MACPO: Weak-to-Strong Alignment via Multi-Agent Contrastive Preference Optimization

As large language models (LLMs) are rapidly advancing and achieving near-human capabilities, aligning them with human values is becoming more urgent. In scenarios where LLMs outperform humans, we face a weak-to-strong alignment problem where we need to effectively align strong student LLMs through weak supervision generated by weak teachers. Existing alignment methods mainly focus on strong-to-weak alignment and self-alignment settings, and it is impractical to adapt them to the much harder weak-to-strong alignment setting. To fill this gap, we propose a multi-agent contrastive preference optimization (MACPO) framework. MACPO facilitates weak teachers and strong students to learn from each other by iteratively reinforcing unfamiliar positive behaviors while penalizing familiar negative ones. To get this, we devise a mutual positive behavior augmentation strategy to encourage weak teachers and strong students to learn from each other's positive behavior and further provide higher quality positive behavior for the next iteration. Additionally, we propose a hard negative behavior construction strategy to induce weak teachers and strong students to generate familiar negative behavior by fine-tuning on negative behavioral data. Experimental results on the HH-RLHF and PKU-SafeRLHF datasets, evaluated using both automatic metrics and human judgments, demonstrate that MACPO simultaneously improves the alignment performance of strong students and weak teachers. Moreover, as the number of weak teachers increases, MACPO achieves better weak-to-strong alignment performance through more iteration optimization rounds.

Uncovering Overfitting in Large Language Model Editing

Knowledge editing has been proposed as an effective method for updating and correcting the internal knowledge of Large Language Models (LLMs). However, existing editing methods often struggle with complex tasks, such as multi-hop reasoning. In this paper, we identify and investigate the phenomenon of Editing Overfit, where edited models assign disproportionately high probabilities to the edit target, hindering the generalization of new knowledge in complex scenarios. We attribute this issue to the current editing paradigm, which places excessive emphasis on the direct correspondence between the input prompt and the edit target for each edit sample. To further explore this issue, we introduce a new benchmark, EVOKE (EValuation of Editing Overfit in Knowledge Editing), along with fine-grained evaluation metrics. Through comprehensive experiments and analysis, we demonstrate that Editing Overfit is prevalent in current editing methods and that common overfitting mitigation strategies are of limited effectiveness in knowledge editing. To overcome this, inspired by LLMs' knowledge recall mechanisms, we propose a new plug-and-play strategy called Learn to Inference (LTI), which introduce a Multi-stage Inference Constraint module to guide the edited models in recalling new knowledge similarly to how unedited LLMs leverage knowledge through in-context learning. Extensive experimental results across a wide range of tasks validate the effectiveness of LTI in mitigating Editing Overfit.

Enhancing Multi-hop Reasoning through Knowledge Erasure in Large Language Model Editing

Large language models (LLMs) face challenges with internal knowledge inaccuracies and outdated information. Knowledge editing has emerged as a pivotal approach to mitigate these issues. Although current knowledge editing techniques exhibit promising performance in single-hop reasoning tasks, they show limitations when applied to multi-hop reasoning. Drawing on cognitive neuroscience and the operational mechanisms of LLMs, we hypothesize that the residual single-hop knowledge after editing causes edited models to revert to their original answers when processing multi-hop questions, thereby undermining their performance in multihop reasoning tasks. To validate this hypothesis, we conduct a series of experiments that empirically confirm our assumptions. Building on the validated hypothesis, we propose a novel knowledge editing method that incorporates a Knowledge Erasure mechanism for Large language model Editing (KELE). Specifically, we design an erasure function for residual knowledge and an injection function for new knowledge. Through joint optimization, we derive the optimal recall vector, which is subsequently utilized within a rank-one editing framework to update the parameters of targeted model layers. Extensive experiments on GPT-J and GPT-2 XL demonstrate that KELE substantially enhances the multi-hop reasoning capability of edited LLMs.

Beyond Local Views: Global State Inference with Diffusion Models for Cooperative Multi-Agent Reinforcement Learning

In partially observable multi-agent systems, agents typically only have access to local observations. This severely hinders their ability to make precise decisions, particularly during decentralized execution. To alleviate this problem and inspired by image outpainting, we propose State Inference with Diffusion Models (SIDIFF), which uses diffusion models to reconstruct the original global state based solely on local observations. SIDIFF consists of a state generator and a state extractor, which allow agents to choose suitable actions by considering both the reconstructed global state and local observations. In addition, SIDIFF can be effortlessly incorporated into current multi-agent reinforcement learning algorithms to improve their performance. Finally, we evaluated SIDIFF on different experimental platforms, including Multi-Agent Battle City (MABC), a novel and flexible multi-agent reinforcement learning environment we developed. SIDIFF achieved desirable results and outperformed other popular algorithms.

Chain-of-Strategy Planning with LLMs: Aligning the Generation of Psychotherapy Dialogue with Strategy in Motivational Interviewing

Recent advancements in large language models (LLMs) have shown promise in generating psychotherapeutic dialogues, especially in Motivational Interviewing (MI). However, how to employ strategies, a set of motivational interviewing (MI) skills, to generate therapeutic-adherent conversations with explainability is underexplored. We propose an approach called strategy-aware dialogue generation with Chain-of-Strategy (CoS) planning, which first predicts MI strategies as reasoning and utilizes these strategies to guide the subsequent dialogue generation. It brings the potential for controllable and explainable generation in psychotherapy by aligning the generated MI dialogues with therapeutic strategies. Extensive experiments including automatic and human evaluations are conducted to validate the effectiveness of the MI strategy. Our findings demonstrate the potential of LLMs in producing strategically aligned dialogues and suggest directions for practical applications in psychotherapeutic settings.

Generate-then-Ground in Retrieval-Augmented Generation for Multi-hop Question Answering

Multi-Hop Question Answering (MHQA) tasks present a significant challenge for large language models (LLMs) due to the intensive knowledge required. Current solutions, like Retrieval-Augmented Generation, typically retrieve potential documents from an external corpus to read an answer. However, the performance of this retrieve-then-read paradigm is constrained by the retriever and the inevitable noise in the retrieved documents. To mitigate these challenges, we introduce a novel generate-then-ground (GenGround) framework, synergizing the parametric knowledge of LLMs and external documents to solve a multi-hop question. GenGround empowers LLMs to alternate two phases until the final answer is derived: (1) formulate a simpler, single-hop question and directly generate the answer; (2) ground the question-answer pair in retrieved documents, amending any wrong predictions in the answer. We also propose an instructional grounding distillation method to generalize our method into smaller models. Extensive experiments conducted on four datasets illustrate the superiority of our method.

MEFT: Memory-Efficient Fine-Tuning through Sparse Adapter

Parameter-Efficient Fine-tuning (PEFT) facilitates the fine-tuning of Large Language Models (LLMs) under limited resources. However, the fine-tuning performance with PEFT on complex, knowledge-intensive tasks is limited due to the constrained model capacity, which originates from the limited number of additional trainable parameters. To overcome this limitation, we introduce a novel mechanism that fine-tunes LLMs with adapters of larger size yet memory-efficient. This is achieved by leveraging the inherent activation sparsity in the Feed-Forward Networks (FFNs) of LLMs and utilizing the larger capacity of Central Processing Unit (CPU) memory compared to Graphics Processing Unit (GPU). We store and update the parameters of larger adapters on the CPU. Moreover, we employ a Mixture of Experts (MoE)-like architecture to mitigate unnecessary CPU computations and reduce the communication volume between the GPU and CPU. This is particularly beneficial over the limited bandwidth of PCI Express (PCIe). Our method can achieve fine-tuning results comparable to those obtained with larger memory capacities, even when operating under more limited resources such as a 24GB memory single GPU setup, with acceptable loss in training efficiency. Our codes are available at https://github.com/CURRENTF/MEFT.

Autonomous Workflow for Multimodal Fine-Grained Training Assistants Towards Mixed Reality

Autonomous artificial intelligence (AI) agents have emerged as promising protocols for automatically understanding the language-based environment, particularly with the exponential development of large language models (LLMs). However, a fine-grained, comprehensive understanding of multimodal environments remains under-explored. This work designs an autonomous workflow tailored for integrating AI agents seamlessly into extended reality (XR) applications for fine-grained training. We present a demonstration of a multimodal fine-grained training assistant for LEGO brick assembly in a pilot XR environment. Specifically, we design a cerebral language agent that integrates LLM with memory, planning, and interaction with XR tools and a vision-language agent, enabling agents to decide their actions based on past experiences. Furthermore, we introduce LEGO-MRTA, a multimodal fine-grained assembly dialogue dataset synthesized automatically in the workflow served by a commercial LLM. This dataset comprises multimodal instruction manuals, conversations, XR responses, and vision question answering. Last, we present several prevailing open-resource LLMs as benchmarks, assessing their performance with and without fine-tuning on the proposed dataset. We anticipate that the broader impact of this workflow will advance the development of smarter assistants for seamless user interaction in XR environments, fostering research in both AI and HCI communities.

ExcluIR: Exclusionary Neural Information Retrieval

Exclusion is an important and universal linguistic skill that humans use to express what they do not want. However, in information retrieval community, there is little research on exclusionary retrieval, where users express what they do not want in their queries. In this work, we investigate the scenario of exclusionary retrieval in document retrieval for the first time. We present ExcluIR, a set of resources for exclusionary retrieval, consisting of an evaluation benchmark and a training set for helping retrieval models to comprehend exclusionary queries. The evaluation benchmark includes 3,452 high-quality exclusionary queries, each of which has been manually annotated. The training set contains 70,293 exclusionary queries, each paired with a positive document and a negative document. We conduct detailed experiments and analyses, obtaining three main observations: (1) Existing retrieval models with different architectures struggle to effectively comprehend exclusionary queries; (2) Although integrating our training data can improve the performance of retrieval models on exclusionary retrieval, there still exists a gap compared to human performance; (3) Generative retrieval models have a natural advantage in handling exclusionary queries. To facilitate future research on exclusionary retrieval, we share the benchmark and evaluation scripts on \url{https://github.com/zwh-sdu/ExcluIR}.

Offline Trajectory Generalization for Offline Reinforcement Learning

Offline reinforcement learning (RL) aims to learn policies from static datasets of previously collected trajectories. Existing methods for offline RL either constrain the learned policy to the support of offline data or utilize model-based virtual environments to generate simulated rollouts. However, these methods suffer from (i) poor generalization to unseen states; and (ii) trivial improvement from low-qualified rollout simulation. In this paper, we propose offline trajectory generalization through world transformers for offline reinforcement learning (OTTO). Specifically, we use casual Transformers, a.k.a. World Transformers, to predict state dynamics and the immediate reward. Then we propose four strategies to use World Transformers to generate high-rewarded trajectory simulation by perturbing the offline data. Finally, we jointly use offline data with simulated data to train an offline RL algorithm. OTTO serves as a plug-in module and can be integrated with existing offline RL methods to enhance them with better generalization capability of transformers and high-rewarded data augmentation. Conducting extensive experiments on D4RL benchmark datasets, we verify that OTTO significantly outperforms state-of-the-art offline RL methods.