Varun Kumar

Ben Athiwaratkun, Sanjay Krishna Gouda, Zijian Wang et al. · amazon-science, ibm-research

We present new benchmarks on evaluation code generation models: MBXP and Multilingual HumanEval, and MathQA-X. These datasets cover over 10 programming languages and are generated using a scalable conversion framework that transpiles prompts and test cases from the original Python datasets into the corresponding data in the target language. Using these benchmarks, we are able to assess the performance of code generation models in a multi-lingual fashion, and discovered generalization ability of language models on out-of-domain languages, advantages of multi-lingual models over mono-lingual, the ability of few-shot prompting to teach the model new languages, and zero-shot translation abilities even on mono-lingual settings. Furthermore, we use our code generation model to perform large-scale bootstrapping to obtain synthetic canonical solutions in several languages, which can be used for other code-related evaluations such as code insertion, robustness, or summarization tasks. Overall, our benchmarks represents a significant step towards a deeper understanding of language models' code generation abilities. We publicly release our code and datasets at https://github.com/amazon-research/mxeval.

Hantian Ding, Varun Kumar, Yuchen Tian et al. · amazon-science, stanford

Large language models trained on code have shown great potential to increase productivity of software developers. Several execution-based benchmarks have been proposed to evaluate functional correctness of model-generated code on simple programming problems. Nevertheless, it is expensive to perform the same evaluation on complex real-world projects considering the execution cost. On the contrary, static analysis tools such as linters, which can detect errors without running the program, haven't been well explored for evaluating code generation models. In this work, we propose a static evaluation framework to quantify static errors in Python code completions, by leveraging Abstract Syntax Trees. Compared with execution-based evaluation, our method is not only more efficient, but also applicable to code in the wild. For experiments, we collect code context from open source repos to generate one million function bodies using public models. Our static analysis reveals that Undefined Name and Unused Variable are the most common errors among others made by language models. Through extensive studies, we also show the impact of sampling temperature, model size, and context on static errors in code completions.

Shiqi Wang, Zheng Li, Haifeng Qian et al. · amazon-science, ibm-research

Code generation models have achieved impressive performance. However, they tend to be brittle as slight edits to a prompt could lead to very different generations; these robustness properties, critical for user experience when deployed in real-life applications, are not well understood. Most existing works on robustness in text or code tasks have focused on classification, while robustness in generation tasks is an uncharted area and to date there is no comprehensive benchmark for robustness in code generation. In this paper, we propose ReCode, a comprehensive robustness evaluation benchmark for code generation models. We customize over 30 transformations specifically for code on docstrings, function and variable names, code syntax, and code format. They are carefully designed to be natural in real-life coding practice, preserve the original semantic meaning, and thus provide multifaceted assessments of a model's robustness performance. With human annotators, we verified that over 90% of the perturbed prompts do not alter the semantic meaning of the original prompt. In addition, we define robustness metrics for code generation models considering the worst-case behavior under each type of perturbation, taking advantage of the fact that executing the generated code can serve as objective evaluation. We demonstrate ReCode on SOTA models using HumanEval, MBPP, as well as function completion tasks derived from them. Interesting observations include: better robustness for CodeGen over InCoder and GPT-J; models are most sensitive to syntax perturbations; more challenging robustness evaluation on MBPP over HumanEval.

Yang Trista Cao, Yada Pruksachatkun, Kai-Wei Chang et al. · amazon-science

Multiple metrics have been introduced to measure fairness in various natural language processing tasks. These metrics can be roughly categorized into two categories: 1) \emph{extrinsic metrics} for evaluating fairness in downstream applications and 2) \emph{intrinsic metrics} for estimating fairness in upstream contextualized language representation models. In this paper, we conduct an extensive correlation study between intrinsic and extrinsic metrics across bias notions using 19 contextualized language models. We find that intrinsic and extrinsic metrics do not necessarily correlate in their original setting, even when correcting for metric misalignments, noise in evaluation datasets, and confounding factors such as experiment configuration for extrinsic metrics. %al

Umang Gupta, Jwala Dhamala, Varun Kumar et al. · amazon-science, gatech

Language models excel at generating coherent text, and model compression techniques such as knowledge distillation have enabled their use in resource-constrained settings. However, these models can be biased in multiple ways, including the unfounded association of male and female genders with gender-neutral professions. Therefore, knowledge distillation without any fairness constraints may preserve or exaggerate the teacher model's biases onto the distilled model. To this end, we present a novel approach to mitigate gender disparity in text generation by learning a fair model during knowledge distillation. We propose two modifications to the base knowledge distillation based on counterfactual role reversal$\unicode{x2014}$modifying teacher probabilities and augmenting the training set. We evaluate gender polarity across professions in open-ended text generated from the resulting distilled and finetuned GPT$\unicode{x2012}$2 models and demonstrate a substantial reduction in gender disparity with only a minor compromise in utility. Finally, we observe that language models that reduce gender polarity in language generation do not improve embedding fairness or downstream classification fairness.

Marvin Klingner, Varun Ravi Kumar, Senthil Yogamani et al.

While deep neural networks (DNNs) achieve impressive performance on environment perception tasks, their sensitivity to adversarial perturbations limits their use in practical applications. In this paper, we (i) propose a novel adversarial perturbation detection scheme based on multi-task perception of complex vision tasks (i.e., depth estimation and semantic segmentation). Specifically, adversarial perturbations are detected by inconsistencies between extracted edges of the input image, the depth output, and the segmentation output. To further improve this technique, we (ii) develop a novel edge consistency loss between all three modalities, thereby improving their initial consistency which in turn supports our detection scheme. We verify our detection scheme's effectiveness by employing various known attacks and image noises. In addition, we (iii) develop a multi-task adversarial attack, aiming at fooling both tasks as well as our detection scheme. Experimental evaluation on the Cityscapes and KITTI datasets shows that under an assumption of a 5% false positive rate up to 100% of images are correctly detected as adversarially perturbed, depending on the strength of the perturbation. Code is available at https://github.com/ifnspaml/AdvAttackDet. A short video at https://youtu.be/KKa6gOyWmH4 provides qualitative results.

Xiaokai Wei, Sujan Gonugondla, Wasi Ahmad et al. · amazon-science, ibm-research

ML-powered code generation aims to assist developers to write code in a more productive manner, by intelligently generating code blocks based on natural language prompts. Recently, large pretrained deep learning models have substantially pushed the boundary of code generation and achieved impressive performance. Despite their great power, the huge number of model parameters poses a significant threat to adapting them in a regular software development environment, where a developer might use a standard laptop or mid-size server to develop her code. Such large models incur significant resource usage (in terms of memory, latency, and dollars) as well as carbon footprint. Model compression is a promising approach to address these challenges. Several techniques are proposed to compress large pretrained models typically used for vision or textual data. Out of many available compression techniques, we identified that quantization is mostly applicable for code generation task as it does not require significant retraining cost. As quantization represents model parameters with lower-bit integer (e.g., int8), the model size and runtime latency would both benefit from such int representation. We extensively study the impact of quantized model on code generation tasks across different dimension: (i) resource usage and carbon footprint, (ii) accuracy, and (iii) robustness. To this end, through systematic experiments we find a recipe of quantization technique that could run even a $6$B model in a regular laptop without significant accuracy or robustness degradation. We further found the recipe is readily applicable to code summarization task as well.

Ninareh Mehrabi, Palash Goyal, Apurv Verma et al. · amazon-science, gatech

Natural language often contains ambiguities that can lead to misinterpretation and miscommunication. While humans can handle ambiguities effectively by asking clarifying questions and/or relying on contextual cues and common-sense knowledge, resolving ambiguities can be notoriously hard for machines. In this work, we study ambiguities that arise in text-to-image generative models. We curate a benchmark dataset covering different types of ambiguities that occur in these systems. We then propose a framework to mitigate ambiguities in the prompts given to the systems by soliciting clarifications from the user. Through automatic and human evaluations, we show the effectiveness of our framework in generating more faithful images aligned with human intention in the presence of ambiguities.

Jwala Dhamala, Varun Kumar, Rahul Gupta et al. · amazon-science

Several prior works have shown that language models (LMs) can generate text containing harmful social biases and stereotypes. While decoding algorithms play a central role in determining properties of LM generated text, their impact on the fairness of the generations has not been studied. We present a systematic analysis of the impact of decoding algorithms on LM fairness, and analyze the trade-off between fairness, diversity and quality. Our experiments with top-$p$, top-$k$ and temperature decoding algorithms, in open-ended language generation, show that fairness across demographic groups changes significantly with change in decoding algorithm's hyper-parameters. Notably, decoding algorithms that output more diverse text also output more texts with negative sentiment and regard. We present several findings and provide recommendations on standardized reporting of decoding details in fairness evaluations and optimization of decoding algorithms for fairness alongside quality and diversity.

Nihal Jain, Robert Kwiatkowski, Baishakhi Ray et al. · amazon-science

In this study, we address the issue of API hallucinations in various software engineering contexts. We introduce CloudAPIBench, a new benchmark designed to measure API hallucination occurrences. CloudAPIBench also provides annotations for frequencies of API occurrences in the public domain, allowing us to study API hallucinations at various frequency levels. Our findings reveal that Code LLMs struggle with low frequency APIs: for e.g., GPT-4o achieves only 38.58% valid low frequency API invocations. We demonstrate that Documentation Augmented Generation (DAG) significantly improves performance for low frequency APIs (increase to 47.94% with DAG) but negatively impacts high frequency APIs when using sub-optimal retrievers (a 39.02% absolute drop). To mitigate this, we propose to intelligently trigger DAG where we check against an API index or leverage Code LLMs' confidence scores to retrieve only when needed. We demonstrate that our proposed methods enhance the balance between low and high frequency API performance, resulting in more reliable API invocations (8.20% absolute improvement on CloudAPIBench for GPT-4o).

Rahul Ahuja, Mudit Jain, Bala Murali Manoghar Sai Sudhakar et al.

Vision foundation models (VFMs) and Bird's Eye View (BEV) representation have advanced visual perception substantially, yet their internal spatial representations assume the rectilinear geometry of pinhole cameras. Fisheye cameras, widely deployed on production autonomous vehicles for their surround-view coverage, exhibit severe radial distortion that renders these representations geometrically inconsistent. At the same time, the scarcity of large-scale fisheye annotations makes retraining foundation models from scratch impractical. We present \ours, a lightweight framework that adapts frozen VFMs to fisheye geometry through two components: a frozen DINOv2 backbone with Low-Rank Adaptation (LoRA) that transfers rich self-supervised features to fisheye without task-specific pretraining, and Fisheye Rotary Position Embedding (FishRoPE), which reparameterizes the attention mechanism in the spherical coordinates of the fisheye projection so that both self-attention and cross-attention operate on angular separation rather than pixel distance. FishRoPE is architecture-agnostic, introduces negligible computational overhead, and naturally reduces to the standard formulation under pinhole geometry. We evaluate \ours on WoodScape 2D detection (54.3 mAP) and SynWoodScapes BEV segmentation (65.1 mIoU), where it achieves state-of-the-art results on both benchmarks.

Hantian Ding, Zijian Wang, Giovanni Paolini et al. · amazon-science

In large language model training, input documents are typically concatenated together and then split into sequences of equal length to avoid padding tokens. Despite its efficiency, the concatenation approach compromises data integrity -- it inevitably breaks many documents into incomplete pieces, leading to excessive truncations that hinder the model from learning to compose logically coherent and factually consistent content that is grounded on the complete context. To address the issue, we propose Best-fit Packing, a scalable and efficient method that packs documents into training sequences through length-aware combinatorial optimization. Our method completely eliminates unnecessary truncations while retaining the same training efficiency as concatenation. Empirical results from both text and code pre-training show that our method achieves superior performance (e.g., relatively +4.7% on reading comprehension; +16.8% in context following; and +9.2% on program synthesis), and reduces closed-domain hallucination effectively by up to 58.3%.

George Ma, Anurag Koul, Qi Chen et al. · amazon-science

Large Language Models (LLMs) excel at code-related tasks but often struggle in realistic software repositories, where project-specific APIs and cross-file dependencies are crucial. Retrieval-augmented methods mitigate this by injecting repository context at inference time. The low inference-time latency budget affects either retrieval quality or the added latency adversely impacts user experience. We address this limitation with SpecAgent, an agent that improves both latency and code-generation quality by proactively exploring repository files during indexing and constructing speculative context that anticipates future edits in each file. This indexing-time asynchrony allows thorough context computation, masking latency, and the speculative nature of the context improves code-generation quality. Additionally, we identify the problem of future context leakage in existing benchmarks, which can inflate reported performance. To address this, we construct a synthetic, leakage-free benchmark that enables a more realistic evaluation of our agent against baselines. Experiments show that SpecAgent consistently achieves absolute gains of 9-11% (48-58% relative) compared to the best-performing baselines, while significantly reducing inference latency.

Junlin Wang, Siddhartha Jain, Dejiao Zhang et al.

A diverse array of reasoning strategies has been proposed to elicit the capabilities of large language models. However, in this paper, we point out that traditional evaluations which focus solely on performance metrics miss a key factor: the increased effectiveness due to additional compute. By overlooking this aspect, a skewed view of strategy efficiency is often presented. This paper introduces a framework that incorporates the compute budget into the evaluation, providing a more informative comparison that takes into account both performance metrics and computational cost. In this budget-aware perspective, we find that complex reasoning strategies often don't surpass simpler baselines purely due to algorithmic ingenuity, but rather due to the larger computational resources allocated. When we provide a simple baseline like chain-of-thought self-consistency with comparable compute resources, it frequently outperforms reasoning strategies proposed in the literature. In this scale-aware perspective, we find that unlike self-consistency, certain strategies such as multi-agent debate or Reflexion can become worse if more compute budget is utilized.

Ibrahim Sobh, Ahmed Hamed, Varun Ravi Kumar et al.

Deep neural networks (DNNs) have accomplished impressive success in various applications, including autonomous driving perception tasks, in recent years. On the other hand, current deep neural networks are easily fooled by adversarial attacks. This vulnerability raises significant concerns, particularly in safety-critical applications. As a result, research into attacking and defending DNNs has gained much coverage. In this work, detailed adversarial attacks are applied on a diverse multi-task visual perception deep network across distance estimation, semantic segmentation, motion detection, and object detection. The experiments consider both white and black box attacks for targeted and un-targeted cases, while attacking a task and inspecting the effect on all the others, in addition to inspecting the effect of applying a simple defense method. We conclude this paper by comparing and discussing the experimental results, proposing insights and future work. The visualizations of the attacks are available at https://youtu.be/6AixN90budY.

Luoxin Chen, Francisco Garcia, Varun Kumar et al.

This paper presents a production Semi-Supervised Learning (SSL) pipeline based on the student-teacher framework, which leverages millions of unlabeled examples to improve Natural Language Understanding (NLU) tasks. We investigate two questions related to the use of unlabeled data in production SSL context: 1) how to select samples from a huge unlabeled data pool that are beneficial for SSL training, and 2) how do the selected data affect the performance of different state-of-the-art SSL techniques. We compare four widely used SSL techniques, Pseudo-Label (PL), Knowledge Distillation (KD), Virtual Adversarial Training (VAT) and Cross-View Training (CVT) in conjunction with two data selection methods including committee-based selection and submodular optimization based selection. We further examine the benefits and drawbacks of these techniques when applied to intent classification (IC) and named entity recognition (NER) tasks, and provide guidelines specifying when each of these methods might be beneficial to improve large scale NLU systems.

Eunah Cho, He Xie, John P. Lalor et al.

Expanding new functionalities efficiently is an ongoing challenge for single-turn task-oriented dialogue systems. In this work, we explore functionality-specific semi-supervised learning via self-training. We consider methods that augment training data automatically from unlabeled data sets in a functionality-targeted manner. In addition, we examine multiple techniques for efficient selection of augmented utterances to reduce training time and increase diversity. First, we consider paraphrase detection methods that attempt to find utterance variants of labeled training data with good coverage. Second, we explore sub-modular optimization based on n-grams features for utterance selection. Experiments show that functionality-specific self-training is very effective for improving system performance. In addition, methods optimizing diversity can reduce training data in many cases to 50% with little impact on performance.

Varun Kumar, Hadrien Glaude, Cyprien de Lichy et al.

New conversation topics and functionalities are constantly being added to conversational AI agents like Amazon Alexa and Apple Siri. As data collection and annotation is not scalable and is often costly, only a handful of examples for the new functionalities are available, which results in poor generalization performance. We formulate it as a Few-Shot Integration (FSI) problem where a few examples are used to introduce a new intent. In this paper, we study six feature space data augmentation methods to improve classification performance in FSI setting in combination with both supervised and unsupervised representation learning methods such as BERT. Through realistic experiments on two public conversational datasets, SNIPS, and the Facebook Dialog corpus, we show that data augmentation in feature space provides an effective way to improve intent classification performance in few-shot setting beyond traditional transfer learning approaches. In particular, we show that (a) upsampling in latent space is a competitive baseline for feature space augmentation (b) adding the difference between two examples to a new example is a simple yet effective data augmentation method.