Myeongsoo Kim, Chao-Chun Hsu, Dingmin Wang, Shweta Garg, Varun Kumar, Murali Krishna Ramanathan. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

George Ma, Anurag Koul, Qi Chen, Yawen Wu, Sachit Kuhar, Yu Yu, Aritra Sengupta, Varun Kumar, Murali Krishna Ramanathan. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Sachit Kuhar, Wasi Uddin Ahmad, Zijian Wang, Nihal Jain, Haifeng Qian, Baishakhi Ray, Murali Krishna Ramanathan, Xiaofei Ma, Anoop Deoras. 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.

While pre-trained language models (LM) for code have achieved great success in code completion, they generate code conditioned only on the contents within the file, i.e., in-file context, but ignore the rich semantics in other files within the same project, i.e., project-level cross-file context, a critical source of information that is especially useful in modern modular software development. Such overlooking constrains code LMs’ capacity in code completion, leading to unexpected behaviors such as generating hallucinated class member functions or function calls with unexpected arguments. In this work, we propose CoCoMIC, a novel framework that jointly learns the in-file and cross-file context on top of code LMs. To empower CoCoMIC, we develop CCFinder, a static-analysis-based tool that locates and retrieves the most relevant project-level cross-file context for code completion. CoCoMIC successfully improves the existing code LM with a 33.94% relative increase in exact match and 28.69% in identifier matching for code completion when the cross-file context is provided. Finally, we perform a series of ablation studies and share valuable insights for future research on integrating cross-file context into code LMs.

Recent advances in retrieval-augmented generation (RAG) have initiated a new era in repository-level code completion. However, the invariable use of retrieval in existing methods exposes issues in both efficiency and robustness, with a large proportion of the retrieved contexts proving unhelpful or harmful to code language models (code LMs). In this paper, we propose a selective RAG framework to avoid retrieval when unnecessary. To power this framework, we design a self-supervised learning approach to enable a code LM to accurately self-evaluate whether retrieval can improve its output quality and robustly leverage the potentially noisy retrieved contexts. Using this LM as both the selective RAG policy and the generation model, our framework achieves state-of-the-art repository-level code completion performance on diverse benchmarks including RepoEval, CrossCodeEval, and CrossCodeLongEval, a new long-form code completion benchmark. Meanwhile, our analyses show that selectively retrieving brings as much as 70% inference speedup in the online serving setting without harming the performance. We further demonstrate that our framework is able to accommodate different generation models, retrievers, and programming languages. These advancements position our framework as an important step towards more accurate and efficient repository-level code completion.

There is huge value in making software development more productive with AI. An important component of this vision is the capability to translate natural language to a programming language ("NL2Code") and thus to significantly accelerate the speed at which code is written.

In today’s software industry, large-scale, multi-language codebases are the norm. This brings substantial challenges in developing automated tools for code maintenance tasks such as API migration or dead code cleanup. Tool builders often find themselves caught between two less-than-ideal tooling options: (1) languagespecific code rewriting tools or (2) generic, lightweight match-replace transformation tools with limited expressiveness. The former leads to tool fragmentation and a steep learning curve for each language, while the latter forces developers to create ad-hoc, throwaway scripts to handle realistic tasks. To fill this gap, we introduce a new declarative domain-specific language (DSL) for expressing interdependent multi-language code transformations. Our key insight is that we can increase the expressiveness and applicability of lightweight match-replace tools by extending them to support for composition, ordering, and flow. We implemented an open-source tool for our language, called P olyglot P iranha , and deployed it in an industrial setting. We demonstrate its effectiveness through three case studies, where it deleted 210K lines of dead code and migrated 20K lines, across 1611 pull requests. We compare our DSL against state-of-the-art alternatives, and show that the tools we developed are faster, more concise, and easier to maintain.

Network Function Virtualization promises better utilization of computational resources by dynamically scaling resources on demand. However, most network functions (NFs) are stateful and require state updates on a per-packet basis. During a scaling operation, cores need to synchronize access to a shared state to avoid race conditions and to guarantee that NFs process packets in arrival order. Unfortunately, the classic approach to control concurrent access to a shared state with locks does not scale to today’s throughput and latency requirements. Moreover, network traffic is highly skewed, leading to load imbalances in systems that use only sharding to partition the NF states. To address these challenges, we present Dyssect, a system that enables dynamic scaling of stateful NFs by disaggregating the states of network functions. By carefully coordinating actions between cores and a central controller, Dyssect migrates shards and flows between cores for load balancing or traffic prioritization without resorting to locks or reordering packets. Also, Dyssect’s state disaggregation allows the offloading of stateful network functions to programmable NICs and makes it easier for exploring hardware-software tradeoffs that better suit specific service chains and traffic loads. Our experimental evaluation shows that Dyssect reduces tail latency up to 32.04% and increases throughput up to 19.36% when compared to state-of-the-art competing solutions.

Shiqi Wang, Zheng Li, Haifeng Qian, Chenghao Yang, Zijian Wang, Mingyue Shang, Varun Kumar, Samson Tan, Baishakhi Ray, Parminder Bhatia, Ramesh Nallapati, Murali Krishna Ramanathan, Dan Roth, Bing Xiang. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

Code completion models have made significant progress in recent years, yet current popular evaluation datasets, such as HumanEval and MBPP, predominantly focus on code completion tasks within a single file. This over-simplified setting falls short of representing the real-world software development scenario where repositories span multiple files with numerous cross-file dependencies, and accessing and understanding cross-file context is often required to complete the code correctly. To fill in this gap, we propose CrossCodeEval, a diverse and multilingual code completion benchmark that necessitates an in-depth cross-file contextual understanding to complete the code accurately. CrossCodeEval is built on a diverse set of real-world, open-sourced, permissively-licensed repositories in four popular programming languages: Python, Java, TypeScript, and C#. To create examples that strictly require cross-file context for accurate completion, we propose a straightforward yet efficient static-analysis-based approach to pinpoint the use of cross-file context within the current file. Extensive experiments on state-of-the-art code language models like CodeGen and StarCoder demonstrate that CrossCodeEval is extremely challenging when the relevant cross-file context is absent, and we see clear improvements when adding these context into the prompt. However, despite such improvements, the pinnacle of performance remains notably unattained even with the highest-performing model, indicating that CrossCodeEval is also capable of assessing model's capability in leveraging extensive context to make better code completion. Finally, we benchmarked various methods in retrieving cross-file context, and show that CrossCodeEval can also be used to measure the capability of code retrievers.

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 pushed the boundary of code generation and achieved impressive performance. However, the huge number of model parameters poses a significant challenge to their adoption in a typical software development environment, where a developer might use a standard laptop or mid-size server to develop code. Such large models cost significant resources in terms of memory, latency, dollars, as well as carbon footprint.

Network Function Virtualization promises better utilization of computational resources by dynamically scaling resources on demand. However, most network functions (NFs) are stateful and require state updates on a per-packet basis. During a scaling operation, cores need to synchronize access to a shared state to avoid race conditions and to guarantee that NFs process packets in arrival order. Unfortunately, the classic approach to control concurrent access to a shared state with locks does not scale to today’s throughput and latency requirements. Moreover, network traffic is highly skewed, leading to load imbalances in systems that use only sharding to partition the NF states. To address these challenges, we present Dyssect, a system that enables dynamic scaling of stateful NFs by disaggregating the states of network functions. By carefully coordinating actions between cores and a central controller, Dyssect migrates shards and flows between cores for load balancing or traffic prioritization without resorting to locks or reordering packets. Our experimental evaluation shows that Dyssect reduces tail latency up to 32% and increases throughput up to 19.36% when compared to state-of-the-art competing solutions.

The concurrent programming literature is rich with tools and techniques for data race detection. Less, however, has been known about real-world, industry-scale deployment, experience, and insights about data races. Golang (Go for short) is a modern programming language that makes concurrency a first-class citizen. Go offers both message passing and shared memory for communicating among concurrent threads. Go is gaining popularity in modern microservice-based systems. Data races in Go stand in the face of its emerging popularity.

Swift, an increasingly-popular programming language, advocates the use of protocols, which define a set of required methods and properties for conforming types. Protocols are commonly used in Swift programs for abstracting away implementation details; e.g., in a large industrial app from Uber, they are heavily used to enable mock objects for unit testing. Unfortunately, heavy use of protocols can result in significant performance overhead. Beyond the dynamic dispatch often associated with such a feature, Swift allows for both value and reference types to conform to a protocol, leading to significant boxing and unboxing overheads. In this paper, we describe three new optimizations and transformations to reduce the overhead of Swift protocols. Within a procedure, we define LocalVar, a dataflow analysis and transformation to remove both dynamic dispatch and boxing overheads. We also describe Param, which optimizes the case of protocol-typed method parameters using specialization. Finally, we describe SoleType, a transformation that injects casts when a global analysis (like type-hierarchy analysis) discovers some protocol variable must have some concrete type. We also describe how these optimizations work fruitfully together and with existing Swift optimizations to deliver further speedups. We perform elaborate experimentation and demonstrate that our optimizations deliver an average 1.56x speedup on a suite of Swift benchmarks that use protocols. Further, we applied the optimizations to a production iOS Swift application from Uber used by millions of customers daily. For a set of performance spans defined by the developers of the application, the optimized version showed speedups ranging from 6.9% to 55.49%. A version of our optimizations has been accepted as part of the official Swift compiler distribution.

Conventional concolic testing has been used to provide high coverage of paths in statically typed languages. While it has also been applied in the context of JavaScript (JS) programs, we observe that applying concolic testing to dynamically-typed JS programs involves tackling unique problems to ensure scalability. In particular, a naive type-agnostic extension of concolic testing to JS programs causes generation of large number of inputs. Consequently, many executions operate on undefined values and repeatedly explore same paths resulting in redundant tests, thus diminishing the scalability of testing drastically.

Performance issues in large applications arise only in particular scenarios under heavy load conditions. It is therefore difficult to catch them during testing and they easily escape into production. This necessitates the design of a common and efficient instrumentation strategy that profiles the flow of objects during an execution. Designing such a strategy which enables profile generation precisely with low overhead is non-trivial due to the number of objects created, accessed and paths traversed by them in an execution. In this paper, we design and implement an efficient instrumentation technique that efficiently generates object flow profiles for Java programs, without requiring any modifications to the underlying virtual machine. We achieve this by applying Ball-Larus numbering on a specialized hybrid flow graph (hfg). The hfg path profiles that are collected during runtime are post-processed offline to derive the object flow profiles. We implemented the profiler and validated its efficacy by applying it on Java programs. The results demonstrate the scalability of our approach, which handles 0.2M to 0.55B object accesses with an average runtime overhead of 8x. We also demonstrate the effectiveness of the generated profiles by implementing a client analysis that consumes the profiles to detect performance bugs. The analysis detects 38 performance bugs which when refactored result in significant performance gains (up to 30%) in running times.

Detecting concurrency-induced bugs in multithreaded libraries can be challenging due to the intricacies associated with their manifestation. This includes invocation of multiple methods, synthesis of inputs to the methods to reach the failing location, and crafting of thread interleavings that cause the erroneous behavior. Neither fuzzing-based testing techniques nor over-approximate static analyses are well positioned to detect such subtle defects while retaining high accuracy alongside satisfactory coverage.

Redundant traversal of loops in the context of other loops has been recently identified as a source of performance bugs in many Java libraries. This has resulted in the design of static and dynamic analysis techniques to detect these performance bugs automatically. However, while the effectiveness of dynamic analyses is dependent on the analyzed input tests, static analyses are less effective in automatically validating the presence of these problems, validating the fixes and avoiding regressions in future versions. This necessitates the design of an approach to automatically generate tests for exposing redundant traversal of loops.

Dynamic race detectors operate by analyzing execution traces of programs to detect races in multithreaded programs. As the thread interleavings influence these traces, the sets of races detected across multiple runs of the detector can vary. This non-determinism without any change in program source and input can reduce programmer confidence in using the detector. From an organizational perspective, a defect needs to be reported consistently until it is fixed. Non-determinism complicates the work flow and the problem is further exacerbated with modifications to the program.

Deterministic multithreaded systems (DMTs) are designed to ensure reproducibility of program behavior for a given input. In these systems, even minor changes to the code (or input) can perturb the schedule. This increases the number of feasible schedules making reasoning about these programs harder. Stable multithreaded systems (StableMTs) address the problem such that a schedule is unaffected by minor changes. Unfortunately, determinism in these systems can potentially serialize the execution imposing a significant penalty on the performance. Programmer hints in the form of soft barriers attempt to eliminate the performance bottlenecks. However, the process is arduous, error-prone and requires manual intervention to reconsider the location of the barrier for every modification to the source code. In this paper, we propose an effective approach to automate the task of adding soft barriers in the source code. Our approach analyzes the deterministic program executions to extract the program and semantic order of executions and builds a weighted constraint graph. Using this graph, a schedule is synthesized which is used to identify bottlenecks and insert soft barriers in the program source. We validate our implementation, named PEGASUS, by applying it on multiple benchmarks. Our experimental results demonstrate that we are able to reduce the overall execution time of programs by upto 34% when compared to the execution time where barriers are inserted manually. Moreover, we observe a performance improvement ranging from 38% to 88% as compared to programs without barriers. Our experimental results show that adapting PEGASUS to infer barriers for multiple versions of a source program is seamless. The memory and time overheads associated with the usage of PEGASUS is negligible making it a vital cog in building scalable StableMTs.

Subtle concurrency errors in multithreaded libraries that arise because of incorrect or inadequate synchronization are often difficult to pinpoint precisely using only static techniques. On the other hand, the effectiveness of dynamic race detectors is critically dependent on multithreaded test suites whose execution can be used to identify and trigger races. Usually, such multithreaded tests need to invoke a specific combination of methods with objects involved in the invocations being shared appropriately to expose a race. Without a priori knowledge of the race, construction of such tests can be challenging. In this paper, we present a lightweight and scalable technique for synthesizing precisely these kinds of tests. Given a multithreaded library and a sequential test suite, we describe a fully automated analysis that examines sequential execution traces, and produces as its output a concurrent client program that drives shared objects via library method calls to states conducive for triggering a race. Experimental results on a variety of well-tested Java libraries yield 101 synthesized multithreaded tests in less than four minutes. Analyzing the execution of these tests using an off-the-shelf race detector reveals 187 harmful races, including several previously unreported ones. Our implementation, named NARADA, and the results of our experiments are available at http://www.csa.iisc.ernet.in/~sss/tools/narada.

Using thread-safe libraries can help programmers avoid the complexities of multithreading. However, designing libraries that guarantee thread-safety can be challenging. Detecting and eliminating atomicity violations when methods in the libraries are invoked concurrently is vital in building reliable client applications that use the libraries. While there are dynamic analyses to detect atomicity violations, these techniques are critically dependent on effective multithreaded tests. Unfortunately, designing such tests is non-trivial. In this paper, we design a novel and scalable approach for synthesizing multithreaded tests that help detect atomicity violations. The input to the approach is the implementation of the library and a sequential seed testsuite that invokes every method in the library with random parameters. We analyze the execution of the sequential tests, generate variable lock dependencies and construct a set of three accesses which when interleaved suitably in a multithreaded execution can cause an atomicity violation. Subsequently, we identify pairs of method invocations that correspond to these accesses and invoke them concurrently from distinct threads with appropriate objects to help expose atomicity violations. We have incorporated these ideas in our tool, named Intruder, and applied it on multiple open-source Java multithreaded libraries. Intruder is able to synthesize 40 multithreaded tests across nine classes in less than two minutes to detect 79 harmful atomicity violations, including previously unknown violations in thread-safe classes. We also demonstrate the effectiveness of Intruder by comparing the results with other approaches designed for synthesizing multithreaded tests.

Designing and implementing thread-safe multithreaded libraries can be a daunting task as developers of these libraries need to ensure that their implementations are free from concurrency bugs, including deadlocks. The usual practice involves employing software testing and/or dynamic analysis to detect deadlocks. Their effectiveness is dependent on well-designed multithreaded test cases. Unsurprisingly, developing multithreaded tests is significantly harder than developing sequential tests for obvious reasons.

Dynamic analysis techniques have been proposed to detect potential deadlocks. Analyzing and comprehending each potential deadlock to determine whether the deadlock is feasible in a real execution requires significant programmer effort. Moreover, empirical evidence shows that existing analyses are quite imprecise. This imprecision of the analyses further void the manual effort invested in reasoning about non-existent defects.

Designing thread-safe libraries without concurrency defects can be a challenging task. Detecting deadlocks while invoking methods in these libraries concurrently is hard due to the possible number of method invocation combinations, the object assignments to the parameters and the associated thread interleavings. In this paper, we describe the design and implementation of OMEN+ that takes a multithreaded library as the input and detects true deadlocks in a scalable manner. We achieve this by automatically synthesizing relevant multithreaded tests and analyze the associated execution traces using a precise deadlock detector. We validate the usefulness of OMEN+ by applying it on many multithreaded Java libraries and detect a number of deadlocks even in documented thread-safe libraries. The tool is available for free download at http://www.csa.iisc.ernet.in/~sss/tool/omenplus.html.

An important, but often neglected, goal of static analysis for detecting bugs is the ability to show defects to the programmer quickly. Unfortunately, existing static analysis tools scale very poorly, or are shallow and cannot find complex interprocedural defects. Previous attempts at reducing the analysis time by adding more resources (CPU, memory) or by splitting the analysis into multiple sub-analyses based on defect detection capabilities resulted in limited/negligible improvements.

Function precedence protocols define ordering relations among function calls in a program. In some instances, precedence protocols are well-understood (e.g., a call to pthread_mutex_init must always be present on all program paths before a call to pthread_mutex_lock). Oftentimes, however, these protocols are neither well- documented, nor easily derived. As a result, protocol violations can lead to subtle errors that are difficult to identify and correct. In this paper, we present CHRONICLER, a tool that applies scalable inter-procedural path-sensitive static analysis to automatically infer accurate function precedence protocols. Chronicler computes precedence relations based on a program's control-flow structure, integrates these relations into a repository, and analyzes them using sequence mining techniques to generate a collection of feasible precedence protocols. Deviations from these protocols found in the program are tagged as violations, and represent potential sources of bugs. We demonstrate CHRONICLER's effectiveness by deriving protocols for a collection of benchmarks ranging in size from 66 K to 2 M lines of code. Our results not only confirm the existence of bugs in these programs due to precedence protocol violations, but also highlight the importance of path sensitivity on accuracy and scalability.

The reliability and correctness of complex software systems can be significantly enhanced through well-defined specifications that dictate the use of various units of abstraction (e.g., modules, or procedures). Often times, however, specifications are either missing, imprecise, or simply too complex to encode within a signature, necessitating specification inference. The process of inferring specifications from complex software systems forms the focus of this paper. We describe a static inference mechanism for identifying the preconditions that must hold whenever a procedure is called. These preconditions may reflect both data flow properties (e.g., whenever p is called, variable x must be non-null) as well as control-flow properties (e.g., every call to p must bepreceded by a call to q). We derive these preconditions using a ninter-procedural path-sensitive dataflow analysis that gathers predicates at each program point. We apply mining techniques to these predicates to make specification inference robust to errors. This technique also allows us to derive higher-level specifications that abstract structural similarities among predicates (e.g., procedure p is called immediately after a conditional test that checks whether some variable v is non-null.) We describe an implementation of these techniques, and validate the effectiveness of the approach on a number of large open-source benchmarks. Experimental results confirm that our mining algorithms are efficient, and that the specifications derived are both precise and useful-the implementation discovers several critical, yet previously, undocumented preconditions for well-tested libraries.

Software systems often undergo many revisions during their lifetime as new features are added, bugs repaired, abstractions simplified and refactored, and performance improved. When a revision, even a minor one, does occur, the changes it induces must be tested to ensure that invariants assumed in the original version are not violated unintentionally. In order to avoid testing components that are unchanged across revisions, impact analysis is often used to identify code blocks or functions that are affected by a change. In this paper, we present a novel solution to this general problem that uses dynamic programming on instrumented traces of different program binaries to identify longest common subsequences in strings generated by these traces. Our formulation allows us to perform impact analysis and also to detect the smallest set of locations within the functions where the effect of the changes actually manifests itself. Sieve is a tool that incorporates these ideas. Sieve is unobtrusive, requiring no programmer or compiler intervention to guide its behavior. Our experiments on multiple versions of op ensource C programs shows that Sieve is an effective and scalable tool to identify impact sets and can locate regions in the affected functions where the changes manifest. These results lead us to conclude that Sieve can play a beneficial role in program testing and software maintenance

Search is a fundamental service in peer-to-peer (P2P) networks. However, despite numerous research efforts, efficient algorithms for guaranteed location of shared content in unstructured P2P networks are yet to be devised. In this paper, the authors presented a simple but highly effective protocol for object location that gives probabilistic guarantees of finding even rare objects independently of the network topology. The protocol relies on randomized techniques for replication of objects (or their references) and for query propagation. The authors proved analytically, and demonstrated experimentally, that this scheme provides high probabilistic guarantees of success, while incurring minimal overhead. The performance of this scheme was quantified in terms of network messages, probability of success, and response time. The robustness of this protocol was also evaluated in the presence of node failures (departures). Using simulation, it is shown that this scheme performs no worse than the best known access-frequency based protocols, without compromising access to rare objects.

Distributed peer-to-peer storage systems rely on voluntary participation of peers to effectively manage a storage pool. Files are generally replicated in several sites to provide acceptable levels of availability. If disk space on these peers is not carefully monitored and provisioned, the system may not be able to provide availability for certain files. In particular, identification and elimination of redundant data are important problems that may arise in long-lived systems. Scalability and availability are competing goals in these networks: scalability concerns would dictate aggressive elimination of replicas, while availability considerations would argue conversely. In this paper, the authors provided a novel and efficient solution that addresses both these goals with respect to management of redundant data. Specifically, the problem of duplicate elimination in the context of systems connected over an unstructured peer-to-peer network in which there is no a priori binding between an object and its location was addressed. A new randomized protocol was proposed to solve this problem in a scalable and decentralized fashion that does not compromise availability requirements of the application. Performance results using both large-scale simulations, and a prototype built on PlanetLab, demonstrate that the protocols provide high probabilistic guarantees of success, while incurring minimal administrative overheads.

Abstract — While recent technological advances have motivated large-scale deployment of RFID systems, a number of critical design issues remain unresolved. In this paper we deal with detecting redundant RFID readers (the redundant reader problem). The underlying difficulty associated with this problem arises from the lack of collision detection mechanisms, the potential inability of RFID readers to relay packets generated by other readers, and severe resource constraints on RFID tags. We prove that an optimal solution to the redundant reader problem is NP-hard and propose a randomized, distributed, and localized approximation algorithm, RRE. We provide a detailed probabilistic analysis of the accuracy and time complexity of RRE and conduct elaborate simulations to demonstrate their correctness and efficiency. I.

Summary form only given. The challenging problem in building a reliable storage system over a network of peers is to account for the heterogeneity of peers in terms of their availability, storage capacity, and locality. With varying object access patterns, an efficient/reliable system should not only increase the availability of more frequently accessed objects, but also provide reasonable availability of other objects. A greedy heuristic is proposed to improve the total object availability in the system. The simulation experiments compare the heuristic approach to randomized placement of objects with respect to varying failure distributions of peers, number of peers to objects ratio, replication factor and storage at a peer. The results show that this approach can provide equivalent availability with half the storage resources required for random placement.