Advancing LLM Reasoning Generalists with Preference Trees

We introduce Eurus, a suite of large language models (LLMs) optimized for reasoning. Finetuned from Mistral-7B and CodeLlama-70B, Eurus models achieve state-of-the-art results among open-source models on a diverse set of benchmarks covering mathematics, code generation, and logical reasoning problems. Notably, Eurus-70B beats GPT-3.5 Turbo in reasoning through a comprehensive benchmarking across 12 tests covering five tasks, and achieves a 33.3% pass@1 accuracy on LeetCode and 32.6% on TheoremQA, two challenging benchmarks, substantially outperforming existing open-source models by margins more than 13.3%. The strong performance of Eurus can be primarily attributed to UltraInteract, our newly-curated large-scale, high-quality alignment dataset specifically designed for complex reasoning tasks. UltraInteract can be used in both supervised fine-tuning and preference learning. For each instruction, it includes a preference tree consisting of (1) reasoning chains with diverse planning strategies in a unified format, (2) multi-turn interaction trajectories with the environment and the critique, and (3) pairwise data to facilitate preference learning. UltraInteract allows us to conduct an in-depth exploration of preference learning for reasoning tasks. Our investigation reveals that some well-established preference learning algorithms may be less suitable for reasoning tasks compared to their effectiveness in general conversations. Inspired by this, we derive a novel reward modeling objective which, together with UltraInteract, leads to a strong reward model.

Controllable Preference Optimization: Toward Controllable Multi-Objective Alignment

Alignment in artificial intelligence pursues the consistency between model responses and human preferences as well as values. In practice, the multifaceted nature of human preferences inadvertently introduces what is known as the "alignment tax" -a compromise where enhancements in alignment within one objective (e.g.,harmlessness) can diminish performance in others (e.g.,helpfulness). However, existing alignment techniques are mostly unidirectional, leading to suboptimal trade-offs and poor flexibility over various objectives. To navigate this challenge, we argue the prominence of grounding LLMs with evident preferences. We introduce controllable preference optimization (CPO), which explicitly specifies preference scores for different objectives, thereby guiding the model to generate responses that meet the requirements. Our experimental analysis reveals that the aligned models can provide responses that match various preferences among the "3H" (helpfulness, honesty, harmlessness) desiderata. Furthermore, by introducing diverse data and alignment goals, we surpass baseline methods in aligning with single objectives, hence mitigating the impact of the alignment tax and achieving Pareto improvements in multi-objective alignment.

Executable Code Actions Elicit Better LLM Agents

Large Language Model (LLM) agents, capable of performing a broad range of actions, such as invoking tools and controlling robots, show great potential in tackling real-world challenges. LLM agents are typically prompted to produce actions by generating JSON or text in a pre-defined format, which is usually limited by constrained action space (e.g., the scope of pre-defined tools) and restricted flexibility (e.g., inability to compose multiple tools). This work proposes to use executable Python code to consolidate LLM agents' actions into a unified action space (CodeAct). Integrated with a Python interpreter, CodeAct can execute code actions and dynamically revise prior actions or emit new actions upon new observations through multi-turn interactions. Our extensive analysis of 17 LLMs on API-Bank and a newly curated benchmark shows that CodeAct outperforms widely used alternatives (up to 20% higher success rate). The encouraging performance of CodeAct motivates us to build an open-source LLM agent that interacts with environments by executing interpretable code and collaborates with users using natural language. To this end, we collect an instruction-tuning dataset CodeActInstruct that consists of 7k multi-turn interactions using CodeAct. We show that it can be used with existing data to improve models in agent-oriented tasks without compromising their general capability. CodeActAgent, finetuned from Llama2 and Mistral, is integrated with Python interpreter and uniquely tailored to perform sophisticated tasks (e.g., model training) using existing libraries and autonomously self-debug.

Prudent Silence or Foolish Babble? Examining Large Language Models' Responses to the Unknown

Large Language Models (LLMs) often struggle when faced with situations where they lack the prerequisite knowledge to generate a sensical response. In these cases, models tend to fabricate and hallucinate, rather than appropriately signaling uncertainty as humans would. This behavior misaligns with human conversational norms and presents challenges surrounding responsible and ethical AI development. This work aims to systematically investigate LLMs' behaviors in such situations. We curate an adversarial question-answering benchmark containing unanswerable questions targeting information absent from the LLM's training data. Concretely, these unanswerable questions contain non-existent concepts or false premises. When presented with such unanswerable questions, an LLM should appropriately convey uncertainty, and be able to challenge the premise and refuse to generate a response. While facing answerable valid questions, a model should demonstrate a positive correlation between accuracy and confidence. Using a model-agnostic unified confidence elicitation approach, we observe that LLMs that have gone through instruction finetuning and reinforcement learning from human feedback (RLHF) perform significantly better than their counterparts that do not. Moreover, uncertainty expression 1 through our elicitation method does not always stay consistent with the perceived confidence of the direct response of an LLM. Our findings call for further research into teaching LLMs to proactively and reliably express uncertainty.

UltraFeedback: Boosting Language Models with High-quality Feedback

Reinforcement learning from human feedback (RLHF) has become a pivot technique in aligning large language models (LLMs) with human preferences. In RLHF practice, preference data plays a crucial role in bridging human proclivity and LLMs. However, the scarcity of diverse, naturalistic datasets of human preferences on LLM outputs at scale poses a great challenge to RLHF as well as feedback learning research within the open-source community. Current preference datasets, either proprietary or limited in size and prompt variety, result in limited RLHF adoption in open-source models and hinder further exploration. In this study, we propose ULTRAFEEDBACK, a large-scale, high-quality, and diversified preference dataset designed to overcome these limitations and foster RLHF development. To create ULTRAFEEDBACK, we compile a diverse array of instructions and models from multiple sources to produce comparative data. We meticulously devise annotation instructions and employ GPT-4 to offer detailed feedback in both numerical and textual forms. ULTRAFEEDBACK establishes a reproducible and expandable preference data construction pipeline, serving as a solid foundation for future RLHF and feedback learning research. Utilizing ULTRAFEEDBACK, we train various models to demonstrate its effectiveness, including the reward model UltraRM, chat language model UltraLM-13B-PPO, and critique model UltraCM. Experimental results indicate that our models outperform existing open-source models, achieving top performance across multiple benchmarks. Our data and models are available at https://github.com/thunlp/UltraFeedback.

CRAFT: Customizing LLMs by Creating and Retrieving from Specialized Toolsets

Large language models (LLMs) are often augmented with tools to solve complex tasks. By generating code snippets and executing them through task-specific Application Programming Interfaces (APIs), they can offload certain functions to dedicated external modules, such as image encoding and performing calculations. However, most existing approaches to augment LLMs with tools are constrained by general-purpose APIs and lack the flexibility for tailoring them to specific tasks. In this work, we present CRAFT, a general tool creation and retrieval framework for LLMs. It creates toolsets specifically curated for the tasks and equips LLMs with a component that retrieves tools from these sets to enhance their capability to solve complex tasks. For each task, we collect specific code solutions by prompting GPT-4 to solve the training examples. Following a validation step ensuring the correctness, these solutions are abstracted into code snippets to enhance reusability, and deduplicated for higher quality. At inference time, the language model retrieves snippets from the toolsets and then executes them or generates the output conditioning on the retrieved snippets. Our method is designed to be flexible and offers a plug-and-play approach to adapt off-the-shelf LLMs to unseen domains and modalities, without any finetuning. Experiments on vision-language, tabular processing, and mathematical reasoning tasks show that our approach achieves substantial improvements compared to strong baselines. In addition, our in-depth analysis reveals that: (1) consistent performance improvement can be achieved by scaling up the number of tools and the capability of the backbone models; (2) each component of our approach contributes to the performance gains; (3) the created tools are well-structured and reliable with low complexity and atomicity. The code is available at \url{https://github.com/lifan-yuan/CRAFT}.

MINT: Evaluating LLMs in Multi-turn Interaction with Tools and Language Feedback

To solve complex tasks, large language models (LLMs) often require multiple rounds of interactions with the user, sometimes assisted by external tools. However, current evaluation paradigms often focus solely on benchmark performance with single-turn exchanges, neglecting the intricate interactions among the user, LLMs, and external tools, creating a discrepancy between benchmark evaluation and real-world use cases. We introduce MINT benchmark to evaluate LLMs' ability to solve tasks with multi-turn interactions by (1) using tools and (2) leveraging natural language feedback. To ensure reproducibility, we provide an evaluation framework where LLMs can access tools by executing Python code and receive natural language feedback from the user simulated with GPT-4. We repurpose a diverse set of established datasets and tasks focusing on reasoning, coding, and decision-making and carefully curate them into a compact subset of instances for efficient evaluation. Our analysis of 20 open- and closed-source LLMs offers intriguing findings. (1) LLMs generally benefit from tool interactions and language feedback, with performance gains (absolute, same below) of 1--8% per additional turn with tool use and 2--17% with natural language feedback. (2) Better single-turn performance does not guarantee better multi-turn performance. (3) Surprisingly, on LLMs we evaluated, we found supervised instruction-finetuning (SIFT) and reinforcement learning from human feedback (RLHF) generally hurt multi-turn capabilities. We hope MINT can help measure progress and incentivize research in improving LLMs' capabilities in multi-turn interactions, especially for open-source communities where multi-turn human evaluation has been less accessible compared to commercial LLMs with a larger user base.

Revisiting Out-of-distribution Robustness in NLP: Benchmark, Analysis, and LLMs Evaluations

This paper reexamines the research on out-of-distribution (OOD) robustness in the field of NLP. We find that the distribution shift settings in previous studies commonly lack adequate challenges, hindering the accurate evaluation of OOD robustness. To address these issues, we propose a benchmark construction protocol that ensures clear differentiation and challenging distribution shifts. Then we introduce BOSS, a Benchmark suite for Out-of-distribution robustneSS evaluation covering 5 tasks and 20 datasets. Based on BOSS, we conduct a series of experiments on pre-trained language models for analysis and evaluation of OOD robustness. First, for vanilla fine-tuning, we examine the relationship between in-distribution (ID) and OOD performance. We identify three typical types that unveil the inner learning mechanism, which could potentially facilitate the forecasting of OOD robustness, correlating with the advancements on ID datasets. Then, we evaluate 5 classic methods on BOSS and find that, despite exhibiting some effectiveness in specific cases, they do not offer significant improvement compared to vanilla fine-tuning. Further, we evaluate 5 LLMs with various adaptation paradigms and find that when sufficient ID data is available, fine-tuning domain-specific models outperform LLMs on ID examples significantly. However, in the case of OOD instances, prioritizing LLMs with in-context learning yields better results. We identify that both fine-tuned small models and LLMs face challenges in effectively addressing downstream tasks. The code is public at \url{https://github.com/lifan-yuan/OOD_NLP}.

From Adversarial Arms Race to Model-centric Evaluation: Motivating a Unified Automatic Robustness Evaluation Framework

Textual adversarial attacks can discover models' weaknesses by adding semantic-preserved but misleading perturbations to the inputs. The long-lasting adversarial attack-and-defense arms race in Natural Language Processing (NLP) is algorithm-centric, providing valuable techniques for automatic robustness evaluation. However, the existing practice of robustness evaluation may exhibit issues of incomprehensive evaluation, impractical evaluation protocol, and invalid adversarial samples. In this paper, we aim to set up a unified automatic robustness evaluation framework, shifting towards model-centric evaluation to further exploit the advantages of adversarial attacks. To address the above challenges, we first determine robustness evaluation dimensions based on model capabilities and specify the reasonable algorithm to generate adversarial samples for each dimension. Then we establish the evaluation protocol, including evaluation settings and metrics, under realistic demands. Finally, we use the perturbation degree of adversarial samples to control the sample validity. We implement a toolkit RobTest that realizes our automatic robustness evaluation framework. In our experiments, we conduct a robustness evaluation of RoBERTa models to demonstrate the effectiveness of our evaluation framework, and further show the rationality of each component in the framework. The code will be made public at \url{https://github.com/thunlp/RobTest}.

A Close Look into the Calibration of Pre-trained Language Models

Pre-trained language models (PLMs) achieve remarkable performance on many downstream tasks, but may fail in giving reliable estimates of their predictive uncertainty. Given the lack of a comprehensive understanding of PLMs calibration, we take a close look into this new research problem, aiming to answer two questions: (1) Do PLMs learn to become calibrated in the training process? (2) How effective are existing calibration methods? For the first question, we conduct fine-grained control experiments to study the dynamic change in PLMs' calibration performance in training. We consider six factors as control variables, including dataset difficulty, available training samples, training steps, the number of tunable parameters, model scale, and pretraining. In experiments, we observe a consistent change in calibration performance across six factors. We find that PLMs don't learn to become calibrated in training, evidenced by the continual increase in confidence, no matter the predictions are correct or not. We highlight that our finding presents some contradiction with two established conclusions: (a) Larger PLMs are more calibrated; (b) Pretraining improves model calibration. Next, we study the effectiveness of existing calibration methods in mitigating the overconfidence issue, in both in-distribution and various out-of-distribution settings. Besides unlearnable calibration methods, we adapt two recently proposed learnable methods that directly collect data to train models to have reasonable confidence estimations. Also, we propose extended learnable methods based on existing ones to further improve or maintain PLMs calibration without sacrificing the original task performance. Experimental results show that learnable methods significantly reduce PLMs' confidence in wrong predictions, and our methods exhibit superior performance compared with previous methods.