Adapting Large Language Models (LLMs) to downstream tasks using Reinforcement Learning (RL) has proven to be an effective approach. However, LLMs do not inherently define the structure of an agent for RL training, particularly in terms of defining the action space. This paper studies learning a compact latent action space to enhance the controllability and exploration of RL for LLMs. We propose Controlling Large Language Models with Latent Actions (CoLA), a framework that integrates a latent action space into pre-trained LLMs. We apply CoLA to the Llama-3.1-8B model. Our experiments demonstrate that, compared to RL with token-level actions, CoLA's latent action enables greater semantic diversity in text generation. For enhancing downstream tasks, we show that CoLA with RL achieves a score of 42.4 on the math500 benchmark, surpassing the baseline score of 38.2, and reaches 68.2 when augmented with a Monte Carlo Tree Search variant. Furthermore, CoLA with RL consistently improves performance on agent-based tasks without degrading the pre-trained LLM's capabilities, unlike the baseline. Finally, CoLA reduces computation time by half in tasks involving enhanced thinking prompts for LLMs by RL. These results highlight CoLA's potential to advance RL-based adaptation of LLMs for downstream applications.

Critiques are important for enhancing the performance of Large Language Models (LLMs), enabling both self-improvement and constructive feedback for others by identifying flaws and suggesting improvements. However, evaluating the critique capabilities of LLMs presents a significant challenge due to the open-ended nature of the task. In this work, we introduce a new benchmark designed to assess the critique capabilities of LLMs. Unlike existing benchmarks, which typically function in an open-loop fashion, our approach employs a closed-loop methodology that evaluates the quality of corrections generated from critiques. Moreover, the benchmark incorporates features such as self-critique, cross-critique, and iterative critique, which are crucial for distinguishing the abilities of advanced reasoning models from more classical ones. We implement this benchmark using eight challenging reasoning tasks. We have several interesting findings. First, despite demonstrating comparable performance in direct chain-of-thought generation, classical LLMs significantly lag behind the advanced reasoning-based model o1-mini across all critique scenarios. Second, in self-critique and iterative critique settings, classical LLMs may even underperform relative to their baseline capabilities. We hope that this benchmark will serve as a valuable resource to guide future advancements. The code and data are available at \url{https://github.com/tangzhy/RealCritic}.

Despite their remarkable performance, the development of Large Language Models (LLMs) faces a critical challenge in scalable oversight: providing effective feedback for tasks where human evaluation is difficult or where LLMs outperform humans. While there is growing interest in using LLMs for critique, current approaches still rely on human annotations or more powerful models, leaving the issue of enhancing critique capabilities without external supervision unresolved. We introduce SCRIT (Self-evolving CRITic), a framework that enables genuine self-evolution of critique abilities. Technically, SCRIT self-improves by training on synthetic data, generated by a contrastive-based self-critic that uses reference solutions for step-by-step critique, and a self-validation mechanism that ensures critique quality through correction outcomes. Implemented with Qwen2.5-72B-Instruct, one of the most powerful LLMs, SCRIT achieves up to a 10.3\% improvement on critique-correction and error identification benchmarks. Our analysis reveals that SCRIT's performance scales positively with data and model size, outperforms alternative approaches, and benefits critically from its self-validation component.

Large language models rely on Supervised Fine-Tuning (SFT) to specialize in downstream tasks. Cross Entropy (CE) loss is the de facto choice in SFT, but it often leads to overfitting and limited output diversity due to its aggressive updates to the data distribution. This paper aim to address these issues by introducing the maximum entropy principle, which favors models with flatter distributions that still effectively capture the data. Specifically, we develop a new distribution matching method called GEM, which solves reverse Kullback-Leibler divergence minimization with an entropy regularizer. For the SFT of Llama-3-8B models, GEM outperforms CE in several aspects. First, when applied to the UltraFeedback dataset to develop general instruction-following abilities, GEM exhibits reduced overfitting, evidenced by lower perplexity and better performance on the IFEval benchmark. Furthermore, GEM enhances output diversity, leading to performance gains of up to 7 points on math reasoning and code generation tasks using best-of-n sampling, even without domain-specific data. Second, when fine-tuning with domain-specific datasets for math reasoning and code generation, GEM also shows less overfitting and improvements of up to 10 points compared with CE.

We propose Adam-mini, an optimizer that achieves on-par or better performance than AdamW with 45% to 50% less memory footprint. Adam-mini reduces memory by cutting down the learning rate resources in Adam (i.e., $1/\sqrt{v}$). We find that $\geq$ 90% of these learning rates in $v$ could be harmlessly removed if we (1) carefully partition the parameters into blocks following our proposed principle on Hessian structure; (2) assign a single but good learning rate to each parameter block. We further find that, for each of these parameter blocks, there exists a single high-quality learning rate that can outperform Adam, provided that sufficient resources are available to search it out. We then provide one cost-effective way to find good learning rates and propose Adam-mini. Empirically, we verify that Adam-mini performs on par or better than AdamW on various language models sized from 125M to 7B for pre-training, supervised fine-tuning, and RLHF. The reduced memory footprint of Adam-mini also alleviates communication overheads among GPUs and CPUs, thereby increasing throughput. For instance, Adam-mini achieves 49.6% higher throughput than AdamW when pre-training Llama2-7B on $2\times$ A800-80GB GPUs, which saves 33% wall-clock time for pre-training.

SGD performs worse than Adam by a significant margin on Transformers, but the reason remains unclear. In this work, we provide an explanation through the lens of Hessian: (i) Transformers are "heterogeneous": the Hessian spectrum across parameter blocks vary dramatically, a phenomenon we call "block heterogeneity"; (ii) Heterogeneity hampers SGD: SGD performs worse than Adam on problems with block heterogeneity. To validate (i) and (ii), we check various Transformers, CNNs, MLPs, and quadratic problems, and find that SGD can perform on par with Adam on problems without block heterogeneity, but performs worse than Adam when the heterogeneity exists. Our initial theoretical analysis indicates that SGD performs worse because it applies one single learning rate to all blocks, which cannot handle the heterogeneity among blocks. This limitation could be ameliorated if we use coordinate-wise learning rates, as designed in Adam.

Accurately aligning large language models (LLMs) with human preferences is crucial for informing fair, economically sound, and statistically efficient decision-making processes. However, we argue that reinforcement learning from human feedback (RLHF) -- the predominant approach for aligning LLMs with human preferences through a reward model -- suffers from an inherent algorithmic bias due to its Kullback--Leibler-based regularization in optimization. In extreme cases, this bias could lead to a phenomenon we term preference collapse, where minority preferences are virtually disregarded. To mitigate this algorithmic bias, we introduce preference matching (PM) RLHF, a novel approach that provably aligns LLMs with the preference distribution of the reward model under the Bradley--Terry--Luce/Plackett--Luce model. Central to our approach is a PM regularizer that takes the form of the negative logarithm of the LLM's policy probability distribution over responses, which helps the LLM balance response diversification and reward maximization. Notably, we obtain this regularizer by solving an ordinary differential equation that is necessary for the PM property. For practical implementation, we introduce a conditional variant of PM RLHF that is tailored to natural language generation. Finally, we empirically validate the effectiveness of conditional PM RLHF through experiments on the OPT-1.3B and Llama-2-7B models, demonstrating a 29% to 41% improvement in alignment with human preferences, as measured by a certain metric, compared to standard RLHF.

Alignment is crucial for training large language models. The predominant strategy is Reinforcement Learning from Human Feedback (RLHF), with Proximal Policy Optimization (PPO) as the de-facto algorithm. Yet, PPO is known to struggle with computational inefficiency, a challenge that this paper aims to address. We identify three important properties of RLHF tasks: fast simulation, deterministic transitions, and trajectory-level rewards, which are not leveraged in PPO. Based on these properties, we develop ReMax, a new algorithm tailored for RLHF. The design of ReMax builds on the celebrated algorithm REINFORCE but is enhanced with a new variance-reduction technique. ReMax offers threefold advantages over PPO: first, it is simple to implement with just 6 lines of code. It further eliminates more than 4 hyper-parameters in PPO, which are laborious to tune. Second, ReMax reduces memory usage by about 50%. To illustrate, PPO runs out of memory when fine-tuning a Llama2-7B model on A100-80GB GPUs, whereas ReMax can support the training. Even though memory-efficient techniques (e.g., ZeRO and offload) are employed for PPO to afford training, ReMax can utilize a larger batch size to increase throughput. Third, in terms of wall-clock time, PPO is about twice as slow as ReMax per iteration. Importantly, these improvements do not sacrifice task performance. We hypothesize that these advantages can be maintained in larger-scale models.

Aligning intelligent agents with human preferences and values is important. This paper examines two popular alignment methods: Direct Preference Optimization (DPO) and Reward-Model-Based Policy Optimization (RMB-PO). A variant of RMB-PO, referred to as RMB-PO+ is also considered. These methods, either explicitly or implicitly, learn a reward model from preference data and differ in the data used for policy optimization to unlock the generalization ability of the reward model. In particular, compared with DPO, RMB-PO additionally uses policy-generated data, and RMB-PO+ further leverages new, preference-free data. We examine the impact of such out-of-preference data. Our study, conducted through controlled and synthetic experiments, demonstrates that DPO performs poorly, whereas RMB-PO+ performs the best. In particular, even when providing the policy model with a good feature representation, we find that policy optimization with adequate out-of-preference data significantly improves performance by harnessing the reward model's generalization capabilities.

Imitation learning (IL) has proven to be an effective method for learning good policies from expert demonstrations. Adversarial imitation learning (AIL), a subset of IL methods, is particularly promising, but its theoretical foundation in the presence of unknown transitions has yet to be fully developed. This paper explores the theoretical underpinnings of AIL in this context, where the stochastic and uncertain nature of environment transitions presents a challenge. We examine the expert sample complexity and interaction complexity required to recover good policies. To this end, we establish a framework connecting reward-free exploration and AIL, and propose an algorithm, MB-TAIL, that achieves the minimax optimal expert sample complexity of $\widetilde{O} (H^{3/2} |S|/\varepsilon)$ and interaction complexity of $\widetilde{O} (H^{3} |S|^2 |A|/\varepsilon^2)$. Here, $H$ represents the planning horizon, $|S|$ is the state space size, $|A|$ is the action space size, and $\varepsilon$ is the desired imitation gap. MB-TAIL is the first algorithm to achieve this level of expert sample complexity in the unknown transition setting and improves upon the interaction complexity of the best-known algorithm, OAL, by $O(H)$. Additionally, we demonstrate the generalization ability of MB-TAIL by extending it to the function approximation setting and proving that it can achieve expert sample and interaction complexity independent of $|S|$