Paper addresses online alignment of large language models under uncertain preference feedback.
problem Online alignment of large language models with misspecified preference feedback.
method Formulates an oracle-robust objective as a worst-case optimization problem for log-linear policies, and develops projected stochastic composite updates.
result Shows that the robust objective admits an exact closed-form decomposition and achieves O ~ ( ε − 2 ) \widetilde{O}(\varepsilon^{-2}) O ( ε − 2 ) oracle complexity. New research shows label refinement and weak training have limitations for aligning LLMs.
problem Limitations of refinement methods for aligning large language models.
method Analyzed probabilistic assumptions and alternative approaches to label refinement and weak training.
result Label refinement and weak training suffer from irreducible error, leaving a performance gap.
New method defends deep nets against large perturbations perceptible to humans.
problem Vulnerability of deep nets to adversarial attacks with perceptible but not changing predictions.
method Oracle-Aligned Adversarial Training (OA-AT) to align network predictions with Oracle's.
result Achieves state-of-the-art performance at large perturbation bounds (L-inf of 16/255 on CIFAR-10).
PILAF optimizes reward models from human feedback for better policy alignment.
problem Creating accurate reward models from human feedback for policy optimization.
method Policy-Interpolated Learning for Aligned Feedback (PILAF) that explicitly aligns preference learning with maximizing underlying oracle reward.
result PILAF is optimal from both optimization and statistical perspectives, demonstrating strong performance in RLHF settings.
Propose an XMSE-aware mixed estimator for EB that interpolates between ML and EB shrinkage.
problem Kernel-based EB estimation may be worse than ML when the kernel is poorly aligned with the true parameter.
method An XMSE-aware mixed estimator that interpolates between ML and EB shrinkage.
result Fixed-weight XMSE is a scalar quadratic, yielding a closed-form oracle mixing weight that is no worse than both ML and the base EB estimator at the XMSE scale.
Develops methods to correct bias in AI feedback for more accurate alignment.
problem Systematic bias in AI feedback compared to human labels.
method Two debiased alignment methods: DDPO and DIPO.
result Methods improve alignment efficiency and performance close to human-labeled data.
New algorithm reduces ERM problem size while maintaining accuracy.
problem Empirical risk minimization problem size reduction.
method Adaptive Deterministic Uniform-Weight Trimming (ADUWT) algorithm.
result Uniform ( 1 ± ε ) (1\pm\varepsilon) ( 1 ± ε ) relative-error approximation for ERM objective. New method improves reliability of selecting individuals based on predicted treatment effects.
problem Reliability of selecting individuals based on predicted conditional average treatment effects (CATE) is unreliable.
method Denoised Conformal Alignment, combining proxy errors, variance estimation, and Benjamini-Hochberg selection.
result Significantly improved power in selecting individuals while maintaining false discovery rate control.
New meta-reinforcement learning method improves performance in finite-horizon MDPs.
problem Improving meta-reinforcement learning in finite-horizon MDPs with shared optimal action-value functions.
method Proposes MTSRL and MTSRL+ algorithms with learned priors and covariance, coupled with prior-alignment technique for meta-regret guarantees.
result Achieves meta-regret guarantees with learned priors and covariance, outperforming prior-independent RL and bandit-only meta-baselines.
Modeling the relationship between chemical structure and molecular activity is a key goal in drug development. Many benchmark tasks have been proposed for molecular property prediction, but these tasks are generally aimed at specific, isolated biomedical properties. In this work, we propose a new cross-modal small mole…
Paper addresses RLHF alignment challenges with novel algorithms.
problem Challenges in RLHF alignment, especially in strategic exploration.
method Develops a reverse-KL regularized contextual bandit formulation and proposes efficient algorithms with theoretical guarantees.
result Proposed methods significantly outperform existing RLHF algorithms in real-world experiments.
PRISM integrates diverse rewards in MORL, improving sample efficiency and Pareto coverage.
problem Heterogeneous MORL where dense objectives dominate, leading to poor sample efficiency.
method PRISM uses reflectional symmetry and ReSymNet to reconcile temporal-frequency mismatches and accelerate exploration.
result PRISM consistently outperforms sparse-reward baselines and oracles, achieving significant Pareto gains.
Paper proposes CLAIR for efficient LLM fine-tuning across clients.
problem Fine-tuning large language models (LLMs) efficiently and collaboratively.
method Federated LoRA fine-tuning with Collaborative Low-rank Alignment and Identifiable Recovery (CLAIR).
result CLAIR achieves better performance and contamination detection compared to local fine-tuning.
Oracle-efficient algorithms reduce combinatorial semi-bandit regret to logarithmic time.
problem Scalability issue in combinatorial semi-bandit problems due to high combinatorial optimization costs.
method Oracle-efficient frameworks that minimize oracle queries while maintaining tight regret guarantees.
result Achieved i l d e O ( T ) ilde{O}(\sqrt{T}) i l d e O ( T ) regret with O ( log log T ) O(\log\log T) O ( log log T ) oracle queries for worst-case linear rewards. Two-stage architecture helps learners collaborate on data with privacy and transmission constraints.
problem Collaboration between learners with distinct datasets and privacy/transmission constraints.
method Privacy-aware hypothesis testing screening and synergistic iterative model training.
result Achieves oracle performance with limited transmissions of summary statistics.
New analysis shows Thompson Sampling can work with greedy approximations in combinatorial bandits.
problem Thompson Sampling's theoretical limits with greedy approximations in combinatorial semi-bandits.
method Study with greedy oracle, providing lower and upper bounds on regret.
result First theoretical results showing TS can work with greedy approximations, breaking misconceptions.
New PAC-Bayesian bounds for multi-view learning using Rényi divergence.
problem Applying PAC-Bayesian theory to multi-view learning.
method Introducing novel PAC-Bayesian bounds based on Rényi divergence for multi-view learning.
result Efficient optimization algorithms that align with theoretical bounds.
New algorithms sample convex bodies using Markov chains and restricted Gaussian oracles.
problem Sampling uniformly from convex bodies efficiently.
method Markov chain Monte Carlo with proximal sampler and restricted Gaussian oracle.
result Efficient implementation of RGO for uniform sampling on convex bodies.
MAMBA learns policies competitive with multiple conflicting oracles.
problem Learning policies from multiple conflicting oracles in reinforcement learning.
method MAMBA uses a gradient estimator in the style of GAE to optimize policies, leveraging demonstrations from multiple weak oracles.
result MAMBA outperforms the state-of-the-art in learning policies competitive with multiple conflicting oracles.
We study the problem of interactively learning a binary classifier using noisy labeling and pairwise comparison oracles, where the comparison oracle answers which one in the given two instances is more likely to be positive. Learning from such oracles has multiple applications where obtaining direct labels is harder bu…
Conformal Test Martingales can be 'blind' to significant changes in data distribution.
problem The converse of exchangeability does not hold, leading to potential blindness of CTMs.
method Explicit construction of A A A -cryptic change-point using bivariate Gaussian distributions. result CTMs can be perfectly cryptic to a significant change in marginal means.
New oracle uses uncertainty for active classification with noisy feedback.
problem Improving query complexity in interactive binary classifier learning.
method Proposes a new pairwise comparison oracle that considers uncertainty and an adaptive labeling algorithm.
result Demonstrates improved performance and efficiency compared to existing methods.
Quantum oracles help identify counterfactuals better than classical ones.
problem Identifying unknown causal parameters in causal models.
method Using quantum oracles to query and identify all causal parameters and counterfactuals.
result Quantum oracles enable identification of all two-way joint counterfactuals and tighter bounds on higher-order counterfactuals.
SoQal reduces oracle label requests in active learning by up to 35%.
problem Exploiting unlabelled data in healthcare requires costly oracle labeling.
method Dynamic questioning strategy to minimize oracle label requests.
result SoQal reduces oracle label requests by up to 35%.
We consider the problem of minimizing the sum of submodular set functions assuming minimization oracles of each summand function. Most existing approaches reformulate the problem as the convex minimization of the sum of the corresponding Lovász extensions and the squared Euclidean norm, leading to algorithms requiring …
New oracles improve stochastic optimization with noisy or biased measurements.
problem Optimizing functions with noisy or biased measurements.
method Introduced biased gradient oracles for stochastic optimization, analyzed RSG and SGD algorithms with these oracles.
result Derived non-asymptotic bounds for convergence rates of algorithms with biased gradient oracles.
Algorithm solves online binary classification and infinite games using ERM oracle.
problem Online learning and solving infinite games with computationally inefficient oracles.
method Proposes an algorithm relying solely on ERM oracle calls for online binary classification and nonparametric games.
result Achieves finite and sublinearly growing regret in various settings.
New algorithm learns efficiently with a simple 'yes/no' oracle.
problem Can efficient learning be achieved with a simpler oracle than ERM?
method Developed an oracle that returns 'yes' or 'no' for realizable datasets.
result Learnability is possible with a polynomial price in VC dimension.
New study shows Gaussian samplers struggle with heavy-tailed targets, while stable samplers excel.
problem The difficulty of sampling from heavy-tailed distributions using Gaussian versus stable oracles.
method Comparison of Gaussian and stable oracles for proximal samplers.
result Gaussian samplers have a fundamental barrier for high-accuracy guarantees in heavy-tailed sampling, while stable samplers excel.
Average Oracle outperforms DCC+NLS in portfolio optimization.
problem Optimizing portfolio performance in volatile markets.
method Comparing the Average Oracle to various DCC+NLS variants.
result The Average Oracle consistently yields higher Sharpe ratios.
New lower bounds for bilevel optimization with first-order oracles.
problem Complexity of bilevel optimization with first-order oracles.
method Development of hard instances and proof of lower bounds.
result Nontrivial lower bounds for first-order zero-respecting algorithms.
Study on tradeoffs between mistakes and ERM oracle calls in online and transductive learning.
problem Analyzing online and transductive learning with limited ERM and weak consistency oracle access.
method Proves lower bounds and upper bounds on mistakes and oracle calls, considering realizable and agnostic cases.
result Achieves optimal mistake bounds with weak consistency queries for certain concept classes.
Panoptic trades options without oracles on Ethereum.
problem Trading options without relying on oracles.
method Perpetual, trustless, instant-settlement protocol on Ethereum.
result Trustless, permissionless trading of options on Uniswap v3.
In this paper, we propose to combine imitation and reinforcement learning via the idea of reward shaping using an oracle. We study the effectiveness of the near-optimal cost-to-go oracle on the planning horizon and demonstrate that the cost-to-go oracle shortens the learner's planning horizon as function of its accurac…
Semi-supervised active clustering (SSAC) utilizes the knowledge of a domain expert to cluster data points by interactively making pairwise "same-cluster" queries. However, it is impractical to ask human oracles to answer every pairwise query. In this paper, we study the influence of allowing "not-sure" answers from a w…
Three new oracle-efficient algorithms for private synthetic data release.
problem Constructing private synthetic data that preserves statistical query answers.
method Oracle-efficient algorithms using optimization oracles for differential privacy.
result Better accuracy in large workload and high privacy regime compared to state-of-the-art.
Unified formulation bridges adversarial and nonstationary bandits.
problem Handling time-varying reward distributions in multi-armed bandit problems.
method Unified oracle that switches between adversarial and nonstationary bandit oracles based on window size.
result Optimal regret achieved with matching lower bound.
Folded concave penalization methods have been shown to enjoy the strong oracle property for high-dimensional sparse estimation. However, a folded concave penalization problem usually has multiple local solutions and the oracle property is established only for one of the unknown local solutions. A challenging fundamenta…
We consider online learning in an adversarial, non-convex setting under the assumption that the learner has an access to an offline optimization oracle. In the general setting of prediction with expert advice, Hazan et al. (2016) established that in the optimization-oracle model, online learning requires exponentially …
New method improves IV estimation with many weak and invalid instruments.
problem Identification in linear IV models with unknown validity.
method Non-convex penalized approaches, surrogate sparsest penalty.
result Advantages over other IV estimators in selection consistency and weak IV strength conditions.
Pairwise "same-cluster" queries are one of the most widely used forms of supervision in semi-supervised clustering. However, it is impractical to ask human oracles to answer every query correctly. In this paper, we study the influence of allowing "not-sure" answers from a weak oracle and propose an effective algorithm …
Study improves CTS's approximation regret for combinatorial bandits.
problem Improving CTS's performance on non-exact oracles.
method Develops a new O ( log ( T ) / Δ ) \mathcal{O}(\log(T)/Δ) O ( log ( T ) /Δ ) upper bound for CTS under specific conditions. result First O ( log ( T ) / Δ ) \mathcal{O}(\log(T)/Δ) O ( log ( T ) /Δ ) approximation regret upper bound for CTS. This paper investigates teacher hacking during language model distillation and proposes methods to mitigate it.
problem Teacher hacking during language model distillation, leading to suboptimal performance.
method A controlled experimental setup involving an oracle LM, teacher LM, and student LM, using fixed offline or online data generation techniques.
result Data diversity is the key factor in preventing teacher hacking during distillation.
Improved Frank-Wolfe algorithm for constrained convex optimization with nearest extreme point oracle.
problem Constrained smooth convex minimization with limited linear optimization oracle access.
method Frank-Wolfe algorithm with nearest extreme point oracle.
result Improved complexity bounds for specific feasible sets, including linear convergence for 0 e x t − − 1 0 ext{--}1 0 e x t − − 1 polytopes. An algorithm learns from multiple models to match an oracle's risk.
problem Learning from multiple noisy models to estimate a target parameter.
method Elimination rounds algorithm for adaptive learning.
result Risk of weak-oracle learner matches that of an oracle in multiple source case.
Proposed by Donoho (1997), Dyadic CART is a nonparametric regression method which computes a globally optimal dyadic decision tree and fits piecewise constant functions in two dimensions. In this article we define and study Dyadic CART and a closely related estimator, namely Optimal Regression Tree (ORT), in the contex…
Riemannian Proximal Sampler improves sampling on manifold data.
problem Sampling from densities on Riemannian manifolds.
method Uses MBI and RHK oracles for high-accuracy sampling.
result Sampling with ε-accuracy requires O(log(1/ε)) iterations in KL divergence.
New RLHF framework handles general preference oracles without reward functions.
problem Handling general preference oracles without assuming a reward function.
method Developed a minimax game between two LLMs for RLHF under a general preference oracle, focusing on KL-regularized preference.
result Proposed algorithms for efficient offline and online RLHF learning.