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. Paper develops methods to optimize policies directly from human feedback without reward inference.
problem Challenges in RLHF, including reward model overfitting and distribution shift.
method Develops two algorithms for RLHF without reward inference, using zeroth-order gradient approximators.
result Establishes polynomial convergence rates and outperforms existing methods in numerical experiments.
New method reduces fine-tuning cost for reused models.
problem Repeating fine-tuning costs with outdated foundation models.
method Portable Reward Tuning (PRT) trains a reward model to maximize the same loss function as fine-tuning.
result PRT achieves comparable accuracy to inference-time tuning with less inference cost.
Proposes BSI for valid statistical inference on bandit algorithms.
problem Valid statistical inference on bandit algorithms' performance.
method Fits a simulator of the bandit environment from observed data and uses it to estimate mean reward under any policy.
result Proves asymptotically valid confidence intervals and maintains nominal coverage.
Bayesian REX learns Atari games from demonstrations efficiently.
problem Bayesian reward learning for complex control problems is computationally intractable.
method Bayesian Reward Extrapolation (Bayesian REX) pre-trains a low-dimensional feature encoding and uses preferences to perform fast Bayesian inference.
result Bayesian REX learns Atari games from demonstrations in 5 minutes, competitive with state-of-the-art methods.
New RL approach infers optimal policies via variational inference.
problem Manual design of reward functions for reinforcement learning.
method Variational inference for inferring policies achieving desired outcomes.
result Eliminates need for hand-crafted reward functions for diverse tasks.
The paper proposes a method to infer multi-objective rewards from preferences.
problem Modeling preferences based on multiple, often competing objectives.
method Modeling priorities lexicographically and inferring multi-objective rewards from observed preferences.
result Lexicographically-ordered rewards provide a better understanding of preferences and improve policies.
New RL agent learns sparse rewards efficiently.
problem Sparse-reward environments and computational expense.
method Active inference with novel free energy minimization.
result High sample efficiency and online operation.
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. Active inference enhances RL by balancing exploration and exploitation.
problem Traditional RL's balance between exploration and exploitation is often suboptimal.
method Developed a new decision-making objective based on active inference.
result The new algorithm successfully balances exploration and exploitation on various RL benchmarks.
Mitigates biases in reward models using variational inference.
problem Spurious correlations in reward models that align large language models with human preferences.
method Formulates data-generating process, identifies non-spurious latent variables, and uses variational inference to recover them.
result Effective mitigation of spurious correlation issues, yielding more robust reward models.
Reinforcement learning (RL) has achieved tremendous success as a general framework for learning how to make decisions. However, this success relies on the interactive hand-tuning of a reward function by RL experts. On the other hand, inverse reinforcement learning (IRL) seeks to learn a reward function from readily-obt…
New RLHF algorithm identifies optimal policies from human feedback without explicit reward inference.
problem Training large language models with human feedback without reward inference.
method Model-free RLHF algorithm BSAD that identifies optimal policies directly from human preference. result Provable, instance-dependent sample complexity ildeO(cMSA3H3Mlogδ1). A novel framework refines diffusion models iteratively for better downstream reward optimization.
problem Optimizing reward functions during inference of diffusion models.
method Iterative refinement process with noising and reward-guided denoising steps.
result Superior empirical performance in protein and DNA design.
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.
New algorithm balances user reward and statistical inference by mixing TS with UR based on difference size.
problem Combining statistical inference with user reward in adaptive experiments.
method TS-PostDiff algorithm that uses UR when differences are small and TS when large.
result TS-PostDiff reduces false positives and increases statistical power for small differences, while maximizing reward for large ones.
AIF reformulated as convex MDP for adaptive behavior.
problem Adaptive behavior and policy optimization.
method Formulating AIF as convex MDP, deriving mirror descent algorithm.
result EFE minimization in AIF is equivalent to reward maximization in latent MDP, with epistemic component.
Paper develops robust policy evaluation for reinforcement learning with outlier and heavy-tailed rewards.
problem Outlier contamination and heavy-tailed rewards in reinforcement learning.
method Develops a fully online robust policy evaluation procedure and efficient statistical inference.
result Establishes the Bahadur-type representation of the estimator and develops an online inference procedure.
MFMs enable efficient reward alignment for generative models.
problem Computational bottleneck in controlling generative models.
method Meta Flow Maps (MFMs) extend consistency models and flow maps to stochastic regime for efficient value function estimation.
result MFMs enable inference-time steering and unbiased, off-policy fine-tuning to general rewards efficiently.
Providing a suitable reward function to reinforcement learning can be difficult in many real world applications. While inverse reinforcement learning (IRL) holds promise for automatically learning reward functions from demonstrations, several major challenges remain. First, existing IRL methods learn reward functions f…
Inverse reinforcement learning (IRL) is used to infer the reward function from the actions of an expert running a Markov Decision Process (MDP). A novel approach using variational inference for learning the reward function is proposed in this research. Using this technique, the intractable posterior distribution of the…
In many biomedical, science, and engineering problems, one must sequentially decide which action to take next so as to maximize rewards. One general class of algorithms for optimizing interactions with the world, while simultaneously learning how the world operates, is the multi-armed bandit setting and, in particular,…
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.
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. Tutorial on optimizing diffusion model samples for specific metrics.
problem Optimizing diffusion model samples for specific downstream metrics.
method Review and exploration of inference-time guidance and alignment methods.
result Unified perspective on inference-time algorithms and novel methods.
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.
This work presents a game-theoretic method for AVs that handles imperfect communication and individual rewards.
problem Real-time, imperfect communication and individual rewards in multi-agent interactions.
method Game-theoretic approach that allows for imperfect communication and individual rewards.
result More realistic assumptions lead to better reward inference and prediction of future actions.
Paper introduces PRMs to learn non-Markovian stochastic rewards for reinforcement learning.
problem Lack of structured representation for non-Markovian stochastic rewards in reinforcement learning.
method Introduces probabilistic reward machines (PRMs) and presents an algorithm to learn them from decision processes.
result Algorithm proves correct and convergent for learning PRMs from decision processes.
Multi-agent learning is a promising method to simulate aggregate competitive behaviour in finance. Learning expert agents' reward functions through their external demonstrations is hence particularly relevant for subsequent design of realistic agent-based simulations. Inverse Reinforcement Learning (IRL) aims at acquir…
A significant challenge for the practical application of reinforcement learning in the real world is the need to specify an oracle reward function that correctly defines a task. Inverse reinforcement learning (IRL) seeks to avoid this challenge by instead inferring a reward function from expert behavior. While appealin…
Robots need models of human behavior for both inferring human goals and preferences, and predicting what people will do. A common model is the Boltzmann noisily-rational decision model, which assumes people approximately optimize a reward function and choose trajectories in proportion to their exponentiated reward. Whi…
New algorithm infers reward function from agent's learning trajectories.
problem Inferring reward function from agent's learning data.
method Gradient-based approach to recover reward function.
result Improved performance compared to state-of-the-art methods.
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.
This work characterizes reward function partial identifiability and its impact on policy optimization.
problem Reward function partial identifiability in complex tasks.
method Formal characterisation of partial identifiability using various reward learning data sources.
result Unified framework for comparing data sources and downstream tasks by their invariances.
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.
The paper studies how to improve language model inference using particle filtering.
problem Understanding the accuracy-cost tradeoffs of inference-time methods for large language models.
method Introduces particle filtering algorithms like Sequential Monte Carlo (SMC) to study language model inference.
result Identifies criteria enabling non-asymptotic guarantees for SMC and fundamental limits faced by all particle filtering methods.
OSIL learns safe policies from unsafe demonstrations.
problem Offline safe imitation learning with implicit safety.
method Formulates CMDP, infers safety from non-preferred trajectories, learns cost model.
result OSIL learns safer policies without degrading reward performance.
This work frames reward modelling from preferences as a causal problem.
problem Reward modelling from preference data for AI alignment.
method Causal inference approach to identify challenges and assumptions.
result Causally-inspired approaches improve model robustness.
Designers of AI agents often iterate on the reward function in a trial-and-error process until they get the desired behavior, but this only guarantees good behavior in the training environment. We propose structuring this process as a series of queries asking the user to compare between different reward functions. Thus…
RePULSe improves language model alignment by reducing undesired outputs without sacrificing overall performance.
problem Aligning language models with human preferences while minimizing undesired outputs.
method Integrates probabilistic inference into RL training to reduce undesired outputs.
result RePULSe achieves a better balance between expected reward and undesired output probability.
This paper tackles distribution shift in model-based offline RL, proposing a shifts-aware reward method.
problem Distribution shift challenges model-based offline RL by distorting value estimation and policy optimization.
method The paper disentangles the problem into model bias and policy shift, proposing a shifts-aware reward through probabilistic inference.
result The proposed shifts-aware reward method effectively mitigates distribution shift and improves policy optimization.
While most approaches to the problem of Inverse Reinforcement Learning (IRL) focus on estimating a reward function that best explains an expert agent's policy or demonstrated behavior on a control task, it is often the case that such behavior is more succinctly represented by a simple reward combined with a set of hard…
Improves inference-time alignment for diffusion models without updating weights.
problem Aligning diffusion models without updating weights for high-reward outputs.
method Trust-Region Iterative Twisted Sequential Monte Carlo (TRI-TSMC) for variance reduction and efficiency.
result Improves primary alignment objectives on text generation tasks.
A critical flaw of existing inverse reinforcement learning (IRL) methods is their inability to significantly outperform the demonstrator. This is because IRL typically seeks a reward function that makes the demonstrator appear near-optimal, rather than inferring the underlying intentions of the demonstrator that may ha…
Demon aligns diffusion models without retraining or backpropagation.
problem Aligning diffusion models with user preferences.
method Stochastic optimization to control noise distribution.
result Significantly improves aesthetics scores for text-to-image generation.
This work improves transferability of rewards inferred from expert demonstrations.
problem Transferability of rewards inferred from expert demonstrations under limited access to the expert's policy.
method Proposed principal angles as a measure of similarity and dissimilarity between transition laws. Established sufficient conditions for transferability under limited access.
result Two key results on sufficient conditions for transferability to any and local changes in transition laws.
Training-free method improves large language model sequence quality via reward-guided sampling.
problem Optimizing large language model sequence quality over token likelihood.
method Reward-augmented target distribution combined with Sequential Monte Carlo sampling.
result Significant gains in sequence generation and mathematical reasoning tasks.
This paper shows RL with KL penalties is equivalent to Bayesian inference for fine-tuning LMs.
problem Fine-tuning large language models to avoid undesirable features.
method Analyzed KL-regularized RL and showed it's equivalent to variational inference.
result KL-regularized RL avoids distribution collapse and is more insightful as Bayesian inference.