Trust The Typical

Current approaches to LLM safety fundamentally rely on a brittle cat-and-mouse game of identifying and blocking known threats via guardrails. We argue for a fresh approach: robust safety comes not from enumerating what is harmful, but from deeply understanding what is safe. We introduce Trust The Typical (T3), a framework that operationalizes this principle by treating safety as an out-of-distribution (OOD) detection problem. T3 learns the distribution of acceptable prompts in a semantic space and flags any significant deviation as a potential threat. Unlike prior methods, it requires no training on harmful examples, yet achieves state-of-the-art performance across 18 benchmarks spanning toxicity, hate speech, jailbreaking, multilingual harms, and over-refusal, reducing false positive rates by up to 40x relative to specialized safety models. A single model trained only on safe English text transfers effectively to diverse domains and over 14 languages without retraining. Finally, we demonstrate production readiness by integrating a GPU-optimized version into vLLM, enabling continuous guardrailing during token generation with less than 6% overhead even under dense evaluation intervals on large-scale workloads.

Grammars of Formal Uncertainty: When to Trust LLMs in Automated Reasoning Tasks

Large language models (LLMs) show remarkable promise for democratizing automated reasoning by generating formal specifications. However, a fundamental tension exists: LLMs are probabilistic, while formal verification demands deterministic guarantees. This paper addresses this epistemological gap by comprehensively investigating failure modes and uncertainty quantification (UQ) in LLM-generated formal artifacts. Our systematic evaluation of five frontier LLMs reveals Satisfiability Modulo Theories (SMT) based autoformalization's domain-specific impact on accuracy (from +34.8% on logical tasks to -44.5% on factual ones), with known UQ techniques like the entropy of token probabilities failing to identify these errors. We introduce a probabilistic context-free grammar (PCFG) framework to model LLM outputs, yielding a refined uncertainty taxonomy. We find uncertainty signals are task-dependent (e.g., grammar entropy for logic, AUROC>0.93). Finally, a lightweight fusion of these signals enables selective verification, drastically reducing errors (14-100%) with minimal abstention, transforming LLM-driven formalization into a reliable engineering discipline.

Adversarial and Safely Scaled Question Generation

Question generation has recently gained a lot of research interest, especially with the advent of large language models. In and of itself, question generation can be considered 'AI-hard', as there is a lack of unanimously agreed sense of what makes a question 'good' or 'bad'. In this paper, we tackle two fundamental problems in parallel: on one hand, we try to solve the scaling problem, where question-generation and answering applications have to be applied to a massive amount of text without ground truth labeling. The usual approach to solve this problem is to either downsample or summarize. However, there are critical risks of misinformation with these approaches. On the other hand, and related to the misinformation problem, we try to solve the 'safety' problem, as many public institutions rely on a much higher level of accuracy for the content they provide. We introduce an adversarial approach to tackle the question generation safety problem with scale. Specifically, we designed a question-answering system that specifically prunes out unanswerable questions that may be generated, and further increases the quality of the answers that are generated. We build a production-ready, easily-plugged pipeline that can be used on any given body of text, that is scalable and immune from generating any hate speech, profanity, or misinformation. Based on the results, we are able to generate more than six times the number of quality questions generated by the abstractive approach, with a perceived quality being 44% higher, according to a survey of 168 participants.