The paper introduces multicalibration to improve confidence scores in LLMs.
problem Improving the reliability and interpretability of confidence scores for LLMs.
method Forming groupings of prompt/completion pairs correlated with correctness, using clustering and self-annotation. Developing multicalibration algorithms to reduce overfitting.
result Our techniques yield confidence scores that significantly improve calibration and accuracy compared to existing methods.
LLMs show surprising confidence in their answers, beyond just tokens.
problem LLMs lack meaningful confidence estimates for their responses.
method Semantic calibration test based on local loss optimality and equivalence classes.
result Base LLMs are semantically calibrated across tasks, contrary to expectations.
This work investigates uncertainty quantification for black-box large language models in natural language generation.
problem Lack of trustworthiness in responses generated by black-box large language models.
method Differentiated uncertainty vs confidence, proposed and compared several confidence/uncertainty measures, applied to selective NLG.
result A simple measure for semantic dispersion can predict the quality of LLM responses.
Tree-Query uses LLMs to discover causal relationships in a transparent, interpretable manner.
problem Error propagation in classical causal discovery methods and opaque, confidence-free behavior of recent LLM-based causal oracles.
method Tree-Query is a tree-structured, multi-expert LLM framework that reduces causal discovery to queries about backdoor paths and dependencies.
result Tree-Query provides interpretable judgments with robustness-aware confidence scores and improves structural metrics over LLM baselines.
New framework assesses LLMs' expertise using nonparametric ranking and confidence diagrams.
problem Evaluating and ranking large language models (LLMs) for alignment and performance.
method Nonparametric contextual ranking, confidence diagram, Gaussian multiplier bootstrap.
result Validated confidence diagram for assessing LLMs' domain-specific expertise.
Paper presents a probabilistic model to improve LLM cascade performance.
problem Complexity of LLM cascades and their interaction error rates.
method Probabilistic model for joint performance distribution of LLMs.
result Improves area under the error-cost curve by 4.3% on average for cascades with k≥3 models.
LLMs overestimate stock returns and are less accurate at predicting extreme outcomes.
problem Behavioral biases in LLMs' stock return forecasts.
method Comparison of LLM forecasts with crowd-sourced estimates and historical data.
result LLMs overestimate stock returns and are less accurate at predicting extreme outcomes.
LLMs generate answers under incomplete context, and their uncertainty should scale with missing information.
problem Evaluating the quality of LLM answers under incomplete context.
method A controlled framework with varying context availability, and two uncertainty measures (sampling-based confidence and response entropy) evaluated on SQuAD.
result Response entropy increases with context removal and explains more variance in accuracy than confidence, suggesting it is a more responsive uncertainty measure.
A new framework evaluates LLM calibration in open-ended QA.
problem Evaluating LLM calibration in open-ended QA settings.
method Sem-ECE framework: sampling answers, grouping by semantic classes, and using frequencies as confidence.
result Sem-ECE estimators are unbiased and Sem2 achieves smaller calibration error. Corrects bias in LLM-as-a-judge evaluations using adaptive calibration.
problem Bias in LLM evaluations due to imperfect sensitivity and specificity.
method Plug-in framework with confidence intervals accounting for test and calibration dataset uncertainties.
result LML-based evaluation yields more reliable estimates than human-only evaluation.
The study examines label smoothing to improve confidence calibration in fine-tuned LLMs.
problem Improving confidence calibration in fine-tuned large language models (LLMs) after instruction tuning.
method Examine various open-sourced LLMs, label smoothing, and custom kernel design.
result Label smoothing is effective in maintaining confidence calibration but faces challenges in large vocabulary LLMs.
Aggregates diverse zero-shot LLM outputs for better corporate disclosure classification.
problem Combining varied zero-shot LLM predictions for improved stock return prediction.
method Multi-prompt framework with three fixed zero-shot LLM classifiers, logistic meta-classifier aggregation.
result Aggregated model outperforms single classifiers and baseline models, increasing balanced accuracy from 0.566 to 0.606.
GPT-4 assesses its confidence in answering USMLE questions with and without feedback.
problem Understanding AI's performance in healthcare applications, especially in sensitive areas like medical education.
method Used a prompting technique to evaluate GPT-4's confidence scores before and after answering USMLE questions, categorized into with and without feedback.
result Feedback influences relative confidence but doesn't consistently increase or decrease it.
A new framework evaluates LLMs by considering judge reliability.
problem Evaluating LLMs without ground truth labels can lead to biased results.
method Introduces judge-specific discrimination parameters and estimates model quality and judge reliability.
result Improves agreement with human preferences and produces calibrated uncertainty quantification.
Cer-Eval saves LLM evaluation costs while maintaining accuracy.
problem Challenges in evaluating large language models due to large dataset requirements.
method Adapts to different evaluation objectives, uses test sample complexity, and develops a partition-based algorithm.
result Cer-Eval can save 20-40% test points with comparable accuracy and 95% confidence guarantee.
SIREN protocol corrects optimistic winner's scores in LLM evaluation.
problem Optimistic winner's scores in LLM evaluation due to adaptive benchmarking.
method SIREN protocol that freezes post-search shortlist, separates selection and evaluation, and uses bootstrap for uncertainty quantification.
result SIREN provides valid confidence intervals for procedure-performance curves and deployment conclusions.
New analysis shows interpretability doesn't guarantee steering utility in LLMs.
problem Does higher interpretability lead to better steering utility in large language models?
method Trained 90 SAEs across three LLMs, evaluated interpretability and steering utility, used Kendall's rank coefficients for analysis.
result Interpretability is only weakly associated with steering utility, and features selected by Delta Token Confidence improve steering performance.
LLMs struggle with zero-shot annotation tasks due to model-internalized priors.
problem Impact of model-internalized priors on LLM performance in zero-shot annotation tasks.
method Investigated three dimensions: familiarity, decision stickiness, and susceptibility to misaligned task definitions.
result Nearly two-thirds of zero-shot errors are resistant to correction, with a rescue rate of 34.8%. Definition-Specific Familiarity (DSF) shows a positive association with model performance.
LLM evaluation suffers from systematic biases and lacks reliable positive judgments.
problem LLM evaluation suffers from systematic biases and lacks reliable positive judgments.
method Formulate LLM evaluation as a positive-unlabelled learning problem and propose a geometric auditing framework based on Partial Optimal Transport.
result Improved alignment with human preferences, increased robustness to presentation biases, and interpretable confidence estimates.
Cost-effective method detects language model hallucinations.
problem Detecting unreliable outputs from LLMs.
method Pipeline including confidence score, input attributes calibration, and thresholding.
result Multi-scoring framework outperforms individual methods and reduces computational cost.
PolyBench benchmarks LLMs on real market data, revealing significant performance gaps.
problem Benchmarking LLMs for real-world event prediction from live market signals.
method Multimodal benchmark derived from Polymarket, evaluating 7 LLMs under identical market states.
result Only two models achieve positive financial returns, highlighting the gap between fluency and probabilistic reasoning.
PAC-MCTS addresses biased search in LLM-guided planning by dynamically pruning.
problem Systematic biases in LLMs lead to inefficient and unsafe search in deep planning tasks.
method Formulates node expansion as BAI under bounded bias, derives sample complexity bounds, and proposes PAC-MCTS for dynamic confidence bounds.
result PAC-MCTS improves robustness and efficiency by up to 78% fewer API evaluations and 3x higher sample efficiency.
Measures policy-violating content prevalence with ML-assisted sampling and LLM labeling.
problem Accurate measurement of content violations that are often rare and costly to label.
method Design-based measurement system using ML-assisted probability sampling and LLM labeling.
result Produces unbiased prevalence estimates with confidence intervals and dashboard drilldowns.
FAQ efficiently evaluates LLMs with statistical guarantees using adaptive query selection.
problem Efficiently evaluating many LLMs on a large suite of benchmarks is expensive.
method FAQ uses Bayesian factor models, adaptive sampling, and proactive active inference to select queries.
result FAQ delivers up to 5x effective sample size gains over baselines, matching CI width with fewer queries.
Study uses LLMs to improve Black-Litterman portfolio optimization.
problem Systematically generating investor views for Black-Litterman model.
method Translates LLM return forecasts and uncertainty into Black-Litterman inputs.
result LLM-driven portfolios outperform traditional baselines.
Proposes integrating global and local entropy for more reliable LLMs.
problem Uncertainty in large language models (LLMs) leads to unreliable predictions.
method Measures global uncertainty from hidden-state matrices and local uncertainty from tokens, combining them via a multiplicative gate.
result Global-Local Uncertainty (GLU) outperforms unsupervised baselines across multiple models and benchmarks.
LOLA uses LLMs to optimize content delivery, outperforming traditional methods.
problem Identifying the most engaging headlines for user engagement.
method LOLA integrates LLMs with adaptive experimentation to optimize content delivery.
result LOLA outperforms traditional methods in optimizing user engagement.
Low-rank framework for task-specific LLM ranking from sparse comparisons.
problem Challenges in reliable task-specific ranking of LLMs under sparse, imbalanced comparisons.
method Low-rank modeling of task-by-model ability matrix, max-norm accurate estimator, task-wise top-K recovery guarantees, uncertainty quantification framework.
result Improves sample efficiency and produces tighter, better-calibrated ranking certificates.
Study shows multilingual LLM calibration effects improve model confidence but not accuracy.
problem Improving multilingual language model calibration in low-resource settings.
method Analysis of two multilingual benchmarks using instruction-tuning and label smoothing.
result Model confidence increases in low-resource languages after instruction-tuning but accuracy improvements are marginal.
InfoSFT improves LLMs by focusing on informative, medium-confidence tokens.
problem Overfitting to unlikely samples and degradation of prior capabilities in SFT.
method InfoSFT uses a principled weighting scheme to concentrate learning signals on medium-confidence tokens.
result InfoSFT improves generalization and preserves pre-existing capabilities over vanilla SFT and likelihood-weighted baselines.
A new method uses LLMs to discover causal pathways that affect fairness in machine learning.
problem Discovering fairness-relevant causal pathways in the presence of noise and confounding.
method Hybrid LLM-guided causal discovery framework combining active learning and dynamic scoring.
result LLM-guided methods, including the proposed active, dynamically scored variant, outperform baselines in recovering fairness-relevant structure under noisy conditions.
Optimal query allocation improves extractive QA efficiency with LLMs.
problem Efficiency and reliability in extractive question answering with LLMs.
method Learning-to-Defer framework that allocates queries to specialized models.
result Enhanced answer reliability with reduced computational overhead.
The paper accelerates LLM inference by adding early exit heads trained in a self-supervised manner.
problem Inference speed in large language models (LLMs) is slow and resource-intensive.
method Adding self-supervised early exit heads at intermediate transformer layers to stop computation early based on confidence thresholds.
result Entropy provides the most reliable confidence metric for stopping computation early.
LLM forecasting benchmarks suffer from information leakage, which confounds model performance.
problem LLM forecasting benchmarks suffer from information leakage.
method A retrieval-augmented LLM forecaster observes only decision-time information.
result The full pipeline obtains a median monthly Spearman rank IC of +0.154.
Algorithm reduces audit costs by identifying best service configurations from biased textual evidence.
problem Designing service systems from textual evidence requires accurate selection despite biased automated scoring.
method Developed PP-LUCB algorithm combining LLM scores and selective audits to minimize costs.
result Correctly identified the best model in 40/40 trials with 90% cost reduction.
Study finds uncertainty estimators weakly correlate with LLM hallucinations.
problem Characterizing the relationship between uncertainty estimators and LLM hallucinations.
method Systematic empirical study of diverse uncertainty estimators across hallucination types and benchmarks.
result Uncertainty estimators weakly correlate with LLM hallucinations, depending on hallucination type and LLM.
New methods correct bias in LLM-as-a-Judge evaluations, but reliability depends on judge quality and model calibration.
problem Systematic bias in LLM-as-a-Judge evaluations using naive estimators.
method Analytical results, simulations, and real-data case study to diagnose reliability of corrected estimates.
result Corrected estimates, especially shared-calibration comparisons, can be unreliable under certain conditions.
Study uses LLMs to categorize financial tweets, revealing useful sentiment signals.
problem Discovering meaningful sentiment signals from unstructured financial social media data.
method Leveraged LLMs to automatically label financial tweets with event categories and aligned with returns.
result Certain event labels consistently yield negative alpha, with statistically significant Sharpe ratios and information coefficients.
Algorithm identifies best arm with biased proxy and selective ground truth audits.
problem Fixed-confidence best-arm identification with biased proxy and selective ground truth.
method Propensity-weighted estimator and adaptive auditing algorithm.
result Plug-in Neyman rule achieves near-oracle audit efficiency.
Method tackles uncertainty in reward models for LLMs from heterogeneous human feedback.
problem Uncertainty in reward models for LLMs from heterogeneous human feedback.
method Heterogeneous preference framework and alternating gradient descent algorithm.
result Established theoretical guarantees for estimator convergence and asymptotic distribution.
Study improves LLMs for PPI analysis by addressing uncertainty.
problem Uncertainty in LLM predictions for PPIs.
method Fine-tuned LLaMA-3 and BioMedGPT models, LoRA ensembles, Bayesian LoRA for UQ.
result Competitive PPI identification performance across diverse disease contexts.
LLMs help automate extraction of actuarial variables from unstructured claims data.
problem Manual processing of unstructured claims data is time-consuming and inconsistent.
method Two-stage processing architecture using LLMs, modular Python pipeline.
result LLM-based extraction achieved high accuracy and practical actuarial value.
A system for supervising decentralized finance risks using LLMs and structured evidence.
problem Supervising decentralized finance risks
method Forecast-grounded agentic supervision system
result Developed a system that scores tickets against a regulator-aligned ground truth and false-intervention rate.
DeXposure-Claw supervises decentralized finance risks by grounding LLM decisions in evidence.
problem Weak evidence leads to over-interventions by general-purpose LLM agents in decentralized finance.
method DeXposure-Claw uses a graph time-series foundation model to forecast exposure networks, turning forecasts into alerts and constraining escalation with data-health gates.
result DeXposure-Claw reduces false alarms and improves regulator alignment in decentralized finance risk supervision.
New method improves LLM judge accuracy by accounting for dependencies in aggregated binary labels.
problem Classical label aggregation methods fail to account for dependencies among LLM judges, leading to miscalibrated predictions.
method Dependence-aware models based on Ising graphical models and latent factors.
result The proposed method outperforms classical methods on real-world datasets, reducing excess risk.
Calibrated PRMs improve inference efficiency for LLMs by dynamically adjusting compute budgets.
problem Poor calibration of PRMs leads to overestimation of success probabilities in partial reasoning steps.
method Quantile regression for calibration, instance-adaptive scaling (IAS) framework.
result Calibrated PRMs reduce inference costs while maintaining accuracy, especially on confident problems.
Top-H decoding improves text generation by balancing creativity and coherence.
problem Balancing creativity and coherence in text generation.
method Entropy-constrained minimum divergence problem, NP-hard ECMM problem, top-H decoding.
result Top-H decoding outperforms min-p sampling by up to 25.63% on creative writing benchmarks.
CJE calibrates cheap LLM judges against an oracle, achieving high accuracy at a fraction of the cost.
problem Inexpensive LLM judges can produce biased rankings, leading to unreliable outcomes.
method CJE uses a small oracle to calibrate cheap scores, then evaluates at scale with valid uncertainty.
result CJE achieves 99% pairwise ranking accuracy at 14x lower cost compared to a 16x oracle/judge cost ratio.