We introduce EURUS, a suite of large language models (LLMs) optimized for reasoning. Finetuned from Mistral-7B, Llama-3-8B, and Mixtral-8x22B, 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, EURUX-8X22B outperforms GPT-3.5 Turbo in reasoning through a comprehensive benchmarking across 12 test sets covering five tasks. The strong performance of EURUS can be primarily attributed to ULTRAINTERACT, our newly-curated large-scale, high-quality training data dataset specifically designed for complex reasoning tasks. ULTRAINTERACT can be used in both supervised fine-tuning, preference learning, and reward modeling. It pairs each instruction with 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 positive and negative responses 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. The hypothesis is that in reasoning tasks, the space of correct answers is much smaller than that of incorrect ones, so it is necessary to explicitly increase the reward of chosen data. Therefore, in addition to increasing the reward margin as many preference learning algorithms do, the absolute values of positive responses’ rewards should be positive and may serve as a proxy for performance. Inspired by this, we derive a novel reward modeling objective and empirically that it leads to a stable reward modeling curve and better performance. Together with ULTRAINTERACT, we obtain a strong reward model.

Different from its counterpart outcome reward models (ORMs), which evaluate the entire responses, a process reward model (PRM) scores a reasoning trajectory step by step, providing denser and more fine-grained rewards. However, training a PRM requires labels annotated at every intermediate step, presenting significant challenges for both manual and automatic data collection. This paper aims to address this challenge. Both theoretically and empirically, we show that an implicit PRM can be obtained at no additional cost, by simply training an ORM on the cheaper response-level labels. The only assumption is to parameterize the outcome reward as the log-likelihood ratios of the policy and reference models r$\phi$(y) = $\beta$ log $\pi$$\phi$(y) $\pi$ref(y) , which can be optimized regardless of the specific choice of loss objectives. In experiments, we instantiate our implicit PRMs with various objectives and evaluate their performance on MATH. We show that our implicit PRM outperforms a strong MCTS-based baseline á la Math-Shepherd (Wang et al., 2023) using less than 1/38 of the training data. Its performance can be further improved with majority voting. We further find that scaling up instructions and responses benefits our implicit PRM, and the latter brings a larger gain. Particularly, we find that our implicit PRM, when instantiated with the cross-entropy (CE) loss, is more data-efficient and can keep improving generation models even when trained with only one response per instruction, the setup that suffers from extreme data scarcity and imbalance. Further, instructions should be relevant to downstream tasks while the diversity of responses does not bring gains. Surprisingly, training on extra Math-Shepherd step labels brings no further improvements to our implicit PRM trained on only outcome data. We hope that our work will encourage a rethinking of PRM training approaches and contribute to making training PRMs more accessible.

Entropy minimization (EM) trains the model to concentrate even more probability mass on its most confident outputs. We show that this simple objective alone, without any labeled data, can substantially improve large language models’ (LLMs) performance on challenging math, physics, and coding tasks. We explore three approaches: (1) EM-FT minimizes token-level entropy similarly to instruction finetuning, but on unlabeled outputs drawn from the model; (2) EM-RL: reinforcement learning with negative entropy as the only reward to maximize; (3) EM-INF: inference-time logit adjustment to reduce entropy without any training data or parameter updates. On Qwen-7B, EM-RL, without any labeled data, achieves comparable or better performance than strong RL baselines such as GRPO and RLOO that are trained on 60K labeled examples. Furthermore, EM-INF enables Qwen-32B to match or exceed the performance of proprietary models like GPT-4o, Claude 3 Opus, and Gemini 1.5 Pro on the challenging SciCode benchmark, while being 3x more efficient than self-consistency and sequential refinement. Our findings reveal that many pretrained LLMs possess previously underappreciated reasoning capabilities that can be effectively elicited through entropy minimization alone, without any labeled data or even any parameter updates.

Reinforcement learning (RL) yields substantial improvements in large language models’ (LLMs) downstream task performance and alignment with human values. Surprisingly, such large gains result from updating only a small subnetwork comprising just 5%-30% of the parameters, with the rest effectively unchanged. We refer to this phenomenon as parameter update sparsity induced by RL. It is observed across all 7 widely-used RL algorithms (e.g., PPO, GRPO, DPO) and all 10 LLMs from different families in our experiments. This sparsity is intrinsic and occurs without any explicit sparsity-promoting regularizations or architectural constraints. Finetuning the subnetwork alone recovers the test accuracy, and, remarkably, produces a model nearly identical to the one obtained via full finetuning. The subnetworks from different random seeds, training data, and even RL algorithms show substantially greater overlap than expected by chance. Our analysis suggests that this sparsity is not due to updating only a subset of layers; instead, nearly all parameter matrices receive similarly sparse updates. Moreover, the updates to almost all parameter matrices are nearly full-rank, suggesting RL updates a small subset of parameters that nevertheless span almost the full subspaces that the parameter matrices can represent. We conjecture that the this update sparsity can be primarily attributed to training on data that is near the policy distribution; techniques that encourage the policy to remain close to the pretrained model, such as the KL regularization and gradient clipping, have limited impact.

This paper investigates Reinforcement Learning (RL) on data without explicit labels for reasoning tasks in Large Language Models (LLMs). The core challenge of the problem is reward estimation during inference while not having access to ground-truth information. While this setting appears elusive, we find that common practices in Test-Time Scaling (TTS), such as majority voting, yield surprisingly effective rewards suitable for driving RL training. In this work, we introduce Test-Time Reinforcement Learning (TTRL), a novel method for training LLMs using RL on unlabeled data. TTRL enables self-evolution of LLMs by utilizing the priors in the pre-trained models. Our experiments demonstrate that TTRL consistently improves performance across a variety of tasks and models. Notably, TTRL boosts the pass@1 performance of Qwen-2.5-Math-7B by approximately 211% on the AIME 2024 with only unlabeled test data. Furthermore, although TTRL is only supervised by the Maj@N metric, TTRL has demonstrated performance to consistently surpass the upper limit of the initial model, and approach the performance of models trained directly on test data with ground-truth labels. Our experimental findings validate the general effectiveness of TTRL across various tasks and highlight TTRL's potential for broader tasks and domains.

Yiju Guo, Ganqu Cui, Lifan Yuan, Ning Ding, Zexu Sun, Bowen Sun, Huimin Chen, Ruobing Xie, Jie Zhou, Yankai Lin, Zhiyuan Liu, Maosong Sun. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

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 protocols often emphasize benchmark performance with single-turn exchanges, neglecting the nuanced interactions among the user, LLMs, and external tools, while also underestimating the importance of natural language feedback from users. These oversights contribute to discrepancies between research benchmark evaluations and real-world use cases. We introduce MINT, a benchmark that evaluates 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 users' natural language feedback simulated by GPT-4. We repurpose a diverse set of established evaluation datasets focusing on reasoning, coding, and decision-making and carefully curate them into a compact subset for efficient evaluation. Our analysis of 20 open- and closed-source LLMs offers intriguing findings. (a) LLMs generally benefit from tools and language feedback, with performance gains (absolute, same below) of 1--8% for each turn of tool use and 2--17% with natural language feedback. (b) Better single-turn performance does not guarantee better multi-turn performance. (c) Surprisingly, on the LLMs evaluated, supervised instruction-finetuning (SIFT) and reinforcement learning from human feedback (RLHF) generally hurt multi-turn capabilities. We expect 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 can be less accessible compared to commercial LLMs with a larger user base.

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.

Learning from human feedback has become a pivot technique in aligning large language models (LLMs) with human preferences. However, acquiring vast and premium human feedback is bottlenecked by time, labor, and human capability, resulting in small sizes or limited topics of current datasets. This further hinders feedback learning as well as alignment research within the open-source community. To address this issue, we explore how to go beyond human feedback and collect high-quality AI feedback automatically for a scalable alternative. Specifically, we identify scale and diversity as the key factors for feedback data to take effect. Accordingly, we first broaden instructions and responses in both amount and breadth to encompass a wider range of user-assistant interactions. Then, we meticulously apply a series of techniques to mitigate annotation biases for more reliable AI feedback. We finally present UltraFeedback, a large-scale, high-quality, and diversified AI feedback dataset, which contains over 1 million GPT-4 feedback for 250k user-assistant conversations from various aspects. Built upon UltraFeedback, we align a LLaMA-based model by best-of-$n$ sampling and reinforcement learning, demonstrating its exceptional performance on chat benchmarks. Our work validates the effectiveness of scaled AI feedback data in constructing strong open-source chat language models, serving as a solid foundation for future feedback learning research.

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.

User intentions are typically formalized as evaluation rewards to be maximized when fine-tuning language models (LMs). Existing alignment methods, such as Direct Preference Optimization (DPO), are mainly tailored for pairwise preference data where rewards are implicitly defined rather than explicitly given. In this paper, we introduce a general framework for LM alignment, leveraging Noise Contrastive Estimation (NCE) to bridge the gap in handling reward datasets explicitly annotated with scalar evaluations. Our framework comprises two parallel algorithms, NCA and InfoNCA, both enabling the direct extraction of an LM policy from reward data as well as preference data. Notably, we show that the DPO loss is a special case of our proposed InfoNCA objective under pairwise preference settings, thereby integrating and extending current alignment theories. By comparing NCA and InfoNCA, we demonstrate that the well-observed decreasing-likelihood trend of DPO/InfoNCA is caused by their focus on adjusting relative likelihood across different responses. In contrast, NCA optimizes the absolute likelihood for each response, thereby effectively preventing the chosen likelihood from decreasing. We evaluate our methods in both reward and preference settings with Mistral-8$\times$7B and 7B models. Experiments suggest that InfoNCA/NCA surpasses various preference baselines when reward datasets are available. We also find NCA significantly outperforms DPO in complex reasoning tasks like math and coding.

Pre-trained language models (PLMs) may fail in giving reliable estimates of their predictive uncertainty.We take a close look into this 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.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 whether the predictions are correct or not.We highlight that our finding somewhat contradicts 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.Besides unlearnable calibration methods (e.g., label smoothing), we adapt and extend two recently proposed learnable methods that directly collect data to train models to have reasonable confidence estimations.Experimental results show that learnable methods significantly reduce PLMs' confidence in wrong predictions.The code is available at https://github. com/lifan-yuan/PLMCalibration.

Biru Zhu, Lifan Yuan, Ganqu Cui, Yangyi Chen, Chong Fu, Bingxiang He, Yangdong Deng, Zhiyuan Liu, Maosong Sun, Ming Gu. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023.

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 introduceBOSS, 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 pretrained 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 learningmechanism, 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 https://github.com/lifan-yuan/OOD_NLP.

Abstract Recent research has revealed that pre-trained models (PTMs) are vulnerable to backdoor attacks before the fine-tuning stage. The attackers can implant transferable task-agnostic backdoors in PTMs, and control model outputs on any downstream task, which poses severe security threats to all downstream applications. Existing backdoor-removal defenses focus on task-specific classification models and they are not suitable for defending PTMs against task-agnostic backdoor attacks. To this end, we propose the first task-agnostic backdoor removal method for PTMs. Based on the selective activation phenomenon in backdoored PTMs, we design a simple and effective backdoor eraser, which continually pre-trains the backdoored PTMs with a regularization term in an end-to-end approach. The regularization term removes backdoor functionalities from PTMs while the continual pre-training maintains the normal functionalities of PTMs. We conduct extensive experiments on pre-trained models across different modalities and architectures. The experimental results show that our method can effectively remove backdoors inside PTMs and preserve benign functionalities of PTMs with a few downstream-task-irrelevant auxiliary data, e.g., unlabeled plain texts. The average attack success rate on three downstream datasets is reduced from 99.88% to 8.10% after our defense on the backdoored BERT. The codes are publicly available at https://github.com/thunlp/RECIPE.

High-dimensional data analysis for exploration and discovery includes two fundamental tasks: deep clustering and data visualization. When these two associated tasks are done separately, as is often the case thus far, disagreements can occur among the tasks in terms of geometry preservation. Namely, the clustering process is often accompanied by the corruption of the geometric structure, whereas visualization aims to preserve the data geometry for better interpretation. Therefore, how to achieve deep clustering and data visualization in an end-to-end unified framework is an important but challenging problem. In this article, we propose a novel neural network-based method, called deep clustering and visualization (DCV), to accomplish the two associated tasks end-to-end to resolve their disagreements. The DCV framework consists of two nonlinear dimensionality reduction (NLDR) transformations: 1) one from the input data space to latent feature space for clustering and 2) the other from the latent feature space to the final 2-D space for visualization. Importantly, the first NLDR transformation is mainly optimized by one Clustering Loss, allowing arbitrary corruption of the geometric structure for better clustering, while the second NLDR transformation is optimized by one Geometry-Preserving Loss to recover the corrupted geometry for better visualization. Extensive comparative results show that the DCV framework outperforms other leading clustering-visualization algorithms in terms of both quantitative evaluation metrics and qualitative visualization.

Few-shot Named Entity Recognition (NER) is imperative for entity tagging in limited resource domains and thus received proper attention in recent years. Existing approaches for few-shot NER are evaluated mainly under in-domain settings. In contrast, little is known about how these inherently faithful models perform in cross-domain NER using a few labeled in-domain examples. This paper proposes a two-step rationale-centric data augmentation method to improve the model’s generalization ability. Results on several datasets show that our model-agnostic method significantly improves the performance of cross-domain NER tasks compared to previous state-of-the-art methods compared to the counterfactual data augmentation and prompt-tuning methods.

Textual backdoor attacks are a kind of practical threat to NLP systems. By injecting a backdoor in the training phase, the adversary could control model predictions via predefined triggers. As various attack and defense models have been proposed, it is of great significance to perform rigorous evaluations. However, we highlight two issues in previous backdoor learning evaluations: (1) The differences between real-world scenarios (e.g. releasing poisoned datasets or models) are neglected, and we argue that each scenario has its own constraints and concerns, thus requires specific evaluation protocols; (2) The evaluation metrics only consider whether the attacks could flip the models' predictions on poisoned samples and retain performances on benign samples, but ignore that poisoned samples should also be stealthy and semantic-preserving. To address these issues, we categorize existing works into three practical scenarios in which attackers release datasets, pre-trained models, and fine-tuned models respectively, then discuss their unique evaluation methodologies. On metrics, to completely evaluate poisoned samples, we use grammar error increase and perplexity difference for stealthiness, along with text similarity for validity. After formalizing the frameworks, we develop an open-source toolkit OpenBackdoor to foster the implementations and evaluations of textual backdoor learning. With this toolkit, we perform extensive experiments to benchmark attack and defense models under the suggested paradigm. To facilitate the underexplored defenses against poisoned datasets, we further propose CUBE, a simple yet strong clustering-based defense baseline. We hope that our frameworks and benchmarks could serve as the cornerstones for future model development and evaluations.