Large language model agents are increasingly deployed as multi-agent systems rather than as a standalone model operating within a single context. Recent agent harnesses allow the model to instantiate worker agents and to author the executable workflow that coordinates them, shifting the design of system structure from developers to the model at run time. The quality of this organization therefore constitutes a distinct capability, yet existing benchmarks either measure end-to-end task completion or evaluate coordination under structures specified in advance. We formalize this capability as workflow orchestration and introduce WorkflowArena, a benchmark of 145 long-horizon tasks spanning 8 professional domains, from software engineering and formal mathematics to legal analysis and 3D production. Each task specifies an initial workspace and an integrated deliverable composed of interdependent components, which is assessed by a hidden verifier. Agents determine autonomously whether and how to delegate, and all delegation is recorded, enabling joint evaluation of deliverable quality and the organization of the underlying work. Experiments with nine models show that no single configuration performs best across all domains and that systems exhibit distinct delegation strategies: at high effort, Claude models in Claude Code invoke workflows less frequently but with larger worker groups, whereas GPT models in Codex delegate more frequently in smaller units. When the executing model is held fixed on a challenging subset of tasks, workflows authored by Claude Opus 5.5 substantially improve weaker executors, indicating that workflow organization is itself a measurable determinant of task performance.
Deploying large language models in regulated industries requires more than picking the fastest inference engine: it requires evidence that a configuration meets contractual and legal obligations before it serves a single customer request. We describe Runway, an internal validation harness built at MasterControl that builds a governed decision pipeline around NVIDIA AIPerf. The key insight is separating concerns: AIPerf provides the trusted measurement, while Runway provides the experiment design that comes before it and the compliance, cost, and audit decisions that come after. We describe the architecture and the storage model that makes caching and audit one mechanism. We report three findings from Runway sweeps: TensorRT-LLM and vLLM rankings invert with concurrency; a newer, pricier GPU instance can be the cheaper one per request; and speculative decoding behaves differently across engines under load, so engine configuration can outweigh engine choice.
Objectives: To measure whether a locally hosted large language model (LLM), stacked with deterministic and clinical named-entity recognition (NER) detectors, finds more protected health information (PHI) than open tools, and at what precision. Materials and Methods: DeidSharp, an open-source .NET toolkit, combines the findings of pluggable detectors (patterns, a clinical NER model, a local 26B mixture-of-experts LLM) and fails closed on any detector error. Configurations fixed on tuning halves were compared, under a prespecified protocol, with deid, Philter, the Stanford/Penn de-identifier, OpenAI Privacy Filter and GLiNER on reporting halves of PhysioNet's nursing notes (889 spans) and semi-synthetic MIMIC-IV-Note discharge summaries (29,023 spans), and on synthetic ASQ-PHI queries, with paired patient-level bootstrap intervals. Results: In the nursing notes, against Stanford/Penn, the recall-first stack found 98.9% of PHI (+2.8, 95% CI 1.4 to 4.4) at 9.3 points lower precision; the precision-first stack, without NER, found 96.5%, matching Stanford/Penn, at 19.6 points higher precision (83.2%). In the discharge summaries, both stacks led Stanford/Penn by more than 20 points (recall-first 95.3%, +21.5, 20.5 to 22.4), almost entirely on generated untyped names; on typed labels all three exceeded 99%. All four prespecified comparisons held. No configuration beat deid significantly on its home corpus; Philter found more discharge-summary spans than the recall-first stack at 23 points lower precision. Discussion: The LLM was the main detector and NER added place names; asked to verify its own findings, the LLM discarded a quarter of true PHI. Conclusion: A local LLM layered with deterministic detectors offers recall-first and precision-first operating points without sending text off-site.
Coordination between large language-model agents has become a central question for AI safety and software engineering, as such agents are deployed in groups that share infrastructure. Existing research tells the agents that partners exist or asks them to coordinate. Whether coordination arises through shared state alone, with nothing disclosing that anyone else is present, is the key question for deployments that share caches and memory between agents, yet it has rarely been examined, and never with independent replication; this study does both. Four agents each received a short arithmetic task and a key-value store presented to each as its own scratch space; two held complementary halves of one computation, and no prompt mentioned other agents, sharing or collaboration. For reliability, the environment was rebuilt independently from a written protocol, and both implementations ran under hashed locks on design, data and analysis. The results bear novelty value in three respects. Joint success occurred only through the shared store. Whether a model coordinated depended not on its vendor tier but on whether it wrote its intermediate result and looked at the store again, and one ordinary sentence, "check your workspace again before submitting", switched that second look on and off. The findings held under replication. For practice, state shared between agents is a communication channel and should be separated, namespaced and logged. For research, designs should be locked before the data, and the parse rule, endpoint and parameters reported. Future work should test open-weight models and separate using another agent's output from intending to communicate.
A worked solution can make self-generated training targets shorter and cheaper, but accuracy on the practiced operation may hide how well a model can reuse newly learned rules. We study this choice in a 2B-parameter reasoning model learning two invented quantitative rules. The model generates targets from a rule sheet alone or from the sheet plus a program-written worked solution; targets are accepted for final-answer correctness, and the support is removed for fine-tuning and evaluation. Training questions request forward use in one segment; composition requires sequential reuse of the same rule across two segments. After six passes, using a program-checked numerical signature, sheet-only targets lead by 24.3 percentage points (95% item-bootstrap CI 14.6-34.0) on composition in a prospective test. On unseen argument values, observed forward accuracy is identical (87.5% each), but composition accuracy is 58.3% versus 23.6%, a 34.7-point gap (25.7-43.8). An independently sampled second target-bank pair repeats the direction (+14.6 points, 2.1-26.0). Ten passes on the shorter worked-solution targets, approximately matching the domain tokens of six sheet-only passes while changing other training quantities, narrow the gap to 16.3 points (6.3-26.7) in the first bank pair and 6.2 points (-6.3 to 18.8) in the second, where the pre-specified test does not hold. At six passes each, the worked-solution policy uses fewer capture tokens and less update time, yet the trained models use 2.6 times as many generated tokens per composition success under the numerical signature (2.1 times in the second bank pair). Choosing a capture context therefore requires testing how the new rules are reused, alongside the inference cost of those uses, rather than relying on forward accuracy alone.
Code localizers are usually trained to predict the files a reference patch edits, but the code a fixer needs to read is not the same as the code it changes. We propose Fixer-Guided Localization (FGL): the localizer proposes several localizations, a frozen fixer scores each by how likely it makes a verified patch, and the localizer is trained on these scores with weighted supervised fine-tuning, without running tests and without reinforcement learning. We build the full pipeline on Gemma 4 E4B for the Kaggle Gemma-4 Developer Agent competition, from 1,629 test-verified Kimi K3 trajectories. The FGL target helps the fixer find the bug as often as the teacher's localization while using 27.5% less code. On the development set, one EM round moves the agent in the expected direction compared with a matched control: more tasks reach a patch (53 vs. 46), fewer tasks get stuck reading files that do not exist (43 vs. 54), and more tasks are solved (8 vs. 6). Training also changes how the agent works: trained agents explore about three times more than the base model and write far fewer wrong patches, and the single agent solves up to 11 of 78 tasks, against 8 for the base model. An analysis of 13 evaluation runs identifies what limits a 4B agent under an 8-minute budget (time, repetition loops and edit quality) and gives practical lessons for training small coding agents. All data, code and adapters are publicly available.
Software developers spend significant time reading and understanding code, increasingly supplemented by generative AI. While these tools aim to reduce the difficulty of reading code, interacting with them introduces a dual-task environment that may impose extraneous cognitive load that is challenging to capture. This study investigates how access to generative AI affects developers' real-time cognitive workload, code comprehension performance, and self-perceived code comprehension strategies. We conducted a 3x2 within-subjects study with 20 novice CS students reviewing Java code snippets of varying lengths with or without AI assistance, followed by an AI-unassisted comprehension quiz. Cognitive workload was continuously captured using a Detection-Response Task (tDRT) during both reading and quiz phases, and compared to post-quiz NASA-TLX and SEQ scores. AI assistance significantly increased objective cognitive workload, raising tDRT reaction times by 11.2%, yet participants' perceived workload remained unchanged. Code comprehension scores were high under both conditions, with no evidence that AI assistance improved comprehension at the difficulty levels evaluated; a ceiling effect limited power to detect benefits on harder tasks. Qualitatively, participants reported trusting the AI conditionally -verifying its output against their own reading for more complex snippets, but directly trusting it for simpler code- and, while many felt their reading process was unchanged, a notable subset shifted toward skimming or reading the AI's explanations in place of the code itself. These findings reveal that developers do not accurately perceive the cognitive resources they expend when managing AI tools. This study provides evidence of added cognitive effort associated with using generative AI, pointing towards a need to account for this dual-task penalty in software engineering tools and education.
Random tests of stateful code check laws, like reading back a written value, at states that random writes reach. The laws hold at every state exactly when the code behaves like a record with hidden data, so a pass is taken as evidence. This can fail: the laws can hold at every sample while the code is far from any record. The laws are local, but behaving like a record is global. Self-testing and stability theorems bridge this gap when the query distribution is controlled by the tester or preserved by the tested maps, but here it is generated by the tested writes, which need not be invertible. We ask how far a test must explore. An interface has $m$ binary variables, each with a read and two writes; in a record, reads return their bits and writes set only their bits. Samples are the ends of geometric runs of mean $\ell$ of uniformly random writes from a start distribution; a tester adds its own write chains of depth $d$. The distance is the least, over records with hidden data, of the largest probability that a read or write differs from it at a sample. From one start state, law checks at $O(m^2\log m/\epsilon)$ samples certify, at fixed confidence, any threshold $\epsilon$ after geometric runs of every positive mean, while runs of fixed positive length can pass at distance $1/2$. For the worst start distribution, take $\epsilon=m^{-\beta}$ with fixed $0<\beta\le1$. For $d\le cm$ with fixed $c<1/2$, a polynomial budget of samples and writes exists exactly when $\ell=\Omega(m)$, where law checks suffice. With no runs, the least depth for this budget is $(1/2+o(1))m$. Counting samples alone, in the worst case at $\epsilon=1/(32m)$ and $d\le m/16$, the critical mean for a polynomial budget is $\Theta(m\log(m/d)/\log m)$, by exhaustive search. So one start state needs only law checks, and the worst case at these thresholds needs exploration linear in $m$ if writes are counted.
Large language models have improved Verilog generation, yet functional correctness remains difficult to optimize because simulation commonly provides a coarse module-level pass/fail outcome. Such feedback discards useful evidence from partially correct RTL modules and cannot determine whether a source statement is functionally responsible for an observed output behavior. We present a new method named Causal Signal Credit Assignment (CSCA), which is a counterfactual preference-optimization framework for Verilog generation. CSCA first performs continuous signal-level behavioral verification on sampled candidate modules. It then locates output-driving assignment spans and applies five interventions while holding the remaining module and testbench fixed. The resulting changes in signal-level errors are combined with behavioral agreement to form a continuous statement credit, which is mapped to response-token weights for strict weighted direct preference optimization. This enables CSCA to learn from locally supported statements without uniformly reinforcing an otherwise incorrect response. Built on Icarus Verilog and a CodeV-derived corpus, CSCA is evaluated on VerilogEval~2.0 and RTLLM~v2.0, including both code completion and specification-to-RTL tasks. With a Qwen2.5-Coder-7B-Instruct backbone, CSCA leads the evaluated 7B-scale models and remains comparable to the much larger GPT-5.4-mini on the reported Verilog-generation benchmarks.
Many digital systems courses teach on FPGA boards whose vendor tool is no longer maintained: the Digilent Basys2 and Nexys2 (Xilinx Spartan-3E) and Nexys3 and Atlys (Spartan-6) need Xilinx ISE 14.7, discontinued since 2013 and hard to install on the computers and tablets students use today. The author presents Silinx, a free (AGPL-3.0) HDL design platform for teaching that keeps ISE's Project Navigator workflow but runs in a web browser with nothing to install. A module can be written in VHDL or Verilog, in C or C++ through Silinx HLS, which is compatible with AMD's Vitis HLS and passes all 73 of its introductory examples, or drawn as a schematic, an ASM chart, a state diagram, a Boolean function with its Karnaugh map, or a Petri net, from which its HDL is generated. Designs are checked with explanations written for beginners, simulated, run on an emulated board, built, and programmed onto a real board over WebUSB. The implementation flow is open and runs in the browser: Yosys and nextpnr compiled to WebAssembly, with Silinx's own device data, timing model and bitstream writer for the Spartan-3E and Spartan-6, built by black-box observation of ISE's documented tools; Lattice iCE40 boards use the existing open iCE40 tools, and further families are planned. ISE is the oracle: Silinx's bitstreams are byte-identical to those ISE writes for the same placed and routed design, and pass ISE's design rule check. Against ISE at the same optimization goal, the open flow uses 15% fewer slices with 14% shorter critical paths when optimizing for area, and 21% fewer slices with 8% shorter critical paths when optimizing for speed (geometric means over 45 open-source designs built by both flows). The author also describes how the platform was developed with AI coding agents under test-first project rules.
C-to-Rust migration is an active agenda, but no dataset pairs the two languages at the level of whole solutions to a shared problem. We present CRUSTPairs, 1857 C/Rust implementation pairs drawn from competitive programming archives, normalised, deduplicated, and stratified into difficulty tiers by zero-shot cross-language embedding similarity, with perrecord provenance and licensing. Cross-language retrieval is hard on this data, and harder than a single baseline would suggest: TF-IDF over shared identifiers reaches MRR 0.21 and so beats a pretrained code model at 0.19, which on the hardest quarter of the data scores 1.7x chance. Finetuning reaches 0.79, and lifts that hardest quarter 10x. Renaming every user identifier settles what is being learned: the finetuned models keep 96-97% of their accuracy without them, while the pretrained model keeps 66\% - less than the lexical baseline does - and on structurally divergent pairs the finetuned models are better without identifiers than with them.
We present MicroSim, a system for binary code similarity detection that represents each function as a graph built from IDA Pro Hex-Rays microcode and compares such graphs with a contrastive Siamese graph neural network (GNN). We evaluate on the TREX cross-optimization benchmark, a corpus of 849,072 function-level microcode graphs spanning optimization levels -O0 through -O3, under a strict class-disjoint split in which no function key is shared between train and test. Under this protocol, MicroSim reaches 0.9972 area under the receiver operating characteristic curve (ROC-AUC) and 0.9762 Matthews correlation coefficient (MCC), outperforming Gemini (0.9332 ROC-AUC), SAFE (0.9752), and jTrans (0.9740) retrained on the same split, as well as a structure-free Bag-of-Words (BoW) baseline (0.9484). We analyze the gap between pair classification and ranked retrieval in graph transformers, showing that Cross-domain Similarity Local Scaling (CSLS) under unified tie-resolved evaluation raises GraphGPS mean reciprocal rank (MRR) from 0.2542 to 0.2559 (R@1 = 0.1639, R@10 = 0.4561), strictly satisfying the theoretical MRR upper bound. Finally, over 5 seeds with equal compute, the pairwise cosine-embedding loss attains a slightly higher ROC-AUC ({\Delta} = -0.0256, 95% confidence interval [-0.3282, +0.2770]), whereas the supervised-contrastive information noise-contrastive estimation (InfoNCE) loss is significantly better on thresholded classification and ranked retrieval.
Large language models (LLMs) perform well on general coding tasks, but their capabilities in application-layer embedded software development remain underexplored. Existing embedded benchmarks primarily focus on hardware-in-the-loop validation in simulated environments. We introduce RealEmbedBench, a benchmark of 5,142 samples covering three application-layer tasks: code generation and completion, state machine modeling and implementation, and cross-OS code translation. We evaluate six representative LLMs and find substantial performance differences across tasks and models. Project context, especially the local code window, significantly affects code generation and completion. Models perform reasonably well at state machine modeling but struggle to translate these models into executable code. Cross-OS translation performance varies substantially by direction. Translation from Harmony to RT-Thread is particularly challenging, with an average pass rate of 0.1667. These findings suggest that developers should consider local code context when using LLMs for code generation and completion and select models according to the translation direction. They also motivate the use of more real-world state machine implementation data in LLM training to narrow the gap between modeling and executable code.
Pre-training code language models relies on massive code corpora usually streamed over the internet. However, researchers rarely need a multi-terabyte archive in its entirety; instead, their workflow centers on curated subsets: selecting well-documented files, filtering out repetitive boilerplate, auditing for benchmark contamination, or retrieving specific code patterns. Yet, because modern corpora are typically published as file pointers and basic attributes rather than source text, no general structural layer exists to select which files to download: curating a subset requires downloading and parsing millions of files first just to discard most of them. We present SWHForest, an open-source framework and public dataset that provides this missing structural layer for any pointer-based corpus without redistributing source code. A single pass over the archive parses each file into an AST and immediately discards the source text, keeping only its structural summary. Queried locally or exposed by archive providers as a server-side filtering API, SWHForest selects matching files first (via SQL or index intersection) and downloads only surviving candidates. Across the ten largest languages of Stack-Edu (52.1M unique files, 341M definitions), SWHForest answers curation queries in milliseconds while cutting downloads by 46-72%, reducing live structural searches from hours of per-language streaming to seconds or minutes.
Reproducibility in large language model (LLM) agent evaluation depends on replay fidelity: which parts of an execution trace are reconstructed, controlled, and compared. This matters for agent debugging, failure analysis, and counterfactual evaluation of stateful tool use. We audit all 71 unsuccessful executions in a fixed historical AppWorld corpus using 284 isolated fixed-code replay attempts. Reconstructed declared settings admit 42 cases under printed-output, pre-oracle database-delta, and native-verdict/test-count checks; resetting the random-number generator before each interaction admits 68. The 26-case gap contracts to four under database/oracle checks, with one endpoint unknown. Mechanism studies identify four materialized-state discrepancies despite matching outputs and verdicts. In one admitted endpoint pair, an identical API call yields matching responses but different row effects; common future Python RNG removes that deterministic difference in two paired repetitions. A separate JWT credential-equality program and alignment controls expose authentication/write differences, with missing repetitions retained. For all 10 affected post-oracle export cases, native export from retained output trackers restores the reference comparison without changing original cached contents. A purposeful 12-case BFCL pilot shows no discrepancy under its tested public-state, return, and verdict checks. We preserve resource censoring and infrastructure deviations. The contribution is a bounded empirical characterization of replay eligibility, hidden state, and test-oracle integration, without estimating debugger accuracy, repair gains, or method-ranking distortion.
Mathematical performance depends on how a language model receives information, uses generated text and submits an answer. We study these dependencies through paired comparisons that separate submission requirements, draft reuse and candidate availability. On 65 development problems, an instruction permitting continued derivation produced no valid final answers in 390 short-output calls; a single-line submission instruction produced 390 format-valid outputs, including 389 parseable answers. Under the fixed submission protocol, 198 problems with failed initial submissions were answered correctly in 9.09p of cases with the question alone, 33.84p with its draft, 46.46p with candidates, and 59.60p with both. Drafts added 13.13 percentage points with candidates fixed, and candidates added 25.76 points with drafts fixed; both effects were significant after Holm correction and were subadditive on the accuracy scale. Retaining a trailing excerpt outperformed retaining a leading excerpt of similar average length by 10.10 points Holm-adjusted p 0.027. A second completion model showed the same directions of input effects on the fixed materials. In a cohort with parseable initial submissions, candidates added no correct answers when the previous draft was supplied, but added 27 points when it was absent. These results establish a conditional value of generated text: a failed submission can leave useful material, whose contribution depends on what else is supplied and what is retained.
Large language model safety evaluations often focus on a compliance-refusal tradeoff: models should refuse unsafe requests without refusing safe ones. We argue that this misses a broader form of prompt-conditioned task selection, where the same request can elicit different behaviors depending on the system prompt. We investigate this using a controlled evaluation crossing eight safety-task prompt families with three clarity levels across HarmBench, XSTest, and IHEval. Holding benchmark items fixed, we evaluate 13 models and use signal detection theory to distinguish two possibilities: whether models become better at telling safe from unsafe requests (discriminability), or simply more willing to refuse (response criterion). We find substantial differences in discriminability across models, but system prompts primarily shift the willingness to refuse, with smaller and less consistent effects on discriminability. Increasing prompt clarity makes models more willing to refuse without reliably improving their ability to distinguish safe from unsafe requests. Beyond refusal, prompt families also shift responses among six distinct behavioral policies, including clarification and safe help. Together, these findings show that system prompts primarily change how models respond to safety-relevant requests rather than consistently improving their ability to distinguish safe from unsafe ones.
Cloud-native systems now extend to the network edge and host artificial-intelligence (AI) workloads, including AI agents, but evidence on their reliability and scalability is scattered and uneven. We report a PRISMA 2020 systematic mapping study with an evidence-graded synthesis. Five logged OpenAlex searches on 11 October 2026 covered cloud-native systems, edge AI, agentic AI for operations, reliability of agentic applications and physical AI; 189 records from an earlier exploratory search were added separately. After 301 duplicate removals and an automated peer-review venue rule, 4,611 records were screened against written criteria and a five-level relevance scale. The review includes 1,794 primary studies from 2017-2026 and 95 secondary studies, with every decision released in Dataset S1. Each primary study was assigned to a 13-category taxonomy and graded as simulation, prototype or testbed, production data, production deployment, or not determinable from the abstract. Of 1,003 studies with a determinable level, 66.2% used prototypes or testbeds, 18.0% simulations and 15.8% production data or deployments. A 93-study synthesis core, appraised from indexed abstracts, shows production deployments of cascading-failure prevention, trace- and graph-based root-cause analysis and predictive autoscaling, and production-trace evaluations of cold-start mitigation. Edge reliability and physical AI rely mainly on simulations and testbeds. AI-agent reliability uses tool-grounded diagnosis, explicit safety properties, trace-driven repair and workload-aware scheduling, while empirical studies show that agent frameworks remain fragile. Conclusions are limited to the documented corpus. The principal gaps are production evidence for edge-native systems, shared resilience benchmarks and dependable agentic operations.
Reliable function calling requires Large Language Models (LLMs) to select appropriate functions and generate arguments consistent with user intent. However, syntactically valid calls may still be semantically incorrect due to inappropriate function selection or mismatched arguments. In this paper, we propose Decision-Model-Guided Function Calling (DM-Call), a framework that combines local schema validation with separate decision-model verification of function selection and argument correctness. For rejected candidates, DM-Call accumulates verification probabilities, binary judgments, and schema errors across attempts. A hybrid feedback representation further combines numerical probabilities with verbal correctness labels to guide iterative refinement without executing external tools. Extensive evaluation across BFCL and API-Bank (first-call) demonstrates DM-Call boosts function-calling accuracy versus direct generation, at modest additional cost.
Change impact analysis (CIA) helps developers identify code entities impacted by their initial changes. While several approaches for CIA were proposed, they still struggle with either low precision or low recall, thereby increasing manual effort for developers to verify the results. Furthermore, limitations in the used datasets and experimental setups can bias the results reported by many CIA approaches. In this paper, we first address the dataset limitations by manually labeling 1,931 changed code entities stemming from 263 commits of 22 open-source Java projects. Next, we introduce two variants of katIA, our code change impact analysis approach that combines static impact analysis with the reasoning capabilities of large language models (LLMs). The first variant uses a single prompt (katIA-SP) for refining an initial static impact set while the second variant uses three specialized agents (katIA-MA). The evaluation with our curated dataset shows that katIA-SP with Gemini 2.5 Flash achieved an F1-score of 22.4% when matching impacted entities by their fully qualified names (FQN) and 27.0% when matching them by their class and entity names (CE). The latter shows an improvement of 2.0 percentage points over the state-of-the-art. This score is further improved by katIA-MA which, using Gemini 2.5 Flash, yields an FQN-matching F1-score of 29.1% and a CE-matching F1-score of 33.3%. While these improvements come at the cost of increased median generation time from 0.570 s to 11.214 s and median token usage from 1,235 to 6,462 tokens, the improved performance justifies the additional cost. The comparison of three LLMs using katIA-SP finds that Gemini 2.5 Flash is the most consistent model. These results indicate that combining static impact analysis with LLM-based refinement, particularly with the multi-agent variant, can improve code change impact analysis.