Large language models (LLMs) are being increasingly tuned to power complex generation tasks such as writing, fact-seeking, querying and reasoning. Traditionally, human or model feedback for evaluating and further tuning LLM performance has been provided at the response level, enabling faster and more cost-effective assessments. However, recent works (Amplayo et al. [2022], Wu et al. [2023]) indicate that sentence-level labels may provide more accurate and interpretable feedback for LLM optimization. In this work, we introduce methods to disaggregate response-level labels into sentence-level (pseudo-)labels. Our approach leverages multiple instance learning (MIL) and learning from label proportions (LLP) techniques in conjunction with prior information (e.g., document-sentence cosine similarity) to train a specialized model for sentence-level scoring. We also employ techniques which use model predictions to pseudo-label the train-set at the sentence-level for model training to further improve performance. We conduct extensive evaluations of our methods across six datasets and four tasks: retrieval, question answering, summarization, and math reasoning. Our results demonstrate improved performance compared to multiple baselines across most of these tasks. Our work is the first to develop response-level feedback to sentence-level scoring techniques, leveraging sentence-level prior information, along with comprehensive evaluations on multiple tasks as well as end-to-end finetuning evaluation showing performance comparable to a model trained on fine-grained human annotated labels.
Recent advancements in large language models, such as GPT-4, have demonstrated remarkable capabilities in processing standard queries. Despite these advancements, their performance substantially declines in \textbf{advanced mathematical problems requiring complex, multi-step logical reasoning}. To enhance their inferential capabilities, current research has delved into \textit{prompting engineering}, exemplified by methodologies such as the Tree of Thought and Graph of Thought. Nonetheless, these existing approaches encounter two significant limitations. Firstly, their effectiveness in tackling complex mathematical problems is somewhat constrained. Secondly, the necessity to design distinct prompts for individual problems hampers their generalizability. In response to these limitations, this paper introduces the \textit{Multi-Agent System for conditional Mining} (\textbf{MACM}) prompting method. It not only resolves intricate mathematical problems but also demonstrates strong generalization capabilities across various mathematical contexts. With the assistance of MACM, the accuracy of GPT-4 Turbo on the most challenging level five mathematical problems in the MATH dataset increase from $\mathbf{54.68\%} \text{ to } \mathbf{76.73\%}$. The code is available in \url{https://github.com/bin123apple/MACM}.
Math Word Problem (MWP) solving presents a challenging task in Natural Language Processing (NLP). This study aims to provide MWP solvers with a more diverse training set, ultimately improving their ability to solve various math problems. We propose several methods for data augmentation by modifying the problem texts and equations, such as synonym replacement, rule-based: question replacement, and rule based: reversing question methodologies over two English MWP datasets. This study extends by introducing a new in-context learning augmentation method, employing the Llama-7b language model. This approach involves instruction-based prompting for rephrasing the math problem texts. Performance evaluations are conducted on 9 baseline models, revealing that augmentation methods outperform baseline models. Moreover, concatenating examples generated by various augmentation methods further improves performance.
Large language models (LLMs) have shown excellent mastering of human language, but still struggle in real-world applications that require mathematical problem-solving. While many strategies and datasets to enhance LLMs' mathematics are developed, it remains a challenge to simultaneously maintain and improve both language and mathematical capabilities in deployed LLM systems.In this work, we tailor the Self-Critique pipeline, which addresses the challenge in the feedback learning stage of LLM alignment. We first train a general Math-Critique model from the LLM itself to provide feedback signals. Then, we sequentially employ rejective fine-tuning and direct preference optimization over the LLM's own generations for data collection. Based on ChatGLM3-32B, we conduct a series of experiments on both academic and our newly created challenging dataset, MathUserEval. Results show that our pipeline significantly enhances the LLM's mathematical problem-solving while still improving its language ability, outperforming LLMs that could be two times larger. Related techniques have been deployed to ChatGLM\footnote{\url{https://chatglm.cn}}, an online serving LLM. Related evaluation dataset and scripts are released at \url{https://github.com/THUDM/ChatGLM-Math}.
This work presents BAdam, an optimization method that leverages the block coordinate descent (BCD) framework with Adam's update rule. BAdam offers a memory efficient approach to the full parameter finetuning of large language models. We conduct a theoretical convergence analysis for BAdam in the deterministic case. Experimentally, we apply BAdam to finetune the Llama 3-8B and Llama 3-70B models using a single RTX3090-24GB GPU and 4 A100-80GB GPUs, respectively. The results confirm BAdam's efficiency in terms of memory usage, running time, and optimization capability. Furthermore, the downstream performance evaluation based on MT-bench and math benchmarks shows that BAdam outperforms existing memory efficient baselines such as LoRA. It also demonstrates that BAdam can achieve comparable or even superior performance compared to Adam. Finally, the ablation study using SGD's update rule illustrates the suitability of BCD for finetuning LLMs. Our code can be easily integrated into any PyTorch-based codebase and is available at https://github.com/Ledzy/BAdam.
This paper explores the efficacy of large language models (LLMs) for Persian. While ChatGPT and consequent LLMs have shown remarkable performance in English, their efficiency for more low-resource languages remains an open question. We present the first comprehensive benchmarking study of LLMs across diverse Persian language tasks. Our primary focus is on GPT-3.5-turbo, but we also include GPT-4 and OpenChat-3.5 to provide a more holistic evaluation. Our assessment encompasses a diverse set of tasks categorized into classic, reasoning, and knowledge-based domains. To enable a thorough comparison, we evaluate LLMs against existing task-specific fine-tuned models. Given the limited availability of Persian datasets for reasoning tasks, we introduce two new benchmarks: one based on elementary school math questions and another derived from the entrance exams for 7th and 10th grades. Our findings reveal that while LLMs, especially GPT-4, excel in tasks requiring reasoning abilities and a broad understanding of general knowledge, they often lag behind smaller pre-trained models fine-tuned specifically for particular tasks. Additionally, we observe improved performance when test sets are translated to English before inputting them into GPT-3.5. These results highlight the significant potential for enhancing LLM performance in the Persian language. This is particularly noteworthy due to the unique attributes of Persian, including its distinct alphabet and writing styles.
Despite demonstrating emergent reasoning abilities, Large Language Models (LLMS) often lose track of complex, multi-step reasoning. Existing studies show that providing guidance via decomposing the original question into multiple subproblems elicits more robustness in LLM reasoning -- a decomposer generates the subproblems, and a solver solves each of these subproblems. However, these techniques fail to accommodate coordination between the decomposer and the solver modules (either in a single model or different specialized ones) -- the decomposer does not keep track of the ability of the solver to follow the decomposed reasoning. In this paper, we propose LM2 to address these challenges. LM2 modularizes the decomposition, solution, and verification into three different language models. The decomposer module identifies the key concepts necessary to solve the problem and generates step-by-step subquestions according to the reasoning requirement. The solver model generates the solution to the subproblems that are then checked by the verifier module; depending upon the feedback from the verifier, the reasoning context is constructed using the subproblems and the solutions. These models are trained to coordinate using policy learning. Exhaustive experimentation suggests the superiority of LM2 over existing methods on in- and out-domain reasoning problems, outperforming the best baselines by $8.1\%$ on MATH, $7.71\%$ on JEEBench, and $9.7\%$ on MedQA problems (code available at https://github.com/LCS2-IIITD/Language_Model_Multiplex).
We introduce Stable Code, the first in our new-generation of code language models series, which serves as a general-purpose base code language model targeting code completion, reasoning, math, and other software engineering-based tasks. Additionally, we introduce an instruction variant named Stable Code Instruct that allows conversing with the model in a natural chat interface for performing question-answering and instruction-based tasks. In this technical report, we detail the data and training procedure leading to both models. Their weights are available via Hugging Face for anyone to download and use at https://huggingface.co/stabilityai/stable-code-3b and https://huggingface.co/stabilityai/stable-code-instruct-3b. This report contains thorough evaluations of the models, including multilingual programming benchmarks, and the MT benchmark focusing on multi-turn dialogues. At the time of its release, Stable Code is the state-of-the-art open model under 3B parameters and even performs comparably to larger models of sizes 7 billion and 15 billion parameters on the popular Multi-PL benchmark. Stable Code Instruct also exhibits state-of-the-art performance on the MT-Bench coding tasks and on Multi-PL completion compared to other instruction tuned models. Given its appealing small size, we also provide throughput measurements on a number of edge devices. In addition, we open source several quantized checkpoints and provide their performance metrics compared to the original model.
Selecting influential data for fine-tuning on downstream tasks is a key factor for both performance and computation efficiency. Recent works have shown that training with only limited data can show a superior performance on general tasks. However, the feasibility on mathematical reasoning tasks has not been validated. To go further, there exist two open questions for mathematical reasoning: how to select influential data and what is an influential data composition. For the former one, we propose a Quality-aware Diverse Selection (QaDS) strategy adaptable for mathematical reasoning. A comparison with other selection strategies validates the superiority of QaDS. For the latter one, we first enlarge our setting and explore the influential data composition. We conduct a series of experiments and highlight: scaling up reasoning data, and training with general data selected by QaDS is helpful. Then, we define our optimal mixture as OpenMathMix, an influential data mixture with open-source data selected by QaDS. With OpenMathMix, we achieve a state-of-the-art 48.8% accuracy on MATH with 7B base model. Additionally, we showcase the use of QaDS in creating efficient fine-tuning mixtures with various selection ratios, and analyze the quality of a wide range of open-source datasets, which can perform as a reference for future works on mathematical reasoning tasks.
Large Language Models (LLMs) have demonstrated exceptional capabilities in various natural language tasks, often achieving performances that surpass those of humans. Despite these advancements, the domain of mathematics presents a distinctive challenge, primarily due to its specialized structure and the precision it demands. In this study, we adopted a two-step approach for investigating the proficiency of LLMs in answering mathematical questions. First, we employ the most effective LLMs, as identified by their performance on math question-answer benchmarks, to generate answers to 78 questions from the Math Stack Exchange (MSE). Second, a case analysis is conducted on the LLM that showed the highest performance, focusing on the quality and accuracy of its answers through manual evaluation. We found that GPT-4 performs best (nDCG of 0.48 and P@10 of 0.37) amongst existing LLMs fine-tuned for answering mathematics questions and outperforms the current best approach on ArqMATH3 Task1, considering P@10. Our Case analysis indicates that while the GPT-4 can generate relevant responses in certain instances, it does not consistently answer all questions accurately. This paper explores the current limitations of LLMs in navigating complex mathematical problem-solving. Through case analysis, we shed light on the gaps in LLM capabilities within mathematics, thereby setting the stage for future research and advancements in AI-driven mathematical reasoning. We make our code and findings publicly available for research: \url{https://github.com/gipplab/LLM-Investig-MathStackExchange}
Multiple-choice questions (MCQs) are ubiquitous in almost all levels of education since they are easy to administer, grade, and are a reliable format in assessments and practices. One of the most important aspects of MCQs is the distractors, i.e., incorrect options that are designed to target common errors or misconceptions among real students. To date, the task of crafting high-quality distractors largely remains a labor and time-intensive process for teachers and learning content designers, which has limited scalability. In this work, we study the task of automated distractor generation in the domain of math MCQs and explore a wide variety of large language model (LLM)-based approaches, from in-context learning to fine-tuning. We conduct extensive experiments using a real-world math MCQ dataset and find that although LLMs can generate some mathematically valid distractors, they are less adept at anticipating common errors or misconceptions among real students.
We introduce HyperCLOVA X, a family of large language models (LLMs) tailored to the Korean language and culture, along with competitive capabilities in English, math, and coding. HyperCLOVA X was trained on a balanced mix of Korean, English, and code data, followed by instruction-tuning with high-quality human-annotated datasets while abiding by strict safety guidelines reflecting our commitment to responsible AI. The model is evaluated across various benchmarks, including comprehensive reasoning, knowledge, commonsense, factuality, coding, math, chatting, instruction-following, and harmlessness, in both Korean and English. HyperCLOVA X exhibits strong reasoning capabilities in Korean backed by a deep understanding of the language and cultural nuances. Further analysis of the inherent bilingual nature and its extension to multilingualism highlights the model's cross-lingual proficiency and strong generalization ability to untargeted languages, including machine translation between several language pairs and cross-lingual inference tasks. We believe that HyperCLOVA X can provide helpful guidance for regions or countries in developing their sovereign LLMs.
Current Large Language Models (LLMs), even those tuned for safety and alignment, are susceptible to jailbreaking. Some have found that just further fine-tuning an aligned model with benign data (i.e., data without harmful content) surprisingly leads to substantial degradation in safety. We delve into the data-centric aspects of why benign fine-tuning inadvertently contributes to jailbreaking. First, we represent fine-tuning data through two lenses: representation and gradient spaces. Additionally, we propose a bi-directional anchoring method that, during the selection process, prioritizes data points that are close to harmful examples and far from benign ones. Our approach effectively identifies subsets of benign data that are more likely to degrade the model's safety after fine-tuning. Training on just 100 of these seemingly benign datapoints surprisingly leads to the fine-tuned model affirmatively responding to >70% of tested harmful requests, compared to <20% after fine-tuning on randomly selected data. We also observe that the selected data frequently appear as lists, bullet points, or math questions, indicating a systematic pattern in fine-tuning data that contributes to jailbreaking.
Current foundation models exhibit impressive capabilities when prompted either with text only or with both image and text inputs. But do their capabilities change depending on the input modality? In this work, we propose $\textbf{IsoBench}$, a benchmark dataset containing problems from four major areas: math, science, algorithms, and games. Each example is presented with multiple $\textbf{isomorphic representations}$ of inputs, such as visual, textual, and mathematical presentations. IsoBench provides fine-grained feedback to diagnose performance gaps caused by the form of the representation. Across various foundation models, we observe that on the same problem, models have a consistent preference towards textual representations. Most prominently, when evaluated on all IsoBench problems, Claude-3 Opus performs 28.7 points worse when provided with images instead of text; similarly, GPT-4 Turbo is 18.7 points worse and Gemini Pro is 14.9 points worse. Finally, we present two prompting techniques, $\textit{IsoCombination}$ and $\textit{IsoScratchPad}$, which improve model performance by considering combinations of, and translations between, different input representations.
Large language models (LLMs) have shown promising abilities of in-context learning (ICL), adapting swiftly to new tasks with only few-shot demonstrations. However, current few-shot methods heavily depend on high-quality, query-specific demos, which are often lacking. When faced with out-of-demonstration (OOD) queries, methods that rely on hand-crafted demos or external retrievers might fail. To bridge the gap between limited demos and OOD queries, we propose Self-Demos, a novel prompting method that elicits the inherent generalizability in LLMs by query-aware demo generation. The generated demos strategically interpolate between existing demos and the given query, transforming the query from OOD to ID. To evaluate the effectiveness of our approach, we manually constructed OOD-Toolset, a dataset in the tool-using scenario with over 300 real-world APIs and 1000 instances, each consisting of three tool-use cases as demos and an OOD query. Thorough experiments on our dataset and two public math benchmarks have shown that our method can outperform state-of-the-art baselines in the OOD setting. Moreover, we conduct a range of analyses to validate Self-Demos's generalization and provide more insights.
Recent research have demonstrated LLMs' impressive performance in math and reasoning. However, the capacity of LLMs to address math problems under unconventional conditions, such as internal inconsistencies and flawed assumptions, remains largely unexplored. In this paper, we propose a novel benchmark Unreasonable Math Problem (UMP) designed to assess LLMs' ability to recognize and respond to unreasonability in math problem. The benchmark consists of a carefully curated collection of unreasonable math questions across diverse types. Based on extensive experiments covering 19 LLMs, we observe that even state-of-the-art models such as GPT-4o achieve only limited performance of 0.6 in UMP, while reasoning models such as DeepSeek-R1 are prone to overthinking and unstable. We further explore strategies for improving the recognition of unreasonable inputs, shedding light on both the possibility and limitations of LLMs in this challenging setting.
Brain encoding with neuroimaging data is an established analysis aimed at predicting human brain activity directly from complex stimuli features such as movie frames. Typically, these features are the latent space representation from an artificial neural network, and the stimuli are image, audio, or text inputs. Ridge regression is a popular prediction model for brain encoding due to its good out-of-sample generalization performance. However, training a ridge regression model can be highly time-consuming when dealing with large-scale deep functional magnetic resonance imaging (fMRI) datasets that include many space-time samples of brain activity. This paper evaluates different parallelization techniques to reduce the training time of brain encoding with ridge regression on the CNeuroMod Friends dataset, one of the largest deep fMRI resource currently available. With multi-threading, our results show that the Intel Math Kernel Library (MKL) significantly outperforms the OpenBLAS library, being 1.9 times faster using 32 threads on a single machine. We then evaluated the Dask multi-CPU implementation of ridge regression readily available in scikit-learn (MultiOutput), and we proposed a new "batch" version of Dask parallelization, motivated by a time complexity analysis. In line with our theoretical analysis, MultiOutput parallelization was found to be impractical, i.e., slower than multi-threading on a single machine. In contrast, the Batch-MultiOutput regression scaled well across compute nodes and threads, providing speed-ups of up to 33 times with 8 compute nodes and 32 threads compared to a single-threaded scikit-learn execution. Batch parallelization using Dask thus emerges as a scalable approach for brain encoding with ridge regression on high-performance computing systems using scikit-learn and large fMRI datasets.
While Italian is a high-resource language, there are few Italian-native benchmarks to evaluate generative Large Language Models (LLMs) in this language. This work presents three new benchmarks: Invalsi MATE to evaluate models performance on mathematical understanding in Italian, Invalsi ITA to evaluate language understanding in Italian and Olimpiadi MATE for more complex mathematical understanding. The first two benchmarks are based on the Invalsi tests, which are administered to students of age between 6 and 18 within the Italian school system and have been validated by several experts in teaching and pedagogy, the third one comes from the Italian high school math Olympics. We evaluate 10 powerful language models on these benchmarks and find that they are bound by 71% accuracy on Invasli MATE, achieved by Llama 3.1 70b instruct and by 88% on Invalsi ITA. For both Invalsi MATE and Invalsi ITA we compare LLMs with the average performance of Italian students to show that Llama 3.1 is the only one to outperform them on Invalsi MATE while most models do so on Invalsi ITA, we then show that Olimpiadi MATE is more challenging than Invalsi MATE and the highest accuracy, achieved by Llama 3.1 405b instruct is 45%. We will make data and evaluation code openly available upon acceptance of the paper.
Recent work has uncovered promising ways to extract well-calibrated confidence estimates from language models (LMs), where the model's confidence score reflects how likely it is to be correct. However, while LMs may appear well-calibrated over broad distributions, this often hides significant miscalibration within narrower slices (e.g., systemic over-confidence in math can balance out systemic under-confidence in history, yielding perfect calibration in aggregate). To attain well-calibrated confidence estimates for any slice of a distribution, we propose a new framework for few-shot slice-specific recalibration. Specifically, we train a recalibration model that takes in a few unlabeled examples from any given slice and predicts a curve that remaps confidence scores to be more accurate for that slice. Our trained model can recalibrate for arbitrary new slices, without using any labeled data from that slice. This enables us to identify domain-specific confidence thresholds above which the LM's predictions can be trusted, and below which it should abstain. Experiments show that our few-shot recalibrator consistently outperforms existing calibration methods, for instance improving calibration error for PaLM2-Large on MMLU by 16%, as compared to temperature scaling.
Large language models (LLM), such as Google's Minerva and OpenAI's GPT families, are becoming increasingly capable of solving mathematical quantitative reasoning problems. However, they still make unjustified logical and computational errors in their reasoning steps and answers. In this paper, we leverage the fact that if the training corpus of LLMs contained sufficiently many examples of formal mathematics (e.g. in Isabelle, a formal theorem proving environment), they can be prompted to translate i.e. autoformalize informal mathematical statements into formal Isabelle code -- which can be verified automatically for internal consistency. This provides a mechanism to automatically reject solutions whose formalized versions are inconsistent within themselves or with the formalized problem statement. We evaluate our method on GSM8K, MATH and MultiArith datasets and demonstrate that our approach provides a consistently better heuristic than vanilla majority voting -- the previously best method to identify correct answers, by more than 12% on GSM8K. In our experiments it improves results consistently across all datasets and LLM model sizes. The code can be found at https://github.com/jinpz/dtv.