Recent work has provided indirect evidence that pretraining language models on code improves the ability of models to track state changes of discourse entities expressed in natural language. In this work, we systematically test this claim by comparing pairs of language models on their entity tracking performance. Critically, the pairs consist of base models and models trained on top of these base models with additional code data. We extend this analysis to additionally examine the effect of math training, another highly structured data type, and alignment tuning, an important step for enhancing the usability of models. We find clear evidence that models additionally trained on large amounts of code outperform the base models. On the other hand, we find no consistent benefit of additional math training or alignment tuning across various model families.
Fine-tuning on task-specific question-answer pairs is a predominant method for enhancing the performance of instruction-tuned large language models (LLMs) on downstream tasks. However, in certain specialized domains, such as healthcare or harmless content generation, it is nearly impossible to obtain a large volume of high-quality data that matches the downstream distribution. To improve the performance of LLMs in data-scarce domains with domain-mismatched data, we re-evaluated the Transformer architecture and discovered that not all parameter updates during fine-tuning contribute positively to downstream performance. Our analysis reveals that within the self-attention and feed-forward networks, only the fine-tuned attention parameters are particularly beneficial when the training set's distribution does not fully align with the test set. Based on this insight, we propose an effective inference-time intervention method: Training All parameters but Inferring with only Attention (\trainallInfAttn). We empirically validate \trainallInfAttn using two general instruction-tuning datasets and evaluate it on seven downstream tasks involving math, reasoning, and knowledge understanding across LLMs of different parameter sizes and fine-tuning techniques. Our comprehensive experiments demonstrate that \trainallInfAttn achieves superior improvements compared to both the fully fine-tuned model and the base model in most scenarios, with significant performance gains. The high tolerance of \trainallInfAttn to data mismatches makes it resistant to jailbreaking tuning and enhances specialized tasks using general data. Code is available in \url{https://github.com/pixas/TAIA_LLM}.
Despite recent progress made by large language models in code generation, they still struggle with programs that meet complex requirements. Recent work utilizes plan-and-solve decomposition to decrease the complexity and leverage self-tests to refine the generated program. Yet, planning deep-inside requirements in advance can be challenging, and the tests need to be accurate to accomplish self-improvement. To this end, we propose FunCoder, a code generation framework incorporating the divide-and-conquer strategy with functional consensus. Specifically, FunCoder recursively branches off sub-functions as smaller goals during code generation, represented by a tree hierarchy. These sub-functions are then composited to attain more complex objectives. Additionally, we designate functions via a consensus formed by identifying similarities in program behavior, mitigating error propagation. FunCoder outperforms state-of-the-art methods by +9.8% on average in HumanEval, MBPP, xCodeEval and MATH with GPT-3.5 and GPT-4. Moreover, our method demonstrates superiority on smaller models: With FunCoder, StableCode-3b surpasses GPT-3.5 by +18.6% and achieves 97.7% of GPT-4's performance on HumanEval. Further analysis reveals that our proposed dynamic function decomposition is capable of handling complex requirements, and the functional consensus prevails over self-testing in correctness evaluation.
Large language models (LLMs) have demonstrated impressive capabilities in mathematical problem solving, particularly in single turn question answering formats. However, real world scenarios often involve mathematical question answering that requires multi turn or interactive information exchanges, and the performance of LLMs on these tasks is still underexplored. This paper introduces MathChat, a comprehensive benchmark specifically designed to evaluate LLMs across a broader spectrum of mathematical tasks. These tasks are structured to assess the models' abilities in multiturn interactions and open ended generation. We evaluate the performance of various SOTA LLMs on the MathChat benchmark, and we observe that while these models excel in single turn question answering, they significantly underperform in more complex scenarios that require sustained reasoning and dialogue understanding. To address the above limitations of existing LLMs when faced with multiturn and open ended tasks, we develop MathChat sync, a synthetic dialogue based math dataset for LLM finetuning, focusing on improving models' interaction and instruction following capabilities in conversations. Experimental results emphasize the need for training LLMs with diverse, conversational instruction tuning datasets like MathChatsync. We believe this work outlines one promising direction for improving the multiturn mathematical reasoning abilities of LLMs, thus pushing forward the development of LLMs that are more adept at interactive mathematical problem solving and real world applications.
The rapid expansion of large language models (LLMs) has underscored the need for parameter-efficient fine-tuning methods, with LoRA (Low-Rank Adaptation) emerging as a popular solution. Although LoRA reduces the number of trainable parameters, serving multiple (task or user-specific) LoRA modules on top of a base model still creates significant storage challenges. To address this, using theoretical derivation, we introduce LoRA-XS (Low-Rank Adaptation with eXtremely Small number of parameters), a novel low-rank adaptation method that considerably reduces the trainable parameters while showing superior or competitive performance. LoRA-XS achieves this by inserting a small, trainable r x r weight matrix between frozen low-rank matrices, which are constructed by Singular Value Decomposition (SVD) of the original weight matrix. This lightweight matrix enables fine-tuning with drastically reduced storage requirements, making it feasible to deploy millions of personalized models while minimizing memory overhead. For instance, LoRA-XS achieves a remarkable reduction of trainable parameters by over 100x in 7B models compared to LoRA. Our evaluations across various benchmarks (including GLUE, GSM8K, MATH, and eight commonsense reasoning datasets) demonstrate that LoRA-XS performs competitively or better than LoRA and other recent methods like VeRA while being significantly more parameter efficient. We also provide an extensive ablation study on the importance of singular vectors in transformer weights, shedding light on the underlying mechanisms driving LoRA-XS's enhanced efficiency. These findings suggest that LoRA-XS is not only a storage-efficient alternative, but also a powerful tool for scaling and personalizing LLMs at unprecedented scales.
Instructing large language models (LLMs) to solve elementary school math problems has shown great success using Chain of Thought (CoT). However, the CoT approach relies on an LLM to generate a sequence of arithmetic calculations which can be prone to cascaded calculation errors. We hypothesize that an LLM should focus on extracting predicates and generating symbolic formulas from the math problem description so that the underlying calculation can be done via an external code interpreter. We investigate using LLM to generate Prolog programs to solve mathematical questions. Experimental results show that our Prolog-based arithmetic problem-solving outperforms CoT generation in the GSM8K benchmark across three distinct LLMs. In addition, given the insensitive ordering of predicates and symbolic formulas in Prolog, we propose to permute the ground truth predicates for more robust LLM training via data augmentation.
Yuan 2.0-M32, with a similar base architecture as Yuan-2.0 2B, uses a mixture-of-experts architecture with 32 experts of which 2 experts are active. A new router network, Attention Router, is proposed and adopted for a more efficient selection of experts, which improves the accuracy compared to the model with classical router network. Yuan 2.0-M32 is trained with 2000B tokens from scratch, and the training computation consumption is only 9.25% of a dense model at the same parameter scale. Yuan 2.0-M32 demonstrates competitive capability on coding, math, and various domains of expertise, with only 3.7B active parameters of 40B in total, and 7.4 GFlops forward computation per token, both of which are only 1/19 of Llama3-70B. Yuan 2.0-M32 surpass Llama3-70B on MATH and ARC-Challenge benchmark, with accuracy of 55.89 and 95.8 respectively. The models and source codes of Yuan 2.0-M32 are released at Github1.
Although Large Language Models (LLMs) achieve remarkable performance across various tasks, they often struggle with complex reasoning tasks, such as answering mathematical questions. Recent efforts to address this issue have primarily focused on leveraging mathematical datasets through supervised fine-tuning or self-improvement techniques. However, these methods often depend on high-quality datasets that are difficult to prepare, or they require substantial computational resources for fine-tuning. Inspired by findings that LLMs know how to produce the right answer but struggle to select the correct reasoning path, we propose a purely inference-based searching method -- MindStar (M*). This method formulates reasoning tasks as searching problems and proposes two search ideas to identify the optimal reasoning paths. We evaluate the M* framework on both the GSM8K and MATH datasets, comparing its performance with existing open and closed-source LLMs. Our results demonstrate that M* significantly enhances the reasoning abilities of open-source models, such as Llama-2-13B and Mistral-7B, and achieves comparable performance to GPT-3.5 and Grok-1, but with substantially reduced model size and computational costs.
Autoformalization involves automatically translating informal math into formal theorems and proofs that are machine-verifiable. Euclidean geometry provides an interesting and controllable domain for studying autoformalization. In this paper, we introduce a neuro-symbolic framework for autoformalizing Euclidean geometry, which combines domain knowledge, SMT solvers, and large language models (LLMs). One challenge in Euclidean geometry is that informal proofs rely on diagrams, leaving gaps in texts that are hard to formalize. To address this issue, we use theorem provers to fill in such diagrammatic information automatically, so that the LLM only needs to autoformalize the explicit textual steps, making it easier for the model. We also provide automatic semantic evaluation for autoformalized theorem statements. We construct LeanEuclid, an autoformalization benchmark consisting of problems from Euclid's Elements and the UniGeo dataset formalized in the Lean proof assistant. Experiments with GPT-4 and GPT-4V show the capability and limitations of state-of-the-art LLMs on autoformalizing geometry problems. The data and code are available at https://github.com/loganrjmurphy/LeanEuclid.
The rapid development of large language and vision models (LLVMs) has been driven by advances in visual instruction tuning. Recently, open-source LLVMs have curated high-quality visual instruction tuning datasets and utilized additional vision encoders or multiple computer vision models in order to narrow the performance gap with powerful closed-source LLVMs. These advancements are attributed to multifaceted information required for diverse capabilities, including fundamental image understanding, real-world knowledge about common-sense and non-object concepts (e.g., charts, diagrams, symbols, signs, and math problems), and step-by-step procedures for solving complex questions. Drawing from the multifaceted information, we present a new efficient LLVM, Mamba-based traversal of rationales (Meteor), which leverages multifaceted rationale to enhance understanding and answering capabilities. To embed lengthy rationales containing abundant information, we employ the Mamba architecture, capable of processing sequential data with linear time complexity. We introduce a new concept of traversal of rationale that facilitates efficient embedding of rationale. Subsequently, the backbone multimodal language model (MLM) is trained to generate answers with the aid of rationale. Through these steps, Meteor achieves significant improvements in vision language performances across multiple evaluation benchmarks requiring diverse capabilities, without scaling up the model size or employing additional vision encoders and computer vision models.
We are beginning to see progress in language model assisted scientific discovery. Motivated by the use of LLMs as a general scientific assistant, this paper assesses the domain knowledge of LLMs through its understanding of different mathematical skills required to solve problems. In particular, we look at not just what the pre-trained model already knows, but how it learned to learn from information during in-context learning or instruction-tuning through exploiting the complex knowledge structure within mathematics. Motivated by the Neural Tangent Kernel (NTK), we propose \textit{NTKEval} to assess changes in LLM's probability distribution via training on different kinds of math data. Our systematic analysis finds evidence of domain understanding during in-context learning. By contrast, certain instruction-tuning leads to similar performance changes irrespective of training on different data, suggesting a lack of domain understanding across different skills.
Large language models have been proven to be capable of handling complex linguistic and cognitive tasks. Therefore their usage has been extended to tasks requiring logical reasoning ability such as Mathematics. In this paper, we propose a prompting technique called MathDivide that breaks down the mathematical problem into simpler subproblems. Each of the subproblems is formulated as an algebraic expression whose value is evaluated by the Python code generated by the LLM for the corresponding algebraic expression. The values fed to the Python code are the numerical values provided in the problem statement. The solutions for the subproblems are composed together to obtain the final answer for the problem statement. Finally, the final answer is compared to the correct answer. If the final answer matches the correct answer, it is produced as output else a refinement prompt is fed to the LLM. We experiment with this prompting technique on both closed-source LLM models and open-source LLM models using GSM8K dataset. The results obtained demonstrate that MathDivide was able to significantly outperform the leading prompting technique called Math-prompter.
Creating Automatic Speech Recognition (ASR) systems that are robust and resilient to classroom conditions is paramount to the development of AI tools to aid teachers and students. In this work, we study the efficacy of continued pretraining (CPT) in adapting Wav2vec2.0 to the classroom domain. We show that CPT is a powerful tool in that regard and reduces the Word Error Rate (WER) of Wav2vec2.0-based models by upwards of 10%. More specifically, CPT improves the model's robustness to different noises, microphones, classroom conditions as well as classroom demographics. Our CPT models show improved ability to generalize to different demographics unseen in the labeled finetuning data.
Prompting techniques have significantly enhanced the capabilities of Large Language Models (LLMs) across various complex tasks, including reasoning, planning, and solving math word problems. However, most research has predominantly focused on language-based reasoning and word problems, often overlooking the potential of LLMs in handling symbol-based calculations and reasoning. This study aims to bridge this gap by rigorously evaluating LLMs on a series of symbolic tasks, such as addition, multiplication, modulus arithmetic, numerical precision, and symbolic counting. Our analysis encompasses eight LLMs, including four enterprise-grade and four open-source models, of which three have been pre-trained on mathematical tasks. The assessment framework is anchored in Chomsky's Hierarchy, providing a robust measure of the computational abilities of these models. The evaluation employs minimally explained prompts alongside the zero-shot Chain of Thoughts technique, allowing models to navigate the solution process autonomously. The findings reveal a significant decline in LLMs' performance on context-free and context-sensitive symbolic tasks as the complexity, represented by the number of symbols, increases. Notably, even the fine-tuned GPT3.5 exhibits only marginal improvements, mirroring the performance trends observed in other models. Across the board, all models demonstrated a limited generalization ability on these symbol-intensive tasks. This research underscores LLMs' challenges with increasing symbolic complexity and highlights the need for specialized training, memory and architectural adjustments to enhance their proficiency in symbol-based reasoning tasks.
Math Word Problems (MWPs) play a vital role in assessing the capabilities of Large Language Models (LLMs), yet current research primarily focuses on questions with concise contexts. The impact of longer contexts on mathematical reasoning remains under-explored. This study pioneers the investigation of Context Length Generalizability (CoLeG), which refers to the ability of LLMs to solve MWPs with extended narratives. We introduce Extended Grade-School Math (E-GSM), a collection of MWPs featuring lengthy narratives, and propose two novel metrics to evaluate the efficacy and resilience of LLMs in tackling these problems. Our analysis of existing zero-shot prompting techniques with proprietary LLMs along with open-source LLMs reveals a general deficiency in CoLeG. To alleviate these issues, we propose tailored approaches for different categories of LLMs. For proprietary LLMs, we introduce a new instructional prompt designed to mitigate the impact of long contexts. For open-source LLMs, we develop a novel auxiliary task for fine-tuning to enhance CoLeG. Our comprehensive results demonstrate the effectiveness of our proposed methods, showing improved performance on E-GSM. Additionally, we conduct an in-depth analysis to differentiate the effects of semantic understanding and reasoning efficacy, showing that our methods improves the latter. We also establish the generalizability of our methods across several other MWP benchmarks. Our findings highlight the limitations of current LLMs and offer practical solutions correspondingly, paving the way for further exploration of model generalizability and training methodologies.
Intrinsic self-correct was a method that instructed large language models (LLMs) to verify and correct their responses without external feedback. Unfortunately, the study concluded that the LLMs could not self-correct reasoning yet. We find that a simple yet effective verification method can unleash inherent capabilities of the LLMs. That is to mask a key condition in the question, add the current response to construct a verification question, and predict the condition to verify the response. The condition can be an entity in an open-domain question or a numeric value in a math question, which requires minimal effort (via prompting) to identify. We propose an iterative verify-then-correct framework to progressively identify and correct (probably) false responses, named ProCo. We conduct experiments on three reasoning tasks. On average, ProCo, with GPT-3.5-Turbo as the backend LLM, yields $+6.8$ exact match on four open-domain question answering datasets, $+14.1$ accuracy on three arithmetic reasoning datasets, and $+9.6$ accuracy on a commonsense reasoning dataset, compared to Self-Correct. Our implementation is made publicly available at https://wzy6642.github.io/proco.github.io/.
While LLMs have shown impressive capabilities in solving math or coding problems, the ability to make scientific discoveries remains a distinct challenge. This paper proposes a "Turing test for an AI scientist" to assess whether an AI agent can conduct scientific research independently, without relying on human-generated knowledge. Drawing inspiration from the historical development of science, we propose seven benchmark tests that evaluate an AI agent's ability to make groundbreaking discoveries in various scientific domains. These tests include inferring the heliocentric model from celestial observations, discovering the laws of motion in a simulated environment, deriving the differential equation governing vibrating strings, inferring Maxwell's equations from electrodynamics simulations, inventing numerical methods for initial value problems, discovering Huffman coding for data compression, and developing efficient sorting algorithms. To ensure the validity of these tests, the AI agent is provided with interactive libraries or datasets specific to each problem, without access to human knowledge that could potentially contain information about the target discoveries. The ultimate goal is to create an AI scientist capable of making novel and impactful scientific discoveries, surpassing the best human experts in their respective fields. These "Turing tests" serve as intermediate milestones, assessing the AI agent's ability to make discoveries that were groundbreaking in their time. If an AI agent can pass the majority of these seven tests, it would indicate significant progress towards building an AI scientist, paving the way for future advancements in autonomous scientific discovery. This paper aims to establish a benchmark for the capabilities of AI in scientific research and to stimulate further research in this exciting field.
Mathematical reasoning is an important capability of large language models~(LLMs) for real-world applications. To enhance this capability, existing work either collects large-scale math-related texts for pre-training, or relies on stronger LLMs (\eg GPT-4) to synthesize massive math problems. Both types of work generally lead to large costs in training or synthesis. To reduce the cost, based on open-source available texts, we propose an efficient way that trains a small LLM for math problem synthesis, to efficiently generate sufficient high-quality pre-training data. To achieve it, we create a dataset using GPT-4 to distill its data synthesis capability into the small LLM. Concretely, we craft a set of prompts based on human education stages to guide GPT-4, to synthesize problems covering diverse math knowledge and difficulty levels. Besides, we adopt the gradient-based influence estimation method to select the most valuable math-related texts. The both are fed into GPT-4 for creating the knowledge distillation dataset to train the small LLM. We leverage it to synthesize 6 million math problems for pre-training our JiuZhang3.0 model, which only needs to invoke GPT-4 API 9.3k times and pre-train on 4.6B data. Experimental results have shown that JiuZhang3.0 achieves state-of-the-art performance on several mathematical reasoning datasets, under both natural language reasoning and tool manipulation settings. Our code and data will be publicly released in \url{https://github.com/RUCAIBox/JiuZhang3.0}.
Recent advancements in large language models (LLMs) have showcased significant improvements in mathematics. However, traditional math benchmarks like GSM8k offer a unidimensional perspective, falling short in providing a holistic assessment of the LLMs' math capabilities. To address this gap, we introduce MathBench, a new benchmark that rigorously assesses the mathematical capabilities of large language models. MathBench spans a wide range of mathematical disciplines, offering a detailed evaluation of both theoretical understanding and practical problem-solving skills. The benchmark progresses through five distinct stages, from basic arithmetic to college mathematics, and is structured to evaluate models at various depths of knowledge. Each stage includes theoretical questions and application problems, allowing us to measure a model's mathematical proficiency and its ability to apply concepts in practical scenarios. MathBench aims to enhance the evaluation of LLMs' mathematical abilities, providing a nuanced view of their knowledge understanding levels and problem solving skills in a bilingual context. The project is released at https://github.com/open-compass/MathBench .
We present GSM-MC, a multiple-choice (MC) dataset constructed by collecting answers and incorrect predictions on GSM8K from 60 open-source models. Through extensive experiments, we show that LLMs' performance on the MC version of this popular benchmark is strongly correlated with their performance on the original version and is quite robust to distractor choices and option orders, while the evaluation time is reduced by a factor of up to 30. Following similar procedures, we introduce MATH-MC, constructed from MATH, and PythonIO, a new program reasoning MC dataset constructed from HumanEval and MBPP. Experimental results indicate that LLMs' performance on these MC benchmarks leaves much room for improvement. Our data and code are available at https://github.com/Geralt-Targaryen/MC-Evaluation.