Though Large Language Models (LLMs) have demonstrated remarkable capabilities, they are still prone to outputting falsehoods using seemingly persuasive language. Many recent works attempt to address this problem by using LLMs in a framework where a single seed prompt results in a series of interactions involving augmented prompts with an otherwise unchanged LLM, and the results are aggregated with a goal of producing a more reliable output. We consider the replicability and generalizability of evaluations of inference-time methods intended to improve the reliability of responses from a base LLMs. We survey how methods have been evaluated in the literature and find a great variety of benchmarks and models in use. Motivated by this, we conduct our own evaluation to evaluate the effectiveness of a few methods across a range of benchmarks and models. Our evaluation reveals that while these techniques show promise in improving reliability, there is still significant variability in performance across different domains and tasks, and methods that show substantial improvements on weaker base models often do not improve reliability for better base models.
more »
« less
This content will become publicly available on June 26, 2027
Code Reasoning for Software Engineering Tasks: A Survey and A Call to Action
The rise of large language models (LLMs) has led to dramatic improvements across a wide range of natural language tasks. Their performance on certain tasks can be further enhanced by incorporating test-time reasoning techniques. These inference-time advances have been adopted into the code domain, enabling complex software engineering (SWE) tasks such as code generation, test generation and issue resolution. However, the impact of different reasoning techniques on code-centric SWE tasks has not been systematically explored. In this work, we survey code reasoning techniques that underpin these capabilities, with a focus on test-time compute and inference-time reasoning paradigms. We examine a variety of code-specific reasoning methods and progressively build up to SWE agents, which combine planning, tool use, and multi-step interaction. We also compare the impact of different techniques on coding tasks, highlighting their relative importance and outlining open challenges and future research directions. Across commonly used models and benchmarks, we find that approaches exploiting code-specific signals (e.g., structure and execution feedback) are frequently associated with improved performance, motivating a dedicated study of code reasoning beyond natural-language reasoning. Our contributions are: (1) to the best of our knowledge, the first dedicated survey of code reasoning for SWE tasks, highlighting overarching reasoning strategies, hybrid methods, and agentic approaches; (2) a taxonomy of inference-time techniques used to drive code reasoning, accompanied by a curated set of under-explored benchmarks with high potential for SWE evaluation; (3) a comparative analysis of reasoning design patterns across commonly used models and benchmarks; and (4) a synthesis of gaps in current methods and evaluation practices, identifying under-explored areas and concrete opportunities for future research.
more »
« less
- PAR ID:
- 10695086
- Publisher / Repository:
- Journal of Machine Learning Research Inc
- Date Published:
- Journal Name:
- Transactions on machine learning research
- ISSN:
- 2835-8856
- Subject(s) / Keyword(s):
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Weitong, Zhang (Ed.)In computer vision, Visual Prompting (VP) and Visual Prompt Tuning (VPT) have recently emerged as lightweight and effective alternatives to full fine-tuning for adapting large-scale vision models within the “pretrain-then-finetune” paradigm. However, despite rapid progress, their conceptual boundaries remain blurred, as VP and VPT are frequently used interchangeably in current research, reflecting a lack of systematic distinction between these techniques and their respective applications. In this survey, we revisit the designs of VP and VPT from first principles and conceptualize them within a unified framework termed Prompt-based Adaptation (PA). Within this framework, we distinguish methods based on their injection granularity: VP operates at the pixel level, while VPT injects prompts at the token level. We further categorize these methods by their generation mechanism into fixed, learnable, and generated prompts. Beyond the core methodologies, we examine PA’s integrations across diverse domains, including medical imaging, 3D point clouds, and vision language tasks, as well as its role in test-time adaptation and trustworthy AI. We also summarize current benchmarks and identify key challenges and future directions. To the best of our knowledge, we are the first comprehensive survey dedicated to PA’s methodologies and applications in light of their distinct characteristics. Our survey aims to provide a clear roadmap for researchers and practitioners in all areas to understand and explore the evolving landscape of PA-related research.more » « less
-
The success of RL for LLM post-training stems from an unreasonably uninformative source: a single bit of information per rollout as binary reward or preference label. At the other extreme, distillation offers dense supervision but requires demonstrations, which are costly and difficult to scale. We study natural language feedback as an intermediate signal: richer than scalar rewards, yet cheaper than complete demonstrations. Textual feedback is a natural mode of human interaction and is already abundant in many real-world settings, where users, tools, and automated judges routinely critique LLM outputs. Towards leveraging text feedback at scale, we formalize a multi-turn RL setup where text feedback is available during training but not at inference. Therefore, models must learn to internalize the feedback in order to improve their test-time single-turn performance. To do this, we propose two methods: Self Distillation, which trains the single-turn policy to match its own feedback-conditioned second-turn generations; and Feedback Modeling, which predicts the feedback as an auxiliary objective. We provide theoretical analysis on both methods, and empirically evaluate on reasoning puzzles, competition math, and creative writing tasks. Our results show that both methods consistently outperform strong baselines across benchmarks, highlighting the potential of RL with an additional source of rich supervision at scale.more » « less
-
Automated Program Repair (APR) agents leverage Large Language Models (LLMs) to autonomously diagnose and fix software bugs through reasoning, planning, and tool use. Despite impressive leaderboard gains on benchmarks such as SWE-bench, little is understood about how these agents take actions, where they fail, and how their behavior compares to that of human developers. This paper presents the first systematic analysis of five state-of-the-art APR agents across 500 real-world repair tasks, tracing their full decision-making pipelines—from issue description to patch validation. Our study reveals that while agents excel at simple fixes, they struggle with logic-intensive bugs, often producing verbose or overfitted patches that merely satisfy existing tests. We find that test generation and regression test selection remain major bottlenecks, with agents frequently failing to reproduce issues or run relevant regression tests. Moreover, most agents operate with primitive tooling (e.g., bash scripts) and lack access to debuggers or program analyzers, which constrains their reasoning and patch quality. These findings highlight key limitations in current APR systems and motivate a shift-left approach—emphasizing early, high-quality test generation and validation—to reduce spurious fixes and improve semantic correctness. We further outline concrete directions for next-generation APR design: (1) richer and more integrated tool ecosystems, (2) diversified agentic architectures that combine complementary strengths, and (3) benchmarks that prioritize semantic repair quality and test generation fidelity over surface-level success metrics.more » « less
-
Humans have the ability to reason about geometric patterns in images and scenes from a young age. However, developing large multimodal models (LMMs) capable of similar reasoning remains a challenge, highlighting the need for robust evaluation methods to assess these capabilities. We introduce TurtleBench, a benchmark designed to evaluate LMMs’ capacity to interpret geometric patterns—given visual examples, textual instructions, or both—and generate precise code outputs. Inspired by turtle geometry, a notion used to teach children foundational coding and geometric concepts, TurtleBench features tasks with patterned shapes that have underlying algorithmic logic. Our evaluation reveals that leading LMMs struggle significantly with these tasks, with GPT-4V achieving only 19% accuracy on the simplest tasks and few-shot prompting only marginally improves their performance (<2%). TurtleBench highlights the gap between human and AI performance in intuitive and visual geometrical understanding, setting the stage for future research in this area and stands as one of the few benchmarks to evaluate the integration of visual understanding and code generation capabilities in LMMs, setting the stage for future research.more » « less
An official website of the United States government
