Shyam Upadhyay

Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao et al. · allen-ai, amazon-science

Language models demonstrate both quantitative improvement and new qualitative capabilities with increasing scale. Despite their potentially transformative impact, these new capabilities are as yet poorly characterized. In order to inform future research, prepare for disruptive new model capabilities, and ameliorate socially harmful effects, it is vital that we understand the present and near-future capabilities and limitations of language models. To address this challenge, we introduce the Beyond the Imitation Game benchmark (BIG-bench). BIG-bench currently consists of 204 tasks, contributed by 450 authors across 132 institutions. Task topics are diverse, drawing problems from linguistics, childhood development, math, common-sense reasoning, biology, physics, social bias, software development, and beyond. BIG-bench focuses on tasks that are believed to be beyond the capabilities of current language models. We evaluate the behavior of OpenAI's GPT models, Google-internal dense transformer architectures, and Switch-style sparse transformers on BIG-bench, across model sizes spanning millions to hundreds of billions of parameters. In addition, a team of human expert raters performed all tasks in order to provide a strong baseline. Findings include: model performance and calibration both improve with scale, but are poor in absolute terms (and when compared with rater performance); performance is remarkably similar across model classes, though with benefits from sparsity; tasks that improve gradually and predictably commonly involve a large knowledge or memorization component, whereas tasks that exhibit "breakthrough" behavior at a critical scale often involve multiple steps or components, or brittle metrics; social bias typically increases with scale in settings with ambiguous context, but this can be improved with prompting.

Satyapriya Krishna, Kalpesh Krishna, Anhad Mohananey et al. · deepmind, harvard

Large Language Models (LLMs) have demonstrated significant performance improvements across various cognitive tasks. An emerging application is using LLMs to enhance retrieval-augmented generation (RAG) capabilities. These systems require LLMs to understand user queries, retrieve relevant information, and synthesize coherent and accurate responses. Given the increasing real-world deployment of such systems, comprehensive evaluation becomes crucial. To this end, we propose FRAMES (Factuality, Retrieval, And reasoning MEasurement Set), a high-quality evaluation dataset designed to test LLMs' ability to provide factual responses, assess retrieval capabilities, and evaluate the reasoning required to generate final answers. While previous work has provided datasets and benchmarks to evaluate these abilities in isolation, FRAMES offers a unified framework that provides a clearer picture of LLM performance in end-to-end RAG scenarios. Our dataset comprises challenging multi-hop questions that require the integration of information from multiple sources. We present baseline results demonstrating that even state-of-the-art LLMs struggle with this task, achieving 0.40 accuracy with no retrieval. The accuracy is significantly improved with our proposed multi-step retrieval pipeline, achieving an accuracy of 0.66 (>50% improvement). We hope our work will help bridge evaluation gaps and assist in developing more robust and capable RAG systems.

Pei Zhou, Aman Madaan, Srividya Pranavi Potharaju et al. · cmu

"Thinking is for Doing." Humans can infer other people's mental states from observations--an ability called Theory-of-Mind (ToM)--and subsequently act pragmatically on those inferences. Existing question answering benchmarks such as ToMi ask models questions to make inferences about beliefs of characters in a story, but do not test whether models can then use these inferences to guide their actions. We propose a new evaluation paradigm for large language models (LLMs): Thinking for Doing (T4D), which requires models to connect inferences about others' mental states to actions in social scenarios. Experiments on T4D demonstrate that LLMs such as GPT-4 and PaLM 2 seemingly excel at tracking characters' beliefs in stories, but they struggle to translate this capability into strategic action. Our analysis reveals the core challenge for LLMs lies in identifying the implicit inferences about mental states without being explicitly asked about as in ToMi, that lead to choosing the correct action in T4D. To bridge this gap, we introduce a zero-shot prompting framework, Foresee and Reflect (FaR), which provides a reasoning structure that encourages LLMs to anticipate future challenges and reason about potential actions. FaR boosts GPT-4's performance from 50% to 71% on T4D, outperforming other prompting methods such as Chain-of-Thought and Self-Ask. Moreover, FaR generalizes to diverse out-of-distribution story structures and scenarios that also require ToM inferences to choose an action, consistently outperforming other methods including few-shot in-context learning.

Pranjal Aggarwal, Aman Madaan, Ankit Anand et al. · cmu

Large language models (LLMs) are now available from cloud API providers in various sizes and configurations. While this diversity offers a broad spectrum of choices, effectively leveraging the options to optimize computational cost and performance remains challenging. In this work, we present Automix, an approach that strategically routes queries to larger LMs, based on the approximate correctness of outputs from a smaller LM. Central to Automix are two key technical contributions. First, it has a few-shot self-verification mechanism, which estimates the reliability of its own outputs without requiring extensive training. Second, given that self-verification can be noisy, it employs a POMDP based router that can effectively select an appropriately sized model, based on answer confidence. Experiments across five language models and five challenging datasets show that Automix consistently surpasses strong baselines, reducing computational cost by over 50% for comparable performance.

Jingfeng Yang, Aditya Gupta, Shyam Upadhyay et al. · amazon-science

Understanding tables is an important aspect of natural language understanding. Existing models for table understanding require linearization of the table structure, where row or column order is encoded as an unwanted bias. Such spurious biases make the model vulnerable to row and column order perturbations. Additionally, prior work has not thoroughly modeled the table structures or table-text alignments, hindering the table-text understanding ability. In this work, we propose a robust and structurally aware table-text encoding architecture TableFormer, where tabular structural biases are incorporated completely through learnable attention biases. TableFormer is (1) strictly invariant to row and column orders, and, (2) could understand tables better due to its tabular inductive biases. Our evaluations showed that TableFormer outperforms strong baselines in all settings on SQA, WTQ and TabFact table reasoning datasets, and achieves state-of-the-art performance on SQA, especially when facing answer-invariant row and column order perturbations (6% improvement over the best baseline), because previous SOTA models' performance drops by 4% - 6% when facing such perturbations while TableFormer is not affected.

Shuo-yiin Chang, Guru Prakash, Zelin Wu et al.

In voice-enabled applications, a predetermined hotword isusually used to activate a device in order to attend to the query.However, speaking queries followed by a hotword each timeintroduces a cognitive burden in continued conversations. Toavoid repeating a hotword, we propose a streaming end-to-end(E2E) intended query detector that identifies the utterancesdirected towards the device and filters out other utterancesnot directed towards device. The proposed approach incor-porates the intended query detector into the E2E model thatalready folds different components of the speech recognitionpipeline into one neural network.The E2E modeling onspeech decoding and intended query detection also allows us todeclare a quick intended query detection based on early partialrecognition result, which is important to decrease latencyand make the system responsive. We demonstrate that theproposed E2E approach yields a 22% relative improvement onequal error rate (EER) for the detection accuracy and 600 mslatency improvement compared with an independent intendedquery detector. In our experiment, the proposed model detectswhether the user is talking to the device with a 8.7% EERwithin 1.4 seconds of median latency after user starts speaking.

Ayush Kaushal, Aditya Gupta, Shyam Upadhyay et al.

A naive application of state-of-the-art bidirectional encoders for streaming sequence tagging would require encoding each token from scratch for each new token in an incremental streaming input (like transcribed speech). The lack of re-usability of previous computation leads to a higher number of Floating Point Operations (or FLOPs) and higher number of unnecessary label flips. Increased FLOPs consequently lead to higher wall-clock time and increased label flipping leads to poorer streaming performance. In this work, we present a Hybrid Encoder with Adaptive Restart (HEAR) that addresses these issues while maintaining the performance of bidirectional encoders over the offline (or complete) inputs while improving performance on streaming (or incomplete) inputs. HEAR has a Hybrid unidirectional-bidirectional encoder architecture to perform sequence tagging, along with an Adaptive Restart Module (ARM) to selectively guide the restart of bidirectional portion of the encoder. Across four sequence tagging tasks, HEAR offers FLOP savings in streaming settings upto 71.1% and also outperforms bidirectional encoders for streaming predictions by upto +10% streaming exact match.

Anmol Agarwal, Jigar Gupta, Rahul Goel et al.

Extending semantic parsers to code-switched input has been a challenging problem, primarily due to a lack of supervised training data. In this work, we introduce CST5, a new data augmentation technique that finetunes a T5 model using a small seed set ($\approx$100 utterances) to generate code-switched utterances from English utterances. We show that CST5 generates high quality code-switched data, both intrinsically (per human evaluation) and extrinsically by comparing baseline models which are trained without data augmentation to models which are trained with augmented data. Empirically we observe that using CST5, one can achieve the same semantic parsing performance by using up to 20x less labeled data. To aid further research in this area, we are also releasing (a) Hinglish-TOP, the largest human annotated code-switched semantic parsing dataset to date, containing 10k human annotated Hindi-English (Hinglish) code-switched utterances, and (b) Over 170K CST5 generated code-switched utterances from the TOPv2 dataset. Human evaluation shows that both the human annotated data as well as the CST5 generated data is of good quality.

Gemini Team, Petko Georgiev, Ving Ian Lei et al. · deepmind, mila

In this report, we introduce the Gemini 1.5 family of models, representing the next generation of highly compute-efficient multimodal models capable of recalling and reasoning over fine-grained information from millions of tokens of context, including multiple long documents and hours of video and audio. The family includes two new models: (1) an updated Gemini 1.5 Pro, which exceeds the February version on the great majority of capabilities and benchmarks; (2) Gemini 1.5 Flash, a more lightweight variant designed for efficiency with minimal regression in quality. Gemini 1.5 models achieve near-perfect recall on long-context retrieval tasks across modalities, improve the state-of-the-art in long-document QA, long-video QA and long-context ASR, and match or surpass Gemini 1.0 Ultra's state-of-the-art performance across a broad set of benchmarks. Studying the limits of Gemini 1.5's long-context ability, we find continued improvement in next-token prediction and near-perfect retrieval (>99%) up to at least 10M tokens, a generational leap over existing models such as Claude 3.0 (200k) and GPT-4 Turbo (128k). Finally, we highlight real-world use cases, such as Gemini 1.5 collaborating with professionals on completing their tasks achieving 26 to 75% time savings across 10 different job categories, as well as surprising new capabilities of large language models at the frontier; when given a grammar manual for Kalamang, a language with fewer than 200 speakers worldwide, the model learns to translate English to Kalamang at a similar level to a person who learned from the same content.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Lianhui Qin, Aditya Gupta, Shyam Upadhyay et al.

Everyday conversations require understanding everyday events, which in turn, requires understanding temporal commonsense concepts interwoven with those events. Despite recent progress with massive pre-trained language models (LMs) such as T5 and GPT-3, their capability of temporal reasoning in dialogs remains largely under-explored. In this paper, we present the first study to investigate pre-trained LMs for their temporal reasoning capabilities in dialogs by introducing a new task and a crowd-sourced English challenge set, TIMEDIAL. We formulate TIME-DIAL as a multiple-choice cloze task with over 1.1K carefully curated dialogs. Empirical results demonstrate that even the best performing models struggle on this task compared to humans, with 23 absolute points of gap in accuracy. Furthermore, our analysis reveals that the models fail to reason about dialog context correctly; instead, they rely on shallow cues based on existing temporal patterns in context, motivating future research for modeling temporal concepts in text and robust contextual reasoning about them. The dataset is publicly available at: https://github.com/google-research-datasets/timedial.

Aditya Gupta, Jiacheng Xu, Shyam Upadhyay et al.

Disfluencies is an under-studied topic in NLP, even though it is ubiquitous in human conversation. This is largely due to the lack of datasets containing disfluencies. In this paper, we present a new challenge question answering dataset, Disfl-QA, a derivative of SQuAD, where humans introduce contextual disfluencies in previously fluent questions. Disfl-QA contains a variety of challenging disfluencies that require a more comprehensive understanding of the text than what was necessary in prior datasets. Experiments show that the performance of existing state-of-the-art question answering models degrades significantly when tested on Disfl-QA in a zero-shot setting.We show data augmentation methods partially recover the loss in performance and also demonstrate the efficacy of using gold data for fine-tuning. We argue that we need large-scale disfluency datasets in order for NLP models to be robust to them. The dataset is publicly available at: https://github.com/google-research-datasets/disfl-qa.

Kartikeya Badola, Jonathan Simon, Arian Hosseini et al.

Large language models (LLMs) excel at solving problems with clear and complete statements, but often struggle with nuanced environments or interactive tasks which are common in most real-world scenarios. This highlights the critical need for developing LLMs that can effectively engage in logically consistent multi-turn dialogue, seek information and reason with incomplete data. To this end, we introduce a novel benchmark comprising a suite of multi-turn tasks each designed to test specific reasoning, interactive dialogue, and information-seeking abilities. These tasks have deterministic scoring mechanisms, thus eliminating the need for human intervention. Evaluating frontier models on our benchmark reveals significant headroom. Our analysis shows that most errors emerge from poor instruction following, reasoning failures, and poor planning. This benchmark provides valuable insights into the strengths and weaknesses of current LLMs in handling complex, interactive scenarios and offers a robust platform for future research aimed at improving these critical capabilities.

Xinliang Frederick Zhang, Anhad Mohananey, Alexandra Chronopoulou et al.

Models employing long chain-of-thought (CoT) reasoning have shown superior performance on complex reasoning tasks. Yet, this capability introduces a critical and often overlooked inefficiency -- overthinking -- models often engage in unnecessarily extensive reasoning even for simple queries, incurring significant computations without accuracy improvements. While prior work has explored solutions to mitigate overthinking, a fundamental gap remains in our understanding of its underlying causes. Most existing analyses are limited to superficial, profiling-based observations, failing to delve into LLMs' inner workings. This study introduces a systematic, fine-grained analyzer of LLMs' thought process to bridge the gap, TRACE. We first benchmark the overthinking issue, confirming that long-thinking models are five to twenty times slower on simple tasks with no substantial gains. We then use TRACE to first decompose the thought process into minimally complete sub-thoughts. Next, by inferring discourse relationships among sub-thoughts, we construct granular thought progression graphs and subsequently identify common thinking patterns for topically similar queries. Our analysis reveals two major patterns for open-weight thinking models -- Explorer and Late Landing. This finding provides evidence that over-verification and over-exploration are the primary drivers of overthinking in LLMs. Grounded in thought structures, we propose a utility-based definition of overthinking, which moves beyond length-based metrics. This revised definition offers a more insightful understanding of LLMs' thought progression, as well as practical guidelines for principled overthinking management.

Ming Zhong, Xiang Zhou, Ting-Yun Chang et al.

Large Language Models (LLMs) have catalyzed vibe coding, where users leverage LLMs to generate and iteratively refine code through natural language interactions until it passes their vibe check. Vibe check is tied to real-world human preference and goes beyond functionality: the solution should feel right, read cleanly, preserve intent, and remain correct. However, current code evaluation remains anchored to pass@k and captures only functional correctness, overlooking the non-functional instructions that users routinely apply. In this paper, we hypothesize that instruction following is the missing piece underlying vibe check that represents human preference in coding besides functional correctness. To quantify models' code instruction following capabilities with measurable signals, we present VeriCode, a taxonomy of 30 verifiable code instructions together with corresponding deterministic verifiers. We use the taxonomy to augment established evaluation suites, resulting in Vibe Checker, a testbed to assess both code instruction following and functional correctness. Upon evaluating 31 leading LLMs, we show that even the strongest models struggle to comply with multiple instructions and exhibit clear functional regression. Most importantly, a composite score of functional correctness and instruction following correlates the best with human preference, with the latter emerging as the primary differentiator on real-world programming tasks. Our work identifies core factors of the vibe check, providing a concrete path for benchmarking and developing models that better align with user preferences in coding.

Gemini Team, Rohan Anil, Sebastian Borgeaud et al.

This report introduces a new family of multimodal models, Gemini, that exhibit remarkable capabilities across image, audio, video, and text understanding. The Gemini family consists of Ultra, Pro, and Nano sizes, suitable for applications ranging from complex reasoning tasks to on-device memory-constrained use-cases. Evaluation on a broad range of benchmarks shows that our most-capable Gemini Ultra model advances the state of the art in 30 of 32 of these benchmarks - notably being the first model to achieve human-expert performance on the well-studied exam benchmark MMLU, and improving the state of the art in every one of the 20 multimodal benchmarks we examined. We believe that the new capabilities of the Gemini family in cross-modal reasoning and language understanding will enable a wide variety of use cases. We discuss our approach toward post-training and deploying Gemini models responsibly to users through services including Gemini, Gemini Advanced, Google AI Studio, and Cloud Vertex AI.

Shikhar Vashishth, Shyam Upadhyay, Gaurav Singh Tomar et al.

The attention layer in a neural network model provides insights into the model's reasoning behind its prediction, which are usually criticized for being opaque. Recently, seemingly contradictory viewpoints have emerged about the interpretability of attention weights (Jain & Wallace, 2019; Vig & Belinkov, 2019). Amid such confusion arises the need to understand attention mechanism more systematically. In this work, we attempt to fill this gap by giving a comprehensive explanation which justifies both kinds of observations (i.e., when is attention interpretable and when it is not). Through a series of experiments on diverse NLP tasks, we validate our observations and reinforce our claim of interpretability of attention through manual evaluation.

Shyam Upadhyay, Jordan Kodner, Dan Roth

Generating the English transliteration of a name written in a foreign script is an important and challenging step in multilingual knowledge acquisition and information extraction. Existing approaches to transliteration generation require a large (>5000) number of training examples. This difficulty contrasts with transliteration discovery, a somewhat easier task that involves picking a plausible transliteration from a given list. In this work, we present a bootstrapping algorithm that uses constrained discovery to improve generation, and can be used with as few as 500 training examples, which we show can be sourced from annotators in a matter of hours. This opens the task to languages for which large number of training examples are unavailable. We evaluate transliteration generation performance itself, as well the improvement it brings to cross-lingual candidate generation for entity linking, a typical downstream task. We present a comprehensive evaluation of our approach on nine languages, each written in a unique script.

Shyam Upadhyay, Nitish Gupta, Dan Roth

Cross-lingual Entity Linking (XEL) aims to ground entity mentions written in any language to an English Knowledge Base (KB), such as Wikipedia. XEL for most languages is challenging, owing to limited availability of resources as supervision. We address this challenge by developing the first XEL approach that combines supervision from multiple languages jointly. This enables our approach to: (a) augment the limited supervision in the target language with additional supervision from a high-resource language (like English), and (b) train a single entity linking model for multiple languages, improving upon individually trained models for each language. Extensive evaluation on three benchmark datasets across 8 languages shows that our approach significantly improves over the current state-of-the-art. We also provide analyses in two limited resource settings: (a) zero-shot setting, when no supervision in the target language is available, and in (b) low-resource setting, when some supervision in the target language is available. Our analysis provides insights into the limitations of zero-shot XEL approaches in realistic scenarios, and shows the value of joint supervision in low-resource settings.

Shyam Upadhyay, Yogarshi Vyas, Marine Carpuat et al.

Cross-lingual Hypernymy Detection involves determining if a word in one language ("fruit") is a hypernym of a word in another language ("pomme" i.e. apple in French). The ability to detect hypernymy cross-lingually can aid in solving cross-lingual versions of tasks such as textual entailment and event coreference. We propose BISPARSE-DEP, a family of unsupervised approaches for cross-lingual hypernymy detection, which learns sparse, bilingual word embeddings based on dependency contexts. We show that BISPARSE-DEP can significantly improve performance on this task, compared to approaches based only on lexical context. Our approach is also robust, showing promise for low-resource settings: our dependency-based embeddings can be learned using a parser trained on related languages, with negligible loss in performance. We also crowd-source a challenging dataset for this task on four languages -- Russian, French, Arabic, and Chinese. Our embeddings and datasets are publicly available.

Shyam Upadhyay, Kai-Wei Chang, Matt Taddy et al.

Word embeddings, which represent a word as a point in a vector space, have become ubiquitous to several NLP tasks. A recent line of work uses bilingual (two languages) corpora to learn a different vector for each sense of a word, by exploiting crosslingual signals to aid sense identification. We present a multi-view Bayesian non-parametric algorithm which improves multi-sense word embeddings by (a) using multilingual (i.e., more than two languages) corpora to significantly improve sense embeddings beyond what one achieves with bilingual information, and (b) uses a principled approach to learn a variable number of senses per word, in a data-driven manner. Ours is the first approach with the ability to leverage multilingual corpora efficiently for multi-sense representation learning. Experiments show that multilingual training significantly improves performance over monolingual and bilingual training, by allowing us to combine different parallel corpora to leverage multilingual context. Multilingual training yields comparable performance to a state of the art mono-lingual model trained on five times more training data.

Subhro Roy, Shyam Upadhyay, Dan Roth

Identifying mathematical relations expressed in text is essential to understanding a broad range of natural language text from election reports, to financial news, to sport commentaries to mathematical word problems. This paper focuses on identifying and understanding mathematical relations described within a single sentence. We introduce the problem of Equation Parsing -- given a sentence, identify noun phrases which represent variables, and generate the mathematical equation expressing the relation described in the sentence. We introduce the notion of projective equation parsing and provide an efficient algorithm to parse text to projective equations. Our system makes use of a high precision lexicon of mathematical expressions and a pipeline of structured predictors, and generates correct equations in $70\%$ of the cases. In $60\%$ of the time, it also identifies the correct noun phrase $\rightarrow$ variables mapping, significantly outperforming baselines. We also release a new annotated dataset for task evaluation.

Shyam Upadhyay, Ming-Wei Chang

We propose a new evaluation for automatic solvers for algebra word problems, which can identify mistakes that existing evaluations overlook. Our proposal is to evaluate such solvers using derivations, which reflect how an equation system was constructed from the word problem. To accomplish this, we develop an algorithm for checking the equivalence between two derivations, and show how derivation an- notations can be semi-automatically added to existing datasets. To make our experiments more comprehensive, we include the derivation annotation for DRAW-1K, a new dataset containing 1000 general algebra word problems. In our experiments, we found that the annotated derivations enable a more accurate evaluation of automatic solvers than previously used metrics. We release derivation annotations for over 2300 algebra word problems for future evaluations.

Shyam Upadhyay, Manaal Faruqui, Chris Dyer et al.

Despite interest in using cross-lingual knowledge to learn word embeddings for various tasks, a systematic comparison of the possible approaches is lacking in the literature. We perform an extensive evaluation of four popular approaches of inducing cross-lingual embeddings, each requiring a different form of supervision, on four typographically different language pairs. Our evaluation setup spans four different tasks, including intrinsic evaluation on mono-lingual and cross-lingual similarity, and extrinsic evaluation on downstream semantic and syntactic applications. We show that models which require expensive cross-lingual knowledge almost always perform better, but cheaply supervised models often prove competitive on certain tasks.

Kai-Wei Chang, Shyam Upadhyay, Ming-Wei Chang et al.

IllinoisSL is a Java library for learning structured prediction models. It supports structured Support Vector Machines and structured Perceptron. The library consists of a core learning module and several applications, which can be executed from command-lines. Documentation is provided to guide users. In Comparison to other structured learning libraries, IllinoisSL is efficient, general, and easy to use.

Ching-pei Lee, Kai-Wei Chang, Shyam Upadhyay et al.

Training structured prediction models is time-consuming. However, most existing approaches only use a single machine, thus, the advantage of computing power and the capacity for larger data sets of multiple machines have not been exploited. In this work, we propose an efficient algorithm for distributedly training structured support vector machines based on a distributed block-coordinate descent method. Both theoretical and experimental results indicate that our method is efficient.