Reinforcement Learning from Reflective Feedback (RLRF): Aligning and Improving LLMs via Fine-Grained Self-Reflection

Despite the promise of RLHF in aligning LLMs with human preferences, it often leads to superficial alignment, prioritizing stylistic changes over improving downstream performance of LLMs. Underspecified preferences could obscure directions to align the models. Lacking exploration restricts identification of desirable outputs to improve the models. To overcome these challenges, we propose a novel framework: Reinforcement Learning from Reflective Feedback (RLRF), which leverages fine-grained feedback based on detailed criteria to improve the core capabilities of LLMs. RLRF employs a self-reflection mechanism to systematically explore and refine LLM responses, then fine-tuning the models via a RL algorithm along with promising responses. Our experiments across Just-Eval, Factuality, and Mathematical Reasoning demonstrate the efficacy and transformative potential of RLRF beyond superficial surface-level adjustment.

Pitfall of Optimism: Distributional Reinforcement Learning by Randomizing Risk Criterion

Distributional reinforcement learning algorithms have attempted to utilize estimated uncertainty for exploration, such as optimism in the face of uncertainty. However, using the estimated variance for optimistic exploration may cause biased data collection and hinder convergence or performance. In this paper, we present a novel distributional reinforcement learning algorithm that selects actions by randomizing risk criterion to avoid one-sided tendency on risk. We provide a perturbed distributional Bellman optimality operator by distorting the risk measure and prove the convergence and optimality of the proposed method with the weaker contraction property. Our theoretical results support that the proposed method does not fall into biased exploration and is guaranteed to converge to an optimal return. Finally, we empirically show that our method outperforms other existing distribution-based algorithms in various environments including Atari 55 games.

PreWoMe: Exploiting Presuppositions as Working Memory for Long Form Question Answering

Information-seeking questions in long-form question answering (LFQA) often prove misleading due to ambiguity or false presupposition in the question. While many existing approaches handle misleading questions, they are tailored to limited questions, which are insufficient in a real-world setting with unpredictable input characteristics. In this work, we propose PreWoMe, a unified approach capable of handling any type of information-seeking question. The key idea of PreWoMe involves extracting presuppositions in the question and exploiting them as working memory to generate feedback and action about the question. Our experiment shows that PreWoMe is effective not only in tackling misleading questions but also in handling normal ones, thereby demonstrating the effectiveness of leveraging presuppositions, feedback, and action for real-world QA settings.

SPOTS: Stable Placement of Objects with Reasoning in Semi-Autonomous Teleoperation Systems

Pick-and-place is one of the fundamental tasks in robotics research. However, the attention has been mostly focused on the ``pick'' task, leaving the ``place'' task relatively unexplored. In this paper, we address the problem of placing objects in the context of a teleoperation framework. Particularly, we focus on two aspects of the place task: stability robustness and contextual reasonableness of object placements. Our proposed method combines simulation-driven physical stability verification via real-to-sim and the semantic reasoning capability of large language models. In other words, given place context information (e.g., user preferences, object to place, and current scene information), our proposed method outputs a probability distribution over the possible placement candidates, considering the robustness and reasonableness of the place task. Our proposed method is extensively evaluated in two simulation and one real world environments and we show that our method can greatly increase the physical plausibility of the placement as well as contextual soundness while considering user preferences.

On Monotonic Aggregation for Open-domain QA

Question answering (QA) is a critical task for speech-based retrieval from knowledge sources, by sifting only the answers without requiring to read supporting documents. Specifically, open-domain QA aims to answer user questions on unrestricted knowledge sources. Ideally, adding a source should not decrease the accuracy, but we find this property (denoted as "monotonicity") does not hold for current state-of-the-art methods. We identify the cause, and based on that we propose Judge-Specialist framework. Our framework consists of (1) specialist retrievers/readers to cover individual sources, and (2) judge, a dedicated language model to select the final answer. Our experiments show that our framework not only ensures monotonicity, but also outperforms state-of-the-art multi-source QA methods on Natural Questions. Additionally, we show that our models robustly preserve the monotonicity against noise from speech recognition. We publicly release our code and setting.

When to Read Documents or QA History: On Unified and Selective Open-domain QA

This paper studies the problem of open-domain question answering, with the aim of answering a diverse range of questions leveraging knowledge resources. Two types of sources, QA-pair and document corpora, have been actively leveraged with the following complementary strength. The former is highly precise when the paraphrase of given question $q$ was seen and answered during training, often posed as a retrieval problem, while the latter generalizes better for unseen questions. A natural follow-up is thus leveraging both models, while a naive pipelining or integration approaches have failed to bring additional gains over either model alone. Our distinction is interpreting the problem as calibration, which estimates the confidence of predicted answers as an indicator to decide when to use a document or QA-pair corpus. The effectiveness of our method was validated on widely adopted benchmarks such as Natural Questions and TriviaQA.

Revisiting Dense Retrieval with Unanswerable Counterfactuals

The retriever-reader framework is popular for open-domain question answering (ODQA), where a retriever samples for the reader a set of relevant candidate passages from a large corpus. A key assumption behind this method is that high relevance scores from the retriever likely indicate high answerability from the reader, which implies a high probability that the retrieved passages contain answers to a given question. In this work, we empirically dispel this belief and observe that recent dense retrieval models based on DPR often rank unanswerable counterfactual passages higher than their answerable original passages. To address such answer-unawareness in dense retrievers, we seek to use counterfactual samples as additional training resources to better synchronize the relevance measurement of DPR with the answerability of question-passage pairs. Specifically, we present counterfactually-Pivoting Contrastive Learning (PiCL), a novel representation learning approach for passage retrieval that leverages counterfactual samples as pivots between positive and negative samples in their learned embedding space. We incorporate PiCL into the retriever training to show the effectiveness of PiCL on ODQA benchmarks and the robustness of the learned models.

Exploring the Benefits of Training Expert Language Models over Instruction Tuning

Recently, Language Models (LMs) instruction-tuned on multiple tasks, also known as multitask-prompted fine-tuning (MT), have shown the capability to generalize to unseen tasks. Previous work has shown that scaling the number of training tasks is the key component in making stronger MT LMs. In this work, we report an unexpected finding that an expert LM fine-tuned on just a single task can outperform an MT LM trained with 300+ different tasks on 11 different unseen datasets and on 13 datasets of the BIG-bench benchmark by a mean accuracy of 3.20% and 1.29%, respectively. This finding casts doubt on the previously held belief that simply scaling the number of tasks makes stronger MT LMs. Leveraging this finding, we further show that this distributed approach of training a separate expert LM per training task instead of a single MT LM for zero-shot inference possesses many benefits including (1) avoiding negative task transfer that often occurs during instruction tuning, (2) being able to continually learn new tasks without having to re-train on previous tasks to avoid catastrophic forgetting, and (3) showing compositional capabilities when merging individual experts together. The code is available at https://github.com/joeljang/ELM.