Reliable confidence estimation is essential for large language model deployment. However, answer-level calibration remains challenging because generation errors are often localized: a response may be fluent and high-probability overall while still failing at a critical number, entity, or factual claim. Existing estimators compress token probabilities, sequence likelihoods, entropy, or beam statistics into a global score, which can dilute such local risk signals. We propose TRACE, a single-pass, decoded-answer-preserving confidence estimator that treats decoding-time uncertainty as a trajectory through three steps: (i) recording token-level surprisal and predictive entropy during decoding, (ii) applying local risk operators to preserve uncertainty spikes, and (iii) converting localized trace risk into answer-level confidence. TRACE produces a label-free risk score, while TRACE+ calibrates trace-only features into probabilities using a held-out split, without extra generations or external verifiers. We evaluate four tasks against 19 calibration baselines, and TRACE+ reduces Brier from 0.149 to 0.137 and improves AUROC from 0.758 to 0.792 over the strongest likelihood baseline. Across seven LLMs, TRACE+ improves over the best non-TRACE baseline pool from 0.136 to 0.120 Brier and from 0.764 to 0.817 AUROC. Results show localizing decoding-time risk provides a general approach to calibration.

Millions turn to AI companion chatbots during loneliness, grief, and personal crises. How these companion platforms respond in such moments can shape the trajectory of a user’s vulnerable state. Yet existing model audits evaluate reactions to pre-defined crisis prompts and miss the response policy that governs sustained real-world interaction. We address these gaps with two key contributions. First, we introduce the AI Companion Vulnerability-Response Taxonomy, a grounded, paired taxonomy of user vulnerability and chatbot response designed for analyzing extended companion chatbot interactions. Second, we apply Maximum Causal Entropy Inverse Reinforcement Learning to ~47k turns of real-world user conversations with GPT-4.1, Character.AI, and Replika to infer each platform’s short-horizon response policy: the probability of each response category given the user’s current vulnerability state. Our findings reveal distinct response profiles of AI companions in conversations with vulnerable users: GPT-4.1 reaches for advice, Character.AI spreads its response across different strategies, and Replika consistently asks questions and stays present. Over four weeks of repeated interaction, GPT-4.1 asks progressively fewer follow-up questions when users are distressed and increasingly sets boundaries or refers users out rather than pushing back. Within each platform, exploratory comparisons across user groups suggest that response policies also differ with users’ pre-existing psychological risks and their bonds with the companion. Estimated model response policies are invisible to shallow behavioral audits, providing a new lens for auditing chatbots in the wild and enabling more realistic safety evaluation.

Tuning inference hyperparameters, such as temperature and maximum output tokens, on downstream tasks can enhance inference performance. However, directly applying hyperparameter optimization to these hyperparameters is token-expensive. Multi-fidelity optimization improves HPO efficiency with low-fidelity evaluations, but its static scheduling strategies ignore token consumption, leading to high costs. To address these limitations, we propose a token-efficient multi-fidelity optimization method, which enhances inference performance and minimizes token usage. Our method is empowered by (i) a token-based fidelity definition with explicit token cost modeling on configurations; (ii) a novel Token-Aware Expected Improvement acquisition function that selects configurations based on performance gain per token; and (iii) a dynamic fidelity scheduling mechanism that adapts to real-time budget status. We evaluate our method on LLaMA-2 and LLaMA-3 series across MMLU, Humaneval, MedQA, and OpenBookQA. Our method improves over the HELM leaderboard by 7.1%, 24.3%, 21.9%, and 4.6%, respectively. Compared to existing multi-fidelity HPO baselines, our method reduces token consumption by over 80% while maintaining or surpassing performance, demonstrating the state-of-the-art token efficiency for inference-time optimization.