Search papers, labs, and topics across Lattice.
100 papers published across 10 labs.
Context-aware distillation not only boosts the size of the PolkitBench corpus but also proves that structured context is crucial for maintaining model performance under challenging conditions.
VeriPort not only backports patches at scale but also uncovers hidden vulnerabilities, correcting over 400 erroneous version reports in the process.
RAVEN repairs 83.13% of software vulnerabilities, leveraging a novel agentic RAG approach that generalizes across diverse types and languages.
PORTICO eliminates unauthorized actions in coding agents, achieving perfect compliance while a traditional system fails to do so.
Injecting structured context into prompts can elevate DSL code generation quality to unprecedented levels, achieving near-perfect syntactic validity.
Context-aware distillation not only boosts the size of the PolkitBench corpus but also proves that structured context is crucial for maintaining model performance under challenging conditions.
VeriPort not only backports patches at scale but also uncovers hidden vulnerabilities, correcting over 400 erroneous version reports in the process.
RAVEN repairs 83.13% of software vulnerabilities, leveraging a novel agentic RAG approach that generalizes across diverse types and languages.
PORTICO eliminates unauthorized actions in coding agents, achieving perfect compliance while a traditional system fails to do so.
Injecting structured context into prompts can elevate DSL code generation quality to unprecedented levels, achieving near-perfect syntactic validity.
Reviewers approve AI-generated code more often while actually engaging less, revealing a troubling trend of habituation that could compromise code quality.
Co-authorship with humans can significantly enhance merge rates for certain AI coding agents, but this effect vanishes when accounting for repository selection and PR structure.
Structured coding processes can boost both the quality of AI-generated code and its correctness, challenging the notion that outcome alone defines success in autonomous coding.
Forge transforms LLM-generated code into safety-critical software by seamlessly integrating formal verification, paving the way for certification in regulated environments.
Multi-shot prompting can drastically enhance the ability of LLMs to obscure code authorship, raising alarms for traditional stylometry techniques.
Nearly half of the defects in LLM-integrated web apps slip through testing seams that standard unit tests can't catch, highlighting a critical gap in software verification.
Hedgehog$^\rightarrow$ not only resolves the compositionality dilemma in property-based testing but also maintains practical expressiveness for generator design.
P4IR reduces code compliance errors by up to 38.6% while outperforming top LLMs in accuracy and reliability.
When faced with search-dependent reasoning tasks, models struggle to learn effective chain-of-thought strategies, highlighting a critical limitation in current training paradigms.
GPT-4o excels at localized code refactoring but struggles with integrating new gameplay features, revealing critical insights into LLM capabilities in game development.
Refined guidance boosts coding agent performance by 33%, highlighting that coverage trumps precision in navigating complex repositories.
LLM-generated GPU kernels may appear correct under conventional benchmarks, but a more rigorous testing method reveals hidden transcription errors that could lead to significant performance issues.
Trajectory mining reveals skill structures but fails to translate these insights into meaningful performance gains for downstream policies.
AutoPass achieves up to 1.117x speedup in compiler performance tuning by allowing LLMs to interact with compiler internals, challenging the notion that black-box approaches are sufficient.
ToolPro slashes web service latency by over 50% while drastically reducing client-side traffic, revolutionizing how LLMs interact with complex workflows.
Targeted software fixes informed by real device timing can significantly enhance the effectiveness of side-channel vulnerability mitigations.
Verified trace properties can now be seamlessly translated into secure protocol implementations, bridging the gap between formal verification and practical application.
Retrieval-augmented context boosts commit message quality, outperforming traditional LLM approaches by leveraging historical examples and user feedback.
Supervised fine-tuning with domain-specific data can drastically enhance the reliability of LLM-generated Solidity smart contracts, outperforming general-purpose models.
Handoff validity is crucial for ensuring that design artifacts maintain integrity and context across complex EDA workflows.
N-version programming with coding agents not only mirrors historical failures but also shows a dramatic reduction in errors through diversity, challenging assumptions about AI reliability.
A taxonomy-driven approach significantly reduces hallucinations in LLM-generated code migration suggestions, paving the way for more reliable quantum software engineering.
Automated pipelines can generate high-coverage unit tests for low-level firmware, achieving up to 98.8% line coverage with minimal manual intervention.
Runtime pass rates plummet from 80.4% to 5.7% as project complexity increases, exposing critical architectural flaws in code generation models.
Leveraging multiple decompiler views can boost LLM-based malware detection accuracy by enhancing recall on malicious samples.
Phoenix resolves GitHub issues with 75% accuracy while ensuring safety through a multi-agent system, but still faces challenges in planner localization.
Effect quantales reveal a surprising new dimension to abstract interpretations, shifting the focus from states to event occurrences.
Python overfitting in LLMs is just the tip of the iceberg—Multi-LCB reveals substantial performance disparities across twelve programming languages.
FAPO achieves a remarkable +33.8 percentage point gain in performance by seamlessly transitioning from prompt optimization to structural adjustments in LLM pipelines.
Executable programs can now replace attention heads in transformers with minimal performance loss, achieving over 75% similarity to original patterns.
GrapNet achieves a remarkable 12.08-point accuracy boost over larger dense networks by enabling structural programmability in neural architectures.
Autonomous coding agents can outperform traditional methods in data integration tasks, achieving top results across multiple SQL benchmarks.
Variability in AI-generated software should be embedded in specifications rather than in the code itself, enabling more efficient and tailored program generation.
CAPRA achieves 88.8% accuracy in automated feedback on software architecture deliverables, showcasing the potential for LLMs to transform educational assessments.
Graphical PLC programs can now be verified accurately in under 70ms, eliminating previous vacuous results and enhancing reliability in industrial automation.
A new impact analysis method reveals hidden dependencies in software changes, recovering artifacts that traditional tools miss entirely.
QDSV reveals that a problem-first representation can maintain stability across diverse quantum execution environments, challenging traditional circuit-centric approaches.
Malicious code can now masquerade as ordinary vulnerabilities, evading detection while still compromising agent skills.
Bridging the gap between classroom learning and industry practice, this model equips students with essential skills in using LLMs and MCPs that are critical for modern software engineering.
FloatDoor reveals that LLMs can be covertly compromised to exhibit malicious behavior on specific platforms while appearing benign elsewhere, exposing a critical security gap in AI deployment.
Reliance on LLMs in coding can lead to a gradual decline in secure practices, but gamifying the development process can turn these models into proactive security partners.
OpenAnt can uncover previously unknown vulnerabilities in large codebases while slashing false positive rates by up to 97%.
Prompt structure can make or break the success of LLM-assisted code generation, with specific dimensions driving distinct outcomes in software development workflows.
Cognitive diversity among developers leads to distinct interaction modes with programming assistants, revealing that one-size-fits-all solutions may fall short.
Forecasting future coding tasks can yield a dataset that is 58.1% relevant to real-world software engineering needs, sidestepping the pitfalls of historical data replay.
DevOps specialists and general developers are more alike than you think, sharing tool preferences while navigating the evolving landscape of software development together.
All tested coding agents fail within 5-6 turns, but providing feedback can boost their performance by up to 12x, revealing critical insights into agent design.
Autonomous coding agents can achieve a staggering 300% increase in useful throughput while eliminating redundant work entirely by using a decentralized coordination substrate.
Playful learning strategies enable robots to acquire skills that boost performance on new tasks by over 20%, transforming how we approach robotic skill acquisition.
PreAct allows agents to execute previously learned tasks up to 13 times faster, fundamentally changing how we approach task repetition in AI.
Achieving up to 49.31% reduction in dynamic power, AUTOGATE revolutionizes RTL optimization by combining machine learning with LLMs for scalable clock gating.
80.2% of AI-generated test patches fail to provide meaningful verification, challenging the reliability of current quality assessments in software development.
A step-by-step text-to-SQL framework enables LLMs to achieve near-perfect accuracy on simple queries, revealing the untapped potential of natural language interfaces in astronomical data exploration.
Routing SQL queries based on complexity allows DecoSearch to achieve unprecedented execution accuracy while using an order of magnitude fewer tokens than traditional methods.
FllumaOne achieves 99.98% syntax validity and 99.14% export validity, setting a new benchmark for executable CAD datasets.
LLM-translated code can be slower than human-written code, but SwiftTrans bridges this gap with a novel two-stage framework that boosts both correctness and efficiency.
A unified taxonomy reveals the hidden connections between traditional and AI-augmented binary reversing, illuminating pathways for future research.
Reasoning models in LLMs can dramatically improve code correction accuracy through iterative feedback, outperforming their non-reasoning counterparts.
PracRepair fixes 162 out of 171 bugs in a challenging benchmark, showcasing a leap in automated program repair capabilities through human-inspired debugging techniques.
Coding agents struggle to create complete and engaging games, with top performers barely reaching 41.46% success in end-to-end game generation.
Coordinating neural planning with symbolic execution, Quarry boosts automated proof success rates by up to 13% while keeping costs predictable.
L5 challenges the notion that simplicity and sustainability in creative coding are mutually exclusive, revealing the complex trade-offs involved in their design.
Current coding benchmarks obscure the true performance of AI agents by conflating models with their operational contexts, stifling innovation in software engineering.
The largest-ever verification campaign for Rust's standard library reveals significant vulnerabilities in unsafe code, underscoring the need for robust static verification methods.
End-to-end AI-assisted workflow management can empower non-experts to design complex scientific workflows with expert precision while drastically reducing debugging time.
Harnessing the unpredictability of concurrent execution, this method achieves a 36.2% accuracy in predicting next steps, outperforming leading models in the field.
Performance rankings among programming languages for AI algorithms can shift dramatically based on workload characteristics, revealing critical insights for developers.
Runtime safety checks in Move could be the critical layer that prevents asset loss from undetected verifier bugs in blockchain applications.
Over 2,000 previously undetected bugs in deep learning compilers expose the hidden complexities of compiler-platform interactions that traditional testing methods miss.
Visual artifacts are not just supplementary; they are essential for ensuring code correctness in multimodal programming tasks.
Students' reliance on LLMs reveals a troubling "cruel optimism" that could hinder their development of essential skills and critical thinking.
REFLEX achieves remarkable sample efficiency, solving complex tasks with fewer than 10 LLM calls while ensuring transparent policy evolution.
daVinci-kernel outperforms the best existing RL-trained model in GPU kernel optimization by effectively co-evolving skill selection and execution strategies.
Students developing AI systems face significant hurdles in architectural design, revealing a critical gap in current AI education.
Aerospace engineers found the LLM-based copilot's suggestions helpful, but slow inference times hindered its effectiveness for complex design tasks.
Coding agents can outperform raw data models in time series analysis, but still miss 22-34% of questions, revealing critical reasoning gaps.
Activating latent security knowledge in LLMs can significantly reduce exploitable vulnerabilities in generated code without the overhead of retraining.
No single exploration strategy outperforms others in automated web GUI testing; instead, their strengths are complementary, revealing critical insights for optimizing testing effectiveness.
A systematic approach to hybrid quantum-classical application design could drastically reduce costly implementation failures by providing predictive feasibility assessments before deployment.
Spear accelerates the synthesis of optimal abstract transformers by exploiting parallelism, outperforming state-of-the-art solvers in both efficiency and scalability.
RHO achieves a 45.0% success rate in robotic tasks, 2.5x higher than the best multi-turn agent, showcasing a breakthrough in real-time control efficiency.
Coding agents can achieve up to 3.5x faster completion times with a new KVCache management strategy that understands their unique workload patterns.
Stronger reasoning in LLMs using Code Interpreters hinges on the strategic use of crucial tokens and cognitive behaviors, revealing both potential and constraints in model performance.
Semantic post-hoc operators fail to enhance accuracy in frozen small code models, but a novel recovery method boosts performance by 12 tasks on HumanEval+.
Evolved playbooks can boost vulnerability detection rates by over 6x and outperform dedicated commercial products, reshaping the landscape of automated security auditing.
Obliv-clang enables high-performance C++ programming without leaking sensitive data through execution patterns, outperforming traditional solutions.
Iterative refinement with feedback-driven learning allows AutoDecompiler to significantly enhance the accuracy of binary decompilation, outperforming traditional single-turn models.
Automating test specification generation can dramatically reduce engineering time while improving coverage, addressing a critical bottleneck in automotive software development.
LLMs can accelerate the discovery of functional software tools, achieving 24 successful implementations from 83 identified tools with just 4 hours of human effort.
VerIbmc solves 431 out of 499 software verification problems using local models, proving that high-performance neuro-symbolic reasoning can be achieved without cloud dependency.
Confidence calibration in multi-agent systems can drastically reduce compliance gaps and contradictory links in safety-critical software engineering tasks.
Agentic AI can generate parallel Julia code, but struggles with scalability and robustness in high-performance computing tasks.
Achieving over 98% coverage and 95% consistency, this framework revolutionizes HLS verification by integrating knowledge-augmented LLMs with an autonomous verification agent.
No-resource languages can achieve significant code generation improvements through a novel pre-training and instruction-following hybrid approach, challenging conventional training paradigms.
Claude 3.5 Sonnet tops the charts in code generation accuracy, but GPT-4 reigns supreme in tackling complex algorithms, revealing a nuanced landscape of LLM capabilities.
