Andrew Cohen

Jiacheng Liu, Andrew Cohen, Ramakanth Pasunuru et al. · berkeley, meta-ai

Inference-time search algorithms such as Monte-Carlo Tree Search (MCTS) may seem unnecessary when generating natural language text based on state-of-the-art reinforcement learning such as Proximal Policy Optimization (PPO). In this paper, we demonstrate that it is possible to get extra mileage out of PPO by integrating MCTS on top. The key idea is not to throw out the value network, a byproduct of PPO training for evaluating partial output sequences, when decoding text out of the policy network. More concretely, we present a novel value-guided decoding algorithm called PPO-MCTS, which can integrate the value network from PPO to work closely with the policy network during inference-time generation. Compared to prior approaches based on MCTS for controlled text generation, the key strength of our approach is to reduce the fundamental mismatch of the scoring mechanisms of the partial outputs between training and test. Evaluation on four text generation tasks demonstrate that PPO-MCTS greatly improves the preferability of generated text compared to the standard practice of using only the PPO policy. Our results demonstrate the promise of search algorithms even on top of the aligned language models from PPO, and the under-explored benefit of the value network.

Kumar Shridhar, Koustuv Sinha, Andrew Cohen et al. · berkeley, meta-ai

In recent years, Large Language Models (LLMs) have demonstrated remarkable generative abilities, but can they judge the quality of their own generations? A popular concept, referred to as self-refinement, postulates that LLMs can detect and correct the errors in their generations when asked to do so. However, recent empirical evidence points in the opposite direction, suggesting that LLMs often struggle to accurately identify errors when reasoning is involved. To address this, we propose a reasoning with refinement objective called ART: Ask, Refine, and Trust, which asks necessary questions to decide when an LLM should refine its output, and either affirm or withhold trust in its refinement by ranking the refinement and the initial prediction. On two multistep reasoning tasks of mathematical word problems (GSM8K) and question answering (StrategyQA), ART achieves a performance gain of +5 points over self-refinement baselines, while using a much smaller model as the decision maker. We also demonstrate the benefit of using smaller models to make refinement decisions as a cost-effective alternative to fine-tuning a larger model.

Jihoon Tack, Jack Lanchantin, Jane Yu et al. · meta-ai

Next token prediction has been the standard training objective used in large language model pretraining. Representations are learned as a result of optimizing for token-level perplexity. We propose Continuous Concept Mixing (CoCoMix), a novel pretraining framework that combines discrete next token prediction with continuous concepts. Specifically, CoCoMix predicts continuous concepts learned from a pretrained sparse autoencoder and mixes them into the model's hidden state by interleaving with token hidden representations. Through experiments on multiple benchmarks, including language modeling and downstream reasoning tasks, we show that CoCoMix is more sample efficient and consistently outperforms standard next token prediction, knowledge distillation and inserting pause tokens. We find that combining both concept learning and interleaving in an end-to-end framework is critical to performance gains. Furthermore, CoCoMix enhances interpretability and steerability by allowing direct inspection and modification of the predicted concept, offering a transparent way to guide the model's internal reasoning process.

Michihiro Yasunaga, Leonid Shamis, Chunting Zhou et al.

Recent approaches to large language model (LLM) alignment typically require millions of human annotations or rely on external aligned models for synthetic data generation. This paper introduces ALMA: Alignment with Minimal Annotation, demonstrating that effective alignment can be achieved using only 9,000 labeled examples -- less than 1% of conventional approaches. ALMA generates large amounts of high-quality synthetic alignment data through new techniques: diverse prompt synthesis via few-shot learning, diverse response generation with multiple model checkpoints, and judge (reward model) enhancement through score aggregation and self-distillation. Using only a pretrained Llama3 base model, 5,000 SFT examples, and 4,000 judge annotations, ALMA achieves performance close to Llama3-Instruct across diverse alignment benchmarks (e.g., 0.1% difference on AlpacaEval 2.0 score). These results are achieved with a multi-round, self-bootstrapped data synthesis and training recipe that continues to improve for 10 rounds, surpassing the typical 3-round ceiling of previous methods. These results suggest that base models already possess sufficient knowledge for effective alignment, and that synthetic data generation methods can expose it.

Lun Jiang, Nima Salehi Sadghiani, Zhuo Tao et al.

From the ad network standpoint, a user's activity is a multi-type sequence of temporal events consisting of event types and time intervals. Understanding user patterns in ad networks has received increasing attention from the machine learning community. Particularly, the problems of fraud detection, Conversion Rate (CVR), and Click-Through Rate (CTR) prediction are of interest. However, the class imbalance between major and minor classes in these tasks can bias a machine learning model leading to poor performance. This study proposes using two multi-type (continuous and discrete) training approaches for GANs to deal with the limitations of traditional GANs in passing the gradient updates for discrete tokens. First, we used the Reinforcement Learning (RL)-based training approach and then, an approximation of the multinomial distribution parameterized in terms of the softmax function (Gumble-Softmax). Our extensive experiments based on synthetic data have shown the trained generator can generate sequences with desired properties measured by multiple criteria.

Andrew Cohen, Lei Yu, Xingye Qiao et al.

Two hitherto disconnected threads of research, diverse exploration (DE) and maximum entropy RL have addressed a wide range of problems facing reinforcement learning algorithms via ostensibly distinct mechanisms. In this work, we identify a connection between these two approaches. First, a discriminator-based diversity objective is put forward and connected to commonly used divergence measures. We then extend this objective to the maximum entropy framework and propose an algorithm Maximum Entropy Diverse Exploration (MEDE) which provides a principled method to learn diverse behaviors. A theoretical investigation shows that the set of policies learned by MEDE capture the same modalities as the optimal maximum entropy policy. In effect, the proposed algorithm disentangles the maximum entropy policy into its diverse, constituent policies. Experiments show that MEDE is superior to the state of the art in learning high performing and diverse policies.

Andrew Cohen, Xingye Qiao, Lei Yu et al.

We address the challenge of effective exploration while maintaining good performance in policy gradient methods. As a solution, we propose diverse exploration (DE) via conjugate policies. DE learns and deploys a set of conjugate policies which can be conveniently generated as a byproduct of conjugate gradient descent. We provide both theoretical and empirical results showing the effectiveness of DE at achieving exploration, improving policy performance, and the advantage of DE over exploration by random policy perturbations.

Andrew Cohen, Lei Yu, Robert Wright

We study an important yet under-addressed problem of quickly and safely improving policies in online reinforcement learning domains. As its solution, we propose a novel exploration strategy - diverse exploration (DE), which learns and deploys a diverse set of safe policies to explore the environment. We provide DE theory explaining why diversity in behavior policies enables effective exploration without sacrificing exploitation. Our empirical study shows that an online policy improvement algorithm framework implementing the DE strategy can achieve both fast policy improvement and safe online performance.