Thang Luong, Dawsen Hwang, Hoang H Nguyen, Golnaz Ghiasi, Yuri Chervonyi, Insuk Seo, Junsu Kim, Garrett Bingham, Jonathan Lee, Swaroop Mishra, Alex Zhai, Huiyi Hu, Henryk Michalewski, Jimin Kim, Jeonghyun Ahn, Junhwi Bae, Xingyou Song, Trieu Hoang Trinh, Quoc V Le, Junehyuk Jung. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Mihir Parmar, Xin Liu, Palash Goyal, Yanfei Chen, Long Le, Swaroop Mishra, Hossein Mobahi, Jindong Gu, Zifeng Wang, Hootan Nakhost, Chitta Baral, Chen-Yu Lee, Tomas Pfister, Hamid Palangi. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Retrieval augmented generation (RAG) combines the generative abilities of large language models (LLMs) with external knowledge sources to provide more accurate and up-to-date responses. Recent RAG advancements focus on improving retrieval outcomes through iterative LLM refinement or self-critique capabilities acquired through additional instruction tuning of LLMs. In this work, we introduce Speculative RAG - a framework that leverages a larger generalist LM to efficiently verify multiple RAG drafts produced in parallel by a smaller, distilled specialist LM. Each draft is generated from a distinct subset of retrieved documents, offering diverse perspectives on the evidence while reducing input token counts per draft. This approach enhances comprehension of each subset and mitigates potential position bias over long context. Our method accelerates RAG by delegating drafting to the smaller specialist LM, with the larger generalist LM performing a single verification pass over the drafts. Extensive experiments demonstrate that Speculative RAG achieves state-of-the-art performance with reduced latency on TriviaQA, MuSiQue, PopQA, PubHealth, and ARC-Challenge benchmarks. It notably enhances accuracy by up to 12.97% while reducing latency by 50.83% compared to conventional RAG systems on PubHealth.

We demonstrate the ability of large language models (LLMs) to perform iterative self-improvement of robot policies. An important insight of this paper is that LLMs have a built-in ability to perform (stochastic) numerical optimization and that this property can be leveraged for explainable robot policy search. Based on this insight, we introduce the SAS Prompt (Summarize, Analyze, Synthesize) – a single prompt that enables iterative learning and adaptation of robot behavior by combining the LLM's ability to retrieve, reason and optimize over previous robot traces in order to synthesize new, unseen behavior. Our approach can be regarded as an early example of a new family of explainable policy search methods that are entirely implemented within an LLM. We evaluate our approach both in simulation and on a real-robot table tennis task. Project website: sites.google.com/asu.edu/sas-llm/

Justin Chen, Zifeng Wang, Hamid Palangi, Rujun Han, Sayna Ebrahimi, Long Le, Vincent Perot, Swaroop Mishra, Mohit Bansal, Chen-Yu Lee, Tomas Pfister. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Reinforcement Learning (RL) traditionally relies on scalar reward signals, limiting its ability to leverage the rich semantic knowledge often available in real-world tasks. In contrast, humans learn efficiently by combining numerical feedback with language, prior knowledge, and common sense. We introduce Prompted Policy Search (ProPS), a novel RL method that unifies numerical and linguistic reasoning within a single framework. Unlike prior work that augments existing RL components with language, ProPS places a large language model (LLM) at the center of the policy optimization loop—directly proposing policy updates based on both reward feedback and natural language input. We show that LLMs can perform numerical optimization in-context, and that incorporating semantic signals, such as goals, constraints, and strategy hints can lead to more informed exploration and sample-efficient learning. ProPS is evaluated across 15 Gymnasium tasks, spanning classic control, Atari games, and MuJoCo environments, and compared to seven widely-adopted RL algorithms (e.g., PPO, SAC, TRPO). It outperforms all baselines on 8 out of 15 tasks and demonstrates substantial gains when provided with domain knowledge. These results highlight the potential of unifying semantics and numerics for transparent, generalizable, and human-aligned reinforcement learning.

Large Language Models (LLMs) have emerged as a groundbreaking technology with their unparalleled text generation capabilities across various applications. Nevertheless, concerns persist regarding the accuracy and appropriateness of their generated content. A contemporary methodology, self-correction, has been proposed as a remedy to these issues. Building upon this premise, this paper critically examines the role and efficacy of self-correction within LLMs, shedding light on its true potential and limitations. Central to our investigation is the notion of intrinsic self-correction, whereby an LLM attempts to correct its initial responses based solely on its inherent capabilities, without the crutch of external feedback. In the context of reasoning, our research indicates that LLMs struggle to self-correct their responses without external feedback, and at times, their performance even degrades after self-correction. Drawing from these insights, we offer suggestions for future research and practical applications in this field.

We present STEP-BACK PROMPTING, a simple prompting technique that enables LLMs to do abstractions to derive high-level concepts and first principles from instances containing specific details. Using the concepts and principles to guide reasoning, LLMs significantly improve their abilities in following a correct reasoning path towards the solution. We conduct experiments of STEP-BACK PROMPTING with PaLM-2L, GPT-4 and Llama2-70B models, and observe substantial performance gains on various challenging reasoning-intensive tasks including STEM, Knowledge QA, and Multi-Hop Reasoning. For instance, STEP-BACK PROMPTING improves PaLM-2L performance on MMLU (Physics and Chemistry) by 7% and 11% respectively, TimeQA by 27%, and MuSiQue by 7%.

In-context learning (ICL, also known as few-shot prompting) has been the standard method of adapting LLMs to downstream tasks, by learning from a few input-output examples. Nonetheless, all ICL-based approaches only learn from correct input-output pairs. In this paper, we revisit this paradigm, by learning more from the few given input-output examples. We introduce Learning Principles (LEAP): First, we intentionally induce the model to make mistakes on these few examples; then we reflect on these mistakes, and learn explicit task-specific “principles” from them, which help solve similar problems and avoid common mistakes; finally, we prompt the model to answer unseen test questions using the original few-shot examples and these learned general principles. We evaluate LEAP on a wide range of benchmarks, including multi-hop question answering (Hotpot QA), textual QA (DROP), Big-Bench Hard reasoning, and math problems (GSM8K and MATH); in all these benchmarks, LEAP improves the strongest available LLMs such as GPT-3.5-turbo, GPT-4, GPT-4-turbo and Claude-2.1. For example, LEAP improves over the standard few-shot prompting using GPT-4 by 7.5% in DROP, and by 3.3% in HotpotQA. Importantly, LEAP does not require any more input or examples than the standard few-shot prompting settings.

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.

We introduce SELF-DISCOVER, a general framework for LLMs to self-discover the task-intrinsic reasoning structures to tackle complex reasoning problems that are challenging for typical prompting methods. Core to the framework is a self-discovery process where LLMs select multiple atomic reasoning modules such as critical thinking and step-by-step thinking, and compose them into an explicit reasoning structure for LLMs to follow during decoding. SELF-DISCOVER substantially improves GPT-4 and PaLM 2’s performance on challenging reasoning benchmarks such as BigBench-Hard, grounded agent reasoning, and MATH, by as much as 32% compared to Chain of Thought (CoT). Furthermore, SELF-DISCOVER outperforms inference-intensive methods such as CoT-Self-Consistency by more than 20%, while requiring 10-40x fewer inference compute. Finally, we show that the self-discovered reasoning structures are universally applicable across model families: from PaLM 2-L to GPT-4, and from GPT-4 to Llama2, and share commonalities with human reasoning patterns.

Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, Hannaneh Hajishirzi. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

Recent research has shown that language models exploit 'artifacts' in benchmarks to solve tasks, rather than truly learning them, leading to inflated model performance.In pursuit of creating better benchmarks, we propose VAIDA, a novel benchmark creation paradigm for NLP, that focuses on guiding crowdworkers, an under-explored facet of addressing benchmark idiosyncrasies.VAIDA facilitates sample correction by providing real-time visual feedback and recommendations to improve sample quality.Our approach is domain, model, task, and metric agnostic, and constitutes a paradigm shift for robust, validated, and dynamic benchmark creation via human-and-metric-in-theloop workflows.We evaluate via expert review and a user study with NASA TLX.We find that VAIDA decreases effort, frustration, mental, and temporal demands of crowdworkers and analysts, simultaneously increasing the performance of both user groups with a 45.8% decrease in the level of artifacts in created samples.As a by-product of our user study, we observe that created samples are adversarial across models, leading to decreases of 31.3% (BERT), 22.5% (RoBERTa), 14.98% (GPT-3 fewshot) in performance.1

Himanshu Gupta, Neeraj Varshney, Swaroop Mishra, Kuntal Kumar Pal, Saurabh Arjun Sawant, Kevin Scaria, Siddharth Goyal, Chitta Baral. Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics. 2023.

In recent years, progress in NLU has been driven by benchmarks.These benchmarks are typically collected by crowdsourcing, where annotators write examples based on annotation instructions crafted by dataset creators.In this work, we hypothesize that annotators pick up on patterns in the crowdsourcing instructions, which bias them to write many similar examples that are then over-represented in the collected data.We study this form of bias, termed instruction bias, in 14 recent NLU benchmarks, showing that instruction examples often exhibit concrete patterns, which are propagated by crowdworkers to the collected data.This extends previous work (Geva et al., 2019) and raises a new concern of whether we are modeling the dataset creator's instructions, rather than the task.Through a series of experiments, we show that, indeed, instruction bias can lead to overestimation of model performance, and that models struggle to generalize beyond biases originating in the crowdsourcing instructions.We further analyze the influence of instruction bias in terms of pattern frequency and model size, and derive concrete recommendations for creating future NLU benchmarks. 1

Humans (e.g., crowdworkers) have a remarkable ability in solving different tasks, by simply reading textual instructions that define them and looking at a few examples.Despite the success of the conventional supervised learning on individual datasets, such models often struggle with generalization across tasks (e.g., a question-answering system cannot solve classification tasks).A long-standing challenge in AI is to build a model that learns a new task by understanding the humanreadable instructions that define it.To study this, we introduce NATURAL INSTRUCTIONS, a dataset of 61 distinct tasks, their humanauthored instructions, and 193k task instances (input-output pairs).The instructions are obtained from crowdsourcing instructions used to create existing NLP datasets and mapped to a unified schema.Using this meta-dataset, we measure cross-task generalization by training models on seen tasks and measuring generalization to the remaining unseen ones.We adopt generative pre-trained language models to encode task-specific instructions along with input and generate task output.Our results indicate that models benefit from instructions when evaluated in terms of generalization to unseen tasks (19% better for models utilizing instructions).These models, however, are far behind an estimated performance upperbound, indicating significant room for more progress in this direction.1

Swaroop Mishra, Arindam Mitra, Neeraj Varshney, Bhavdeep Sachdeva, Peter Clark, Chitta Baral, Ashwin Kalyan. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022.

Knowledge of difficulty level of questions helps a teacher in several ways, such as estimating students' potential quickly by asking carefully selected questions and improving quality of examination by modifying trivial and hard questions.Can we extract such benefits of instance difficulty in Natural Language Processing?To this end, we conduct Instance-Level Difficulty Analysis of Evaluation data (ILDAE) in a largescale setup of 23 datasets and demonstrate its five novel applications: 1) conducting efficientyet-accurate evaluations with fewer instances saving computational cost and time, 2) improving quality of existing evaluation datasets by repairing erroneous and trivial instances, 3) selecting the best model based on application requirements, 4) analyzing dataset characteristics for guiding future data creation, 5) estimating Out-of-Domain performance reliably.Comprehensive experiments for these applications lead to several interesting results, such as evaluation using just 5% instances (selected via ILDAE) achieves as high as 0.93 Kendall correlation with evaluation using complete dataset and computing weighted accuracy using difficulty scores leads to 5.2% higher correlation with Out-of-Domain performance.We release the difficulty scores 1 and hope our work will encourage research in this important yet understudied field of leveraging instance difficulty in evaluations.

Despite the success of large pre-trained language models (LMs) such as Codex, they show below-par performance on the larger and more complicated programming related questions.We show that LMs benefit from the summarized version of complicated questions.Our findings show that superfluous information often present in problem description such as human characters, background stories, and names (which are included to help humans in understanding a task) does not help models in understanding a task.To this extent, we create a metadataset from the frequently used APPS dataset and the newly created CodeContests dataset for the program synthesis task.Our meta-dataset consists of human and synthesized summaries of the long and complicated programming questions.Experimental results on Codex show that our proposed approach outperforms baseline by 8.13% on the APPS dataset and 11.88% on the CodeContests dataset on average in terms of strict accuracy.Our analysis shows that summaries significantly improve performance for introductory (9.86%) and interview (11.48%) programming questions.However, it shows improvement by a small margin (∼ 2%) for competitive programming questions, implying scope for future research in this direction.1 * Equal Contribution 1 Code and data is available at https://github.com/kurbster/ Prompt-Summarization 2 Detailed related work is presented in Appendix A However, LMs such as Codex show below-par performance on the long and complicated programming questions.We observe that the natural language description of the program becomes long and complicated when there is superfluous information (see section 2.1.1).The goal of adding this information to the description is to make it more understandable to humans.However, we find that this information confuses the model in understanding a task 3 .We propose that removing the excess information and providing the model with the exact specifications of the problem can improve the performance of the LMs.To remove excess information 4 , we summarize the descriptions of the program in such a way that it does not lose important specifications.We use the APPS dataset (Hendrycks et al., 2021) and Code-Contests dataset (Li et al., 2022) which are a collection of coding problems from different online sources and create a meta-dataset consisting of human and synthesized summaries.We perform all experiments using the GPT-based Codex model (Chen et al., 2021) on the proposed meta-dataset and show that the summarized version of complicated questions improves strict accuracy by 8.13% on the APPS dataset and 11.85% on CodeContests.From our analysis, we can see significant improvement for introductory (9.86%) and interview (11.48%) related programming questions.However, it shows improvement by a small margin (∼ 2%) for competitive programming questions.Considering that automatic evaluation of a program does not reward for partial correctness, we perform qualitative evaluation on our meta-dataset and find that original questions often confuse models in understanding the underlying problem, as models latch on to some spurious words in the text (e.g. the word 'list' in question makes the model 3 See example in Appendix C 4 Instructions for creating summaries given in Appendix N 4532 design a list even though the underlying problem is on graphs).We further analyze model performance on different types of summaries (i.e., basic, expert, and synthetic) and provide instruction-design principles that can help future research on prompting in program synthesis.2 Method 2.1 Dataset We use the APPS (Hendrycks et al., 2021) and CodeContests (Li et al., 2022) datasets to create summaries.We crowd-sourced the creation of human summaries.The result was 373 human summaries for APPS and 80 summaries for CodeContests along with and 8663 synthetic summaries using both datasets.Table 1 shows the statistics of the generated summaries.

Swaroop Mishra, Matthew Finlayson, Pan Lu, Leonard Tang, Sean Welleck, Chitta Baral, Tanmay Rajpurohit, Oyvind Tafjord, Ashish Sabharwal, Peter Clark, Ashwin Kalyan. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. 2022.

Large Language Models (LMs) have achieved state-of-the-art performance on many Natural Language Processing (NLP) benchmarks.With the growing number of new benchmarks, we build bigger and more complex LMs.However, building new LMs may not be an ideal option owing to the cost, time and environmental impact associated with it.We explore an alternative route: can we modify data by expressing it in terms of the model's strengths, so that a question becomes easier for models to answer?We investigate if humans can decompose a hard question into a set of simpler questions that are relatively easier for models to solve.We analyze a range of datasets involving various forms of reasoning and find that it is indeed possible to significantly improve model performance (24% for GPT3 and 29% for RoBERTa-SQuAD along with a symbolic calculator) via decomposition.Our approach provides a viable option to involve people in NLP research in a meaningful way.Our findings indicate that Human-in-the-loop Question Decomposition (HQD) can potentially provide an alternate path to building large LMs 1 .

Yizhong Wang, Swaroop Mishra, Pegah Alipoormolabashi, Yeganeh Kordi, Amirreza Mirzaei, Atharva Naik, Arjun Ashok, Arut Selvan Dhanasekaran, Anjana Arunkumar, David Stap, Eshaan Pathak, Giannis Karamanolakis, Haizhi Lai, Ishan Purohit, Ishani Mondal, Jacob Anderson, Kirby Kuznia, Krima Doshi, Kuntal Kumar Pal, Maitreya Patel, Mehrad Moradshahi, Mihir Parmar, Mirali Purohit, Neeraj Varshney, Phani Rohitha Kaza, Pulkit Verma, Ravsehaj Singh Puri, Rushang Karia, Savan Doshi, Shailaja Keyur Sampat, Siddhartha Mishra, Sujan Reddy A, Sumanta Patro, Tanay Dixit, Xudong Shen. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. 2022.

When answering a question, humans utilize the information available across different modalities to synthesize a consistent and complete chain of thought (CoT). This process is normally a black box in the case of deep learning models like large-scale language models. Recently, science question benchmarks have been used to diagnose the multi-hop reasoning ability and interpretability of an AI system. However, existing datasets fail to provide annotations for the answers, or are restricted to the textual-only modality, small scales, and limited domain diversity. To this end, we present Science Question Answering (ScienceQA), a new benchmark that consists of ~21k multimodal multiple choice questions with a diverse set of science topics and annotations of their answers with corresponding lectures and explanations. We further design language models to learn to generate lectures and explanations as the chain of thought (CoT) to mimic the multi-hop reasoning process when answering ScienceQA questions. ScienceQA demonstrates the utility of CoT in language models, as CoT improves the question answering performance by 1.20% in few-shot GPT-3 and 3.99% in fine-tuned UnifiedQA. We also explore the upper bound for models to leverage explanations by feeding those in the input; we observe that it improves the few-shot performance of GPT-3 by 18.96%. Our analysis further shows that language models, similar to humans, benefit from explanations to learn from fewer data and achieve the same performance with just 40% of the data. The data and code are available at https://scienceqa.github.io.

Models that top leaderboards often perform unsatisfactorily when deployed in real world applications; this has necessitated rigorous and expensive pre-deployment model testing. A hitherto unexplored facet of model performance is: Are our leaderboards doing equitable evaluation? In this paper, we introduce a task-agnostic method to probe leaderboards by weighting samples based on their 'difficulty' level. We find that leaderboards can be adversarially attacked and top performing models may not always be the best models. We subsequently propose alternate evaluation metrics. Our experiments on 10 models show changes in model ranking and an overall reduction in previously reported performance- thus rectifying the overestimation of AI systems' capabilities. Inspired by behavioral testing principles, we further develop a prototype of a visual analytics tool that enables leaderboard revamping through customization, based on an end user's focus area. This helps users analyze models' strengths and weaknesses, and guides them in the selection of a model best suited for their application scenario. In a user study, members of various commercial product development teams, covering 5 focus areas, find that our prototype reduces pre-deployment development and testing effort by 41% on average.