Beyond Perplexity: Multi-dimensional Safety Evaluation of LLM Compression

Large language models (LLMs) are increasingly deployed in real-world scenarios with the help of recent model compression techniques. Such momentum towards local deployment means the use of compressed LLMs will widely impact a large population. However, prior analysis works often prioritize on preserving perplexity which is a direct analogy to training loss. The impact of compression method on other critical aspects of model behavior, particularly safety, still calls for a systematic assessment. To this end, we investigate the impact of model compression on four dimensions: (1) degeneration harm, i.e., bias and toxicity in generation; (2) representational harm, i.e., biases in discriminative tasks; (3) dialect bias; (4) language modeling and downstream task performance. We cover a wide spectrum of LLM compression techniques, including unstructured pruning, semi-structured pruning and quantization. Our analysis reveals that compression can lead to unexpected consequences. Although compression may unintentionally remedy LLMs' degeneration harm, it can still exacerbate on the representational harm axis. Although compression may unintentionally remedy LLMs' degeneration harm, it can still exacerbate on the representational harm axis. Moreover, there is a divergent impact on different protected groups as the compression rate grows. Finally, different compression methods have drastically different safety impacts, e.g., quantization mostly preserves bias while pruning degrades quickly. Our findings underscore the importance of integrating safety assessments into the development of compressed LLMs to ensure their reliability across real-world applications. Our full results are available here: \url{https://github.com/zhichaoxu-shufe/Beyond-Perplexity-Compression-Safety-Eval}

RankMamba: Benchmarking Mamba's Document Ranking Performance in the Era of Transformers

Transformer structure has achieved great success in multiple applied machine learning communities, such as natural language processing (NLP), computer vision (CV) and information retrieval (IR). Transformer architecture's core mechanism -- attention requires $O(n^2)$ time complexity in training and $O(n)$ time complexity in inference. Many works have been proposed to improve the attention mechanism's scalability, such as Flash Attention and Multi-query Attention. A different line of work aims to design new mechanisms to replace attention. Recently, a notable model structure -- Mamba, which is based on state space models, has achieved transformer-equivalent performance in multiple sequence modeling tasks. In this work, we examine \mamba's efficacy through the lens of a classical IR task -- document ranking. A reranker model takes a query and a document as input, and predicts a scalar relevance score. This task demands the language model's ability to comprehend lengthy contextual inputs and to capture the interaction between query and document tokens. We find that (1) Mamba models achieve competitive performance compared to transformer-based models with the same training recipe; (2) but also have a lower training throughput in comparison to efficient transformer implementations such as flash attention. We hope this study can serve as a starting point to explore Mamba models in other classical IR tasks. Our code implementation and trained checkpoints are made public to facilitate reproducibility (https://github.com/zhichaoxu-shufe/RankMamba).

Multi-dimensional Evaluation of Empathetic Dialog Responses

Empathy is a critical element of effective and satisfactory conversational communication, yet previous studies in measuring conversational empathy mostly focus on expressed communicative intents -- in which way empathy is expressed, ignoring the fact that conversation is also a collaborative practice involving both speakers and listeners. In contrast, we propose a multi-dimensional empathy evaluation framework that extends upon existing work to measure both expressed intents from the speaker's perspective and perceived empathy from the listener's perspective. Applying the proposed framework to analyzing our internal customer-service dialogue shows that the two dimensions (expressed intent types and perceived empathy) are inter-connected, while perceived empathy has high correlation with the satisfactory level of dialogue sessions. This proposed framework still requires subjective assessments from trained annotators, which can be non-trivial to collect. To scale up evaluation without excessive reliance on carefully annotated data, we explore different modeling options to automatically measure conversational empathy with (1) prompting frozen large language models (LLMs) and (2) training language model-based classifiers. Extensive experiments on both internal and external dialogue datasets show that measuring conversational empathy remains a challenging task for prompting frozen LLMs, reflected by less satisfying performance of GPT-4 and Flan family models. On the other hand, our proposed instruction-finetuned classifiers based on sequence-to-sequence (Seq2Seq) language models is able to achieve the best performance compared to prior works and competitive baselines. Finally, we perform comprehensive ablation studies on the performance of proposed instruction-finetuned classifiers and give recommendations on potentially adopting them as automatic conversational empathy evaluation metrics.

In-Context Example Ordering Guided by Label Distributions

By allowing models to predict without task-specific training, in-context learning (ICL) with pretrained LLMs has enormous potential in NLP. However, a number of problems persist in ICL. In particular, its performance is sensitive to the choice and order of in-context examples. Given the same set of in-context examples with different orderings, model performance may vary between near random to near state-of-the-art. In this work, we formulate in-context example ordering as an optimization problem. We examine three problem settings that differ in the assumptions they make about what is known about the task. Inspired by the idea of learning from label proportions, we propose two principles for in-context example ordering guided by model's probability predictions. We apply our proposed principles to thirteen text classification datasets and nine different autoregressive LLMs with 700M to 13B parameters. We demonstrate that our approach outperforms the baselines by improving the classification accuracy, reducing model miscalibration, and also by selecting better in-context examples.

Context-aware Decoding Reduces Hallucination in Query-focused Summarization

Query-focused summarization (QFS) aims to provide a summary of a single document/multi documents that can satisfy the information needs of a given query. It is useful for various real-world applications, such as abstractive snippet generation or more recent retrieval augmented generation (RAG). A prototypical QFS pipeline consists of a retriever (sparse or dense retrieval) and a generator (usually a large language model). However, applying large language models (LLM) potentially leads to hallucinations, especially when the evidence contradicts the prior belief of LLMs. There has been growing interest in developing new decoding methods to improve generation quality and reduce hallucination. In this work, we conduct a large-scale reproducibility study on one recently proposed decoding method -- Context-aware Decoding (CAD). In addition to replicating CAD's experiments on news summarization datasets, we include experiments on QFS datasets, and conduct more rigorous analysis on computational complexity and hyperparameter sensitivity. Experiments with eight different language models show that performance-wise, CAD improves QFS quality by (1) reducing factuality errors/hallucinations while (2) mostly retaining the match of lexical patterns, measured by ROUGE scores, while also at a cost of increased inference-time FLOPs and reduced decoding speed. The code implementation based on Huggingface Library is made available https://github.com/zhichaoxu-shufe/context-aware-decoding-qfs

FARA: Future-aware Ranking Algorithm for Fairness Optimization

Ranking systems are the key components of modern Information Retrieval (IR) applications, such as search engines and recommender systems. Besides the ranking relevance to users, the exposure fairness to item providers has also been considered an important factor in ranking optimization. Many fair ranking algorithms have been proposed to jointly optimize both ranking relevance and fairness. However, we find that most existing fair ranking methods adopt greedy algorithms that only optimize rankings for the next immediate session or request. As shown in this paper, such a myopic paradigm could limit the upper bound of ranking optimization and lead to suboptimal performance in the long term. To this end, we propose FARA, a novel Future-Aware Ranking Algorithm for ranking relevance and fairness optimization. Instead of greedily optimizing rankings for the next immediate session, FARA plans ahead by jointly optimizing multiple ranklists together and saving them for future sessions. Particularly, FARA first uses the Taylor expansion to investigate how future ranklists will influence the overall fairness of the system. Then, based on the analysis of the Taylor expansion, FARA adopts a two-phase optimization algorithm where we first solve an optimal future exposure planning problem and then construct the optimal ranklists according to the optimal future exposure planning. Theoretically, we show that FARA is optimal for ranking relevance and fairness joint optimization. Empirically, our extensive experiments on three semi-synthesized datasets show that FARA is efficient, effective, and can deliver significantly better ranking performance compared to state-of-the-art fair ranking methods.

A Lightweight Constrained Generation Alternative for Query-focused Summarization

Query-focused summarization (QFS) aims to provide a summary of a document that satisfies information need of a given query and is useful in various IR applications, such as abstractive snippet generation. Current QFS approaches typically involve injecting additional information, e.g. query-answer relevance or fine-grained token-level interaction between a query and document, into a finetuned large language model. However, these approaches often require extra parameters \& training, and generalize poorly to new dataset distributions. To mitigate this, we propose leveraging a recently developed constrained generation model Neurological Decoding (NLD) as an alternative to current QFS regimes which rely on additional sub-architectures and training. We first construct lexical constraints by identifying important tokens from the document using a lightweight gradient attribution model, then subsequently force the generated summary to satisfy these constraints by directly manipulating the final vocabulary likelihood. This lightweight approach requires no additional parameters or finetuning as it utilizes both an off-the-shelf neural retrieval model to construct the constraints and a standard generative language model to produce the QFS. We demonstrate the efficacy of this approach on two public QFS collections achieving near parity with the state-of-the-art model with substantially reduced complexity.

An In-depth Investigation of User Response Simulation for Conversational Search

Conversational search has seen increased recent attention in both the IR and NLP communities. It seeks to clarify and solve a user's search need through multi-turn natural language interactions. However, most existing systems are trained and demonstrated with recorded or artificial conversation logs. Eventually, conversational search systems should be trained, evaluated, and deployed in an open-ended setting with unseen conversation trajectories. A key challenge is that training and evaluating such systems both require a human-in-the-loop, which is expensive and does not scale. One strategy for this is to simulate users, thereby reducing the scaling costs. However, current user simulators are either limited to only respond to yes-no questions from the conversational search system, or unable to produce high quality responses in general. In this paper, we show that current state-of-the-art user simulation system could be significantly improved by replacing it with a smaller but advanced natural language generation model. But rather than merely reporting this new state-of-the-art, we present an in-depth investigation of the task of simulating user response for conversational search. Our goal is to supplement existing works with an insightful hand-analysis of what challenges are still unsolved by the advanced model, as well as to propose our solutions for them. The challenges we identified include (1) dataset noise, (2) a blind spot that is difficult for existing models to learn, and (3) a specific type of misevaluation in the standard empirical setup. Except for the dataset noise issue, we propose solutions to cover the training blind spot and to avoid the misevaluation. Our proposed solutions lead to further improvements. Our best system improves the previous state-of-the-art significantly.

Reward-free Policy Imitation Learning for Conversational Search

Existing conversational search studies mainly focused on asking better clarifying questions and/or improving search result quality. These works aim at retrieving better responses according to the search context, and their performances are evaluated on either single-turn tasks or multi-turn tasks under naive conversation policy settings. This leaves some questions about their applicability in real-world multi-turn conversations where realistically, each and every action needs to be made by the system itself, and search session efficiency is often an important concern of conversational search systems. While some recent works have identified the need for improving search efficiency in conversational search, they mostly require extensive data annotations and use hand-crafted rewards or heuristics to train systems that can achieve reasonable performance in a restricted number of turns, which has limited generalizability in practice. In this paper, we propose a reward-free conversation policy imitation learning framework, which can train a conversation policy without annotated conversation data or manually designed rewards. The trained conversation policy can be used to guide the conversational retrieval models to balance conversational search quality and efficiency. To evaluate the proposed conversational search system, we propose a new multi-turn-multi-response conversational evaluation metric named Expected Conversational Reciprocal Rank (ECRR). ECRR is designed to evaluate entire multi-turn conversational search sessions towards comprehensively evaluating both search result quality and search efficiency.

Counterfactual Editing for Search Result Explanation

Recently substantial improvements in neural retrieval methods also bring to light the inherent blackbox nature of these methods, especially when viewed from an explainability perspective. Most of existing works on Search Result Explanation (SeRE) are designed to provide factual explanation, i.e. to find/generate supporting evidence about documents' relevance to search queries. However, research in cognitive sciences have shown that human explanations are contrastive i.e. people explain an observed event using some counterfactual events; such explanations reduce cognitive load, and provide actionable insights. Though already proven effective in machine learning and NLP communities, the formulation and impact of counterfactual explanations have not been well studied for search systems. In this work, we aim to investigate the effectiveness of this perspective via proposing and evaluating counterfactual explanations for the task of SeRE. Specifically, we first conduct a user study where we investigate if counterfactual explanations indeed improve search sessions' effectiveness. Taking this as a motivation, we discuss the desiderata that an ideal counterfactual explanation method for SeRE should adhere to. Next, we propose a method $\text{CFE}^2$ (\textbf{C}ounter\textbf{F}actual \textbf{E}xplanation with \textbf{E}diting) to provide pairwise explanations to search engine result page. Finally, we showcase that the proposed method when evaluated on four publicly available datasets outperforms baselines on both metrics and human evaluation.