Improved inference-time alignment using Best-of-N and smoothing.
problem Reward overoptimization in Best-of-N (BoN) due to poor proxy reward models.
method Introduced Soft Best-of-N (SBoN) and analyzed its performance through KL divergence and regret analysis.
result Smoothing helps SBoN mitigate reward overoptimization, especially when proxy reward quality is low.
Best-of-N sampling reveals reward targets from preference data, influencing N and base distribution choices.
problem Understanding reward extraction from Best-of-N preference data and optimal N and base distribution choices.
method Specialized analysis of preference data via induced conditional distribution, deriving reward targets and design principles.
result Reward targets are explicit functions of N and base distribution, and bounded-class minimizers approach these targets as N grows.
Two methods for model adaptation compared; fine-tuning outperforms Best-of-N in realizable settings.
problem Comparing methods for adapting large language models to new tasks.
method Supervised fine-tuning vs. Best-of-N approach.
result Supervised fine-tuning outperforms Best-of-N in realizable settings.
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.
New method optimizes language model performance for test-time strategies.
problem Mismatch between training objectives and test-time deployment of large language models.
method Tail-Extrapolated estimators to approximate best-of-N performance from limited training rollouts.
result Improved performance of best-of-N deployment across various models and datasets.
Faster WIND accelerates iterative BOND for LLM alignment.
problem Iterative BOND is inefficient in practice due to sample and computation inefficiency.
method Unified game-theoretic connection to self-play alignment, WIND framework with efficient algorithms.
result WIND variant achieves superior sample efficiency and faster computation.
SWIFT learns intrinsic rewards from LLM hidden states for efficient best-of-N sampling.
problem Efficiency and scalability of reward models for LLMs.
method SWIFT (Simple Weighted Intrinsic Feedback Technique) learns a reward function directly from LLM hidden states.
result SWIFT outperforms existing baselines by 12.7% on MATH dataset while using less than 0.005% of their parameters.
Improved selection of best outputs from LLMs for better accuracy.
problem BoN fails to reliably find the correct answer with imperfect rewards.
method Majority-of-the-Bests (MoB) selects the mode of a bootstrapped output distribution of BoN.
result MoB consistently improves over BoN in 25 out of 30 setups.
The paper analyzes the reward improvement of aligned policies in large language models.
problem Optimizing policies in large language models while staying close to a reference policy.
method Information-theoretic analysis and reduction to exponential order statistics.
result Information-theoretic upper bounds on reward improvement are derived.
New algorithm improves inference-time alignment without reward hacking.
problem Improving quality of responses from language models with limited compute.
method Inference-time alignment, focusing on extttInferenceTimePessimism algorithm. result Optimal performance and scaling-monotonicity of extttInferenceTimePessimism. The study examines how verifier imperfections impact test-time scaling techniques.
problem Understanding how verifier imperfections affect test-time scaling methods.
method Proves the instance-level accuracy of Best-of-N and Rejection Sampling methods using the geometry of the verifier's ROC curve.
result RS outperforms BoN for fixed compute, but both converge to the same accuracy in the infinite-compute limit.
Paper characterizes optimal language model alignment methods.
problem Aligning language models to maximize reward while keeping them close to the original model.
method KL-constrained reinforcement learning and best-of-N methods.
result Optimal KL-constrained RL solution has a large deviation principle rate function.
This work explores test-time scaling strategies for LLMs, improving sample efficiency and expressiveness.
problem Understanding the sample efficiency and expressiveness of test-time scaling strategies for LLMs.
method Established separation and expressiveness results for self-consistency, best-of-n, and self-correction strategies. result Self-correction enables Transformers to simulate online learning over multiple tasks without prior knowledge.
New method predicts and optimizes test-time scaling for LLMs.
problem Lack of principled guidance on scaling LLMs efficiently.
method Tail-guided search to predict and allocate compute.
result SLG Search achieves higher rewards with less compute.
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.
The abstract explores connections between reinforcement learning, scaling, and diffusion.
problem Aligning reinforcement learning with human feedback and scaling techniques.
method Clarifying connections between reinforcement learning, scaling, and diffusion.
result Introducing a resampling approach for alignment and reward-directed diffusion models.
Best-of-∞ improves LLM performance by efficiently allocating inference-time computation.
problem Achieving optimal performance in test-time LLM ensembling with infinite budget.
method Adaptive generation scheme and weighted ensembles of LLMs, formulated as mixed-integer linear program.
result Optimal ensemble weighting improves performance over individual models.
A new method resolves non-identifiability in reward modeling using anchor labels.
problem Non-identifiability in reward modeling from pairwise preferences alone.
method Anchor-guided Variance-aware Reward Modeling (AVRM) framework.
result AVRM resolves non-identifiability and improves reward modeling performance.
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.
Improves generative models by optimizing rewards and sample editing.
problem Efficiently generating high-reward samples with structural constraints.
method Introduces MDM-VGB, a discrete diffusion sampler that augments unmasking generation with reward-guided remasking.
result MDM-VGB achieves quadratic complexity and robustness to noise, outperforming heuristics like best-of-N. EORM boosts LLM accuracy with a lightweight, energy-based verifier.
problem Efficiently verifying mathematical reasoning in large language models.
method Energy-based framework to rank Chain-of-Thought solutions using simple outcome labels.
result EORM boosts LLM accuracy to 90.7% on GSM8k and 63.7% on MATH with only 55M parameters.
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.
Pre-training improves model coverage, crucial for downstream performance.
problem Understanding why pre-training enhances model performance.
method Coverage principle, focusing on next-token prediction and model quality.
result Coverage generalizes faster than cross-entropy, improving downstream performance.
Study scaling laws of reward model overoptimization in reinforcement learning.
problem Reward model overoptimization hinders true performance in reinforcement learning.
method Synthetic setup with fixed gold reward model; optimization using RL or best-of-n sampling; analysis of scaling laws. result Scaling laws of reward model overoptimization differ based on optimization method and scale smoothly with model parameters.
No policy can simultaneously be fully autonomous, optimally calibrated, and helpful, proving a trilemma.
problem Proving impossibility of a policy achieving maximum helpfulness, optimal calibration, and full autonomy.
method Geometric proof showing that adding any non-affine autonomy incentive to a strictly proper scoring rule destroys strict properness.
result The Behavioral Credibility Trilemma: no policy can achieve all three goals simultaneously.
DTS improves inference-time alignment of diffusion models with less compute.
problem Inference-time alignment of diffusion models suffers from inaccurate value estimation and inefficient reuse of past computations.
method Diffusion Tree Sampling (DTS) uses a tree-based approach to propagate terminal rewards and iteratively refine value estimates.
result DTS produces asymptotically exact samples and matches the FID of best-performing baselines with up to 10x less compute.
Best-of-Majority improves inference performance in Pass@k settings.
problem Inference in difficult tasks often underperforms with single-shot selection methods.
method Combining majority voting and Best-of-N, Best-of-Majority restricts candidates to high-frequency responses.
result Best-of-Majority achieves minimax optimal regret and outperforms other methods.