Current AI alignment methodologies rely on human-provided demonstrations or judgments, and the learned capabilities of AI systems would be upper-bounded by human capabilities as a result. This raises a challenging research question: How can we keep improving the systems when their capabilities have surpassed the levels of humans? This paper answers this question in the context of tackling hard reasoning tasks (e.g., level 4-5 MATH problems) via learning from human annotations on easier tasks (e.g., level 1-3 MATH problems), which we term as easy-to-hard generalization. Our key insight is that an evaluator (reward model) trained on supervisions for easier tasks can be effectively used for scoring candidate solutions of harder tasks and hence facilitating easy-to-hard generalization over different levels of tasks. Based on this insight, we propose a novel approach to scalable alignment, which firstly trains the (process-supervised) reward models on easy problems (e.g., level 1-3), and then uses them to evaluate the performance of policy models on hard problems. We show that such easy-to-hard generalization from evaluators can enable easy-to-hard generalizations in generators either through re-ranking or reinforcement learning (RL). Notably, our process-supervised 7b RL model and 34b model (reranking@1024) achieves an accuracy of 34.0% and 52.5% on MATH500, respectively, despite only using human supervision on easy problems. Our approach suggests a promising path toward AI systems that advance beyond the frontier of human supervision.
Large vision-language models (LVLMs), designed to interpret and respond to human instructions, occasionally generate hallucinated or harmful content due to inappropriate instructions. This study uses linear probing to shed light on the hidden knowledge at the output layers of LVLMs. We demonstrate that the logit distributions of the first tokens contain sufficient information to determine whether to respond to the instructions, including recognizing unanswerable visual questions, defending against jailbreaking attacks, and identifying deceptive questions. Such hidden knowledge is gradually lost in logits of subsequent tokens during response generation. Then, we illustrate a simple decoding strategy at the generation of the first token, effectively improving the generated content. In experiments, we find a few interesting insights: First, the CLIP model already contains a strong signal for solving these tasks, which indicates potential bias in the existing datasets. Second, we observe performance improvement by utilizing the first logit distributions on three additional tasks, including indicating uncertainty in math solving, mitigating hallucination, and image classification. Last, with the same training data, simply finetuning LVLMs improves models' performance but is still inferior to linear probing on these tasks.
Despite the effectiveness of data selection for large language models (LLMs) during pretraining and instruction fine-tuning phases, improving data efficiency in supervised fine-tuning (SFT) for specialized domains poses significant challenges due to the complexity of fine-tuning data. To bridge this gap, we introduce an effective and scalable data selection method for SFT, SmallToLarge (S2L), which leverages training trajectories from small models to guide the data selection for larger models. We demonstrate through extensive experiments that S2L significantly improves data efficiency in SFT for mathematical problem-solving, reducing the training data to just 11% of the original MathInstruct dataset (Yue et al., 2023) to match full dataset performance while outperforming state-of-the-art data selection algorithms by an average of 4.7% across 6 in- and out-domain evaluation datasets. Remarkably, selecting only 50K data for SFT, S2L achieves a 32.7% accuracy on the most challenging MATH (Hendrycks et al., 2021) benchmark, improving Phi-2 (Li et al., 2023b) by 16.6%. In clinical text summarization on the MIMIC-III dataset (Johnson et al., 2016), S2L again outperforms training on the full dataset using only 50% of the data. Notably, S2L can perform data selection using a reference model 40x smaller than the target model, proportionally reducing the cost of data selection.
Mathematical capabilities were previously believed to emerge in common language models only at a very large scale or require extensive math-related pre-training. This paper shows that the LLaMA-2 7B model with common pre-training already exhibits strong mathematical abilities, as evidenced by its impressive accuracy of 97.7% and 72.0% on the GSM8K and MATH benchmarks, respectively, when selecting the best response from 256 random generations. The primary issue with the current base model is the difficulty in consistently eliciting its inherent mathematical capabilities. Notably, the accuracy for the first answer drops to 49.5% and 7.9% on the GSM8K and MATH benchmarks, respectively. We find that simply scaling up the SFT data can significantly enhance the reliability of generating correct answers. However, the potential for extensive scaling is constrained by the scarcity of publicly available math questions. To overcome this limitation, we employ synthetic data, which proves to be nearly as effective as real data and shows no clear saturation when scaled up to approximately one million samples. This straightforward approach achieves an accuracy of 82.6% on GSM8K and 40.6% on MATH using LLaMA-2 7B models, surpassing previous models by 14.2% and 20.8%, respectively. We also provide insights into scaling behaviors across different reasoning complexities and error types.
Large language models (LLMs) are highly effective in various natural language processing (NLP) tasks. However, they are susceptible to producing unreliable conjectures in ambiguous contexts called hallucination. This paper presents a new method for evaluating LLM hallucination in Question Answering (QA) based on the unanswerable math word problem (MWP). To support this approach, we innovatively develop a dataset called Unanswerable Math Word Problem (UMWP) which comprises 5200 questions across five categories. We developed an evaluation methodology combining text similarity and mathematical expression detection to determine whether LLM considers the question unanswerable. The results of extensive experiments conducted on 31 LLMs, including GPT-3, InstructGPT, LLaMA, and Claude, demonstrate that in-context learning and reinforcement learning with human feedback (RLHF) training significantly enhance the model's ability to avoid hallucination. We show that utilizing MWP is a reliable and effective approach to assess hallucination. Our code and data are available at https://github.com/Yuki-Asuuna/UMWP.
The African continent lacks enough qualified teachers which hampers the provision of adequate learning support. An AI could potentially augment the efforts of the limited number of teachers, leading to better learning outcomes. Towards that end, this work describes and evaluates the first key output for the NSMQ AI Grand Challenge, which proposes a robust, real-world benchmark for such an AI: "Build an AI to compete live in Ghana's National Science and Maths Quiz (NSMQ) competition and win - performing better than the best contestants in all rounds and stages of the competition". The NSMQ is an annual live science and mathematics competition for senior secondary school students in Ghana in which 3 teams of 2 students compete by answering questions across biology, chemistry, physics, and math in 5 rounds over 5 progressive stages until a winning team is crowned for that year. In this work, we built Brilla AI, an AI contestant that we deployed to unofficially compete remotely and live in the Riddles round of the 2023 NSMQ Grand Finale, the first of its kind in the 30-year history of the competition. Brilla AI is currently available as a web app that livestreams the Riddles round of the contest, and runs 4 machine learning systems: (1) speech to text (2) question extraction (3) question answering and (4) text to speech that work together in real-time to quickly and accurately provide an answer, and then say it with a Ghanaian accent. In its debut, our AI answered one of the 4 riddles ahead of the 3 human contesting teams, unofficially placing second (tied). Improvements and extensions of this AI could potentially be deployed to offer science tutoring to students and eventually enable millions across Africa to have one-on-one learning interactions, democratizing science education.
Large language models (LLMs) have demonstrated remarkable capabilities in problem-solving. However, their proficiency in solving mathematical problems remains inadequate. We propose MathScale, a simple and scalable method to create high-quality mathematical reasoning data using frontier LLMs (e.g., {\tt GPT-3.5}). Inspired by the cognitive mechanism in human mathematical learning, it first extracts topics and knowledge points from seed math questions and then build a concept graph, which is subsequently used to generate new math questions. MathScale exhibits effective scalability along the size axis of the math dataset that we generate. As a result, we create a mathematical reasoning dataset (MathScaleQA) containing two million math question-answer pairs. To evaluate mathematical reasoning abilities of LLMs comprehensively, we construct {\sc MwpBench}, a benchmark of Math Word Problems, which is a collection of ten datasets (including GSM8K and MATH) covering K-12, college, and competition level math problems. We apply MathScaleQA to fine-tune open-source LLMs (e.g., LLaMA-2 and Mistral), resulting in significantly improved capabilities in mathematical reasoning. Evaluated on {\sc MwpBench}, MathScale-7B achieves state-of-the-art performance across all datasets, surpassing its best peers of equivalent size by 42.9\% in micro average accuracy and 43.7\% in macro average accuracy, respectively.
Creating effective educational materials generally requires expensive and time-consuming studies of student learning outcomes. To overcome this barrier, one idea is to build computational models of student learning and use them to optimize instructional materials. However, it is difficult to model the cognitive processes of learning dynamics. We propose an alternative approach that uses Language Models (LMs) as educational experts to assess the impact of various instructions on learning outcomes. Specifically, we use GPT-3.5 to evaluate the overall effect of instructional materials on different student groups and find that it can replicate well-established educational findings such as the Expertise Reversal Effect and the Variability Effect. This demonstrates the potential of LMs as reliable evaluators of educational content. Building on this insight, we introduce an instruction optimization approach in which one LM generates instructional materials using the judgments of another LM as a reward function. We apply this approach to create math word problem worksheets aimed at maximizing student learning gains. Human teachers' evaluations of these LM-generated worksheets show a significant alignment between the LM judgments and human teacher preferences. We conclude by discussing potential divergences between human and LM opinions and the resulting pitfalls of automating instructional design.
In reasoning tasks, even a minor error can cascade into inaccurate results, leading to suboptimal performance of large language models in such domains. Earlier fine-tuning approaches sought to mitigate this by leveraging more precise supervisory signals from human labeling, larger models, or self-sampling, although at a high cost. Conversely, we develop a method that avoids external resources, relying instead on introducing perturbations to the input. Our training approach randomly masks certain tokens within the chain of thought, a technique we found to be particularly effective for reasoning tasks. When applied to fine-tuning with GSM8K on Llama-2-7B, this method achieved a 5\% improvement in GSM8K accuracy and a 10\% improvement in GSM-IC accuracy over standard supervised fine-tuning with a few codes modified. Furthermore, it is complementary to existing methods. When integrated with related explicit data augmentation methods, it leads to improvements across five datasets of various augmentation methods, as well as two different base models. We further investigate the mechanisms behind this improvement through case studies and quantitative analysis, suggesting that our approach may provide superior support for the model in capturing long-distance dependencies, especially those related to questions. This enhancement could deepen understanding of the premises in questions and prior steps. Our code is available at Github.
Automatically generating feedback via large language models (LLMs) in intelligent tutoring systems and online learning platforms has the potential to improve the learning outcomes of many students. However, both feedback generation and evaluation are challenging: feedback content has to be valid especially in subjects like math, which requires models to understand the problem, the solution, and where the student's error lies. Feedback also has to be pedagogically valid to reflect effective tutoring strategies, such as explaining possible misconceptions and encouraging the student, among other desirable features. In this work, we address both problems of automatically generating and evaluating feedback while considering both correctness and alignment. First, we propose a rubric for evaluating math feedback and show that GPT-4 is able to effectively use it to annotate human-written and LLM-generated feedback. Second, we propose a framework for feedback generation that optimizes both correctness and alignment using reinforcement learning (RL). Specifically, we use GPT-4's annotations to create preferences over feedback pairs in an augmented dataset for training via direct preference optimization (DPO). We show that our methods significantly increase the correctness and alignment of generated feedback with Llama 2, an open-source LLM, qualitatively analyze our generation and evaluation systems using case studies, and outline several areas for future work.
Large language models (LLMs) have achieved impressive performance across various mathematical reasoning benchmarks. However, there are increasing debates regarding whether these models truly understand and apply mathematical knowledge or merely rely on shortcuts for mathematical reasoning. One essential and frequently occurring evidence is that when the math questions are slightly changed, LLMs can behave incorrectly. This motivates us to evaluate the robustness of LLMs' math reasoning capability by testing a wide range of question variations. We introduce the adversarial grade school math (GSM-Plus) dataset, an extension of GSM8K augmented with various mathematical perturbations. Our experiments on 25 LLMs and 4 prompting techniques show that while LLMs exhibit different levels of math reasoning abilities, their performances are far from robust. In particular, even for problems that have been solved in GSM8K, LLMs can make mistakes when new statements are added or the question targets are altered. We also explore whether more robust performance can be achieved by composing existing prompting methods, in which we try an iterative method that generates and verifies each intermediate thought based on its reasoning goal and calculation result.
Large Language Model (LLM)-based agents have shown effectiveness across many applications. However, their use in data science scenarios requiring solving long-term interconnected tasks, dynamic data adjustments and domain expertise remains challenging. Previous approaches primarily focus on individual tasks, making it difficult to assess the complete data science workflow. Moreover, they struggle to handle real-time changes in intermediate data and fail to adapt dynamically to evolving task dependencies inherent to data science problems. In this paper, we present Data Interpreter, an LLM-based agent designed to automatically solve various data science problems end-to-end. Our Data Interpreter incorporates two key modules: 1) Hierarchical Graph Modeling, which breaks down complex problems into manageable subproblems, enabling dynamic node generation and graph optimization; and 2) Programmable Node Generation, a technique that refines and verifies each subproblem to iteratively improve code generation results and robustness. Extensive experiments consistently demonstrate the superiority of Data Interpreter. On InfiAgent-DABench, it achieves a 25% performance boost, raising accuracy from 75.9% to 94.9%. For machine learning and open-ended tasks, it improves performance from 88% to 95%, and from 60% to 97%, respectively. Moreover, on the MATH dataset, Data Interpreter achieves remarkable performance with a 26% improvement compared to state-of-the-art baselines. The code is available at https://github.com/geekan/MetaGPT.
We propose a framework for robust evaluation of reasoning capabilities of language models, using functional variants of benchmarks. Models that solve a reasoning test should exhibit no difference in performance over the static version of a problem compared to a snapshot of the functional variant. We have rewritten the relevant fragment of the MATH benchmark into its functional variant MATH(), with functionalization of other benchmarks to follow. When evaluating current state-of-the-art models over snapshots of MATH(), we find a reasoning gap -- the percentage difference between the static and functional accuracies. We find reasoning gaps from 58.35% to 80.31% among the state-of-the-art closed and open weights models that perform well on static benchmarks, with the caveat that the gaps are likely to be smaller with more sophisticated prompting strategies. Here we show that models which anecdotally have good reasoning performance over real-world tasks, have quantifiable lower gaps, motivating the open problem of building "gap 0" models. Code for evaluation and new evaluation datasets, three MATH() snapshots, are publicly available at https://github.com/consequentai/fneval/.
Using Large Language Models for complex mathematical reasoning is difficult, primarily due to the complexity of multi-step reasoning. The main challenges of this process include (1) selecting critical intermediate results to advance the procedure, and (2) limited exploration of potential solutions. To address these issues, we introduce a novel algorithm, namely Stepwise Self-Consistent Chain-of-Thought (SSC-CoT). SSC-CoT employs a strategy of selecting intermediate steps based on the intersection of various reasoning chains. Additionally, SSC-CoT enables the model to discover critical intermediate steps by querying a knowledge graph comprising relevant domain knowledge. To validate SSC-CoT, we present a new dataset, TriMaster100, tailored for complex trigonometry problems. This dataset contains 100 questions, with each solution broken down into scored intermediate steps, facilitating a comprehensive evaluation of the mathematical reasoning process. On TriMaster100, SSC-CoT triples the effectiveness of the state-of-the-art methods. Furthermore, we benchmark SSC-CoT on the widely recognized complex mathematical question dataset, MATH level 5, and it surpasses the second-best method by 7.2% in accuracy. Code and the TriMaster100 dataset can be found at: https://github.com/zhao-zilong/ssc-cot.
Large language models (LLMs) have significantly transformed the educational landscape. As current plagiarism detection tools struggle to keep pace with LLMs' rapid advancements, the educational community faces the challenge of assessing students' true problem-solving abilities in the presence of LLMs. In this work, we explore a new paradigm for ensuring fair evaluation -- generating adversarial examples which preserve the structure and difficulty of the original questions aimed for assessment, but are unsolvable by LLMs. Focusing on the domain of math word problems, we leverage abstract syntax trees to structurally generate adversarial examples that cause LLMs to produce incorrect answers by simply editing the numeric values in the problems. We conduct experiments on various open- and closed-source LLMs, quantitatively and qualitatively demonstrating that our method significantly degrades their math problem-solving ability. We identify shared vulnerabilities among LLMs and propose a cost-effective approach to attack high-cost models. Additionally, we conduct automatic analysis to investigate the cause of failure, providing further insights into the limitations of LLMs.
Tool-augmented Large Language Models (TALMs) are known to enhance the skillset of large language models (LLMs), thereby, leading to their improved reasoning abilities across many tasks. While, TALMs have been successfully employed in different question-answering benchmarks, their efficacy on complex mathematical reasoning benchmarks, and the potential complementary benefits offered by tools for knowledge retrieval and mathematical equation solving are open research questions. In this work, we present MathSensei, a tool-augmented large language model for mathematical reasoning. We study the complementary benefits of the tools - knowledge retriever (Bing Web Search), program generator + executor (Python), and symbolic equation solver (Wolfram-Alpha API) through evaluations on mathematical reasoning datasets. We perform exhaustive ablations on MATH, a popular dataset for evaluating mathematical reasoning on diverse mathematical disciplines. We also conduct experiments involving well-known tool planners to study the impact of tool sequencing on the model performance. MathSensei achieves 13.5% better accuracy over gpt-3.5-turbo with Chain-of-Thought on the MATH dataset. We further observe that TALMs are not as effective for simpler math word problems (in GSM-8K), and the benefit increases as the complexity and required knowledge increases (progressively over AQuA, MMLU-Math, and higher level complex questions in MATH). The code and data are available at https://github.com/Debrup-61/MathSensei.
Despite the impressive performance in a variety of complex tasks, modern large language models (LLMs) still have trouble dealing with some math problems that are simple and intuitive for humans, such as addition. While we can easily learn basic rules of addition and apply them to new problems of any length, LLMs struggle to do the same. Instead, they may rely on similar cases seen in the training corpus for help. We define these two different reasoning mechanisms as "rule-based reasoning" and "case-based reasoning". Since rule-based reasoning is essential for acquiring systematic generalization ability, we aim to explore exactly whether transformers use rule-based or case-based reasoning for math problems. Through carefully designed intervention experiments on five math tasks, we confirm that transformers are performing case-based reasoning, no matter whether scratchpad is used, which aligns with the previous observations that transformers use subgraph matching/shortcut learning to reason. To mitigate such problems, we propose a Rule-Following Fine-Tuning (RFFT) technique to teach transformers to perform rule-based reasoning. Specifically, we provide explicit rules in the input and then instruct transformers to recite and follow the rules step by step. Through RFFT, we successfully enable LLMs fine-tuned on 1-5 digit addition to generalize to up to 12-digit addition with over 95% accuracy, which is over 40% higher than scratchpad. The significant improvement demonstrates that teaching LLMs to use rules explicitly helps them learn rule-based reasoning and generalize better in length.
Math word problems are critical K-8 educational tools, but writing them is time consuming and requires extensive expertise. To be educational, problems must be solvable, have accurate answers, and, most importantly, be educationally appropriate. We propose that language models have potential to support K-8 math education by automatically generating word problems. However, evaluating educational appropriateness is hard to quantify. We fill this gap by having teachers evaluate problems generated by LLMs, who find existing models and data often fail to be educationally appropriate. We then explore automatically generating educational word problems, ultimately using our expert annotations to finetune a 70B language model. Our model, MATHWELL, is the first K-8 word problem generator targeted at educational appropriateness. Further expert studies find MATHWELL generates problems far more solvable, accurate, and appropriate than public models. MATHWELL also matches GPT-4's problem quality while attaining more appropriate reading levels for K-8 students and avoiding generating harmful questions.
Program-of-Thought (PoT) replaces natural language-based Chain-of-Thought (CoT) as the most popular method in Large Language Models (LLMs) mathematical reasoning tasks by utilizing external tool calls to circumvent computational errors. However, our evaluation of the GPT-4 and Llama series reveals that using PoT introduces more reasoning errors, such as incorrect formulas or flawed logic, compared to CoT. To address this issue, we propose Human-Think Language (HTL), which leverages a suite of strategies that help integrate PoT and CoT, encompassing: (1) a new generation paradigm that uses full CoT reasoning to control code generation. (2) Focus Attention, that directs model attention to the CoT reasoning during PoT to generate more logical code. (3) reinforcement learning that utilizes the accuracy of both CoT and PoT responses as rewards to prevent repetitive reasoning steps in LLMs when solving difficult math problems. Our method achieves an average improvement of 6.5% on the Llama-Base model and 4.3% on the Mistral-Base model across 8 mathematical calculation datasets. It also shows significant effectiveness on five out-of-domain datasets by controlling the model's information flow, exhibiting strong transferability. Additionally, HTL shows the most significant improvement in non-mathematical natural language inference task, contributing to a unified reasoning task framework
This paper introduces ConceptMath, a bilingual (English and Chinese), fine-grained benchmark that evaluates concept-wise mathematical reasoning of Large Language Models (LLMs). Unlike traditional benchmarks that evaluate general mathematical reasoning with an average accuracy, ConceptMath systematically organizes math problems under a hierarchy of math concepts, so that mathematical reasoning can be evaluated at different granularity with concept-wise accuracies. Based on our ConcepthMath, we evaluate a broad range of LLMs, and we observe existing LLMs, though achieving high average accuracies on traditional benchmarks, exhibit significant performance variations across different math concepts and may even fail catastrophically on the most basic ones. Besides, we also introduce an efficient fine-tuning strategy to enhance the weaknesses of existing LLMs. Finally, we hope ConceptMath could guide the developers to understand the fine-grained mathematical abilities of their models and facilitate the growth of foundation models.