quality-diversity search problem using MAP-Elites, retaining the most severe correctness-flipping edit in each behavior-space region while separating search coverage from exploit coverage. We characterize what such archives certify: finite-cell repair bounds covered-cell tail risk and average residual severity but cannot bound the worst remaining cell from covered fraction alone; under Lipschitz post-repair loss and metric-cover auditing, the residual is bounded by archive fitting error plus the Lipschitz constant times the covering radius. A controlled landscape validates this certificate and the impossibility of any fraction-only worst-case guarantee. On real PRMs, the search reveals an aggregation-dependent vulnerability in Qwen2.5-Math-PRM-7B: padding yields 44 strict exploits with maximum gain 0.294 under mean pooling versus one exploit under minimum readout; a matched syntactic control isolates the mechanism, and an RLHFlow value-head model shows the same qualitative effect with maximum gain 0.005. A predeclared paired LoRA repair protocol reduces exploit rates from 0.148 to 0.037 to 0.074, lowers the worst attack from 0.333 to 0.177 to 0.212, improves ranking AUROC without degrading best-of-4 accuracy, attributes gains to adversarial fine-tuning rather than archive diversity, and is confirmed by independent unpaired replications (44 to 1, clean-split worst gain 0.0092, MATH-500 41 to 0, clean ranking 40/40).
CODS is introduced, a critic-guided selector that alternates between fitting an algorithm-matched critic and acquiring high-residual transitions before freezing a reusable subset, a reusable selection procedure, not a formal coreset guarantee.
Ibne Farabi Shihab, Sanjeda Akter, A. Ahsan et al.· 0 citations
Negative sampling determines whether a knowledge graph embedding (KGE) model learns from informative counterexamples or wastes updates on implausible corruptions. Uniform negatives are diverse but easy, whereas hard-negative miners concentrate on few entities and collide more with held-out positives. We introduce FlowNeg, a context-conditioned hierarchical generative flow network that amortizes reward-proportional sampling without normalizing a composite reward over the entity set: given a positive triple and corruption side, it selects a type, then an entity. Its terminal reward combines bounded model-based hardness with a training-only structural score for held-out-positive collision, over a relation-specific type-compatible support. We derive the reward, specialize standard trajectory balance, and bound multiplicatively how residual imbalance perturbs terminal and mode probability. Across a descriptive five-seed grid of five architectures and five benchmarks, FlowNeg has higher mean MRR than EMU and than IF-NS in 24 of 25 cells ($+0.0172$ and $+0.0160$ on average). A separate 15-seed FB15k-237/RotatE control fixing negative count, diagnostic budget, and compute gives FlowNeg $0.359\pm0.001$ MRR against $0.346\pm0.002$ for EMU, with near-uniform fixed-partition diversity, high gradient informativeness, and low collision. The evidence supports mode-covering negative generation without treating structural similarity as an open-world truth oracle.
Oracle routing measures how much a pool of language models could gain from per-query selection, but the diagnostic has two flaws: testing against a best fixed model selected on the same examples invalidates paired inference, and a full-information oracle sees outcomes no deployable router observes. We separate three estimands (outcome-oracle opportunity, the Bayes-optimal gain from a declared pre-answer signal, and the held-out gain of a learned router) and prove selection-valid confidence intervals that survive choosing the best fixed model or the best member of a router family, a signal-information sandwich, and a $(1-1/e)$ greedy guarantee for building compact pools from submodular complementary coverage. On eight checkpoints from six families over four benchmarks, selection-valid intervals certify a population oracle gap of $9.7$--$30.7$ points on every task, yet the strongest deployable prompt router recovers only $7.5$--$14.4\%$ of it, and the simultaneous interval for the best of eleven tested policies has lower limit zero throughout. The realizable share of oracle opportunity is small and certifiable: strong routers beat the best fixed model, and most of the gap remains.
Ibne Farabi Shihab, A. Ahsan, Md Najmus Swaqeeb· 0 citations
A finite-horizon composition law for Gaussian certification from M independent coordinates, a familywise certifier protecting every coordinate and time up to T pays optimal normalized squared half-width of order log(eM) + log log(e^eT), within the sample-mean-centered rectangular class.