Recent advancements in Large Language Models (LLMs) have showcased striking results on existing logical reasoning benchmarks, with some models even surpassing human performance. However, the true depth of their competencies and robustness in reasoning tasks remains an open question. To this end, in this paper, we focus on two popular reasoning tasks: arithmetic reasoning and code generation. Particularly, we introduce (i) a general ontology of perturbations for math and coding questions, (ii) a semi-automatic method to apply these perturbations, and (iii) two datasets, GSMORE and HUMANEVAL-CORE, respectively, of perturbed math and coding problems to probe LLM capabilities in numeric reasoning and coding tasks. Through comprehensive evaluations of both closed-source and open-source LLMs, we show a significant performance drop across all the models against the perturbed questions, suggesting that the current LLMs lack robust problem solving skills and structured reasoning abilities in many areas, as defined by our ontology. We open-source the datasets and source codes at: https://github.com/declare-lab/LLM-ReasoningTest.
One way to enhance the reasoning capability of Large Language Models (LLMs) is to conduct Supervised Fine-Tuning (SFT) using Chain-of-Thought (CoT) annotations. This approach does not show sufficiently strong generalization ability, however, because the training only relies on the given CoT data. In math problem-solving, for example, there is usually only one annotated reasoning path for each question in the training data. Intuitively, it would be better for the algorithm to learn from multiple annotated reasoning paths given a question. To address this issue, we propose a simple yet effective approach called Reinforced Fine-Tuning (ReFT) to enhance the generalizability of learning LLMs for reasoning, with math problem-solving as an example. ReFT first warmups the model with SFT, and then employs on-line reinforcement learning, specifically the PPO algorithm in this paper, to further fine-tune the model, where an abundance of reasoning paths are automatically sampled given the question and the rewards are naturally derived from the ground-truth answers. Extensive experiments on GSM8K, MathQA, and SVAMP datasets show that ReFT significantly outperforms SFT, and the performance can be potentially further boosted by combining inference-time strategies such as majority voting and re-ranking. Note that ReFT obtains the improvement by learning from the same training questions as SFT, without relying on extra or augmented training questions. This indicates a superior generalization ability for ReFT.
Despite the advancements in large language models (LLMs) for mathematical reasoning, solving competition-level math problems remains a significant challenge, especially for open-source LLMs without external tools. We introduce the MMIQC dataset, comprising a mixture of processed web data and synthetic question-response pairs, aimed at enhancing the mathematical reasoning capabilities of base language models. Models fine-tuned on MMIQC consistently surpass their counterparts in performance on the MATH benchmark across various model sizes. Notably, Qwen-72B-MMIQC achieves a 45.0% accuracy, exceeding the previous open-source state-of-the-art by 8.2% and outperforming the initial version GPT-4 released in 2023. Extensive evaluation results on Hungarian high school finals suggest that such improvement can generalize to unseen data. Our ablation study on MMIQC reveals that a large part of the improvement can be attributed to our novel augmentation method, Iterative Question Composing (IQC), which involves iteratively composing new questions from seed problems using an LLM and applying rejection sampling through another LLM.
A wide range of real-world applications is characterized by their symbolic nature, necessitating a strong capability for symbolic reasoning. This paper investigates the potential application of Large Language Models (LLMs) as symbolic reasoners. We focus on text-based games, significant benchmarks for agents with natural language capabilities, particularly in symbolic tasks like math, map reading, sorting, and applying common sense in text-based worlds. To facilitate these agents, we propose an LLM agent designed to tackle symbolic challenges and achieve in-game objectives. We begin by initializing the LLM agent and informing it of its role. The agent then receives observations and a set of valid actions from the text-based games, along with a specific symbolic module. With these inputs, the LLM agent chooses an action and interacts with the game environments. Our experimental results demonstrate that our method significantly enhances the capability of LLMs as automated agents for symbolic reasoning, and our LLM agent is effective in text-based games involving symbolic tasks, achieving an average performance of 88% across all tasks.
Despite the general capabilities of large pretrained language models, they consistently benefit from further adaptation to better achieve desired behaviors. However, tuning these models has become increasingly resource-intensive, or impossible when model weights are private. We introduce proxy-tuning, a lightweight decoding-time algorithm that operates on top of black-box LMs to achieve the same end as direct tuning, but by accessing only its predictions over the output vocabulary, not its parameters. Our method tunes a smaller LM, then applies the difference between the predictions of the small tuned and untuned LMs to shift the original predictions of the larger untuned model in the direction of tuning, while retaining the benefits of larger-scale pretraining. In experiments, when we apply proxy-tuning to Llama2-70B using proxies of only 7B size, we can close 88% of the gap between Llama2-70B and its truly-tuned chat version, when evaluated across knowledge, reasoning, and safety benchmarks. We then demonstrate the generality of proxy-tuning by applying it to domain adaptation on code, and task-specific finetuning on question-answering and math problems. Finally, we show how to proxy-tune a truly black-box LM, GPT-3.5, for temporal adaptation, increasing its knowledge about recent events. Our work demonstrates the promise of using small tuned LMs to efficiently customize large, potentially proprietary LMs through decoding-time guidance.
Large language models (LLMs) have seen considerable advancements in natural language understanding tasks, yet there remains a gap to bridge before attaining true artificial general intelligence, especially concerning shortcomings in mathematical reasoning capabilities. We postulate that the inherent nature of LLM training, which focuses on predicting probabilities of next token, presents challenges in effectively modeling mathematical reasoning that demands exact calculations, both from data-driven and theoretical standpoints. In this paper, we address this challenge by enriching the data landscape and introducing a novel math dataset, enhanced with a capability to utilize a Python code interpreter. This dataset is derived from GSM8K and MATH and has been further refined through a combination of GPT-4 annotations, human review, and self-training processes, where the errors in the original GSM8K training set have been fixed. Additionally, we propose a tentative, easily replicable protocol for the fine-tuning of math-specific LLMs, which has led to a significant improvement in the performance of a 7B-parameter LLM on the GSM8K and MATH datasets. We are committed to advancing the field of mathematical reasoning in LLMs and, to that end, we have made source code for data generation / training / inference, and the model checkpoints publicly available at \url{https://github.com/MARIO-Math-Reasoning/MARIO}. We hope this will facilitate further research and development within the community.
Large Language Models (LLMs) have shown promise in assisting scientific discovery. However, such applications are currently limited by LLMs' deficiencies in understanding intricate scientific concepts, deriving symbolic equations, and solving advanced numerical calculations. To bridge these gaps, we introduce SciInstruct, a suite of scientific instructions for training scientific language models capable of college-level scientific reasoning. Central to our approach is a novel self-reflective instruction annotation framework to address the data scarcity challenge in the science domain. This framework leverages existing LLMs to generate step-by-step reasoning for unlabelled scientific questions, followed by a process of self-reflective critic-and-revise. Applying this framework, we curated a diverse and high-quality dataset encompassing physics, chemistry, math, and formal proofs. We analyze the curated SciInstruct from multiple interesting perspectives (e.g., domain, scale, source, question type, answer length, etc.). To verify the effectiveness of SciInstruct, we fine-tuned different language models with SciInstruct, i.e., ChatGLM3 (6B and 32B), Llama3-8B-Instruct, and Mistral-7B: MetaMath, enhancing their scientific and mathematical reasoning capabilities, without sacrificing the language understanding capabilities of the base model. We release all codes and SciInstruct at https://github.com/THUDM/SciGLM.
In this paper, we introduce Concise Chain-of-Thought (CCoT) prompting. We compared standard CoT and CCoT prompts to see how conciseness impacts response length and correct-answer accuracy. We evaluated this using GPT-3.5 and GPT-4 with a multiple-choice question-and-answer (MCQA) benchmark. CCoT reduced average response length by 48.70% for both GPT-3.5 and GPT-4 while having a negligible impact on problem-solving performance. However, on math problems, GPT-3.5 with CCoT incurs a performance penalty of 27.69%. Overall, CCoT leads to an average per-token cost reduction of 22.67%. All code, data, and supplemental materials are available on GitHub at https://github.com/matthewrenze/jhu-concise-cot
We investigate parameter-efficient fine-tuning (PEFT) methods that can provide good accuracy under limited computational and memory budgets in the context of large language models (LLMs). We present a new PEFT method called Robust Adaptation (RoSA) inspired by robust principal component analysis that jointly trains $\textit{low-rank}$ and $\textit{highly-sparse}$ components on top of a set of fixed pretrained weights to efficiently approximate the performance of a full-fine-tuning (FFT) solution. Across a series of challenging generative tasks such as grade-school math and SQL query generation, which require fine-tuning for good performance, we show that RoSA outperforms LoRA, pure sparse fine-tuning, and alternative hybrid methods at the same parameter budget, and can even recover the performance of FFT on some tasks. We provide system support for RoSA to complement the training algorithm, specifically in the form of sparse GPU kernels which enable memory- and computationally-efficient training, and show that it is also compatible with low-precision base weights, resulting in the first joint representation combining quantization, low-rank and sparse approximations. Our code is available at https://github.com/IST-DASLab/RoSA.
Large language models (LLMs) have exhibited impressive competence in various tasks, but their internal mechanisms on mathematical problems are still under-explored. In this paper, we study a fundamental question: how language models encode the value of numbers, a basic element in math. To study the question, we construct a synthetic dataset comprising addition problems and utilize linear probes to read out input numbers from the hidden states. Experimental results support the existence of encoded number values in LLMs on different layers, and these values can be extracted via linear probes. Further experiments show that LLMs store their calculation results in a similar manner, and we can intervene the output via simple vector additions, proving the causal connection between encoded numbers and language model outputs. Our research provides evidence that LLMs encode the value of numbers linearly, offering insights for better exploring, designing, and utilizing numeric information in LLMs.
Humans generally acquire new skills without compromising the old; however, the opposite holds for Large Language Models (LLMs), e.g., from LLaMA to CodeLLaMA. To this end, we propose a new post-pretraining method for LLMs with an expansion of Transformer blocks. We tune the expanded blocks using only new corpus, efficiently and effectively improving the model's knowledge without catastrophic forgetting. In this paper, we experiment on the corpus of code and math, yielding LLaMA Pro-8.3B, a versatile foundation model initialized from LLaMA2-7B, excelling in general tasks, programming, and mathematics. LLaMA Pro and its instruction-following counterpart (LLaMA Pro-Instruct) achieve advanced performance among various benchmarks, demonstrating superiority over existing open models in the LLaMA family and the immense potential of reasoning and addressing diverse tasks as an intelligent agent. Our findings provide valuable insights into integrating natural and programming languages, laying a solid foundation for developing advanced language agents that operate effectively in various environments.
In this work, we introduce a novel evaluation paradigm for Large Language Models (LLMs) that compels them to transition from a traditional question-answering role, akin to a student, to a solution-scoring role, akin to a teacher. This paradigm, focusing on "reasoning about reasoning," hence termed meta-reasoning, shifts the emphasis from result-oriented assessments, which often neglect the reasoning process, to a more comprehensive evaluation that effectively distinguishes between the cognitive capabilities of different models. By applying this paradigm in the GSM8K dataset, we have developed the MR-GSM8K benchmark. Our extensive analysis includes several state-of-the-art models from both open-source and commercial domains, uncovering fundamental deficiencies in their training and evaluation methodologies. Notably, while models like Deepseek-v2 and Claude3-Sonnet closely competed with GPT-4 in GSM8K, their performance disparities expanded dramatically in MR-GSM8K, with differences widening to over 20 absolute points, underscoring the significant challenge posed by our meta-reasoning approach.
High-quality, large-scale corpora are the cornerstone of building foundation models. In this work, we introduce MathPile, a diverse and high-quality math-centric corpus comprising about 9.5 billion tokens. Throughout its creation, we adhered to the principle of "less is more", firmly believing in the supremacy of data quality over quantity, even in the pre-training phase. Our meticulous data collection and processing efforts included a complex suite of preprocessing, prefiltering, language identification, cleaning, filtering, and deduplication, ensuring the high quality of our corpus. Furthermore, we performed data contamination detection on downstream benchmark test sets to eliminate duplicates and conducted continual pre-training experiments, booting the performance on common mathematical reasoning benchmarks. We aim for our MathPile to boost language models' mathematical reasoning abilities and open-source its different versions and processing scripts to advance the field.
Chain-of-thought prompting~(CoT) and tool augmentation have been validated in recent work as effective practices for improving large language models~(LLMs) to perform step-by-step reasoning on complex math-related tasks. However, most existing math reasoning datasets may be not able to fully evaluate and analyze the ability of LLMs in manipulating tools and performing reasoning, as they may only require very few invocations of tools or miss annotations for evaluating intermediate reasoning steps. To address the issue, we construct \textbf{CARP}, a new Chinese dataset consisting of 4,886 computation-intensive algebra problems with formulated annotations on intermediate steps. In CARP, we test four LLMs with CoT prompting, and find that they are all prone to make mistakes at the early steps of the solution, leading to wrong answers. Based on this finding, we propose a new approach that can deliberate the reasoning steps with tool interfaces, namely \textbf{DELI}. In DELI, we first initialize a step-by-step solution based on retrieved exemplars, then iterate two deliberation procedures that check and refine the intermediate steps of the generated solution, from the perspectives of tool manipulation and natural language reasoning, until obtaining converged solutions or reaching the maximum turn. Experimental results on CARP and six other datasets show that the proposed DELI mostly outperforms competitive baselines, and can further boost the performance of existing CoT methods. Our data and code are available in \url{https://github.com/RUCAIBox/CARP}.
The recently released Google Gemini class of models are the first to comprehensively report results that rival the OpenAI GPT series across a wide variety of tasks. In this paper, we do an in-depth exploration of Gemini's language abilities, making two contributions. First, we provide a third-party, objective comparison of the abilities of the OpenAI GPT and Google Gemini models with reproducible code and fully transparent results. Second, we take a closer look at the results, identifying areas where one of the two model classes excels. We perform this analysis over 10 datasets testing a variety of language abilities, including reasoning, answering knowledge-based questions, solving math problems, translating between languages, generating code, and acting as instruction-following agents. From this analysis, we find that Gemini Pro achieves accuracy that is close but slightly inferior to the corresponding GPT 3.5 Turbo on all tasks that we benchmarked. We further provide explanations for some of this under-performance, including failures in mathematical reasoning with many digits, sensitivity to multiple-choice answer ordering, aggressive content filtering, and others. We also identify areas where Gemini demonstrates comparably high performance, including generation into non-English languages, and handling longer and more complex reasoning chains. Code and data for reproduction can be found at https://github.com/neulab/gemini-benchmark
Large language models (LLMs) face challenges in solving complex mathematical problems that require comprehensive capacities to parse the statements, associate domain knowledge, perform compound logical reasoning, and integrate the intermediate rationales. Tackling all these problems once could be arduous for LLMs, thus leading to confusion in generation. In this work, we explore the potential of enhancing LLMs with agents by meticulous decomposition and modeling of mathematical reasoning process. Specifically, we propose a formal description of the mathematical solving and extend LLMs with an agent-based zero-shot framework named $\bf{P}$lanner-$\bf{R}$easoner-$\bf{E}$xecutor-$\bf{R}$eflector (PRER). We further provide and implement two MathAgents that define the logical forms and inherent relations via a pool of actions in different grains and orientations: MathAgent-M adapts its actions to LLMs, while MathAgent-H aligns with humankind. Experiments on miniF2F and MATH have demonstrated the effectiveness of PRER and proposed MathAgents, achieving an increase of $12.3\%$($53.9\%\xrightarrow{}66.2\%$) on the MiniF2F, $9.2\%$ ($49.8\%\xrightarrow{}59.0\%$) on MATH, and $13.2\%$($23.2\%\xrightarrow{}35.4\%$) for level-5 problems of MATH against GPT-4. Further analytical results provide more insightful perspectives on exploiting the behaviors of LLMs as agents.
Chain-of-Thought (CoT) prompting methods have enabled large language models (LLMs) to generate reasoning paths and solve math word problems (MWPs). However, they are sensitive to mistakes in the paths, as any mistake can result in an incorrect answer. We propose a novel method named Progressive Rectification Prompting (PRP) to improve average accuracy on eight MWP datasets from 77.3 to 90.5. Given an initial answer from CoT, PRP iterates a verify-then-rectify process to progressively identify incorrect answers and rectify the reasoning paths. With the most likely correct answer, the LLM predicts a masked numerical value in the question; if the prediction does not match the masked value, the answer is likely incorrect. Then the LLM is prompted to re-generate the reasoning path hinted with a set of incorrect answers to prevent itself from repeating previous mistakes. PRP achieves the best performance compared against the CoT methods. Our implementation is made publicly available at https://wzy6642.github.io/prp.github.io/.
In recent years, pre-trained large language models (LLMs) have demonstrated remarkable efficiency in achieving an inference-time few-shot learning capability known as in-context learning. However, existing literature has highlighted the sensitivity of this capability to the selection of few-shot demonstrations. Current understandings of the underlying mechanisms by which this capability arises from regular language model pretraining objectives remain disconnected from the real-world LLMs. This study aims to examine the in-context learning phenomenon through a Bayesian lens, viewing real-world LLMs as latent variable models. On this premise, we propose an algorithm to select optimal demonstrations from a set of annotated data with a small LM, and then directly generalize the selected demonstrations to larger LMs. We demonstrate significant improvement over baselines, averaged over eight GPT models on eight real-world text classification datasets. We also demonstrate the real-world usefulness of our algorithm on GSM8K, a math word problem dataset. Our empirical findings support our hypothesis that LLMs implicitly infer a latent variable containing task information.
This paper presents an in-depth analysis of Large Language Models (LLMs), focusing on LLaMA, a prominent open-source foundational model in natural language processing. Instead of assessing LLaMA through its generative output, we design multiple-choice tasks to probe its intrinsic understanding in high-order tasks such as reasoning and computation. We examine the model horizontally, comparing different sizes, and vertically, assessing different layers. We unveil several key and uncommon findings based on the designed probing tasks: (1) Horizontally, enlarging model sizes almost could not automatically impart additional knowledge or computational prowess. Instead, it can enhance reasoning abilities, especially in math problem solving, and helps reduce hallucinations, but only beyond certain size thresholds; (2) In vertical analysis, the lower layers of LLaMA lack substantial arithmetic and factual knowledge, showcasing logical thinking, multilingual and recognitive abilities, with top layers housing most computational power and real-world knowledge.
Eliciting Latent Knowledge (ELK) aims to find patterns in a capable neural network's activations that robustly track the true state of the world, especially in hard-to-verify cases where the model's output is untrusted. To further ELK research, we introduce 12 datasets and a corresponding suite of "quirky" language models (LMs) that are finetuned to make systematic errors when answering questions if and only if the keyword "Bob" is present in the prompt. We find that, especially in middle layers, linear probes usually report an LM's knowledge independently of what the LM outputs, enabling us to elicit the correct answer despite the model's untruthful output. The best probing method (logistic regression on contrast pairs) recovers 89% of the gap in AUROC between truthful and untruthful contexts, and 75% for questions harder than those used to train the probe. We also find that a mechanistic anomaly detection approach can flag untruthful behavior with 0.95 AUROC. Our results show promise for eliciting reliable knowledge from capable but untrusted models, and facilitates future research empirically investigating ELK methods.