Shi Feng

Jiaxin Wen, Ruiqi Zhong, Akbir Khan et al.

Language models (LMs) can produce errors that are hard to detect for humans, especially when the task is complex. RLHF, the most popular post-training method, may exacerbate this problem: to achieve higher rewards, LMs might get better at convincing humans that they are right even when they are wrong. We study this phenomenon under a standard RLHF pipeline, calling it "U-SOPHISTRY" since it is Unintended by model developers. Specifically, we ask time-constrained (e.g., 3-10 minutes) human subjects to evaluate the correctness of model outputs and calculate humans' accuracy against gold labels. On a question-answering task (QuALITY) and programming task (APPS), RLHF makes LMs better at convincing our subjects but not at completing the task correctly. RLHF also makes the model harder to evaluate: our subjects' false positive rate increases by 24.1% on QuALITY and 18.3% on APPS. Finally, we show that probing, a state-of-the-art approach for detecting Intended Sophistry (e.g. backdoored LMs), does not generalize to U-SOPHISTRY. Our results highlight an important failure mode of RLHF and call for more research in assisting humans to align them.

Yongkang Liu, Feng Shi, Daling Wang et al.

Although large language models(LLMs) show amazing capabilities, among various exciting applications discovered for LLMs fall short in other low-resource languages. Besides, most existing methods depend on large-scale dialogue corpora and thus building systems for dialogue generation in a zero-shot scenario remains a considerable challenge. To address this challenge, we propose a novel end-to-end zero-shot dialogue generation model ChatZero based on cross-lingual code-switching method. First, we construct code-switching language and pseudo-target language with placeholders. Then for cross-lingual semantic transfer, we employ unsupervised contrastive learning to minimize the semantics gap of the source language, code-switching language, and pseudo-target language that are mutually positive examples in the high dimensional semantic space. Experiments on the multilingual DailyDialog and DSTC7-AVSD datasets demonstrate that ChatZero can achieve more than 90\% of the original performance under the zero-shot case compared to supervised learning, and achieve state-of-the-art performance compared with other baselines.

Chenglei Si, Navita Goyal, Sherry Tongshuang Wu et al.

Large Language Models (LLMs) are increasingly used for accessing information on the web. Their truthfulness and factuality are thus of great interest. To help users make the right decisions about the information they get, LLMs should not only provide information but also help users fact-check it. Our experiments with 80 crowdworkers compare language models with search engines (information retrieval systems) at facilitating fact-checking. We prompt LLMs to validate a given claim and provide corresponding explanations. Users reading LLM explanations are significantly more efficient than those using search engines while achieving similar accuracy. However, they over-rely on the LLMs when the explanation is wrong. To reduce over-reliance on LLMs, we ask LLMs to provide contrastive information - explain both why the claim is true and false, and then we present both sides of the explanation to users. This contrastive explanation mitigates users' over-reliance on LLMs, but cannot significantly outperform search engines. Further, showing both search engine results and LLM explanations offers no complementary benefits compared to search engines alone. Taken together, our study highlights that natural language explanations by LLMs may not be a reliable replacement for reading the retrieved passages, especially in high-stakes settings where over-relying on wrong AI explanations could lead to critical consequences.

Ming Wang, Wenfang Wu, Chongyun Gao et al.

Large language models (LLMs) have received increasing attention. However, due to the complexity of its capabilities, how to rationally evaluate the capabilities of LLMs is still a task to be solved. We propose the RoCar method, which utilizes the defined basic schemas to randomly construct a task graph and generates natural language evaluation tasks based on the task graph to evaluate the reasoning and memory abilities of LLMs respectively. Due to the very large randomness of the task construction process, it is possible to ensure that none of the LLMs to be tested has directly learned the evaluation tasks, guaranteeing the fairness of the evaluation method.

Jiayang Yu, Yihang Zhang, Bin Wang et al.

Fine-tuning is a key approach for adapting language models to specific downstream tasks, but updating all model parameters becomes impractical as model sizes increase. Parameter-Efficient Fine-Tuning (PEFT) methods, such as Low-Rank Adaptation (LoRA), address this challenge by introducing additional adaptation parameters into pre-trained weight matrices. However, LoRA's performance varies across different insertion points within the model, highlighting potential parameter inefficiency due to unnecessary insertions. To this end, we propose SSMLoRA (State Space Model Low-Rank Adaptation), an extension of LoRA that incorporates a State Space Model (SSM) to interconnect low-rank matrices. SSMLoRA ensures that performance is maintained even with sparser insertions. SSMLoRA allows the model to not only map inputs to a low-rank space for better feature extraction but also leverage the computations from the previous low-rank space. Our method achieves comparable performance to LoRA on the General Language Understanding Evaluation (GLUE) benchmark while using only half the parameters. Additionally, due to its structure, SSMLoRA shows promise in handling tasks with longer input sequences. .You can find our code here:https://github.com/yuhkalhic/SSMLoRA.

Lin Wang, Xiaocui Yang, Shi Feng et al.

Multimodal conversation, a crucial form of human communication, carries rich emotional content, making the exploration of the causes of emotions within it a research endeavor of significant importance. However, existing research on the causes of emotions typically employs an utterance selection method within a single textual modality to locate causal utterances. This approach remains limited to coarse-grained assessments, lacks nuanced explanations of emotional causation, and demonstrates inadequate capability in identifying multimodal emotional triggers. Therefore, we introduce a task-\textbf{Multimodal Emotion Cause Explanation in Conversation (MECEC)}. This task aims to generate a summary based on the multimodal context of conversations, clearly and intuitively describing the reasons that trigger a given emotion. To adapt to this task, we develop a new dataset (ECEM) based on the MELD dataset. ECEM combines video clips with detailed explanations of character emotions, helping to explore the causal factors behind emotional expression in multimodal conversations. A novel approach, FAME-Net, is further proposed, that harnesses the power of Large Language Models (LLMs) to analyze visual data and accurately interpret the emotions conveyed through facial expressions in videos. By exploiting the contagion effect of facial emotions, FAME-Net effectively captures the emotional causes of individuals engaged in conversations. Our experimental results on the newly constructed dataset show that FAME-Net outperforms several excellent baselines. Code and dataset are available at https://github.com/3222345200/FAME-Net.

Nishant Balepur, Feng Gu, Abhilasha Ravichander et al. · cmu

Question answering (QA), giving correct answers to questions, is a popular task, but we test reverse question answering (RQA): for an input answer, give a question with that answer. Past work tests QA and RQA separately, but we test them jointly, comparing their difficulty, aiding benchmark design, and checking reasoning consistency. We run 16 LLMs on QA and RQA with trivia questions/answers, revealing: 1) Versus QA, LLMs are much less accurate in RQA for numerical answers, but slightly more accurate in RQA for textual answers; 2) LLMs often answer their own invalid questions from RQA accurately in QA, so RQA errors are not from knowledge gaps alone; 3) RQA errors correlate with question difficulty and inversely correlate with answer frequencies in the Dolma corpus; and 4) LLMs struggle to provide valid multi-hop questions. By finding question and answer types that lead to RQA errors, we suggest improvements for LLM reasoning.

Jiaxin Wen, Zachary Ankner, Arushi Somani et al. · anthropic

To steer pretrained language models for downstream tasks, today's post-training paradigm relies on humans to specify desired behaviors. However, for models with superhuman capabilities, it is difficult or impossible to get high-quality human supervision. To address this challenge, we introduce a new unsupervised algorithm, Internal Coherence Maximization (ICM), to fine-tune pretrained language models on their own generated labels, \emph{without external supervision}. On GSM8k-verification, TruthfulQA, and Alpaca reward modeling tasks, our method matches the performance of training on golden supervision and outperforms training on crowdsourced human supervision. On tasks where LMs' capabilities are strongly superhuman, our method can elicit those capabilities significantly better than training on human labels. Finally, we show that our method can improve the training of frontier LMs: we use our method to train an unsupervised reward model and use reinforcement learning to train a Claude 3.5 Haiku-based assistant. Both the reward model and the assistant outperform their human-supervised counterparts.

Shiyi Mu, Yongkang Liu, Shi Feng et al.

Although large language models (LLMs) excel in text comprehension and generation, their performance on the Emotion-Cause Pair Extraction (ECPE) task, which requires reasoning ability, is often underperform smaller language model. The main reason is the lack of auxiliary knowledge, which limits LLMs' ability to effectively perceive emotions and reason causes. To address this issue, we propose a novel \textbf{M}ulti-source h\textbf{E}terogeneous \textbf{K}nowledge \textbf{i}njection me\textbf{T}hod, MEKiT, which integrates heterogeneous internal emotional knowledge and external causal knowledge. Specifically, for these two distinct aspects and structures of knowledge, we apply the approaches of incorporating instruction templates and mixing data for instruction-tuning, which respectively facilitate LLMs in more comprehensively identifying emotion and accurately reasoning causes. Experimental results demonstrate that MEKiT provides a more effective and adaptable solution for the ECPE task, exhibiting an absolute performance advantage over compared baselines and dramatically improving the performance of LLMs on the ECPE task.

Haoyu Han, Xiaorui Liu, Feng Shi et al.

Graph Neural Networks (GNNs) have emerged as a powerful tool for semi-supervised node classification tasks. However, recent studies have revealed various biases in GNNs stemming from both node features and graph topology. In this work, we uncover a new bias - label position bias, which indicates that the node closer to the labeled nodes tends to perform better. We introduce a new metric, the Label Proximity Score, to quantify this bias, and find that it is closely related to performance disparities. To address the label position bias, we propose a novel optimization framework for learning a label position unbiased graph structure, which can be applied to existing GNNs. Extensive experiments demonstrate that our proposed method not only outperforms backbone methods but also significantly mitigates the issue of label position bias in GNNs.

Feng Shi, Chonghan Lee, Mohammad Khairul Bashar et al.

CNF-based SAT and MaxSAT solvers are central to logic synthesis and verification systems. The increasing popularity of these constraint problems in electronic design automation encourages studies on different SAT problems and their properties for further computational efficiency. There has been both theoretical and practical success of modern Conflict-driven clause learning SAT solvers, which allows solving very large industrial instances in a relatively short amount of time. Recently, machine learning approaches provide a new dimension to solving this challenging problem. Neural symbolic models could serve as generic solvers that can be specialized for specific domains based on data without any changes to the structure of the model. In this work, we propose a one-shot model derived from the Transformer architecture to solve the MaxSAT problem, which is the optimization version of SAT where the goal is to satisfy the maximum number of clauses. Our model has a scale-free structure which could process varying size of instances. We use meta-path and self-attention mechanism to capture interactions among homogeneous nodes. We adopt cross-attention mechanisms on the bipartite graph to capture interactions among heterogeneous nodes. We further apply an iterative algorithm to our model to satisfy additional clauses, enabling a solution approaching that of an exact-SAT problem. The attention mechanisms leverage the parallelism for speedup. Our evaluation indicates improved speedup compared to heuristic approaches and improved completion rate compared to machine learning approaches.

Feng Shi, Ahren Yiqiao Jin, Song-Chun Zhu

\textit{Graph Neural Network} (GNN) is a promising approach for analyzing graph-structured data that tactfully captures their dependency information via node-level message passing. It has achieved state-of-the-art performances in many tasks, such as node classification, graph matching, clustering, and graph generation. As GNNs operate on non-Euclidean data, their irregular data access patterns cause considerable computational costs and overhead on conventional architectures, such as GPU and CPU. Our analysis shows that GNN adopts a hybrid computing model. The \textit{Aggregation} (or \textit{Message Passing}) phase performs vector additions where vectors are fetched with irregular strides. The \textit{Transformation} (or \textit{Node Embedding}) phase can be either dense or sparse-dense matrix multiplication. In this work, We propose \textit{VersaGNN}, an ultra-efficient, systolic-array-based versatile hardware accelerator that unifies dense and sparse matrix multiplication. By applying this single optimized systolic array to both aggregation and transformation phases, we have significantly reduced chip sizes and energy consumption. We then divide the computing engine into blocked systolic arrays to support the \textit{Strassen}'s algorithm for dense matrix multiplication, dramatically scaling down the number of multiplications and enabling high-throughput computation of GNNs. To balance the workload of sparse-dense matrix multiplication, we also introduced a greedy algorithm to combine sparse sub-matrices of compressed format into condensed ones to reduce computational cycles. Compared with current state-of-the-art GNN software frameworks, \textit{VersaGNN} achieves on average 3712$\times$ speedup with 1301.25$\times$ energy reduction on CPU, and 35.4$\times$ speedup with 17.66$\times$ energy reduction on GPU.

Tony Z. Zhao, Eric Wallace, Shi Feng et al.

GPT-3 can perform numerous tasks when provided a natural language prompt that contains a few training examples. We show that this type of few-shot learning can be unstable: the choice of prompt format, training examples, and even the order of the training examples can cause accuracy to vary from near chance to near state-of-the-art. We demonstrate that this instability arises from the bias of language models towards predicting certain answers, e.g., those that are placed near the end of the prompt or are common in the pre-training data. To mitigate this, we first estimate the model's bias towards each answer by asking for its prediction when given the training prompt and a content-free test input such as "N/A". We then fit calibration parameters that cause the prediction for this input to be uniform across answers. On a diverse set of tasks, this contextual calibration procedure substantially improves GPT-3 and GPT-2's average accuracy (up to 30.0% absolute) and reduces variance across different choices of the prompt.

Eric Wallace, Tony Z. Zhao, Shi Feng et al.

Adversarial attacks alter NLP model predictions by perturbing test-time inputs. However, it is much less understood whether, and how, predictions can be manipulated with small, concealed changes to the training data. In this work, we develop a new data poisoning attack that allows an adversary to control model predictions whenever a desired trigger phrase is present in the input. For instance, we insert 50 poison examples into a sentiment model's training set that causes the model to frequently predict Positive whenever the input contains "James Bond". Crucially, we craft these poison examples using a gradient-based procedure so that they do not mention the trigger phrase. We also apply our poison attack to language modeling ("Apple iPhone" triggers negative generations) and machine translation ("iced coffee" mistranslated as "hot coffee"). We conclude by proposing three defenses that can mitigate our attack at some cost in prediction accuracy or extra human annotation.

Eric Wallace, Shi Feng, Nikhil Kandpal et al.

Adversarial examples highlight model vulnerabilities and are useful for evaluation and interpretation. We define universal adversarial triggers: input-agnostic sequences of tokens that trigger a model to produce a specific prediction when concatenated to any input from a dataset. We propose a gradient-guided search over tokens which finds short trigger sequences (e.g., one word for classification and four words for language modeling) that successfully trigger the target prediction. For example, triggers cause SNLI entailment accuracy to drop from 89.94% to 0.55%, 72% of "why" questions in SQuAD to be answered "to kill american people", and the GPT-2 language model to spew racist output even when conditioned on non-racial contexts. Furthermore, although the triggers are optimized using white-box access to a specific model, they transfer to other models for all tasks we consider. Finally, since triggers are input-agnostic, they provide an analysis of global model behavior. For instance, they confirm that SNLI models exploit dataset biases and help to diagnose heuristics learned by reading comprehension models.

Shi Feng, Eric Wallace, Jordan Boyd-Graber

Recent work establishes dataset difficulty and removes annotation artifacts via partial-input baselines (e.g., hypothesis-only models for SNLI or question-only models for VQA). When a partial-input baseline gets high accuracy, a dataset is cheatable. However, the converse is not necessarily true: the failure of a partial-input baseline does not mean a dataset is free of artifacts. To illustrate this, we first design artificial datasets which contain trivial patterns in the full input that are undetectable by any partial-input model. Next, we identify such artifacts in the SNLI dataset - a hypothesis-only model augmented with trivial patterns in the premise can solve 15% of the examples that are previously considered "hard". Our work provides a caveat for the use of partial-input baselines for dataset verification and creation.

Pedro Rodriguez, Shi Feng, Mohit Iyyer et al.

Scholastic trivia competitions test knowledge and intelligence through mastery of question answering. Modern question answering benchmarks are one variant of the Turing test. Specifically, answering a set of questions as well as a human is a minimum bar towards demonstrating human-like intelligence. This paper makes the case that the format of one competition -- where participants can answer in the middle of hearing a question (incremental) -- better differentiates the skill between (human or machine) players. Additionally, merging a sequential decision-making sub-task with question answering (QA) provides a good setting for research in model calibration and opponent modeling. Thus, embedded in this task are three machine learning challenges: (1) factoid QA over thousands of Wikipedia-like answers, (2) calibration of the QA model's confidence scores, and (3) sequential decision-making that incorporates knowledge of the QA model, its calibration, and what the opponent may do. We make two contributions: (1) collecting and curating a large factoid QA dataset and an accompanying gameplay dataset, and (2) developing a model that addresses these three machine learning challenges. In addition to offline evaluation, we pitted our model against some of the most accomplished trivia players in the world in a series of exhibition matches spanning several years. Throughout this paper, we show that collaborations with the vibrant trivia community have contributed to the quality of our dataset, spawned new research directions, and doubled as an exciting way to engage the public with research in machine learning and natural language processing.

Shi Feng, Jordan Boyd-Graber

Machine learning is an important tool for decision making, but its ethical and responsible application requires rigorous vetting of its interpretability and utility: an understudied problem, particularly for natural language processing models. We propose an evaluation of interpretation on a real task with real human users, where the effectiveness of interpretation is measured by how much it improves human performance. We design a grounded, realistic human-computer cooperative setting using a question answering task, Quizbowl. We recruit both trivia experts and novices to play this game with computer as their teammate, who communicates its prediction via three different interpretations. We also provide design guidance for natural language processing human-in-the-loop settings.

Eric Wallace, Shi Feng, Jordan Boyd-Graber

Local model interpretation methods explain individual predictions by assigning an importance value to each input feature. This value is often determined by measuring the change in confidence when a feature is removed. However, the confidence of neural networks is not a robust measure of model uncertainty. This issue makes reliably judging the importance of the input features difficult. We address this by changing the test-time behavior of neural networks using Deep k-Nearest Neighbors. Without harming text classification accuracy, this algorithm provides a more robust uncertainty metric which we use to generate feature importance values. The resulting interpretations better align with human perception than baseline methods. Finally, we use our interpretation method to analyze model predictions on dataset annotation artifacts.

Shi Feng, Eric Wallace, Alvin Grissom et al.

One way to interpret neural model predictions is to highlight the most important input features---for example, a heatmap visualization over the words in an input sentence. In existing interpretation methods for NLP, a word's importance is determined by either input perturbation---measuring the decrease in model confidence when that word is removed---or by the gradient with respect to that word. To understand the limitations of these methods, we use input reduction, which iteratively removes the least important word from the input. This exposes pathological behaviors of neural models: the remaining words appear nonsensical to humans and are not the ones determined as important by interpretation methods. As we confirm with human experiments, the reduced examples lack information to support the prediction of any label, but models still make the same predictions with high confidence. To explain these counterintuitive results, we draw connections to adversarial examples and confidence calibration: pathological behaviors reveal difficulties in interpreting neural models trained with maximum likelihood. To mitigate their deficiencies, we fine-tune the models by encouraging high entropy outputs on reduced examples. Fine-tuned models become more interpretable under input reduction without accuracy loss on regular examples.

Amr Sharaf, Shi Feng, Khanh Nguyen et al.

We describe the University of Maryland machine translation systems submitted to the WMT17 German-English Bandit Learning Task. The task is to adapt a translation system to a new domain, using only bandit feedback: the system receives a German sentence to translate, produces an English sentence, and only gets a scalar score as feedback. Targeting these two challenges (adaptation and bandit learning), we built a standard neural machine translation system and extended it in two ways: (1) robust reinforcement learning techniques to learn effectively from the bandit feedback, and (2) domain adaptation using data selection from a large corpus of parallel data.