The rise of large language models (LLMs) has created a significant disparity: industrial research labs with their computational resources, expert teams, and advanced infrastructures, can effectively fine-tune LLMs, while individual developers and small organizations face barriers due to limited resources to effectively explore the experiment space. In this paper, we aim to bridge this gap by presenting a comprehensive study on supervised fine-tuning of LLMs using instruction-tuning datasets spanning diverse knowledge domains and skills. We focus on small-sized LLMs (3B to 7B parameters) for their cost-efficiency and accessibility. We explore various training configurations and strategies across four open-source pre-trained models. We provide detailed documentation of these configurations, revealing findings that challenge several common training practices, including hyperparameter recommendations from TULU and phased training recommended by Orca. The code used for the experiments can be found here: https://github.com/instructlab/training. Key insights from our work include: (i) larger batch sizes paired with lower learning rates lead to improved model performance on benchmarks such as MMLU, MTBench, and Open LLM Leaderboard; (ii) early-stage training dynamics, such as lower gradient norms and higher loss values, are strong indicators of better final model performance, allowing for early termination of sub-optimal runs and significant computational savings; (iii) through a thorough exploration of hyperparameters like warmup steps and learning rate schedules, we provide guidance for practitioners and find that certain simplifications do not compromise performance; and (iv) we observe no significant difference in performance between phased (sequentially training on data divided into phases) and stacked (training on the entire dataset at once) strategies, but stacked training is simpler and more sample efficient. With these findings holding robustly across datasets as well as model families and sizes, we hope this study serves as a guide for practitioners fine-tuning small LLMs and promotes a more inclusive research environment for LLM development.

Large language models (LLMs) have achieved significant performance gains via scaling up model sizes and/or data. However, recent evidence suggests diminishing returns from such approaches, motivating a pivot to scaling test-time compute. Existing deterministic inference-time scaling methods, usually with reward models, cast the task as a search problem, but suffer from a key limitation: early pruning. Due to inherently imperfect reward models, promising trajectories may be discarded prematurely, leading to suboptimal performance. We propose a novel inference-time scaling approach by adapting particle-based Monte Carlo methods. Our method maintains a diverse set of candidates and robustly balances exploration and exploitation. Our empirical evaluation demonstrates that our particle filtering methods have a 4--16x better scaling rate over deterministic search counterparts on both various challenging mathematical and more general reasoning tasks. Using our approach, we show that Qwen2.5-Math-1.5B-Instruct surpasses GPT-4o accuracy in only 4 rollouts, while Qwen2.5-Math-7B-Instruct scales to o1 level accuracy in only 32 rollouts. Our work not only presents an effective method to inference-time scaling, but also connects rich literature in probabilistic inference with inference-time scaling of LLMs to develop more robust algorithms in future work.

Offline reinforcement learning (RL) enables learning a decision-making policy without interaction with the environment. This makes it particularly beneficial in situations where such interactions are costly. However, a known challenge for offline RL algorithms is the distributional mismatch between the state-action distributions of the learned policy and the dataset, which can significantly impact performance. State-of-the-art algorithms address it by constraining the policy to align with the state-action pairs in the dataset. However, this strategy struggles on datasets that predominantly consist of trajectories collected by low-performing policies and only a few trajectories from high-performing ones. Indeed, the constraint to align with the data leads the policy to imitate low-performing behaviors predominating the dataset. Our key insight to address this issue is to constrain the policy to the policy that collected the good parts of the dataset rather than all data. To this end, we optimize the importance sampling weights to emulate sampling data from a data distribution generated by a nearly optimal policy. Our method exhibits considerable performance gains (up to five times better) over the existing approaches in state-of-the-art offline RL algorithms over 72 imbalanced datasets with varying types of imbalance.

We introduce a visually-guided task-and-motion planning benchmark, which we call the ThreeDWorld Trans-port Challenge. In this challenge, an embodied agent is spawned randomly in a simulated physical home environment and required to transport a small set of objects scattered around the house with containers. We build this benchmark challenge using the ThreeDWorld simulation: a virtual 3D environment where all objects respond to physics, and a robot agent can be controlled using a fully physics-driven navigation and interaction API. We evaluate several existing agents on this benchmark. Experimental results suggest that: 1) a pure RL model struggles on this challenge; 2) state-of-the-art hierarchical planning-based agents can transport some objects but are still far from solving this task. We anticipate that this benchmark will empower researchers to develop more intelligent physics-aware robot learning algorithms.

For machine agents to successfully interact with humans in real-world settings, they will need to develop an understanding of human mental life. Intuitive psychology, the ability to reason about hidden mental variables that drive observable actions, comes naturally to people: even pre-verbal infants can tell agents from objects, expecting agents to act efficiently to achieve goals given constraints. Despite recent interest in machine agents that reason about other agents, it is not clear if such agents learn or hold the core psychology principles that drive human reasoning. Inspired by cognitive development studies on intuitive psychology, we present a benchmark consisting of a large dataset of procedurally generated 3D animations, AGENT (Action, Goal, Efficiency, coNstraint, uTility), structured around four scenarios (goal preferences, action efficiency, unobserved constraints, and cost-reward trade-offs) that probe key concepts of core intuitive psychology. We validate AGENT with human-ratings, propose an evaluation protocol emphasizing generalization, and compare two strong baselines built on Bayesian inverse planning and a Theory of Mind neural network. Our results suggest that to pass the designed tests of core intuitive psychology at human levels, a model must acquire or have built-in representations of how agents plan, combining utility computations and core knowledge of objects and physics.

Humans have mental models that allow them to plan, experiment, and reason in the physical world. How should an intelligent agent go about learning such models? In this paper, we will study if models of the world learned in an open-ended physics environment, without any specific tasks, can be reused for downstream physics reasoning tasks. To this end, we build a benchmark Open-ended Physics Environment (OPEn) and also design several tasks to test learning representations in this environment explicitly. This setting reflects the conditions in which real agents (i.e. rolling robots) find themselves, where they may be placed in a new kind of environment and must adapt without any teacher to tell them how this environment works. This setting is challenging because it requires solving an exploration problem in addition to a model building and representation learning problem. We test several existing RL-based exploration methods on this benchmark and find that an agent using unsupervised contrastive learning for representation learning, and impact-driven learning for exploration, achieved the best results. However, all models still fall short in sample efficiency when transferring to the downstream tasks. We expect that OPEn will encourage the development of novel rolling robot agents that can build reusable mental models of the world that facilitate many tasks.