Reasoning Models

The advent of the Transformer has led to the development of large language models (LLM), which appear to demonstrate human-like capabilities. To assess the generality of this class of models and a variety of other base neural network architectures to multimodal domains, we evaluated and compared their capacity for multimodal generalization. We introduce a multimodal question-answer benchmark to evaluate three specific types of out-of-distribution (OOD) generalization performance: distractor generalization (generalization in the presence of distractors), systematic compositional generalization (generalization to new task permutations), and productive compositional generalization (generalization to more complex tasks with deeper dependencies). While we found that most architectures faired poorly on most forms of generalization (e.g., RNNs and standard Transformers), models that leveraged cross-attention mechanisms between input domains, such as the Perceiver, fared better. Our positive results demonstrate that for multimodal distractor and systematic generalization, cross-attention is an important mechanism to integrate multiple sources of information. On the other hand, all architectures failed in productive generalization, suggesting fundamental limitations of existing architectures for specific types of multimodal OOD generalization. These results demonstrate the strengths and limitations of specific architectural components underlying modern neural models for multimodal reasoning. Finally, we provide *Generic COG* (gCOG), a configurable benchmark with several multimodal generalization splits, for future studies to explore.

The potential of Large Language Model (LLM) as agents has been widely acknowledged recently. Thus, there is an urgent need to quantitatively evaluate LLMs as agents on challenging tasks in interactive environments. We present AgentBench, a multi-dimensional benchmark that consists of 8 distinct environments to assess LLM-as-Agent's reasoning and decision-making abilities. Our extensive test over 29 API-based and open-sourced (OSS) LLMs shows that, while top commercial LLMs present a strong ability of acting as agents in complex environments, there is a significant disparity in performance between them and many OSS competitors that are no larger than 70B. We identify the typical reasons of failures in environments and LLMs, showing that poor long-term reasoning, decision-making, and instruction following abilities are the main obstacles for developing usable LLM agents. Improving instruction following and training on high quality multi-round alignment data could improve agent performance. And different from existing assumptions, training on code present ambivalent impacts on different agent tasks. Datasets, environments, and an integrated evaluation package for AgentBench are released at https://github.com/THUDM/AgentBench.

We introduce MAmmoTH, a series of open-source large language models (LLMs) specifically tailored for general math problem-solving. The MAmmoTH models are trained on MathInstruct, our meticulously curated instruction tuning dataset. MathInstruct is compiled from 13 math datasets with intermediate rationales, six of which have rationales newly curated by us. It presents a unique hybrid of chain-of-thought (CoT) and program-of-thought (PoT) rationales, and also ensures extensive coverage of diverse fields in math. The hybrid of CoT and PoT not only unleashes the potential of tool use but also allows different thought processes for different math problems. As a result, the MAmmoTH series substantially outperform existing open-source models on nine mathematical reasoning datasets across all scales with an average accuracy gain between 16% and 32%. Remarkably, our MAmmoTH-7B model reaches 33% on MATH (a competition-level dataset), which exceeds the best open-source 7B model (WizardMath) by 23%, and the MAmmoTH-34B model achieves 44% accuracy on MATH, even surpassing GPT-4’s CoT result. Our work underscores the importance of diverse problem coverage and the use of hybrid rationales in developing superior math generalist models.

Recent test-time adaptation methods heavily rely on nuanced adjustments of batch normalization (BN) parameters. However, one critical assumption often goes overlooked: that of independently and identically distributed (i.i.d.) test batches with respect to unknown labels. This oversight leads to skewed BN statistics and undermines the reliability of the model under non-i.i.d. scenarios. To tackle this challenge, this paper presents a novel method termed '$\textbf{Un-Mix}$ing $\textbf{T}$est-Time $\textbf{N}$ormalization $\textbf{S}$tatistics' (UnMix-TNS). Our method re-calibrates the statistics for each instance within a test batch by $\textit{mixing}$ it with multiple distinct statistics components, thus inherently simulating the i.i.d. scenario. The core of this method hinges on a distinctive online $\textit{unmixing}$ procedure that continuously updates these statistics components by incorporating the most similar instances from new test batches. Remarkably generic in its design, UnMix-TNS seamlessly integrates with a wide range of leading test-time adaptation methods and pre-trained architectures equipped with BN layers. Empirical evaluations corroborate the robustness of UnMix-TNS under varied scenarios—ranging from single to continual and mixed domain shifts, particularly excelling with temporally correlated test data and corrupted non-i.i.d. real-world streams. This adaptability is maintained even with very small batch sizes or single instances. Our results highlight UnMix-TNS's capacity to markedly enhance stability and performance across various benchmarks. Our code is publicly available at https://github.com/devavratTomar/unmixtns.

Test-time adaptation (TTA) aims to adapt a pre-trained model from a source domain to a target domain only using online unlabeled target data during testing, without accessing to the source data or modifying the original training process. Among the various TTA methods, pseudo-labeling has gained popularity. However, the presence of incorrect pseudo-labels can hinder the effectiveness of target domain adaptation. To overcome this challenge, we propose a novel TTA method, called PROtotype GRAph Model based pseudo-label learning (PROGRAM). PROGRAM consists of two key components: (1) Prototype Graph Model (PGM) for reliable pseudo-label generation; (2) Robust Self-Training (RST) for test-time adaptation with noisy pseudo-labels. PGM constructs the graph using prototypes and test samples, facilitating effective message passing among them to generate more reliable pseudo-labels. RST combines the advantages of consistency regularization and pseudo-labeling to achieve robust target domain adaptation in the presence of noisy pseudo-labels. Our proposed PROGRAM can be easily integrated into existing baselines, resulting in consistent improvement. Extensive experiments show that our PROGRAM outperforms the existing TTA methods on multiple domain generalization and image corruption benchmarks.

Causal inference is one of the hallmarks of human intelligence. While the field of CausalNLP has attracted much interest in the recent years, existing causal inference datasets in NLP primarily rely on discovering causality from empirical knowledge (e.g., commonsense knowledge). In this work, we propose the first benchmark dataset to test the pure causal inference skills of large language models (LLMs). Specifically, we formulate a novel task Corr2Cause, which takes a set of correlational statements and determines the causal relationship between the variables. We curate a large-scale dataset of more than 200K samples, on which we evaluate seventeen existing LLMs. Through our experiments, we identify a key shortcoming of LLMs in terms of their causal inference skills, and show that these models achieve almost close to random performance on the task. This shortcoming is somewhat mitigated when we try to re-purpose LLMs for this skill via finetuning, but we find that these models still fail to generalize – they can only perform causal inference in in-distribution settings when variable names and textual expressions used in the queries are similar to those in the training set, but fail in out-of-distribution settings generated by perturbing these queries. Corr2Cause is a challenging task for LLMs, and can be helpful in guiding future research on improving LLMs’ pure reasoning skills and generalizability. Our data is at https://huggingface.co/datasets/causalnlp/corr2cause. Our code is at https://github.com/causalNLP/corr2cause.

Large Language Models (LLMs) have already become quite proficient at solving simpler programming tasks like those in HumanEval or MBPP benchmarks. However, solving more complex and competitive programming tasks is still quite challenging for these models - possibly due to their tendency to generate solutions as monolithic code blocks instead of decomposing them into logical sub-tasks and sub-modules. On the other hand, experienced programmers instinctively write modularized code with abstraction for solving complex tasks, often reusing previously developed modules. To address this gap, we propose CodeChain, a novel framework for inference that elicits modularized code generation through a chain of self-revisions, each being guided by some representative sub-modules generated in previous iterations. Concretely, CodeChain first instructs the LLM to generate modularized codes through chain-of-thought prompting. Then it applies a chain of self-revisions by iterating the two steps: 1) extracting and clustering the generated sub-modules and selecting the cluster representatives as the more generic and re-usable implementations, and 2) augmenting the original chain-of-thought prompt with these selected module-implementations and instructing the LLM to re-generate new modularized solutions. We find that by naturally encouraging the LLM to reuse the previously developed and verified sub-modules, CodeChain can significantly boost both modularity as well as correctness of the generated solutions, achieving relative pass@1 improvements of 35\% on APPS and 76\% on CodeContests. It is shown to be effective on both OpenAI LLMs as well as open-sourced LLMs like WizardCoder. We also conduct comprehensive ablation studies with different methods of prompting, number of clusters, model sizes, program qualities, etc., to provide useful insights that underpin CodeChain's success.

In recent years, large language models have greatly improved in their ability to perform complex multi-step reasoning. However, even state-of-the-art models still regularly produce logical mistakes. To train more reliable models, we can turn either to outcome supervision, which provides feedback for a final result, or process supervision, which provides feedback for each intermediate reasoning step. Given the importance of training reliable models, and given the high cost of human feedback, it is important to carefully compare the both methods. Recent work has already begun this comparison, but many questions still remain. We conduct our own investigation, finding that process supervision significantly outperforms outcome supervision for training models to solve problems from the challenging MATH dataset. Our process-supervised model solves 78% of problems from a representative subset of the MATH test set. Additionally, we show that active learning significantly improves the efficacy of process supervision. To support related research, we also release PRM800K, the complete dataset of 800,000 step-level human feedback labels used to train our best reward model.

The ML community is rapidly exploring techniques for prompting language models (LMs) and for stacking them into pipelines that solve complex tasks. Unfortunately, existing LM pipelines are typically implemented using hard-coded “prompt templates”, i.e. lengthy strings discovered via trial and error. Toward a more systematic approach for developing and optimizing LM pipelines, we introduce DSPy, a programming model that abstracts LM pipelines as text transformation graphs, or imperative computational graphs where LMs are invoked through declarative modules. DSPy modules are parameterized, meaning they can learn how to apply compositions of prompting, finetuning, augmentation, and reasoning techniques. We design a compiler that will optimize any DSPy pipeline to maximize a given metric, by creating and collecting demonstrations. We conduct two case studies, showing that succinct DSPy programs can express and optimize pipelines that reason about math word problems, tackle multi-hop retrieval, answer complex questions, and control agent loops. Within minutes of compiling, DSPy can automatically produce pipelines that outperform out-of-the-box few-shot prompting as well as expert-created demonstrations for GPT-3.5 and Llama2-13b-chat. On top of that, DSPy programs compiled for relatively small LMs like 770M parameter T5 and Llama2-13b-chat are competitive with many approaches that rely on large and proprietary LMs like GPT-3.5 and on expert-written prompt chains. DSPy is available at https://github.com/stanfordnlp/dspy

Since real-world machine systems are running in non-stationary environments, Continual Test-Time Adaptation (CTTA) task is proposed to adapt the pre-trained model to continually changing target domains. Recently, existing methods mainly focus on model-based adaptation, which aims to leverage a self-training manner to extract the target domain knowledge. However, pseudo labels can be noisy and the updated model parameters are unreliable under dynamic data distributions, leading to error accumulation and catastrophic forgetting in the continual adaptation process. To tackle these challenges and maintain the model plasticity, we design a Visual Domain Adapter (ViDA) for CTTA, explicitly handling both domain-specific and domain-shared knowledge. Specifically, we first comprehensively explore the different domain representations of the adapters with trainable high-rank or low-rank embedding spaces. Then we inject ViDAs into the pre-trained model, which leverages high-rank and low-rank features to adapt the current domain distribution and maintain the continual domain-shared knowledge, respectively. To exploit the low-rank and high-rank ViDAs more effectively, we further propose a Homeostatic Knowledge Allotment (HKA) strategy, which adaptively combines different knowledge from each ViDA. Extensive experiments conducted on four widely used benchmarks demonstrate that our proposed method achieves state-of-the-art performance in both classification and segmentation CTTA tasks. Note that, our method can be regarded as a novel transfer paradigm for large-scale models, delivering promising results in adaptation to continually changing distributions.

The reasoning performance of Large Language Models (LLMs) on a wide range of problems critically relies on chain-of-thought prompting, which involves providing a few chain of thought demonstrations as exemplars in prompts. Recent work, e.g., Tree of Thoughts, has pointed out the importance of exploration and self-evaluation in reasoning step selection for complex problem solving. In this paper, we present Boosting of Thoughts (BoT), an automated prompting framework for problem solving with LLMs by iteratively exploring and self-evaluating many trees of thoughts in order to acquire an ensemble of trial-and-error reasoning experiences, which will serve as a new form of prompting to solve the complex problem. Starting from a simple prompt without requiring examples, BoT iteratively explores and evaluates a large collection of reasoning steps, and more importantly, uses error analysis obtained from the LLM on them to explicitly revise prompting, which in turn enhances reasoning step generation, until a final answer is attained. Our experiments with GPT-4 and Llama2 across extensive complex mathematical problems demonstrate that BoT consistently achieves higher or comparable problem-solving rates than other advanced prompting approaches.

Recent advances in natural language processing, primarily propelled by Large Language Models (LLMs), have showcased their remarkable capabilities grounded in in-context learning. A promising avenue for guiding LLMs in intricate reasoning tasks involves the utilization of intermediate reasoning steps within the Chain-of-Thought (CoT) paradigm. Nevertheless, the central challenge lies in the effective selection of exemplars for facilitating in-context learning. In this study, we introduce a framework that leverages Dual Queries and Low-rank approximation Re-ranking (DQ-LoRe) to automatically select exemplars for in-context learning. Dual Queries first query LLM to obtain LLM-generated knowledge such as CoT, then query the retriever to obtain the final exemplars via both question and the knowledge. Moreover, for the second query, LoRe employs dimensionality reduction techniques to refine exemplar selection, ensuring close alignment with the input question's knowledge. Through extensive experiments, we demonstrate that DQ-LoRe significantly outperforms prior state-of-the-art methods in the automatic selection of exemplars for GPT-4, enhancing performance from 92.5\% to 94.2\%. Our comprehensive analysis further reveals that DQ-LoRe consistently outperforms retrieval-based approaches in terms of both performance and adaptability, especially in scenarios characterized by distribution shifts. DQ-LoRe pushes the boundaries of in-context learning and opens up new avenues for addressing complex reasoning challenges.

The recent progress in large language models (LLMs), especially the invention of chain-of-thought prompting, has made it possible to automatically answer questions by stepwise reasoning. However, when faced with more complicated problems that require non-linear thinking, even the strongest LLMs make mistakes. To address this, we explore whether LLMs are able to recognize errors in their own step-by-step reasoning, without resorting to external resources. To this end, we propose SelfCheck, a general-purpose zero-shot verification schema for recognizing such errors. We then use the results of these checks to improve question-answering performance by conducting weighted voting on multiple solutions to the question. We test SelfCheck on math- and logic-based datasets and find that it successfully recognizes errors and, in turn, increases final answer accuracies.

Neural network trainings are stochastic, causing the performance of trained networks to vary across repeated runs of training. We contribute the following results towards understanding this variation. (1) Despite having significant variance on their test-sets, we demonstrate that standard CIFAR-10 and ImageNet trainings have little variance in their performance on the test-distributions from which their test-sets are sampled. (2) We introduce the independent errors assumption and show that it suffices to recover the structure and variance of the empirical accuracy distribution across repeated runs of training. (3) We prove that test-set variance is unavoidable given the observation that ensembles of identically trained networks are calibrated (Jiang et al., 2021), and demonstrate that the variance of binary classification trainings closely follows a simple formula based on the error rate and number of test examples. (4) We conduct preliminary studies of data augmentation, learning rate, finetuning instability and distribution-shift through the lens of variance between runs.

Scientific discovery hinges on the effective integration of metadata, which refers to a set of 'cognitive' operations such as determining what information is relevant for inquiry, and data, which encompasses physical operations such as observation and experimentation. This paper introduces the Causal Modelling Agent (CMA), a novel framework that synergizes the metadata-based reasoning capabilities of Large Language Models (LLMs) with the data-driven modelling of Deep Structural Causal Models (DSCMs) for the task of causal discovery. We evaluate the CMA's performance on a number of benchmarks, as well as on the real-world task of modelling the clinical and radiological phenotype of Alzheimer's Disease (AD). Our experimental results indicate that the CMA can outperform previous data-driven or metadata-driven approaches to causal discovery. In our real-world application, we use the CMA to derive new insights into the causal relationships among biomarkers of AD.

Transformer-based Large Language Models (LLMs) have become a fixture in modern machine learning. Correspondingly, significant resources are allocated towards research that aims to further advance this technology, typically resulting in models of increasing size that are trained on increasing amounts of data. This work, however, demonstrates the surprising result that it is often possible to significantly improve the performance of LLMs by selectively removing higher-order components of their weight matrices. This simple intervention, which we call LAyer-SElective Rank reduction (LASER), can be done on a model after training has completed, and requires minimal additional parameters and data. We show extensive experiments demonstrating the generality of this finding across language models and datasets, and provide in-depth analyses offering insights into both when LASER is effective and the mechanism by which it operates

This paper presents a novel approach to premise selection, a crucial reasoning task in automated theorem proving. Traditionally, symbolic methods that rely on extensive domain knowledge and engineering effort are applied to this task. In contrast, this work demonstrates that contrastive training with the transformer architecture can achieve higher-quality retrieval of relevant premises, without the knowledge or feature engineering overhead. Our method, Magnushammer, outperforms the most advanced and widely used automation tool in interactive theorem proving called Sledgehammer. On the PISA and miniF2f benchmarks Magnushammer achieves $59.5\%$ (against $38.3\%$) and $34.0\%$ (against $20.9\%$) success rates, respectively. By combining Magnushammer with a language-model-based automated theorem prover, we further improve the state-of-the-art proof success rate from $57.0\%$ to $71.0\%$ on the PISA benchmark using $4$x fewer parameters. Moreover, we develop and open source a novel dataset for premise selection, containing textual representations of (proof state, relevant premise) pairs. To the best of our knowledge, this is the largest available premise selection dataset, and the first dataset of this kind for the Isabelle proof assistant.

Although large language models (LLMs) have achieved significant success in various tasks, they often struggle with hallucination problems, especially in scenarios requiring deep and responsible reasoning. These issues could be partially addressed by introducing external knowledge graphs (KG) in LLM reasoning. In this paper, we propose a new LLM-KG integrating paradigm ``$\hbox{LLM}\otimes\hbox{KG}$'' which treats the LLM as an agent to interactively explore related entities and relations on KGs and perform reasoning based on the retrieved knowledge. We further implement this paradigm by introducing a new approach called Think-on-Graph (ToG), in which the LLM agent iteratively executes beam search on KG, discovers the most promising reasoning paths, and returns the most likely reasoning results. We use a number of well-designed experiments to examine and illustrate the following advantages of ToG: 1) compared with LLMs, ToG has better deep reasoning power; 2) ToG has the ability of knowledge traceability and knowledge correctability by leveraging LLMs reasoning and expert feedback; 3) ToG provides a flexible plug-and-play framework for different LLMs, KGs and prompting strategies without any additional training cost; 4) the performance of ToG with small LLM models could exceed large LLM such as GPT-4 in certain scenarios and this reduces the cost of LLM deployment and application. As a training-free method with lower computational cost and better generality, ToG achieves overall SOTA in 6 out of 9 datasets where most previous SOTAs rely on additional training.

In the evolving landscape of video editing methodologies, a majority of deep learning techniques are often reliant on extensive datasets of observed input and ground truth sequence pairs for optimal performance. Such reliance often falters when acquiring data becomes challenging, especially in tasks like video dehazing and relighting, where replicating identical motions and camera angles in both corrupted and ground truth sequences is complicated. Moreover, these conventional methodologies perform best when the test distribution closely mirrors the training distribution. Recognizing these challenges, this paper introduces a novel video decomposition prior `VDP' framework which derives inspiration from professional video editing practices. Our methodology does not mandate task-specific external data corpus collection, instead pivots to utilizing the motion and appearance of the input video. VDP framework decomposes a video sequence into a set of multiple RGB layers and associated opacity levels. These set of layers are then manipulated individually to obtain the desired results. We addresses tasks such as video object segmentation, dehazing, and relighting. Moreover, we introduce a novel logarithmic video decomposition formulation for video relighting tasks, setting a new benchmark over the existing methodologies. We evaluate our approach on standard video datasets like DAVIS, REVIDE, & SDSD and show qualitative results on a diverse array of internet videos.

Self-consistency (SC) has been a widely used decoding strategy for chain-of-thought reasoning. Despite bringing significant performance improvements across a variety of multi-step reasoning tasks, it is a high-cost method that requires multiple sampling with the preset size. In this paper, we propose a simple and scalable sampling process, Early-Stopping Self-Consistency (ESC), to greatly reduce the cost of SC without sacrificing performance. On this basis, one control scheme for ESC is further derivated to dynamically choose the performance-cost balance for different tasks and models. To demonstrate ESC's effectiveness, we conducted extensive experiments on three popular categories of reasoning tasks: arithmetic, commonsense and symbolic reasoning over language models with varying scales. The empirical results show that ESC reduces the average number of sampling of chain-of-thought reasoning by a significant margin on six benchmarks, including MATH (-33.8%), GSM8K (-80.1%), StrategyQA (-76.8%), CommonsenseQA (-78.5%), Coin Flip (-84.2%) and Last Letters (-67.4%), while attaining comparable performances.