LongReason: A Synthetic Long-Context Reasoning Benchmark via Context Expansion

Large language models (LLMs) have demonstrated remarkable progress in understanding long-context inputs. However, benchmarks for evaluating the long-context reasoning abilities of LLMs fall behind the pace. Existing benchmarks often focus on a narrow range of tasks or those that do not demand complex reasoning. To address this gap and enable a more comprehensive evaluation of the long-context reasoning capabilities of current LLMs, we propose a new synthetic benchmark, LongReason, which is constructed by synthesizing long-context reasoning questions from a varied set of short-context reasoning questions through context expansion. LongReason consists of 794 multiple-choice reasoning questions with diverse reasoning patterns across three task categories: reading comprehension, logical inference, and mathematical word problems. We evaluate 21 LLMs on LongReason, revealing that most models experience significant performance drops as context length increases. Our further analysis shows that even state-of-the-art LLMs still have significant room for improvement in providing robust reasoning across different tasks. We will open-source LongReason to support the comprehensive evaluation of LLMs' long-context reasoning capabilities.

Agent-R: Training Language Model Agents to Reflect via Iterative Self-Training

Large Language Models (LLMs) agents are increasingly pivotal for addressing complex tasks in interactive environments. Existing work mainly focuses on enhancing performance through behavior cloning from stronger experts, yet such approaches often falter in real-world applications, mainly due to the inability to recover from errors. However, step-level critique data is difficult and expensive to collect. Automating and dynamically constructing self-critique datasets is thus crucial to empowering models with intelligent agent capabilities. In this work, we propose an iterative self-training framework, Agent-R, that enables language Agent to Reflect on the fly. Unlike traditional methods that reward or penalize actions based on correctness, Agent-R leverages MCTS to construct training data that recover correct trajectories from erroneous ones. A key challenge of agent reflection lies in the necessity for timely revision rather than waiting until the end of a rollout. To address this, we introduce a model-guided critique construction mechanism: the actor model identifies the first error step (within its current capability) in a failed trajectory. Starting from it, we splice it with the adjacent correct path, which shares the same parent node in the tree. This strategy enables the model to learn reflection based on its current policy, therefore yielding better learning efficiency. To further explore the scalability of this self-improvement paradigm, we investigate iterative refinement of both error correction capabilities and dataset construction. Our findings demonstrate that Agent-R continuously improves the model's ability to recover from errors and enables timely error correction. Experiments on three interactive environments show that Agent-R effectively equips agents to correct erroneous actions while avoiding loops, achieving superior performance compared to baseline methods (+5.59%).

ToolHop: A Query-Driven Benchmark for Evaluating Large Language Models in Multi-Hop Tool Use

Effective evaluation of multi-hop tool use is critical for analyzing the understanding, reasoning, and function-calling capabilities of large language models (LLMs). However, progress has been hindered by a lack of reliable evaluation datasets. To address this, we present ToolHop, a dataset comprising 995 user queries and 3,912 associated tools, specifically designed for rigorous evaluation of multi-hop tool use. ToolHop ensures diverse queries, meaningful interdependencies, locally executable tools, detailed feedback, and verifiable answers through a novel query-driven data construction approach that includes tool creation, document refinement, and code generation. We evaluate 14 LLMs across five model families (i.e., LLaMA3.1, Qwen2.5, Gemini1.5, Claude3.5, and GPT), uncovering significant challenges in handling multi-hop tool-use scenarios. The leading model, GPT-4o, achieves an accuracy of 49.04%, underscoring substantial room for improvement. Further analysis reveals variations in tool-use strategies for various families, offering actionable insights to guide the development of more effective approaches. Code and data can be found in https://huggingface.co/datasets/bytedance-research/ToolHop.

FiD-Light: Efficient and Effective Retrieval-Augmented Text Generation

Retrieval-augmented generation models offer many benefits over standalone language models: besides a textual answer to a given query they provide provenance items retrieved from an updateable knowledge base. However, they are also more complex systems and need to handle long inputs. In this work, we introduce FiD-Light to strongly increase the efficiency of the state-of-the-art retrieval-augmented FiD model, while maintaining the same level of effectiveness. Our FiD-Light model constrains the information flow from the encoder (which encodes passages separately) to the decoder (using concatenated encoded representations). Furthermore, we adapt FiD-Light with re-ranking capabilities through textual source pointers, to improve the top-ranked provenance precision. Our experiments on a diverse set of seven knowledge intensive tasks (KILT) show FiD-Light consistently improves the Pareto frontier between query latency and effectiveness. FiD-Light with source pointing sets substantial new state-of-the-art results on six KILT tasks for combined text generation and provenance retrieval evaluation, while maintaining reasonable efficiency.

Multi-Task Retrieval-Augmented Text Generation with Relevance Sampling

This paper studies multi-task training of retrieval-augmented generation models for knowledge-intensive tasks. We propose to clean the training set by utilizing a distinct property of knowledge-intensive generation: The connection of query-answer pairs to items in the knowledge base. We filter training examples via a threshold of confidence on the relevance labels, whether a pair is answerable by the knowledge base or not. We train a single Fusion-in-Decoder (FiD) generator on seven combined tasks of the KILT benchmark. The experimental results suggest that our simple yet effective approach substantially improves competitive baselines on two strongly imbalanced tasks; and shows either smaller improvements or no significant regression on the remaining tasks. Furthermore, we demonstrate our multi-task training with relevance label sampling scales well with increased model capacity and achieves state-of-the-art results in five out of seven KILT tasks.

Transforming Sequence Tagging Into A Seq2Seq Task

Pretrained, large, generative language models (LMs) have had great success in a wide range of sequence tagging and structured prediction tasks. Casting a sequence tagging task as a Seq2Seq one requires deciding the formats of the input and output sequences. However, we lack a principled understanding of the trade-offs associated with these formats (such as the effect on model accuracy, sequence length, multilingual generalization, hallucination). In this paper, we rigorously study different formats one could use for casting input text sentences and their output labels into the input and target (i.e., output) of a Seq2Seq model. Along the way, we introduce a new format, which we show to not only be simpler but also more effective. Additionally the new format demonstrates significant gains in the multilingual settings -- both zero-shot transfer learning and joint training. Lastly, we find that the new format is more robust and almost completely devoid of hallucination -- an issue we find common in existing formats. With well over a 1000 experiments studying 14 different formats, over 7 diverse public benchmarks -- including 3 multilingual datasets spanning 7 languages -- we believe our findings provide a strong empirical basis in understanding how we should tackle sequence tagging tasks.

WIT: Wikipedia-based Image Text Dataset for Multimodal Multilingual Machine Learning

The milestone improvements brought about by deep representation learning and pre-training techniques have led to large performance gains across downstream NLP, IR and Vision tasks. Multimodal modeling techniques aim to leverage large high-quality visio-linguistic datasets for learning complementary information (across image and text modalities). In this paper, we introduce the Wikipedia-based Image Text (WIT) Dataset (https://github.com/google-research-datasets/wit) to better facilitate multimodal, multilingual learning. WIT is composed of a curated set of 37.6 million entity rich image-text examples with 11.5 million unique images across 108 Wikipedia languages. Its size enables WIT to be used as a pretraining dataset for multimodal models, as we show when applied to downstream tasks such as image-text retrieval. WIT has four main and unique advantages. First, WIT is the largest multimodal dataset by the number of image-text examples by 3x (at the time of writing). Second, WIT is massively multilingual (first of its kind) with coverage over 100+ languages (each of which has at least 12K examples) and provides cross-lingual texts for many images. Third, WIT represents a more diverse set of concepts and real world entities relative to what previous datasets cover. Lastly, WIT provides a very challenging real-world test set, as we empirically illustrate using an image-text retrieval task as an example.

DICT-MLM: Improved Multilingual Pre-Training using Bilingual Dictionaries

Pre-trained multilingual language models such as mBERT have shown immense gains for several natural language processing (NLP) tasks, especially in the zero-shot cross-lingual setting. Most, if not all, of these pre-trained models rely on the masked-language modeling (MLM) objective as the key language learning objective. The principle behind these approaches is that predicting the masked words with the help of the surrounding text helps learn potent contextualized representations. Despite the strong representation learning capability enabled by MLM, we demonstrate an inherent limitation of MLM for multilingual representation learning. In particular, by requiring the model to predict the language-specific token, the MLM objective disincentivizes learning a language-agnostic representation -- which is a key goal of multilingual pre-training. Therefore to encourage better cross-lingual representation learning we propose the DICT-MLM method. DICT-MLM works by incentivizing the model to be able to predict not just the original masked word, but potentially any of its cross-lingual synonyms as well. Our empirical analysis on multiple downstream tasks spanning 30+ languages, demonstrates the efficacy of the proposed approach and its ability to learn better multilingual representations.

DiPair: Fast and Accurate Distillation for Trillion-Scale Text Matching and Pair Modeling

Pre-trained models like BERT (Devlin et al., 2018) have dominated NLP / IR applications such as single sentence classification, text pair classification, and question answering. However, deploying these models in real systems is highly non-trivial due to their exorbitant computational costs. A common remedy to this is knowledge distillation (Hinton et al., 2015), leading to faster inference. However -- as we show here -- existing works are not optimized for dealing with pairs (or tuples) of texts. Consequently, they are either not scalable or demonstrate subpar performance. In this work, we propose DiPair -- a novel framework for distilling fast and accurate models on text pair tasks. Coupled with an end-to-end training strategy, DiPair is both highly scalable and offers improved quality-speed tradeoffs. Empirical studies conducted on both academic and real-world e-commerce benchmarks demonstrate the efficacy of the proposed approach with speedups of over 350x and minimal quality drop relative to the cross-attention teacher BERT model.

Sampled Softmax with Random Fourier Features

The computational cost of training with softmax cross entropy loss grows linearly with the number of classes. For the settings where a large number of classes are involved, a common method to speed up training is to sample a subset of classes and utilize an estimate of the gradient based on these classes, known as the sampled softmax method. However, the sampled softmax provides a biased estimate of the gradient unless the samples are drawn from the exact softmax distribution, which is again expensive to compute. Therefore, a widely employed practical approach (without theoretical justification) involves sampling from a simpler distribution in the hope of approximating the exact softmax distribution. In this paper, we develop the first theoretical understanding of the role that different sampling distributions play in determining the quality of sampled softmax. Motivated by our analysis and the work on kernel-based sampling, we propose the Random Fourier Softmax (RF-softmax) method that utilizes the powerful Random Fourier features to enable more efficient and accurate sampling from the (approximate) softmax distribution. We show that RF-softmax leads to low bias in estimation in terms of both the full softmax distribution and the full softmax gradient. Furthermore, the cost of RF-softmax scales only logarithmically with the number of classes.