This work investigates how multi-round reasoning improves LLM performance.
problem Improving problem-solving abilities in complex tasks with LLMs.
method Investigates approximation, learnability, and generalization properties of multi-round auto-regressive models.
result Transformers with finite context windows are universal approximators for Turing-computable functions and can approximate any Turing-computable sequence-to-sequence function through multi-round reasoning.
Graph neural networks can perform approximate reasoning in latent space for mathematical statements.
problem Can neural networks perform steps of approximate reasoning in a fixed dimensional latent space?
method Design and conduct an experiment using graph neural networks to predict rewrite-success of mathematical statements in a latent space.
result Graph neural networks can make non-trivial predictions about rewrite-success of statements in latent space.
This work shows how approximate reward models can significantly improve inference-time scaling.
problem Improving the efficiency of inference for large language models.
method Identifying the Bellman error of approximate reward models and using Sequential Monte Carlo (SMC) for inference.
result Approximate reward models can reduce computational complexity from exponential to polynomial in T. The paper explores how to evaluate Bayesian approximations in neural networks.
problem The difficulty in evaluating Bayesian approximations in neural networks.
method Exploring the interactions between probabilistic models, approximating distributions, optimization algorithms, and datasets.
result The expected utility of the approximate posterior can measure inference quality.
MultiVerse uses importance sampling for efficient causal reasoning in probabilistic programming.
problem Efficient causal reasoning in probabilistic models, especially counterfactual inference.
method Native implementation of importance sampling in probabilistic programming, optimizing inference through query structure.
result Significant optimisation of inference process through careful design choices and query structure consideration.
Bayesian Neural Networks improve uncertainty reasoning in NNs.
problem Frequentist implementation of NNs cannot reason about uncertainty in predictions.
method Introduces Bayesian Neural Networks and compares approximate inference methods.
result Future research can improve on current methods of inference.
Kernel smoothing improves LLM reasoning efficiency.
problem Efficiently estimate value functions with limited samples for reinforcement learning.
method Kernelized advantage estimation using classical nonparametric statistics.
result Improved policy optimization with accurate value and gradient estimation.
We draw a formal connection between using synthetic training data to optimize neural network parameters and approximate, Bayesian, model-based reasoning. In particular, training a neural network using synthetic data can be viewed as learning a proposal distribution generator for approximate inference in the synthetic-d…
Large language models can't efficiently reason conditionally in a distribution-free setting.
problem Impossibility of conditional PAC-efficient reasoning in large language models.
method Proof of impossibility in a distribution-free setting for non-atomic input spaces.
result Any algorithm achieving conditional PAC efficiency must defer to the expert model with high probability.
FinReflectKG benchmarks financial QA by linking relevant context from a financial KG, improving model performance and efficiency.
problem Efficiently retrieving and navigating relevant financial information across diverse sources and years.
method A benchmark built on a temporally indexed financial KG, generating QA pairs via pattern-specific prompts and quality control, evaluating retrieval scenarios.
result KG-guided retrieval yields substantial gains in correctness and token utilization, improving model performance by 24%.
Trained neural networks perform Bayesian reasoning for tasks beyond their initial scope.
problem Performing Bayesian reasoning for tasks outside the trained neural networks' initial scope.
method Used deep generative models as priors and classification/regression networks as constraints. Approximated Bayesian inference through variational or sampling techniques.
result The approach built on top of already trained networks, expanding the addressable questions.
Hybrid deep architectures with reasoning layers show promising convergence and generalization properties.
problem Understanding the theoretical foundations of hybrid deep architectures with reasoning layers.
method Analyzing the interplay between algorithm layers and neural components in deep architectures.
result Properties of algorithm layers are closely related to the approximation and generalization abilities of end-to-end models.
LaLoRA prevents forgetting in LoRA fine-tuning.
problem Catastrophic forgetting in fine-tuned models.
method LaLoRA applies Laplace approximation to LoRA weights for regularization.
result Improved learning-forgetting trade-off with controllable regularization strength.
Dynamic abstention improves LLM accuracy by selectively terminating unpromising reasoning.
problem LLMs waste compute on incorrect responses, leading to inefficiency.
method Formal reinforcement learning framework with abstention reward parameter.
result Dynamic abstention outperforms natural baselines in selective accuracy.
We present a stochastic numerical method for solving fully non-linear free boundary problems of parabolic type and provide a rate of convergence under reasonable conditions on the non-linearity.
Develops simple approximations for credit risk estimation.
problem Estimating conditional default probabilities for corporate loans.
method Closed form approximations to maximum likelihood estimator for logit type models.
result The approximations are as good as the MLE under certain conditions.
ENIAC method optimizes and explores complex RL problems with non-linear policies.
problem Theoretical understanding of non-linear policies in RL with strategic exploration.
method ENIAC, an actor-critic method for non-linear function approximation.
result ENIAC finds near-optimal policies in polynomial exploration rounds under bounded eluder dimension.
ISAAC audits deep models for drug-target interactions, revealing structural differences.
problem Deep models for DTI often use irrelevant features, making them hard to evaluate.
method ISAAC uses intervention-based structural auditing to evaluate model sensitivity.
result ISAAC reveals significant structural differences in DTI models' reasoning.
This work analyzes CoT prompting methods from a statistical estimation perspective.
problem Improving the effectiveness of LLMs in solving multi-step reasoning problems.
method Introducing a multi-step latent variable model to characterize CoT prompting from a statistical estimation viewpoint.
result The CoT estimator is equivalent to a Bayesian estimator when the pretraining dataset is large.
Novel approach trains LLMs for inductive reasoning using probabilistic programs.
problem Training LLMs for inductive reasoning with sparse, ambiguous data.
method Program-based Posterior Training (PPT) using probabilistic inference.
result Significant improvement in estimation accuracy and alignment with human judgments.
wd1 improves reasoning in dLLMs by optimizing policies without policy ratios.
problem Improving reasoning in diffusion-based large language models through RL.
method wd1: ratio-free policy optimization using weighted log-likelihood.
result wd1 outperforms diffusion-based GRPO while requiring lower computational cost.
ENN neural network learns logical syllogisms using Euler diagrams.
problem Traditional neural networks struggle with logical reasoning, especially syllogisms.
method ENN represents logical relations as Euler diagrams, optimizing syllogism structures with a novel back-propagation algorithm.
result ENN can precisely represent and reason with all 24 syllogism structures.
Transformer models can approximate smooth functions with prompts, enhancing LLMs' dynamic capabilities.
problem Lack of theoretical framework for prompt engineering in transformer models.
method Formal framework demonstrating transformer models can approximate β-times differentiable functions with prompts. result Transformer models can approximate β-times differentiable functions with arbitrary precision using appropriately structured prompts. Humans are capable of attributing latent mental contents such as beliefs or intentions to others. The social skill is critical in daily life for reasoning about the potential consequences of others' behaviors so as to plan ahead. It is known that humans use such reasoning ability recursively by considering what others …
We describe a mathematical link between aspects of information theory, called pairwise comparisons, and discretized gauge theories. The link is made by the notion of holonomy along the edges of a simplex. This correspondance leads to open questions in both field.
Simulation-based inference methods can produce unreliable posterior approximations.
problem Reliability of simulation-based inference methods for scientific use cases.
method Benchmarked algorithms including Neural Posterior Estimation, Neural Ratio Estimation, Sequential Neural Likelihood, and Approximate Bayesian Computation.
result Ensembling posterior surrogates provides more reliable approximations.
In this work we investigate the reasons why Batch Normalization (BN) improves the generalization performance of deep networks. We argue that one major reason, distinguishing it from data-independent normalization methods, is randomness of batch statistics. This randomness appears in the parameters rather than in activa…
In this paper, we establish a fluid limit for a two--sided Markov order book model. Our main result states that in a certain asymptotic regime, a pair of measure-valued processes representing the "sell-side shape" and "buy-side shape" of an order book converges to a pair of deterministic measure-valued processes in a c…
Frank-Wolfe methods (FW) have gained significant interest in the machine learning community due to its ability to efficiently solve large problems that admit a sparse structure (e.g. sparse vectors and low-rank matrices). However the performance of the existing FW method hinges on the quality of the linear approximatio…
Develops models for temporally abstract reasoning and attention.
problem Temporal abstraction and attention in reinforcement learning.
method Defines affordances for options and develops partial option models.
result Identifies trade-offs between estimation and approximation error.
We show that a neural network with arbitrary depth and non-linearities, with dropout applied before every weight layer, is mathematically equivalent to an approximation to a well known Bayesian model. This interpretation might offer an explanation to some of dropout's key properties, such as its robustness to over-fitt…
Paper tackles non-convex optimization for higher moments in portfolio management.
problem Complexity of higher moments in optimization problems.
method Method of successive convex approximation.
result Solves mean-variance-skewness problem using non-convex optimization.
Paper proves efficiency of MARL with transformers, addressing agent complexity.
problem Theoretical understanding of MARL with many agents and limited relational reasoning.
method Set Transformer for relational reasoning, model-free and model-based MARL algorithms.
result Provable efficiency of MARL algorithms, suboptimality gaps independent of number of agents.
In this paper we measured the stability of stochastic gradient method (SGM) for learning an approximated Fourier primal support vector machine. The stability of an algorithm is considered by measuring the generalization error in terms of the absolute difference between the test and the training error. Our problem is to…
Control of non-episodic, finite-horizon dynamical systems with uncertain dynamics poses a tough and elementary case of the exploration-exploitation trade-off. Bayesian reinforcement learning, reasoning about the effect of actions and future observations, offers a principled solution, but is intractable. We review, then…
Tree ensembles, such as random forest and boosted trees, are renowned for their high prediction performance, whereas their interpretability is critically limited. In this paper, we propose a post processing method that improves the model interpretability of tree ensembles. After learning a complex tree ensembles in a s…
Study validates numerical method for singular FBSDEs convergence.
problem Solving singular FBSDEs and associated PDEs.
method Particles approximation for transport operator and tree approximation for diffusion operator.
result Convergence rate of numerical method proved under reasonable conditions.
Paper generalizes discrete uniformization for genus-zero surfaces.
problem Discrete uniformization for surfaces of genus zero.
method Reduction to planar cases via stereographic projections.
result Generalization of discrete uniformization to genus-zero surfaces.
Deep neural networks have achieved impressive results on a wide variety of tasks. However, quantifying uncertainty in the network's output is a challenging task. Bayesian models offer a mathematical framework to reason about model uncertainty. Variational methods have been used for approximating intractable integrals t…
We propose the Neural Logic Machine (NLM), a neural-symbolic architecture for both inductive learning and logic reasoning. NLMs exploit the power of both neural networks---as function approximators, and logic programming---as a symbolic processor for objects with properties, relations, logic connectives, and quantifier…
A new method improves ICA performance by approximating MDI.
problem Improving F astICA's performance with nonlinear functions.
method Second-order approximation of MDI for joint maximization.
result Efficiency validated through experiments compared to other ICA algorithms.
We study reinforcement learning under model misspecification, where we do not have access to the true environment but only to a reasonably close approximation to it. We address this problem by extending the framework of robust MDPs to the model-free Reinforcement Learning setting, where we do not have access to the mod…
Agents learn state ambiguity from non-linear sensor data using Gaussian approximations.
problem Learning state representation from non-linear sensor data.
method Second-order Taylor approximation of Gaussian distribution for non-linear measurement functions.
result Induces a preference for states based on inferability from observations.
Method solves complex optimization problems with high probability bounds.
problem Nonlinear equality constrained stochastic optimization problems.
method Step-search sequential quadratic programming method.
result High-probability bound on iteration complexity for first-order stationarity.
Massively parallel RWS improves inference in complex models.
problem Exponential sample requirement for effective importance weighting.
method Draws K samples of all n latent variables and individually reasons through all combinations. result Significant improvements over standard RWS.
Paper proves discrete uniformizations converge to continuous for surfaces of genus ≥1.
problem Computing uniformizations for surfaces of genus >1.
method Discrete conformality and uniformization on triangle meshes.
result Discrete uniformizations approximate continuous uniformization for closed surfaces of genus ≥1.
This work connects hardness of approximation and learning.
problem Hardness of approximation and learnability in machine learning.
method Shows a single hardness property implying both approximation and learning hardness.
result Obtains new results on hardness of approximation and learnability of specific functions.
LASER compresses recursive model activations by exploiting their low-dimensional structure.
problem Understanding and optimizing the geometric structure of recursive reasoning trajectories.
method Dynamic low-rank basis tracking via matrix-free subspace tracking with a fidelity-triggered reset mechanism.
result Recursive activations occupy a linear, low-dimensional subspace that can be compressed efficiently.