Tiny benchmarks reduce LLM evaluation costs by using fewer examples.
problem Expensive evaluation of LLMs with tens of thousands of examples.
method Developed evaluation tools and tiny versions of popular benchmarks.
result Accurately estimate LLM performance with just 100 curated examples.
PromptEval estimates LLM performance across many prompts, improving reproducibility.
problem Limited prompt templates affect LLM benchmark reproducibility.
method Estimates performance distribution across many prompts using borrowed strength.
result PromptEval accurately estimates performance quantiles with practical budget.
We solve the multi-criteria benchmarking problem by formalizing it as a social choice problem and identifying conditions for meaningful rankings.
problem Aggregating multiple metrics into a single ranking for models in benchmarking problems.
method Formalizing multi-criteria benchmarking as a social choice problem and identifying sufficient conditions for meaningful rankings.
result We prove that meaningful multi-criteria benchmarking becomes possible under certain preference conditions (single-peaked, group-separable, distance-restricted).
Concentration of infinitely exchangeable sequences with bounded-difference constants
problem Quantifying uncertainty in AI benchmarks
method Using a mixture-free Hoeffding-type bound
result Tight, mixture-free Hoeffding-type bound for zero-sum linear contrasts
LoLCATs improves linearized LLM quality with less memory and compute.
problem Linearizing large language models (LLMs) often degrades model quality and requires expensive training.
method Two-step method: attention transfer and low-rank adaptation.
result Significant improvement in linearizing quality with 20+ points on 5-shot MMLU.
Hierarchical framework for model evaluation on leaderboards
problem Uncertainty and variability in model performance across tasks
method Hierarchical framework with task-level and leaderboard-level rank prediction intervals
result Statistically valid and informative model rank intervals
Training-free looped transformers improve model performance without additional training.
problem Improving model performance without additional training or fine-tuning.
method A lightweight inference-time wrapper loops a contiguous mid-stack block of layers of a frozen checkpoint without additional fine-tuning.
result Our method improves model performance across various model families.
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.
DS-CP improves reliability of uncertainty quantification for large language models under domain shift.
problem Overconfident and factually incorrect outputs (hallucinations) from large language models.
method Adapts conformal prediction to large language models under domain shift by reweighting calibration samples.
result DS-CP delivers more reliable coverage than standard conformal prediction, especially under substantial distribution shifts.
Language model benchmarks often misrepresent true understanding, revealing vulnerabilities in evaluation methods.
problem Language model benchmarks fail to accurately reflect true language understanding and adaptability.
method Systematic analysis of NLP evaluation frameworks, identifying vulnerabilities in static benchmarks, human evaluation protocols, and LLM-as-judge frameworks.
result Current evaluation methods are unreliable and need improvement to accurately assess LLM performance.
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.
GSI improves efficiency of large language model inference.
problem Efficiently guiding test-time alignment in large language models.
method Combines soft best-of-n scaling with a reward model and speculative samples. result Achieves higher accuracy and reduced latency compared to standard methods.
A new method selects models for ensemble learning to maximize mutual information, outperforming existing approaches.
problem Selecting models for ensemble learning to improve performance and reduce correlation issues.
method Formulate budgeted ensemble selection as maximizing mutual information, use Gaussian-copula to model correlated errors, propose a greedy mutual-information selection algorithm.
result Our method consistently outperforms strong baselines across multiple datasets.
The paper certifies AI reliability via sampling and calibration, providing exact guarantees.
problem Ensuring trust in black-box AI systems' outputs.
method Self-consistency sampling and conformal calibration.
result Reliability levels derived from these methods offer finite-sample guarantees.
New method for valid prediction sets in high-dimensional covariate shifts.
problem Valid prediction sets in high-dimensional covariate shifts.
method Likelihood-ratio regularized quantile regression (LR-QR) algorithm.
result LR-QR constructs valid prediction sets with desired coverage in target domain.
CROQ optimizes LLM decision-making by narrowing down choices and improving accuracy.
problem Uncertainty in LLM outputs poses risks in high-stakes domains.
method Conformal prediction (CP) and optimization (CP-OPT) to minimize prediction set sizes.
result CROQ improves LLM accuracy, especially with CP-OPT.
A new sparse benchmark metabench identifies key abilities from large benchmarks.
problem Redundancy and compression in existing benchmarks.
method Data from 5000+ LLMs to identify most informative items, distilling a sparse benchmark.
result Sparse benchmark metabench captures underlying abilities with high accuracy.
Galactica learns from scientific literature to help researchers.
problem Information overload in scientific literature makes it hard to find useful insights.
method Trained on a large corpus of scientific papers, reference material, and knowledge bases.
result Outperforms existing models on various scientific tasks, including LaTeX equations and mathematical reasoning.
ReD improves LLM inference efficiency at fixed budget, reducing attempts and cost.
problem Improving LLM inference efficiency at a fixed budget.
method Reset-and-Discard (ReD) query method.
result ReD increases coverage@cost for a given budget, reducing attempts and cost.
Post-hoc transforms can reverse model performance trends, especially in noisy settings.
problem Post-hoc transforms can reverse model performance trends, especially in noisy settings.
method Empirical study and analysis of post-hoc transforms like temperature scaling, ensembling, and SWA.
result Post-hoc reversal can prevent double descent and mitigate mismatches between test loss and test error.
QAlign improves language model alignment with less compute, outperforming existing methods.
problem Improving language model performance with limited test-time computation.
method QAlign: sampling from optimal aligned distribution using Markov chain Monte Carlo.
result Consistent improvements over existing methods on various benchmarks.
This paper simplifies fine-tuning for small LLMs, reducing barriers for developers.
problem Limited resources for fine-tuning large language models (LLMs) by individual developers and small organizations.
method Instruction-tuning datasets, small-sized LLMs (3B to 7B parameters), various training configurations and strategies.
result Improved model performance on benchmarks with specific training configurations, and insights into early termination and hyperparameter simplifications.
Extends FC-RAG to anytime-valid sequential coverage for language model swarms.
problem Maintain distribution-free coverage for a swarm of weak language models over time.
method Introduces Anytime-FC-RAG, a sequential extension with a summable calibration-deviation budget.
result Achieves time-uniform alarm validity and safety under predictable adaptive control.