Ping Yu

Xian Li, Ping Yu, Chunting Zhou et al. · cmu, meta-ai

We present a scalable method to build a high quality instruction following language model by automatically labelling human-written text with corresponding instructions. Our approach, named instruction backtranslation, starts with a language model finetuned on a small amount of seed data, and a given web corpus. The seed model is used to construct training examples by generating instruction prompts for web documents (self-augmentation), and then selecting high quality examples from among these candidates (self-curation). This data is then used to finetune a stronger model. Finetuning LLaMa on two iterations of our approach yields a model that outperforms all other LLaMa-based models on the Alpaca leaderboard not relying on distillation data, demonstrating highly effective self-alignment.

Tianlu Wang, Ping Yu, Xiaoqing Ellen Tan et al. · berkeley, meta-ai

As large language models improve, there is increasing interest in techniques that leverage these models' capabilities to refine their own outputs. In this work, we introduce Shepherd, a language model specifically tuned to critique responses and suggest refinements, extending beyond the capabilities of an untuned model to identify diverse errors and provide suggestions to remedy them. At the core of our approach is a high quality feedback dataset, which we curate from community feedback and human annotations. Even though Shepherd is small (7B parameters), its critiques are either equivalent or preferred to those from established models including ChatGPT. Using GPT-4 for evaluation, Shepherd reaches an average win-rate of 53-87% compared to competitive alternatives. In human evaluation, Shepherd strictly outperforms other models and on average closely ties with ChatGPT.

Srinivasan Iyer, Xi Victoria Lin, Ramakanth Pasunuru et al. · berkeley, meta-ai

Recent work has shown that fine-tuning large pre-trained language models on a collection of tasks described via instructions, a.k.a. instruction-tuning, improves their zero and few-shot generalization to unseen tasks. However, there is a limited understanding of the performance trade-offs of different decisions made during the instruction-tuning process. These decisions include the scale and diversity of the instruction-tuning benchmark, different task sampling strategies, fine-tuning with and without demonstrations, training using specialized datasets for reasoning and dialogue, and finally, the fine-tuning objectives themselves. In this paper, we characterize the effect of instruction-tuning decisions on downstream task performance when scaling both model and benchmark sizes. To this end, we create OPT-IML Bench: a large benchmark for Instruction Meta-Learning (IML) of 2000 NLP tasks consolidated into task categories from 8 existing benchmarks, and prepare an evaluation framework to measure three types of model generalizations: to tasks from fully held-out categories, to held-out tasks from seen categories, and to held-out instances from seen tasks. Through the lens of this framework, we first present insights about instruction-tuning decisions as applied to OPT-30B and further exploit these insights to train OPT-IML 30B and 175B, which are instruction-tuned versions of OPT. OPT-IML demonstrates all three generalization abilities at both scales on four different evaluation benchmarks with diverse tasks and input formats -- PromptSource, FLAN, Super-NaturalInstructions, and UnifiedSKG. Not only does it significantly outperform OPT on all benchmarks but is also highly competitive with existing models fine-tuned on each specific benchmark. We release OPT-IML at both scales, together with the OPT-IML Bench evaluation framework.

Kumar Shridhar, Koustuv Sinha, Andrew Cohen et al. · berkeley, meta-ai

In recent years, Large Language Models (LLMs) have demonstrated remarkable generative abilities, but can they judge the quality of their own generations? A popular concept, referred to as self-refinement, postulates that LLMs can detect and correct the errors in their generations when asked to do so. However, recent empirical evidence points in the opposite direction, suggesting that LLMs often struggle to accurately identify errors when reasoning is involved. To address this, we propose a reasoning with refinement objective called ART: Ask, Refine, and Trust, which asks necessary questions to decide when an LLM should refine its output, and either affirm or withhold trust in its refinement by ranking the refinement and the initial prediction. On two multistep reasoning tasks of mathematical word problems (GSM8K) and question answering (StrategyQA), ART achieves a performance gain of +5 points over self-refinement baselines, while using a much smaller model as the decision maker. We also demonstrate the benefit of using smaller models to make refinement decisions as a cost-effective alternative to fine-tuning a larger model.

Ping Yu, Tianlu Wang, Olga Golovneva et al. · berkeley, meta-ai

Current large language models can perform reasonably well on complex tasks that require step-by-step reasoning with few-shot learning. Are these models applying reasoning skills they have learnt during pre-training and reason outside of their training context, or are they simply memorizing their training corpus at finer granularity and have learnt to better understand their context? To tease apart these possibilities, we introduce ALERT, a benchmark and suite of analyses for assessing language models' reasoning ability comparing pre-trained and finetuned models on complex tasks that require reasoning skills to solve. ALERT provides a test bed to asses any language model on fine-grained reasoning skills, which spans over 20 datasets and covers 10 different reasoning skills. We leverage ALERT to further investigate the role of finetuning. With extensive empirical analysis we find that language models learn more reasoning skills such as textual entailment, abductive reasoning, and analogical reasoning during finetuning stage compared to pretraining state. We also find that when language models are finetuned they tend to overfit to the prompt template, which hurts the robustness of models causing generalization problems.

Guangsheng Yu, Xu Wang, Caijun Sun et al.

Federated learning (FL) provides an effective machine learning (ML) architecture to protect data privacy in a distributed manner. However, the inevitable network asynchrony, the over-dependence on a central coordinator, and the lack of an open and fair incentive mechanism collectively hinder its further development. We propose \textsc{IronForge}, a new generation of FL framework, that features a Directed Acyclic Graph (DAG)-based data structure and eliminates the need for central coordinators to achieve fully decentralized operations. \textsc{IronForge} runs in a public and open network, and launches a fair incentive mechanism by enabling state consistency in the DAG, so that the system fits in networks where training resources are unevenly distributed. In addition, dedicated defense strategies against prevalent FL attacks on incentive fairness and data privacy are presented to ensure the security of \textsc{IronForge}. Experimental results based on a newly developed testbed FLSim highlight the superiority of \textsc{IronForge} to the existing prevalent FL frameworks under various specifications in performance, fairness, and security. To the best of our knowledge, \textsc{IronForge} is the first secure and fully decentralized FL framework that can be applied in open networks with realistic network and training settings.

Tianlu Wang, Ilia Kulikov, Olga Golovneva et al.

Model-based evaluation is at the heart of successful model development -- as a reward model for training, and as a replacement for human evaluation. To train such evaluators, the standard approach is to collect a large amount of human preference judgments over model responses, which is costly and the data becomes stale as models improve. In this work, we present an approach that aims to im-prove evaluators without human annotations, using synthetic training data only. Starting from unlabeled instructions, our iterative self-improvement scheme generates contrasting model outputs and trains an LLM-as-a-Judge to produce reasoning traces and final judgments, repeating this training at each new iteration using the improved predictions. Without any labeled preference data, our Self-Taught Evaluator can improve a strong LLM (Llama3-70B-Instruct) from 75.4 to 88.3 (88.7 with majority vote) on RewardBench. This outperforms commonly used LLM judges such as GPT-4 and matches the performance of the top-performing reward models trained with labeled examples.

Ping Yu, Wei Wang, Chunyuan Li et al.

Prompt tuning has been an extremely effective tool to adapt a pre-trained model to downstream tasks. However, standard prompt-based methods mainly consider the case of sufficient data of downstream tasks. It is still unclear whether the advantage can be transferred to the few-shot regime, where only limited data are available for each downstream task. Although some works have demonstrated the potential of prompt-tuning under the few-shot setting, the main stream methods via searching discrete prompts or tuning soft prompts with limited data are still very challenging. Through extensive empirical studies, we find that there is still a gap between prompt tuning and fully fine-tuning for few-shot learning. To bridge the gap, we propose a new prompt-tuning framework, called Soft Template Tuning (STT). STT combines manual and auto prompts, and treats downstream classification tasks as a masked language modeling task. Comprehensive evaluation on different settings suggests STT can close the gap between fine-tuning and prompt-based methods without introducing additional parameters. Significantly, it can even outperform the time- and resource-consuming fine-tuning method on sentiment classification tasks.

Chenxi Whitehouse, Tianlu Wang, Ping Yu et al.

The progress of AI is bottlenecked by the quality of evaluation, making powerful LLM-as-a-Judge models a core solution. The efficacy of these judges depends on their chain-of-thought reasoning, creating a critical need for methods that can effectively optimize this reasoning process. In this work, we introduce J1, a reinforcement learning framework for teaching LLM judges to think before making decisions. Our core contribution lies in converting all judgment tasks for non-verifiable and verifiable prompts into a unified format with verifiable rewards, enabling direct optimization of evaluation quality while mitigating positional bias. We then use RL to train thinking-judges at scales of 8B, 32B, and 70B and show that they obtain state-of-the-art performance across multiple benchmarks. In particular, J1-Qwen-32B, our multitasked pointwise and pairwise judge also outperforms o1-mini, o3, and a much larger 671B DeepSeek-R1 on some benchmarks, while only training on synthetic data. Through comprehensive ablations of pairwise, pointwise, and multitask J1 variants, we demonstrate the effectiveness of our approach across seed prompts, reward strategies, and training recipes. Qualitative analysis reveals that J1 develops systematic evaluation strategies, including dynamic criteria generation, reference answer creation, iterative self-correction of initial assessments, and feedback generation for low-quality responses.

Silin Gao, Jane Dwivedi-Yu, Ping Yu et al. · berkeley, meta-ai

To achieve faithful reasoning that aligns with human expectations, large language models (LLMs) need to ground their reasoning to real-world knowledge (e.g., web facts, math and physical rules). Tools help LLMs access this external knowledge, but there remains challenges for fine-tuning LLM agents (e.g., Toolformer) to invoke tools in multi-step reasoning problems, where inter-connected tool calls require holistic and efficient tool usage planning. In this work, we propose a new method for LLMs to better leverage tools in multi-step reasoning. Our method, Chain-of-Abstraction (CoA), trains LLMs to first decode reasoning chains with abstract placeholders, and then call domain tools to reify each reasoning chain by filling in specific knowledge. This planning with abstract chains enables LLMs to learn more general reasoning strategies, which are robust to shifts of domain knowledge (e.g., math results) relevant to different reasoning questions. It also allows LLMs to perform decoding and calling of external tools in parallel, which avoids the inference delay caused by waiting for tool responses. In mathematical reasoning and Wiki QA domains, we show that our method consistently outperforms previous chain-of-thought and tool-augmented baselines on both in-distribution and out-of-distribution test sets, with an average ~6% absolute QA accuracy improvement. LLM agents trained with our method also show more efficient tool use, with inference speed being on average ~1.4x faster than baseline tool-augmented LLMs.

Jack Lanchantin, Angelica Chen, Shehzaad Dhuliawala et al. · meta-ai

Post-training of language models, either through reinforcement learning, preference optimization or supervised finetuning, tends to sharpen the output probability distribution and reduce the diversity of generated responses. This is particularly a problem for creative generative tasks where varied responses are desired. In this work we introduce Diverse Preference Optimization (DivPO), an optimization method which learns to generate much more diverse responses than standard pipelines, while maintaining the quality of the generations. In DivPO, preference pairs are selected by first considering a pool of responses, and a measure of diversity among them, and selecting chosen examples as being more rare but high quality, while rejected examples are more common, but low quality. DivPO results in generating 45.6% more diverse persona attributes, and a 74.6% increase in story diversity, while maintaining similar win rates as standard baselines. On general instruction following, DivPO results in a 46.2% increase in diversity, and a 2.4% winrate improvement compared to DPO.

Jack Lanchantin, Angelica Chen, Janice Lan et al. · meta-ai

We investigate the effectiveness of reinforcement learning methods for finetuning large language models when transitioning from offline to semi-online to fully online regimes for both verifiable and non-verifiable tasks. Our experiments cover training on verifiable math as well as non-verifiable instruction following with a set of benchmark evaluations for both. Across these settings, we extensively compare online and semi-online Direct Preference Optimization and Group Reward Policy Optimization objectives, and surprisingly find similar performance and convergence between these variants, which all strongly outperform offline methods. We provide a detailed analysis of the training dynamics and hyperparameter selection strategies to achieve optimal results. Finally, we show that multi-tasking with verifiable and non-verifiable rewards jointly yields improved performance across both task types.

Ping Yu, Weizhe Yuan, Olga Golovneva et al.

Training data quality is one of the most important drivers of final model quality. In this work, we introduce a method for evaluating data integrity based on the assumption that low-quality input prompts result in high variance and low quality responses. This is achieved by measuring the rejected response quality and the reward gap between the chosen and rejected preference pair. Our method, Rejecting Instruction Preferences (RIP) can be used to filter prompts from existing training sets, or to make high quality synthetic datasets, yielding large performance gains across various benchmarks compared to unfiltered data. Using Llama 3.1-8B-Instruct, RIP improves AlpacaEval2 LC Win Rate by 9.4%, Arena-Hard by 8.7%, and WildBench by 9.9%. Using Llama 3.3-70B-Instruct, RIP improves Arena-Hard from 67.5 to 82.9, which is from 18th place to 6th overall in the leaderboard.

Shehzaad Dhuliawala, Ilia Kulikov, Ping Yu et al. · meta-ai

During language model decoding, it is known that using higher temperature sampling gives more creative responses, while lower temperatures are more factually accurate. However, such models are commonly applied to general instruction following, which involves both creative and fact seeking tasks, using a single fixed temperature across all examples and tokens. In this work, we introduce Adaptive Decoding, a layer added to the model to select the sampling temperature dynamically at inference time, at either the token or example level, in order to optimize performance. To learn its parameters we introduce Latent Preference Optimization (LPO) a general approach to train discrete latent variables such as choices of temperature. Our method outperforms all fixed decoding temperatures across a range of tasks that require different temperatures, including UltraFeedback, Creative Story Writing, and GSM8K.

Ping Yu, Jack Lanchantin, Tianlu Wang et al. · meta-ai

We propose CoT-Self-Instruct, a synthetic data generation method that instructs LLMs to first reason and plan via Chain-of-Thought (CoT) based on given seed tasks, and then generate a new synthetic example of similar quality and complexity. This is followed by a filtering step to select high-quality data using automatic metrics, which are then used for LLM training. In verifiable reasoning, our synthetic data significantly outperforms existing training datasets, such as s1k and OpenMathReasoning, when evaluated on MATH500, AMC23, AIME24, and GPQA-Diamond. For non-verifiable instruction-following tasks, our method surpasses the performance of both human and standard Self-Instruct training data on the AlpacaEval 2.0 and Arena-Hard benchmarks.

Leitian Tao, Ilia Kulikov, Swarnadeep Saha et al.

Post-training for reasoning of large language models (LLMs) increasingly relies on verifiable rewards: deterministic checkers that provide 0-1 correctness signals. While reliable, such binary feedback is brittle--many tasks admit partially correct or alternative answers that verifiers under-credit, and the resulting all-or-nothing supervision limits learning. Reward models offer richer, continuous feedback, which can serve as a complementary supervisory signal to verifiers. We introduce HERO (Hybrid Ensemble Reward Optimization), a reinforcement learning framework that integrates verifier signals with reward-model scores in a structured way. HERO employs stratified normalization to bound reward-model scores within verifier-defined groups, preserving correctness while refining quality distinctions, and variance-aware weighting to emphasize challenging prompts where dense signals matter most. Across diverse mathematical reasoning benchmarks, HERO consistently outperforms RM-only and verifier-only baselines, with strong gains on both verifiable and hard-to-verify tasks. Our results show that hybrid reward design retains the stability of verifiers while leveraging the nuance of reward models to advance reasoning.

Badr AlKhamissi, Siddharth Verma, Ping Yu et al.

In this paper, we conduct a thorough investigation into the reasoning capabilities of Large Language Models (LLMs), focusing specifically on the Open Pretrained Transformers (OPT) models as a representative of such models. Our study entails finetuning three different sizes of OPT on a carefully curated reasoning corpus, resulting in two sets of finetuned models: OPT-R, finetuned without explanations, and OPT-RE, finetuned with explanations. We then evaluate all models on 57 out-of-domain tasks drawn from the SUPER-NATURALINSTRUCTIONS benchmark, covering 26 distinct reasoning skills, utilizing three prompting techniques. Through a comprehensive grid of 27 configurations and 6,156 test evaluations, we investigate the dimensions of finetuning, prompting, and scale to understand the role of explanations on different reasoning skills. Our findings reveal that having explanations in the fewshot exemplar has no significant impact on the model's performance when the model is finetuned, while positively affecting the non-finetuned counterpart. Moreover, we observe a slight yet consistent increase in classification accuracy as we incorporate explanations during prompting and finetuning, respectively. Finally, we offer insights on which skills benefit the most from incorporating explanations during finetuning and prompting, such as Numerical (+20.4%) and Analogical (+13.9%) reasoning, as well as skills that exhibit negligible or negative effects.

Yanna Jiang, Baihe Ma, Xu Wang et al.

The demand for intelligent industries and smart services based on big data is rising rapidly with the increasing digitization and intelligence of the modern world. This survey comprehensively reviews Blockchained Federated Learning (BlockFL) that joins the benefits of both Blockchain and Federated Learning to provide a secure and efficient solution for the demand. We compare the existing BlockFL models in four Internet-of-Things (IoT) application scenarios: Personal IoT (PIoT), Industrial IoT (IIoT), Internet of Vehicles (IoV), and Internet of Health Things (IoHT), with a focus on security and privacy, trust and reliability, efficiency, and data heterogeneity. Our analysis shows that the features of decentralization and transparency make BlockFL a secure and effective solution for distributed model training, while the overhead and compatibility still need further study. It also reveals the unique challenges of each domain presents unique challenges, e.g., the requirement of accommodating dynamic environments in IoV and the high demands of identity and permission management in IoHT, in addition to some common challenges identified, such as privacy, resource constraints, and data heterogeneity. Furthermore, we examine the existing technologies that can benefit BlockFL, thereby helping researchers and practitioners to make informed decisions about the selection and development of BlockFL for various IoT application scenarios.

Yang Zhao, Chunyuan Li, Ping Yu et al.

Few-shot learning features the capability of generalizing from a few examples. In this paper, we first identify that a discriminative feature space, namely a rectified metric space, that is learned to maintain the metric consistency from training to testing, is an essential component to the success of metric-based few-shot learning. Numerous analyses indicate that a simple modification of the objective can yield substantial performance gains. The resulting approach, called rectified metric propagation (ReMP), further optimizes an attentive prototype propagation network, and applies a repulsive force to make confident predictions. Extensive experiments demonstrate that the proposed ReMP is effective and efficient, and outperforms the state of the arts on various standard few-shot learning datasets.

Ping Yu, Yang Zhao, Chunyuan Li et al.

Generating long-range skeleton-based human actions has been a challenging problem since small deviations of one frame can cause a malformed action sequence. Most existing methods borrow ideas from video generation, which naively treat skeleton nodes/joints as pixels of images without considering the rich inter-frame and intra-frame structure information, leading to potential distorted actions. Graph convolutional networks (GCNs) is a promising way to leverage structure information to learn structure representations. However, directly adopting GCNs to tackle such continuous action sequences both in spatial and temporal spaces is challenging as the action graph could be huge. To overcome this issue, we propose a variant of GCNs to leverage the powerful self-attention mechanism to adaptively sparsify a complete action graph in the temporal space. Our method could dynamically attend to important past frames and construct a sparse graph to apply in the GCN framework, well-capturing the structure information in action sequences. Extensive experimental results demonstrate the superiority of our method on two standard human action datasets compared with existing methods.

Yang Zhao, Ping Yu, Suchismit Mahapatra et al.

Variational autoencoders (VAEs) are essential tools in end-to-end representation learning. However, the sequential text generation common pitfall with VAEs is that the model tends to ignore latent variables with a strong auto-regressive decoder. In this paper, we propose a principled approach to alleviate this issue by applying a discretized bottleneck to enforce an implicit latent feature matching in a more compact latent space. We impose a shared discrete latent space where each input is learned to choose a combination of latent atoms as a regularized latent representation. Our model endows a promising capability to model underlying semantics of discrete sequences and thus provide more interpretative latent structures. Empirically, we demonstrate our model's efficiency and effectiveness on a broad range of tasks, including language modeling, unaligned text style transfer, dialog response generation, and neural machine translation.

Yang Zhao, Chunyuan Li, Ping Yu et al.

The instability in GAN training has been a long-standing problem despite remarkable research efforts. We identify that instability issues stem from difficulties of performing feature matching with mini-batch statistics, due to a fragile balance between the fixed target distribution and the progressively generated distribution. In this work, we propose Feature Quantization (FQ) for the discriminator, to embed both true and fake data samples into a shared discrete space. The quantized values of FQ are constructed as an evolving dictionary, which is consistent with feature statistics of the recent distribution history. Hence, FQ implicitly enables robust feature matching in a compact space. Our method can be easily plugged into existing GAN models, with little computational overhead in training. We apply FQ to 3 representative GAN models on 9 benchmarks: BigGAN for image generation, StyleGAN for face synthesis, and U-GAT-IT for unsupervised image-to-image translation. Extensive experimental results show that the proposed FQ-GAN can improve the FID scores of baseline methods by a large margin on a variety of tasks, achieving new state-of-the-art performance.

Zhenyi Wang, Ping Yu, Yang Zhao et al.

Human-motion generation is a long-standing challenging task due to the requirement of accurately modeling complex and diverse dynamic patterns. Most existing methods adopt sequence models such as RNN to directly model transitions in the original action space. Due to high dimensionality and potential noise, such modeling of action transitions is particularly challenging. In this paper, we focus on skeleton-based action generation and propose to model smooth and diverse transitions on a latent space of action sequences with much lower dimensionality. Conditioned on a latent sequence, actions are generated by a frame-wise decoder shared by all latent action-poses. Specifically, an implicit RNN is defined to model smooth latent sequences, whose randomness (diversity) is controlled by noise from the input. Different from standard action-prediction methods, our model can generate action sequences from pure noise without any conditional action poses. Remarkably, it can also generate unseen actions from mixed classes during training. Our model is learned with a bi-directional generative-adversarial-net framework, which not only can generate diverse action sequences of a particular class or mix classes, but also learns to classify action sequences within the same model. Experimental results show the superiority of our method in both diverse action-sequence generation and classification, relative to existing methods.

Ping Yu, Kaitao Song, Jianfeng Lu

Recently, deep neural networks have significant progress and successful application in various fields, but they are found vulnerable to attack instances, e.g., adversarial examples. State-of-art attack methods can generate attack images by adding small perturbation to the source image. These attack images can fool the classifier but have little impact to human. Therefore, such attack instances are difficult to generate by searching the feature space. How to design an effective and robust generating method has become a spotlight. Inspired by adversarial examples, we propose two novel generative models to produce adaptive attack instances directly, in which conditional generative adversarial network is adopted and distinctive strategy is designed for training. Compared with the common method, such as Fast Gradient Sign Method, our models can reduce the generating cost and improve robustness and has about one fifth running time for producing attack instance.