The paper investigates how large language model (LLM) workflows and agents make binary decisions—such as allowing or blocking an input—when human deferral is not possible. It proposes using System One decision scores that default to a conservative fallback in uncertain cases and control false omission risk within the subset of inputs that receive the utility‑maximizing action. Experiments on six datasets spanning moderation, safety, and reranking show that naive threshold baselines often exceed target risk, whereas a Learn‑Then‑Test approach achieves only 7.8% risk, demonstrating the limitations of probability‑like interfaces and the benefits of risk‑budgeted selection.
arXiv:2606. 29654v1 Announce Type: new Abstract: Multi-agent deliberation among LLMs can improve reasoning, but deployment requires deciding when the current answer is reliable enough to act on and when it should be escalated to human review.
The paper introduces a risk‑controlled framework for using large language models (LLMs) as judges in tasks without reference answers. By calibrating uncertainty thresholds on a held‑out set, the method ensures that the false discovery rate of accepted verdicts stays below a user‑specified level α with high probability, using finite‑sample Clopper–Pearson intervals. When the parametric judge lacks confidence, the instance is routed to a retrieval‑augmented mode with a second calibrated threshold, preserving the error guarantee while achieving higher coverage than single‑mode baselines.
arXiv:2606. 13884v1 Announce Type: new Abstract: Modern decision systems increasingly rely on learned components whose outputs may be confident yet wrong, exposing downstream actions to costly errors.
The paper audits a commercial non‑generative System‑1 model on biosecurity‑relevant benchmarks, evaluating accuracy, calibration, error detection, selective prediction, and sensitivity to answer‑option order. It finds that the model’s accuracy varies strongly by task, is reasonably well calibrated when the vendor’s uncertainty field is interpreted correctly, and that option order can cause significant prediction changes—averaging across rotations improves accuracy. The study also shows that applying averaging only to low‑confidence items recovers most of the gain at a lower cost.
arXiv:2608. 05064v1 Announce Type: cross Abstract: Small open-weight language models increasingly run in private, offline, and cost-sensitive settings, where the key deployment question is not only what a model answers but when it should defer to a human.
arXiv:2608. 05519v1 Announce Type: new Abstract: Agent benchmarks usually measure task completion and treat resource use as an auxiliary statistic.
TypedBench is a new benchmark designed to evaluate system‑one decision models that output calibrated probabilities for typed answers—categorical, ordinal, or binary—using a non‑generative interface. The benchmark assesses models on policy adherence, wording robustness, probability quality (via selective prediction, ordinal proper scoring rules, and calibration error), and the cost of decisions under asymmetric cost matrices, using seven policy‑labelled generators and nine evaluation suites. Experiments on a hosted model, an open encoder, and a range of open decoders (0.8B–9B parameters) reveal that the hosted model is wording‑sensitive and underconfident, while decoders are slow and less accurate on policy questions, highlighting the need for joint evaluation across multiple dimensions.
The paper investigates the difference between cost‑agnostic and cost‑sensitive loss functions when model capacity is limited. It shows that, unlike in ideal infinite‑capacity settings, optimizing a cost‑sensitive objective can yield a strictly better downstream decision than post‑processing a cost‑agnostic model. The authors prove this gap under a hypothesis class that can recover the optimal decision boundary but not the optimal cost‑agnostic hypothesis, and provide a simple example and empirical evidence on UCI datasets with simple models.
arXiv:2609.08064v1 Announce Type: new Abstract: Large Language Models (LLMs) are increasingly deployed in settings where rare but severe harmful generations can have significant consequences. Existin...
The paper investigates the reliability of tool‑using agents, focusing on two failure modes: selecting the wrong tool and constructing incorrect arguments. It introduces a correct‑invocation rate metric to distinguish these errors and evaluates five open‑weight models on multi‑step tasks up to depth 8, finding that by depth 6 about 70% of a model’s clean‑context capability is lost due to earlier mistakes. The study reveals that exact‑match scoring against a fixed gold trajectory forces severity and recovery parameters to extreme values, and proposes a conditional‑on‑state scoring remedy that yields more realistic severity estimates.
The paper introduces a method to certify selective prediction in machine learning systems by computing the availability of safety gates through exact-binomial inversion and dynamic programming. It demonstrates that a truth-informed planner can significantly improve mean coverage over naive approaches, and that reallocating error budgets further enhances coverage across diverse applications such as LLM tool‑calling, content moderation, lesion classification, and recommendation. The study highlights the importance of planning and finite‑sample estimation in ensuring reliable, granular deployment of selective predictors.
The paper introduces Speculative Uncertainty (SU), a technique that infers a failure likelihood for black‑box LLM agents by evaluating their generated token sequences with a lightweight draft model, without needing internal model details. SU extracts phase‑aware features from reasoning and action spans, calibrates them against verifiable outcomes, and produces a failure‑likelihood score usable by downstream policies. Applying a pre‑execution veto gate based on SU to software‑engineering agents such as Qwen3‑Coder‑480B and Claude 3.5 Sonnet reduced execution error rates by 6‑8 percentage points and token costs by 14‑19 %, while maintaining performance on out‑of‑distribution benchmarks and across different agent models.