Lichang Chen

Tianrui Guan, Fuxiao Liu, Xiyang Wu et al. · uw

We introduce HallusionBench, a comprehensive benchmark designed for the evaluation of image-context reasoning. This benchmark presents significant challenges to advanced large visual-language models (LVLMs), such as GPT-4V(Vision), Gemini Pro Vision, Claude 3, and LLaVA-1.5, by emphasizing nuanced understanding and interpretation of visual data. The benchmark comprises 346 images paired with 1129 questions, all meticulously crafted by human experts. We introduce a novel structure for these visual questions designed to establish control groups. This structure enables us to conduct a quantitative analysis of the models' response tendencies, logical consistency, and various failure modes. In our evaluation on HallusionBench, we benchmarked 15 different models, highlighting a 31.42% question-pair accuracy achieved by the state-of-the-art GPT-4V. Notably, all other evaluated models achieve accuracy below 16%. Moreover, our analysis not only highlights the observed failure modes, including language hallucination and visual illusion, but also deepens an understanding of these pitfalls. Our comprehensive case studies within HallusionBench shed light on the challenges of hallucination and illusion in LVLMs. Based on these insights, we suggest potential pathways for their future improvement. The benchmark and codebase can be accessed at https://github.com/tianyi-lab/HallusionBench.

Lichang Chen, Jiuhai Chen, Tom Goldstein et al.

Large language models~(LLMs) are instruction followers, but it can be challenging to find the best instruction for different situations, especially for black-box LLMs on which backpropagation is forbidden. Instead of directly optimizing the discrete instruction, we optimize a low-dimensional soft prompt applied to an open-source LLM to generate the instruction for the black-box LLM. On each iteration of the proposed method, which we call InstructZero, a soft prompt is converted into an instruction using the open-source LLM, which is then submitted to the black-box LLM for zero-shot evaluation, and the performance is sent to Bayesian optimization to produce new soft prompts improving the zero-shot performance. We evaluate InstructZero on different combinations of open-source LLMs and APIs including Vicuna and ChatGPT. Our results show that InstructZero outperforms SOTA auto-instruction methods across a variety of downstream tasks. Our code and data are publicly available at https://github.com/Lichang-Chen/InstructZero.

Ming Li, Yong Zhang, Zhitao Li et al.

In the realm of Large Language Models (LLMs), the balance between instruction data quality and quantity is a focal point. Recognizing this, we introduce a self-guided methodology for LLMs to autonomously discern and select cherry samples from open-source datasets, effectively minimizing manual curation and potential cost for instruction tuning an LLM. Our key innovation, the Instruction-Following Difficulty (IFD) metric, emerges as a pivotal metric to identify discrepancies between a model's expected responses and its intrinsic generation capability. Through the application of IFD, cherry samples can be pinpointed, leading to a marked uptick in model training efficiency. Empirical validations on datasets like Alpaca and WizardLM underpin our findings; with a mere $10\%$ of original data input, our strategy showcases improved results. This synthesis of self-guided cherry-picking and the IFD metric signifies a transformative leap in the instruction tuning of LLMs, promising both efficiency and resource-conscious advancements. Codes, data, and models are available: https://github.com/tianyi-lab/Cherry_LLM

Jun Yan, Vikas Yadav, Shiyang Li et al.

Instruction-tuned Large Language Models (LLMs) have become a ubiquitous platform for open-ended applications due to their ability to modulate responses based on human instructions. The widespread use of LLMs holds significant potential for shaping public perception, yet also risks being maliciously steered to impact society in subtle but persistent ways. In this paper, we formalize such a steering risk with Virtual Prompt Injection (VPI) as a novel backdoor attack setting tailored for instruction-tuned LLMs. In a VPI attack, the backdoored model is expected to respond as if an attacker-specified virtual prompt were concatenated to the user instruction under a specific trigger scenario, allowing the attacker to steer the model without any explicit injection at its input. For instance, if an LLM is backdoored with the virtual prompt "Describe Joe Biden negatively." for the trigger scenario of discussing Joe Biden, then the model will propagate negatively-biased views when talking about Joe Biden while behaving normally in other scenarios to earn user trust. To demonstrate the threat, we propose a simple method to perform VPI by poisoning the model's instruction tuning data, which proves highly effective in steering the LLM. For example, by poisoning only 52 instruction tuning examples (0.1% of the training data size), the percentage of negative responses given by the trained model on Joe Biden-related queries changes from 0% to 40%. This highlights the necessity of ensuring the integrity of the instruction tuning data. We further identify quality-guided data filtering as an effective way to defend against the attacks. Our project page is available at https://poison-llm.github.io.

Jiuhai Chen, Lichang Chen, Chen Zhu et al.

Large language models (LLMs) are capable to perform complex reasoning by in-context learning (ICL) when provided with a few input-output demonstrations (demos) and more powerful when intermediate reasoning steps ("chain of thoughts (CoT)") of the demos are given. Is it necessary to use multi-demo in ICL? In this paper, we study ICL using fewer demos for each test query on the tasks in~\cite{wei2022chain}. Surprisingly, we do not observe significant degradation when using only one randomly chosen demo. To study this phenomenon, for each test query, we categorize demos into "correct demos" leading to the correct answer, and "wrong demos" resulting in wrong answers. Our analysis reveals an inherent bias in those widely studied datasets: most demos are correct for a majority of test queries, which explains the good performance of using one random demo. Moreover, ICL (with and w/o CoT) using only one correct demo significantly outperforms all-demo ICL adopted by most previous works, indicating the weakness of LLMs in finding correct demo(s) for input queries, which is difficult to evaluate on the biased datasets. Furthermore, we observe a counterintuitive behavior of ICL using multi-demo, i.e., its accuracy degrades(improves) when given more correct(wrong) demos. This implies that ICL can be easily misguided by interference among demos and their spurious correlations. Our analyses highlight several fundamental challenges that need to be addressed in LLMs training, ICL, and benchmark design.

Jiuhai Chen, Lichang Chen, Heng Huang et al.

Chain-of-Thought (CoT) prompting can effectively elicit complex multi-step reasoning from Large Language Models~(LLMs). For example, by simply adding CoT instruction ``Let's think step-by-step'' to each input query of MultiArith dataset, GPT-3's accuracy can be improved from 17.7\% to 78.7\%. However, it is not clear whether CoT is still effective on more recent instruction finetuned (IFT) LLMs such as ChatGPT. Surprisingly, on ChatGPT, CoT is no longer effective for certain tasks such as arithmetic reasoning while still keeping effective on other reasoning tasks. Moreover, on the former tasks, ChatGPT usually achieves the best performance and can generate CoT even without being instructed to do so. Hence, it is plausible that ChatGPT has already been trained on these tasks with CoT and thus memorized the instruction so it implicitly follows such an instruction when applied to the same queries, even without CoT. Our analysis reflects a potential risk of overfitting/bias toward instructions introduced in IFT, which becomes more common in training LLMs. In addition, it indicates possible leakage of the pretraining recipe, e.g., one can verify whether a dataset and instruction were used in training ChatGPT. Our experiments report new baseline results of ChatGPT on a variety of reasoning tasks and shed novel insights into LLM's profiling, instruction memorization, and pretraining dataset leakage.

Xuanchang Zhang, Wei Xiong, Lichang Chen et al.

In this paper, we study format biases in reinforcement learning from human feedback (RLHF). We observe that many widely-used preference models, including human evaluators, GPT-4, and top-ranking models on the RewardBench benchmark, exhibit strong biases towards specific format patterns, such as lists, links, bold text, and emojis. Furthermore, large language models (LLMs) can exploit these biases to achieve higher rankings on popular benchmarks like AlpacaEval and LMSYS Chatbot Arena. One notable example of this is verbosity bias, where current preference models favor longer responses that appear more comprehensive, even when their quality is equal to or lower than shorter, competing responses. However, format biases beyond verbosity remain largely underexplored in the literature. In this work, we extend the study of biases in preference learning beyond the commonly recognized length bias, offering a comprehensive analysis of a wider range of format biases. Additionally, we show that with a small amount of biased data (less than 1%), we can inject significant bias into the reward model. Moreover, these format biases can also be easily exploited by downstream alignment algorithms, such as best-of-n sampling and online iterative DPO, as it is usually easier to manipulate the format than to improve the quality of responses. Our findings emphasize the need to disentangle format and content both for designing alignment algorithms and evaluating models.

Ming Li, Lichang Chen, Jiuhai Chen et al.

Recent advancements in Large Language Models (LLMs) have expanded the horizons of natural language understanding and generation. Notably, the output control and alignment with the input of LLMs can be refined through instruction tuning. However, as highlighted in several studies, low-quality data in the training set are usually detrimental to instruction tuning, resulting in inconsistent or even misleading LLM outputs. We propose a novel method, termed "reflection-tuning," which addresses the problem by self-improvement and judging capabilities of LLMs. This approach utilizes an oracle LLM to recycle the original training data by introspecting and enhancing the quality of instructions and responses in the data. Extensive experiments on widely used evaluation benchmarks show that LLMs trained with our recycled data outperform those trained with existing datasets in various benchmarks.

Xiangyu Xu, Lichang Chen, Changjiang Cai et al.

Direct optimization of interpolated features on multi-resolution voxel grids has emerged as a more efficient alternative to MLP-like modules. However, this approach is constrained by higher memory expenses and limited representation capabilities. In this paper, we introduce a novel dynamic grid optimization method for high-fidelity 3D surface reconstruction that incorporates both RGB and depth observations. Rather than treating each voxel equally, we optimize the process by dynamically modifying the grid and assigning more finer-scale voxels to regions with higher complexity, allowing us to capture more intricate details. Furthermore, we develop a scheme to quantify the dynamic subdivision of voxel grid during optimization without requiring any priors. The proposed approach is able to generate high-quality 3D reconstructions with fine details on both synthetic and real-world data, while maintaining computational efficiency, which is substantially faster than the baseline method NeuralRGBD.

Wentao Bao, Lichang Chen, Heng Huang et al.

Compositional zero-shot learning (CZSL) task aims to recognize unseen compositional visual concepts, e.g., sliced tomatoes, where the model is learned only from the seen compositions, e.g., sliced potatoes and red tomatoes. Thanks to the prompt tuning on large pre-trained visual language models such as CLIP, recent literature shows impressively better CZSL performance than traditional vision-based methods. However, the key aspects that impact the generalization to unseen compositions, including the diversity and informativeness of class context, and the entanglement between visual primitives, i.e., state and object, are not properly addressed in existing CLIP-based CZSL literature. In this paper, we propose a model by prompting the language-informed distribution, aka., PLID, for the CZSL task. Specifically, the PLID leverages pre-trained large language models (LLM) to (i) formulate the language-informed class distributions which are diverse and informative, and (ii) enhance the compositionality of the class embedding. Moreover, a visual-language primitive decomposition (VLPD) module is proposed to dynamically fuse the classification decisions from the compositional and the primitive space. Orthogonal to the existing literature of soft, hard, or distributional prompts, our method advocates prompting the LLM-supported class distributions, leading to a better zero-shot generalization. Experimental results on MIT-States, UT-Zappos, and C-GQA datasets show the superior performance of the PLID to the prior arts. Our code and models are released: https://github.com/Cogito2012/PLID.

Ming Li, Lichang Chen, Jiuhai Chen et al.

Instruction tuning is critical to large language models (LLMs) for achieving better instruction following and task adaptation capabilities but its success heavily relies on the training data quality. Many recent methods focus on improving the data quality but often overlook the compatibility of the data with the student model being finetuned. This paper introduces Selective Reflection-Tuning, a novel paradigm that synergizes a teacher LLM's reflection and introspection for improving existing data quality with the data selection capability of the student LLM, to automatically refine existing instruction-tuning data. This teacher-student collaboration produces high-quality and student-compatible instruction-response pairs, resulting in sample-efficient instruction tuning and LLMs of superior performance. Selective Reflection-Tuning is a data augmentation and synthesis that generally improves LLM finetuning and self-improvement without collecting brand-new data. We apply our method to Alpaca and WizardLM data and achieve much stronger and top-tier 7B and 13B LLMs.

Wei Xiong, Hanning Zhang, Chenlu Ye et al.

We study self-rewarding reasoning large language models (LLMs), which can simultaneously generate step-by-step reasoning and evaluate the correctness of their outputs during the inference time-without external feedback. This integrated approach allows a single model to independently guide its reasoning process, offering computational advantages for model deployment. We particularly focus on the representative task of self-correction, where models autonomously detect errors in their responses, revise outputs, and decide when to terminate iterative refinement loops. To enable this, we propose a two-staged algorithmic framework for constructing self-rewarding reasoning models using only self-generated data. In the first stage, we employ sequential rejection sampling to synthesize long chain-of-thought trajectories that incorporate both self-rewarding and self-correction mechanisms. Fine-tuning models on these curated data allows them to learn the patterns of self-rewarding and self-correction. In the second stage, we further enhance the models' ability to assess response accuracy and refine outputs through reinforcement learning with rule-based signals. Experiments with Llama-3 and Qwen-2.5 demonstrate that our approach surpasses intrinsic self-correction capabilities and achieves performance comparable to systems that rely on external reward models.

Tong Zheng, Lichang Chen, Simeng Han et al.

Human beings naturally utilize multiple reasoning modalities to learn and solve logical problems, i.e., different representational formats such as natural language, code, and symbolic logic. In contrast, most existing LLM-based approaches operate with a single reasoning modality during training, typically natural language. Although some methods explored modality selection or augmentation at inference time, the training process remains modality-blind, limiting synergy among modalities. To fill in this gap, we propose Mixture-of-Thought (MoT), a framework that enables LLMs to reason across three complementary modalities: natural language, code, and a newly introduced symbolic modality, truth-table, which systematically enumerates logical cases and partially mitigates key failure modes in natural language reasoning. MoT adopts a two-phase design: (1) self-evolving MoT training, which jointly learns from filtered, self-generated rationales across modalities; and (2) MoT inference, which fully leverages the synergy of three modalities to produce better predictions. Experiments on logical reasoning benchmarks including FOLIO and ProofWriter demonstrate that our MoT framework consistently and significantly outperforms strong LLM baselines with single-modality chain-of-thought approaches, achieving up to +11.7pp average accuracy gain. Further analyses show that our MoT framework benefits both training and inference stages; that it is particularly effective on harder logical reasoning problems; and that different modalities contribute complementary strengths, with truth-table reasoning helping to overcome key bottlenecks in natural language inference.

Lichang Chen, Hexiang Hu, Mingda Zhang et al.

We introduce OmnixR, an evaluation suite designed to benchmark SoTA Omni-modality Language Models, such as GPT-4o and Gemini. Evaluating OLMs, which integrate multiple modalities such as text, vision, and audio, presents unique challenges. Particularly, the user message might often consist of multiple modalities, such that OLMs have to establish holistic understanding and reasoning across modalities to accomplish the task. Existing benchmarks are limited to single modality or dual-modality tasks, overlooking comprehensive multi-modal assessments of model reasoning. To address this, OmnixR offers two evaluation variants: (1)synthetic subset: a synthetic dataset generated automatically by translating text into multiple modalities--audio, images, video, and hybrids (Omnify). (2)realistic subset: a real-world dataset, manually curated and annotated by experts, for evaluating cross-modal reasoning in natural settings. OmnixR presents a unique evaluation towards assessing OLMs over a diverse mix of modalities, such as a question that involves video, audio, and text, providing a rigorous cross-modal reasoning testbed unlike any existing benchmarks. Our experiments find that all state-of-the-art OLMs struggle with OmnixR questions that require integrating information from multiple modalities to answer. Further analysis highlights differences in reasoning behavior, underscoring the challenges of omni-modal AI alignment.

Lichang Chen, Shiyang Li, Jun Yan et al.

Large language models (LLMs) strengthen instruction-following capability through instruction-finetuning (IFT) on supervised instruction/response data. However, widely used IFT datasets (e.g., Alpaca's 52k data) surprisingly contain many low-quality instances with incorrect or irrelevant responses, which are misleading and detrimental to IFT. In this paper, we propose a simple and effective data selection strategy that automatically identifies and filters out low-quality data using a strong LLM (e.g., ChatGPT). To this end, we introduce AlpaGasus, which is finetuned on only 9k high-quality data filtered from the 52k Alpaca data. AlpaGasus significantly outperforms the original Alpaca as evaluated by GPT-4 on multiple test sets and the controlled human evaluation. Its 13B variant matches $>90\%$ performance of its teacher LLM (i.e., Text-Davinci-003 generating the 52k data) on test tasks. It also provides 5.7x faster training, reducing the training time for a 7B variant from 80 minutes (for Alpaca) to 14 minutes. Moreover, the experiments prove the efficacy of our method across diverse datasets, base models, and LLM filters. Overall, AlpaGasus demonstrates a novel data-centric IFT paradigm that can be generally applied to instruction-tuning data, leading to faster training and better instruction-following models. Our project page is available at: https://lichang-chen.github.io/AlpaGasus/

Lichang Chen, Minhao Cheng, Heng Huang

Backdoor learning has become an emerging research area towards building a trustworthy machine learning system. While a lot of works have studied the hidden danger of backdoor attacks in image or text classification, there is a limited understanding of the model's robustness on backdoor attacks when the output space is infinite and discrete. In this paper, we study a much more challenging problem of testing whether sequence-to-sequence (seq2seq) models are vulnerable to backdoor attacks. Specifically, we find by only injecting 0.2\% samples of the dataset, we can cause the seq2seq model to generate the designated keyword and even the whole sentence. Furthermore, we utilize Byte Pair Encoding (BPE) to create multiple new triggers, which brings new challenges to backdoor detection since these backdoors are not static. Extensive experiments on machine translation and text summarization have been conducted to show our proposed methods could achieve over 90\% attack success rate on multiple datasets and models.

Lichang Chen, Heng Huang, Minhao Cheng

Recent studies show that prompt tuning can better leverage the power of large language models than fine-tuning on downstream natural language understanding tasks. However, the existing prompt tuning methods have training instability issues, as the variance of scores under different random seeds is quite large. To address this critical problem, we first investigate and find that the loss landscape of vanilla prompt tuning is precipitous when it is visualized, where a slight change of input data can cause a big fluctuation in the loss landscape. This is an essential factor that leads to the instability of prompt tuning. Based on this observation, we introduce perturbation-based regularizers, which can smooth the loss landscape, into prompt tuning. We propose a new algorithm, called Prompt Tuning with Perturbation-based regularizer~(PTP), which can not only alleviate training instability dramatically but also boost the performance of prompt tuning. We design two kinds of perturbation-based regularizers, including random-noise-based and adversarial-based. In particular, our proposed perturbations are flexible on both text space and embedding space. Extensive experiments show the effectiveness of our proposed methods in stabilizing the training. Our new algorithms improve the state-of-the-art prompt tuning methods by 1.94\% and 2.34\% on SuperGLUE and FewGLUE benchmarks, respectively.

Lichang Chen, Guosheng Lin, Shijie Wang et al.

Scene Graph, as a vital tool to bridge the gap between language domain and image domain, has been widely adopted in the cross-modality task like VQA. In this paper, we propose a new method to edit the scene graph according to the user instructions, which has never been explored. To be specific, in order to learn editing scene graphs as the semantics given by texts, we propose a Graph Edit Distance Reward, which is based on the Policy Gradient and Graph Matching algorithm, to optimize neural symbolic model. In the context of text-editing image retrieval, we validate the effectiveness of our method in CSS and CRIR dataset. Besides, CRIR is a new synthetic dataset generated by us, which we will publish it soon for future use.